Academia.eduAcademia.edu

Outline

Title
Abstract
Figures
Summary (Short)
Outline of the Implementation Method
Experimental Setting
Tagging Results
Experimental Results
Framework for Thai Morphological Analysis
Experimental Settings
Results
Error Analysis
Comparison Against Existing Results
Early Results in Open Test
Discussion and Related Work
Conclusion
Dataset Profiles
Discussion and Conclusions
Conclusions
Related Work
Results and Discussion
Discussion
Comparing and Combining Methods for
Distribution Analysis Approaches
Methods for Reweighting the Expanded Query Terms
Selecting the Reweighting Method
Parameter Study
Click-Through Data
Experimental Protocol
The Explsa Methods Comparison
Conclusion and Discussion
Method to Select Nuggets
Experiment Results
Conclusions and Future Work
Introduction
Experimental
References

A Language Independent Approach for Recognizing Textual Entailment

Profile image of Adrian IfteneAdrian Iftene

Advances in Natural Language Processing and …

Abstract
sparkles

AI

This paper presents a language-independent method for recognizing textual entailment (RTE), which determines whether the meaning of one text fragment can be inferred from another. The approach incorporates complex grammar rules, lemmatization, and machine learning algorithms to enhance the accuracy of entailment classification across various applications within natural language processing, such as information retrieval and question answering.

Figures (113)
Additionally, we built a system to acquire the extra Background Knowledge needec and applied complex grammar rules for rephrasing in English. These grammar rule: can be translated into any language, and we will see how this is done for Romanian. Fig. 1. The process requires an initial pre-processing, followed by the execution of a mair module. This uses the output of the first phase and obtains in the end an output that is used a: input for two machine learning algorithms. The machine learning algorithms eventually classify all pairs
Additionally, we built a system to acquire the extra Background Knowledge needec and applied complex grammar rules for rephrasing in English. These grammar rule: can be translated into any language, and we will see how this is done for Romanian. Fig. 1. The process requires an initial pre-processing, followed by the execution of a mair module. This uses the output of the first phase and obtains in the end an output that is used a: input for two machine learning algorithms. The machine learning algorithms eventually classify all pairs
Table 1. Background Knowledge Subsequently, we use this file with snippets and some previously set patterns with the aim of identifying relations between NEs. For all languages envisaged, we can apply the same format of patterns used for English and, additionally, we must add rules taking into account the specifics of the language. The patterns were initially built from the rules used to extract definitions in a Question Answering system. In order to reduce the search space and the complexity of rules, the latter have been categorized in six types:
Table 1. Background Knowledge Subsequently, we use this file with snippets and some previously set patterns with the aim of identifying relations between NEs. For all languages envisaged, we can apply the same format of patterns used for English and, additionally, we must add rules taking into account the specifics of the language. The patterns were initially built from the rules used to extract definitions in a Question Answering system. In order to reduce the search space and the complexity of rules, the latter have been categorized in six types:
Table 5. Patterns built for Location, Date and Organization answers types
Table 5. Patterns built for Location, Date and Organization answers types
Fig. 1. The logical structure of the text. To save space we omit the description of the segments 23 to 32 of PE methods The comparison between the Precision and the Recall of different logica methods presented in this paper reported to Short and Long manual segmenta. tion is given in the Fig.3. and Fig.4..
Fig. 1. The logical structure of the text. To save space we omit the description of the segments 23 to 32 of PE methods The comparison between the Precision and the Recall of different logica methods presented in this paper reported to Short and Long manual segmenta. tion is given in the Fig.3. and Fig.4..
Table 1. The results of manual segmentation in Short and Long variants. The results of LTT and PE methods To save space we omit the description of the segments 23 to 32 of PE methods.
Table 1. The results of manual segmentation in Short and Long variants. The results of LTT and PE methods To save space we omit the description of the segments 23 to 32 of PE methods.
Fig. 2. ArcInt and ArcReal structure of the text.
Fig. 2. ArcInt and ArcReal structure of the text.
Fig. 3. Comparison of Logical methods with Short segmentation
Fig. 3. Comparison of Logical methods with Short segmentation
Fig. 4. Comparison of Logical methods with Long segmentation
Fig. 4. Comparison of Logical methods with Long segmentation
Fig. 5. Comparison of Dynamic Progr. methods with Short segmentation
Fig. 5. Comparison of Dynamic Progr. methods with Short segmentation
Fig. 6. Comparison of Dynamic Progr. methods with Long segmentation 24 Tatar D., Tamaianu-Morita E., Mihis A. and Lupsa D.
Fig. 6. Comparison of Dynamic Progr. methods with Long segmentation 24 Tatar D., Tamaianu-Morita E., Mihis A. and Lupsa D.
Fig. 7. Precision and Recall for Logical methods (Short summary as standard) The precision and the recall ‘of the different methods when the manuall summaries are considered are presented in Fig.7-Fig.10. a a - --c” -- SN Um as -- iiares: a a: aaa ian mae a aa ae The summary determined by LTT method on initial text is formed by the sentences: {1,4,6,9—10, 12, 16, 18, 27, 29, 33, 36, 40, 44—45, 47, 50, 54, 57, 59, 62}. The presence of pairs 9-10 and 44-45 in summary is caused of the equal scores for the sentences 9,10 and 44,45 (see Fig.1.). The summary determined by LTT method on the text with Anaphora solved is formed by the sentences: {1,5 — 6,8, 11, 14, 16, 18, 20, 29, 31, 33, 36,38, 40, 44, 47, 50, 52, 54, 57, 60, 62, 65}.
Fig. 7. Precision and Recall for Logical methods (Short summary as standard) The precision and the recall ‘of the different methods when the manuall summaries are considered are presented in Fig.7-Fig.10. a a - --c” -- SN Um as -- iiares: a a: aaa ian mae a aa ae The summary determined by LTT method on initial text is formed by the sentences: {1,4,6,9—10, 12, 16, 18, 27, 29, 33, 36, 40, 44—45, 47, 50, 54, 57, 59, 62}. The presence of pairs 9-10 and 44-45 in summary is caused of the equal scores for the sentences 9,10 and 44,45 (see Fig.1.). The summary determined by LTT method on the text with Anaphora solved is formed by the sentences: {1,5 — 6,8, 11, 14, 16, 18, 20, 29, 31, 33, 36,38, 40, 44, 47, 50, 52, 54, 57, 60, 62, 65}.
Fig. 8. Precision and Recall for Logical methods (Long summary as standard)
Fig. 8. Precision and Recall for Logical methods (Long summary as standard)
Fig. 9. Precision and Recall for Dynamic Programming method (Short summary as standard)
Fig. 9. Precision and Recall for Dynamic Programming method (Short summary as standard)
Fig. 10. Precision and Recall for Dynamic Programming method (Long summary as standard)
Fig. 10. Precision and Recall for Dynamic Programming method (Long summary as standard)
Table 1. Oracle Performances of N-best Parse Trees For generating N-best candidate parse trees, we adopted Klein’s vertical and horizon- tal markovization [3]. We simply call it Klein’s markovization. As for vertical mark- ovization level (v) and horizontal level (1), we selected as v = 2 and h = 2. Table 1 shows the oracle performance for N-best parse trees of Klein’s markovization for different Ns (1, 50, 100, 200, 300, and 500), compared with the result of PCFG. Klein’s markovization showed a large difference from PCFG’s result when N is 1. As Nis larger, the performance difference becomes smaller.
Table 1. Oracle Performances of N-best Parse Trees For generating N-best candidate parse trees, we adopted Klein’s vertical and horizon- tal markovization [3]. We simply call it Klein’s markovization. As for vertical mark- ovization level (v) and horizontal level (1), we selected as v = 2 and h = 2. Table 1 shows the oracle performance for N-best parse trees of Klein’s markovization for different Ns (1, 50, 100, 200, 300, and 500), compared with the result of PCFG. Klein’s markovization showed a large difference from PCFG’s result when N is 1. As Nis larger, the performance difference becomes smaller.
Figure 1. Parsing accuracy of Collins-LA for different ks, and log-likelihood curve. 52 Seung-Hoon N., Meixun J., In-Su K. and Jong-Hyeok L.
Figure 1. Parsing accuracy of Collins-LA for different ks, and log-likelihood curve. 52 Seung-Hoon N., Meixun J., In-Su K. and Jong-Hyeok L.
Figure 3. Three center-head binarizations are presented using a) Klein’s markovization, b) Klein’s markovization with more horizontal markovization, and c) Klein’s markovizatoin with Johnson’s parent annotation. We used b) and c) for evaluating PCFG-LA.
Figure 3. Three center-head binarizations are presented using a) Klein’s markovization, b) Klein’s markovization with more horizontal markovization, and c) Klein’s markovizatoin with Johnson’s parent annotation. We used b) and c) for evaluating PCFG-LA.
Table 2. The final parsing accuracy of Collins-LA and PCFG-LA. As shown in Table 2, Collins-LA is comparable to PCFG-LA (1=1, h=2), without showing much difference. Our result of PCFG-LA is less-performed than the result of Matsuzaki’s one. The reason for this is that the initial quality of top-N parse trees is different. In Matsuzaki’s case, top-1 parsing accuracy of the initial parse trees reaches about 80% performance, which is much more superior to about 77% of our top-1 parse tree. We further evaluated PCFG-LA using the backbone grammar based on the binarization with Johnson’s parent annotation (v=2, h=2). Clearly, PCFG-LA with parent annotation (v=2, h=2) is superior to Collins-LA and PCFG-LA (v=1, h=2). Note that PCFG-LA with parent annotation uses a more complex backbone grammar, but Collins-LA uses simple PCFG grammar of (v=1, 4=0). The grammar used in Collins-LA is more simple than even PCFG-LA (v=1, h=2). Regarding this, the cur- rent result of Collins-LA can be more improved when adopting the elaborated back- bone grammar. Although some additional experiments are necessary, this experiment identifies that the formalism of Collins’ head-driven model provides a useful frame- work as PCFG does for latent annotation. This is a sufficiently notable result in the way of the research of latent annotation.
Table 2. The final parsing accuracy of Collins-LA and PCFG-LA. As shown in Table 2, Collins-LA is comparable to PCFG-LA (1=1, h=2), without showing much difference. Our result of PCFG-LA is less-performed than the result of Matsuzaki’s one. The reason for this is that the initial quality of top-N parse trees is different. In Matsuzaki’s case, top-1 parsing accuracy of the initial parse trees reaches about 80% performance, which is much more superior to about 77% of our top-1 parse tree. We further evaluated PCFG-LA using the backbone grammar based on the binarization with Johnson’s parent annotation (v=2, h=2). Clearly, PCFG-LA with parent annotation (v=2, h=2) is superior to Collins-LA and PCFG-LA (v=1, h=2). Note that PCFG-LA with parent annotation uses a more complex backbone grammar, but Collins-LA uses simple PCFG grammar of (v=1, 4=0). The grammar used in Collins-LA is more simple than even PCFG-LA (v=1, h=2). Regarding this, the cur- rent result of Collins-LA can be more improved when adopting the elaborated back- bone grammar. Although some additional experiments are necessary, this experiment identifies that the formalism of Collins’ head-driven model provides a useful frame- work as PCFG does for latent annotation. This is a sufficiently notable result in the way of the research of latent annotation.
Table 1. Evaluation of tagging results We can see that the lexicalized HMM outperforms bi-gram HMM significantly both on known words and unknown words tagging. sentences which have no more than 40 words. Among all the words in test sentences 8.17% are unknown words. _ In order to further confirm the performance of lexicalized HMM, we test it on English Penn tree bank which is much bigger than CTB. We use 39832 sentences from section 02-21 as training data and 2416 sentences from section 23 as test data. We get an accuracy of 96.77% which is close to state of the art of English POS tag- ging. So we can think 93.84% tagging accuracy on CTB is capable as a baseline in our following comparison experiments.
Table 1. Evaluation of tagging results We can see that the lexicalized HMM outperforms bi-gram HMM significantly both on known words and unknown words tagging. sentences which have no more than 40 words. Among all the words in test sentences 8.17% are unknown words. _ In order to further confirm the performance of lexicalized HMM, we test it on English Penn tree bank which is much bigger than CTB. We use 39832 sentences from section 02-21 as training data and 2416 sentences from section 23 as test data. We get an accuracy of 96.77% which is close to state of the art of English POS tag- ging. So we can think 93.84% tagging accuracy on CTB is capable as a baseline in our following comparison experiments.
Table 2. Tagging and parsing results Table 3. Tagging and parsing results
Table 2. Tagging and parsing results Table 3. Tagging and parsing results
Table 4. High frequency tagging error types on known words
Table 4. High frequency tagging error types on known words
Among the outputs of the tagger, only unknown words’ POS tags are kept and known words are re-tagged by the parser. So in this section, we simply follow this strategy. For a known word, the possible POS tags are limited to those which have been seen in training data for that word. Table 3 shows the experimental results.
Among the outputs of the tagger, only unknown words’ POS tags are kept and known words are re-tagged by the parser. So in this section, we simply follow this strategy. For a known word, the possible POS tags are limited to those which have been seen in training data for that word. Table 3 shows the experimental results.
Table 5. Tagging and parsing results Inspired by the above analysis, a straightforward idea is just re-tag the word “f/J” with the parser and keep the POS tags of other words unchanged. Table 5 shows the experimental results. As we expected, the tagging accuracy is increased from 97.20% to 97.55%. However, the F1 score of parsing is decreased from 83.16%to 82.94%. It seems that better tagging results do not always lead to better parsing results. The reason for this interesting phenomenon is currently beyond our knowledge and fur- ther investigation is necessary.
Table 5. Tagging and parsing results Inspired by the above analysis, a straightforward idea is just re-tag the word “f/J” with the parser and keep the POS tags of other words unchanged. Table 5 shows the experimental results. As we expected, the tagging accuracy is increased from 97.20% to 97.55%. However, the F1 score of parsing is decreased from 83.16%to 82.94%. It seems that better tagging results do not always lead to better parsing results. The reason for this interesting phenomenon is currently beyond our knowledge and fur- ther investigation is necessary.
Table 6. Tagging and parsing results Comparing table 2 with table 6, we can see that Collins’ model is superior to lexi- calized HMM on unknown words POS tagging because the accuracy is increased to 61.63% from 57.25%. There are two differences on tagging between Collins’ model and the lexicalized HMM. First, HMM generates a word’s tag conditioning on infor- mation of its previous word, while Collins’ model generates a word’s tag condition- ing on information of the head word which the word depends on. Second, Collins’ model uses higher level syntax structure information additionally. We think these two differences account for the improvement on unknown words POS tagging.
Table 6. Tagging and parsing results Comparing table 2 with table 6, we can see that Collins’ model is superior to lexi- calized HMM on unknown words POS tagging because the accuracy is increased to 61.63% from 57.25%. There are two differences on tagging between Collins’ model and the lexicalized HMM. First, HMM generates a word’s tag conditioning on infor- mation of its previous word, while Collins’ model generates a word’s tag condition- ing on information of the head word which the word depends on. Second, Collins’ model uses higher level syntax structure information additionally. We think these two differences account for the improvement on unknown words POS tagging.
Table 7. Tagging and parsing results
Table 7. Tagging and parsing results
Table 1. Experimental results for the development set
Table 1. Experimental results for the development set
Table 2. Experimental results for the development set is observed (6th row) with the introduction of the Symbol’, ’Digit’, Length’ and ’*FunctionWord’ features. The use of lexicon as the features of ME model further improves the accuracy by 1.8% (7th row). Accuracy of the POS tagger rises to 84.2% (8th row) by including the noun, verb and adjective inflections. Finally, an overall accuracy of 84. 6% i is obtained with the inclusion of ’RareWord’ feature.
Table 2. Experimental results for the development set is observed (6th row) with the introduction of the Symbol’, ’Digit’, Length’ and ’*FunctionWord’ features. The use of lexicon as the features of ME model further improves the accuracy by 1.8% (7th row). Accuracy of the POS tagger rises to 84.2% (8th row) by including the noun, verb and adjective inflections. Finally, an overall accuracy of 84. 6% i is obtained with the inclusion of ’RareWord’ feature.
Table 3. Overall evaluation results of the POS tagger for the development set Clearly, the results of Table 3 show the effectiveness of employing the various strategies for handling the unknown words with the increase in the accuracy values. The ME model exhibits an accuracy of 88.9% with the lexicon, named entity recognizer and unknown word features. The system has demonstrated Evaluation results of the POS tagger by including the various mechanisms for handling the unknown words are presented in Table 3 for the development set. The table also shows the results of the baseline model.
Table 3. Overall evaluation results of the POS tagger for the development set Clearly, the results of Table 3 show the effectiveness of employing the various strategies for handling the unknown words with the increase in the accuracy values. The ME model exhibits an accuracy of 88.9% with the lexicon, named entity recognizer and unknown word features. The system has demonstrated Evaluation results of the POS tagger by including the various mechanisms for handling the unknown words are presented in Table 3 for the development set. The table also shows the results of the baseline model.
Table 4. Experimental results of the system for the test set The performance of the proposed ME based POS tagger has been compared with a HMM based POS tagger [13], in which trigram model was considered. Additional context dependent features were considered for the emission proba- bilities in this HMM based POS tagger . Also, the lexicon [10], named entity recognizer [9] and unknown word features were considered in a similar way in order to handle the unknown word problems. This system has been trained and tested with the same data. Evaluation results of the HMM based POS tagger is presented in Table 5.
Table 4. Experimental results of the system for the test set The performance of the proposed ME based POS tagger has been compared with a HMM based POS tagger [13], in which trigram model was considered. Additional context dependent features were considered for the emission proba- bilities in this HMM based POS tagger . Also, the lexicon [10], named entity recognizer [9] and unknown word features were considered in a similar way in order to handle the unknown word problems. This system has been trained and tested with the same data. Evaluation results of the HMM based POS tagger is presented in Table 5.
Table 5. Evaluation results of the HMM based POS tagger for the test set
Table 5. Evaluation results of the HMM based POS tagger for the test set
Table 6. Evaluation results of the POS tagger for the unknown words in the test set
Table 6. Evaluation results of the POS tagger for the unknown words in the test set
Fig. 1. Lattice produced from an unsegmented string
Fig. 1. Lattice produced from an unsegmented string
Note that the forward and backward probabilities are calculated with recursion similar to the Viterbi algorithm, except using sum function instead of max func- tion. We can apply the Forward-Backward algorithm to obtain p, for a given node y. Let a, be a forward probability and 8, be a backward probability reach- ing node y from the left and right, respectively. The marginal probability of a particular node y can be calculated by:
Note that the forward and backward probabilities are calculated with recursion similar to the Viterbi algorithm, except using sum function instead of max func- tion. We can apply the Forward-Backward algorithm to obtain p, for a given node y. Let a, be a forward probability and 8, be a backward probability reach- ing node y from the left and right, respectively. The marginal probability of a particular node y can be calculated by:
(b) The suspicious node indicated by the dashed line box are identified, and the un- certainty range which has no potential nodes in the paths is located
(b) The suspicious node indicated by the dashed line box are identified, and the un- certainty range which has no potential nodes in the paths is located
(d) Lattice with new candidate word hypotheses. The correct path can be recoverd ir the expanded search space
(d) Lattice with new candidate word hypotheses. The correct path can be recoverd ir the expanded search space
Table 1. Details of dataset used in our experiments Since our problem formulation is based on a log-linear model, a variety of features can be incorporated in the model. Table 2 provides the details of features used in our experiments. We broadly classified character types into 6 categories, including symbol, alphabet, number, ordinal number, Thai and other character types. Note that a word w, is considered to be a rare word, if it occurs less than 2 times in the training set.
Table 1. Details of dataset used in our experiments Since our problem formulation is based on a log-linear model, a variety of features can be incorporated in the model. Table 2 provides the details of features used in our experiments. We broadly classified character types into 6 categories, including symbol, alphabet, number, ordinal number, Thai and other character types. Note that a word w, is considered to be a rare word, if it occurs less than 2 times in the training set.
Table 2. Features are based on three-level, first-order Markov assumption. Each node y contains morphological information of word surface w, POS p and sub-POS p’. fx (-) is operated on yt-1 — Yt 4.2 Results
Table 2. Features are based on three-level, first-order Markov assumption. Each node y contains morphological information of word surface w, POS p and sub-POS p’. fx (-) is operated on yt-1 — Yt 4.2 Results
Table 3. Results of word segmentation Table 4. Results of word segmentation and POS tagging
Table 3. Results of word segmentation Table 4. Results of word segmentation and POS tagging
€ is used for identifying suspicious nodes. The baseline method cannot produce acceptable results. These results indicate how difficult the morphological analysis in ORCHID corpus is. The standard search which performs a single-pass Viterbi search can boost the performance with a reasonable gap. The best performance (except Recallunknown) can be obtained by using the two-pass search algorithm with ¢ = 0.5. The two-pass search algorithm improves the performance of the standard search by 3.23 F in word segmentation and 2.92 F in the combination of word segmentation and POS tagging. The two-pass search algorithm yields the best Recallunknown With ¢€ = 0.8. The more the value of € is set, the more the suspicious nodes are identified. The new word hypotheses are subsequently generated, which in turn increase noises. In practice, the optimal value of € can be selected by cross-validation.
€ is used for identifying suspicious nodes. The baseline method cannot produce acceptable results. These results indicate how difficult the morphological analysis in ORCHID corpus is. The standard search which performs a single-pass Viterbi search can boost the performance with a reasonable gap. The best performance (except Recallunknown) can be obtained by using the two-pass search algorithm with ¢ = 0.5. The two-pass search algorithm improves the performance of the standard search by 3.23 F in word segmentation and 2.92 F in the combination of word segmentation and POS tagging. The two-pass search algorithm yields the best Recallunknown With ¢€ = 0.8. The more the value of € is set, the more the suspicious nodes are identified. The new word hypotheses are subsequently generated, which in turn increase noises. In practice, the optimal value of € can be selected by cross-validation.
Table 5. Results without each feature produced by Two-Pass,-=o.5 case is when unknown words are mixed with various character types. Since OR- CHID corpus was derived from technical papers, the number of technical terms and jargons is relatively high. For example, the surface ‘Gao.65A1I035AS’ is seg- mented into smaller tokens, ‘Gao’, ‘.’, ‘65’, ‘AI’, ‘035’ and ‘AS’, according to their character types. In this case, it is not difficult to resolve, because we can use some regular expressions to detect domain-specific word surfaces. The sec- ond case which makes Thai morphological analysis much more difficult is when unknown words are formed from known words, and each of them is otherwise found independently in the corpus. This case can be thought of as the problem of compound words. To deal with the problem, we need to find a set of features that can capture the compounding process.
Table 5. Results without each feature produced by Two-Pass,-=o.5 case is when unknown words are mixed with various character types. Since OR- CHID corpus was derived from technical papers, the number of technical terms and jargons is relatively high. For example, the surface ‘Gao.65A1I035AS’ is seg- mented into smaller tokens, ‘Gao’, ‘.’, ‘65’, ‘AI’, ‘035’ and ‘AS’, according to their character types. In this case, it is not difficult to resolve, because we can use some regular expressions to detect domain-specific word surfaces. The sec- ond case which makes Thai morphological analysis much more difficult is when unknown words are formed from known words, and each of them is otherwise found independently in the corpus. This case can be thought of as the problem of compound words. To deal with the problem, we need to find a set of features that can capture the compounding process.
Table 2. Performance comparison: features derived by different criteria
Table 2. Performance comparison: features derived by different criteria
Table 3. Performance improvement with features derived from various unsupervised criterio1 combinations using feature formula (5)
Table 3. Performance improvement with features derived from various unsupervised criterio1 combinations using feature formula (5)
Table 4. Comparisons of the best existing results and ours in data of Bakeoff-3 (F-scores)
Table 4. Comparisons of the best existing results and ours in data of Bakeoff-3 (F-scores)
Table 5. Statistical significance: comparisons of the best closed results of Bakeoff-3 and ours
Table 5. Statistical significance: comparisons of the best closed results of Bakeoff-3 and ours
* Using AV features from People’s Daily text. > Using features from external dictionary. Table 6. Performances using external unlabeled data or lexicon
* Using AV features from People’s Daily text. > Using features from external dictionary. Table 6. Performances using external unlabeled data or lexicon
Table 1. Dataset matrix (fictional values) for 2-grain (2 target classes) task. Dataset elements are target classes (e.g. senses of word), training and test instances, features and feature values. In classification tasks such as weather prediction (e.g. sunny/rain, hail/snow), target class is predicted using either binary (0/1), numerical (0..1) or set-based (0,1,n) features as probabilistic pointers to those target classes. 1. Word sense disambiguation (WSD) datasets. Word sense disambiguation is typically based on a combination of contextual-statistical methods (e.g. n-grams or parts of speech from around target word). 1-grams are single words around target word from the whole target word instance, 2-grams are consecutive words in the whole target word instance, parts of speech are POS tags from the local context around target word (see Listing 1).
Table 1. Dataset matrix (fictional values) for 2-grain (2 target classes) task. Dataset elements are target classes (e.g. senses of word), training and test instances, features and feature values. In classification tasks such as weather prediction (e.g. sunny/rain, hail/snow), target class is predicted using either binary (0/1), numerical (0..1) or set-based (0,1,n) features as probabilistic pointers to those target classes. 1. Word sense disambiguation (WSD) datasets. Word sense disambiguation is typically based on a combination of contextual-statistical methods (e.g. n-grams or parts of speech from around target word). 1-grams are single words around target word from the whole target word instance, 2-grams are consecutive words in the whole target word instance, parts of speech are POS tags from the local context around target word (see Listing 1).
Table 2. Average accuracies of classifiers at two different types of gensets. Table 3. Classifier results on different class grains of protein dataset.
Table 2. Average accuracies of classifiers at two different types of gensets. Table 3. Classifier results on different class grains of protein dataset.
» Classifier Strengths
» Classifier Strengths
Defining Relative Strengths of Classifiers at Randomly Generated Datasets 115 Figure 1. Three-dimensional space for representing classifier formations (= relative strengths) with P, N and nrfeat as dimensions. C stands for 'corner' (e.g. C4 is high-P but low-N, C2 is high-N but low-P) and the region remaining between corners we call 'midfield'. S hown are three types of formations: (a) single pear-shaped formation is called 'yin-yang formation’ with one classifier taking the midfield (and the whole of C1 regardless of nrfeat), the other the P and N facets, (b) dotted line in the middle is called "hemispherical formation' with one c P facet (C4) and the other N facet (C2), (c) two horizontal lines denote 'layered layers), with one classifier taking both low and high nrfeats (regardless of P and assifier taking formation' (or ) and the other middle nrfeat range. (Cf. Discussion and [11] for another account on such 1-2-1 phenomena). We refer to these types of formations below when describing classifier performance differences.”
Defining Relative Strengths of Classifiers at Randomly Generated Datasets 115 Figure 1. Three-dimensional space for representing classifier formations (= relative strengths) with P, N and nrfeat as dimensions. C stands for 'corner' (e.g. C4 is high-P but low-N, C2 is high-N but low-P) and the region remaining between corners we call 'midfield'. S hown are three types of formations: (a) single pear-shaped formation is called 'yin-yang formation’ with one classifier taking the midfield (and the whole of C1 regardless of nrfeat), the other the P and N facets, (b) dotted line in the middle is called "hemispherical formation' with one c P facet (C4) and the other N facet (C2), (c) two horizontal lines denote 'layered layers), with one classifier taking both low and high nrfeats (regardless of P and assifier taking formation' (or ) and the other middle nrfeat range. (Cf. Discussion and [11] for another account on such 1-2-1 phenomena). We refer to these types of formations below when describing classifier performance differences.”
Table 2. Translation rules for (NP+LV) structure (Partial) [2] The Translation of (VP+LV) Structure:
Table 2. Translation rules for (NP+LV) structure (Partial) [2] The Translation of (VP+LV) Structure:
Table3. Translation rules for (NP+LV) structure (Partial)
Table3. Translation rules for (NP+LV) structure (Partial)
If VN is also a verbal noun in Chinese, then VN can directly be adopted as a verb. However, we have to pay attention to the treatment of modifiers of VN. We classify the modifiers into the following three types and propose the translation rules (table5). Tabled. Translation rules for (VN-o LV) structure for pure noun (Partial)
If VN is also a verbal noun in Chinese, then VN can directly be adopted as a verb. However, we have to pay attention to the treatment of modifiers of VN. We classify the modifiers into the following three types and propose the translation rules (table5). Tabled. Translation rules for (VN-o LV) structure for pure noun (Partial)
Table 5. Translation rules for (VN-o LV) structure for verbal noun (Partial)
Table 5. Translation rules for (VN-o LV) structure for verbal noun (Partial)
Table 6. Translation rules for (ADJ+LV), (ADV+LV) and (ONO+LV) structures (Partial) Translation of the Light Verb Constructions in Japanese-Chinese Machine Translation 145
Table 6. Translation rules for (ADJ+LV), (ADV+LV) and (ONO+LV) structures (Partial) Translation of the Light Verb Constructions in Japanese-Chinese Machine Translation 145
The statistics shows that the number of Japanese-Chinese translation patterns is considerably less than the number of Japanese-English translation patterns. The difference can be analyzed as follows. 1. One Japanese-Chinese translation pattern corresponds to several Japanese-English translation patterns. We observed the following two points as the major reasons: Table 7. The comparison of the number of translation patterns with Japanese-English In this section, we made comparisons of the Japanese-Chinese translation patterns for LVCs mentioned in section 2 with the Japanese-English translation patterns listed in Goi-Taikei--A Japanese Lexicon[14](Table 7).
The statistics shows that the number of Japanese-Chinese translation patterns is considerably less than the number of Japanese-English translation patterns. The difference can be analyzed as follows. 1. One Japanese-Chinese translation pattern corresponds to several Japanese-English translation patterns. We observed the following two points as the major reasons: Table 7. The comparison of the number of translation patterns with Japanese-English In this section, we made comparisons of the Japanese-Chinese translation patterns for LVCs mentioned in section 2 with the Japanese-English translation patterns listed in Goi-Taikei--A Japanese Lexicon[14](Table 7).
Fig. 1. outline of jaw/Chinese First, a Japanese input sentence is morphologically and syntactically analyzed using Ibuki, a Japanese syntactical structure analysis system developed in our laboratory, and a dependency tree is constructed. Each node of the tree corresponds to one charac conten simatta eristic Japanese linguistic unit (bunsetsu), which is a phrase including one word and any number of function words (or none). For example: kowasarete kamo siremasen ga (however, (it) may have been destroyed) is one bunsetsu, composed of one content word (kowasu: destroy) and several function words reru: be done to), (tesimau: have done), (ta: past), (kamosiremasenn: may be) and (ga: however). In this step, an Input Tree (IT) is created.
Fig. 1. outline of jaw/Chinese First, a Japanese input sentence is morphologically and syntactically analyzed using Ibuki, a Japanese syntactical structure analysis system developed in our laboratory, and a dependency tree is constructed. Each node of the tree corresponds to one charac conten simatta eristic Japanese linguistic unit (bunsetsu), which is a phrase including one word and any number of function words (or none). For example: kowasarete kamo siremasen ga (however, (it) may have been destroyed) is one bunsetsu, composed of one content word (kowasu: destroy) and several function words reru: be done to), (tesimau: have done), (ta: past), (kamosiremasenn: may be) and (ga: however). In this step, an Input Tree (IT) is created.
(1-1) Example lof Japanese pattern for base type (1-2) The corresponding translation rules for Chinese of Example | Table 9. Japanese expression patterns and translation rules for Chinese
(1-1) Example lof Japanese pattern for base type (1-2) The corresponding translation rules for Chinese of Example | Table 9. Japanese expression patterns and translation rules for Chinese
(2-2) The corresponding translation rules for Chinese of Example2
(2-2) The corresponding translation rules for Chinese of Example2
In this way, we implemented the translation rules in jaw/Chinese; next, we carried out closed and open tests using the system. Closed tests were conducted on 30 sentences containing LVCs; all of the translations results were correct. Open tests were carried out using 100 sentences, and the results are shown in Table10. Both of the test shows that the rules is applicable in real MT system and the method gave translation results with a high accuracy, compared with commercially available MT software.
In this way, we implemented the translation rules in jaw/Chinese; next, we carried out closed and open tests using the system. Closed tests were conducted on 30 sentences containing LVCs; all of the translations results were correct. Open tests were carried out using 100 sentences, and the results are shown in Table10. Both of the test shows that the rules is applicable in real MT system and the method gave translation results with a high accuracy, compared with commercially available MT software.
Table10. Evaluation of jaw/Chinese open-test experiments 5 Conclusions
Table10. Evaluation of jaw/Chinese open-test experiments 5 Conclusions
Table 1. Multiple Mapping into Gurmukhi Script 3.4 Word-Boundary Mismatch
Table 1. Multiple Mapping into Gurmukhi Script 3.4 Word-Boundary Mismatch
and Vava &[v] (usage Als] +%:+<[v] = H [sv]). 4.2 Shahmukhi Script
and Vava &[v] (usage Als] +%:+<[v] = H [sv]). 4.2 Shahmukhi Script
Table 3. Shahmukhi Non-Aspirated Consonents Mapping Table 4. Aspirate Consonants (AC) Mapping
Table 3. Shahmukhi Non-Aspirated Consonents Mapping Table 4. Aspirate Consonants (AC) Mapping
Table 5. Shahmukhi Long Vowels Mapping Table 6. Shahmukhi Short Vowels Mapping
Table 5. Shahmukhi Long Vowels Mapping Table 6. Shahmukhi Short Vowels Mapping
Table 7. Mapping of other Diacritical Marks or Symbols
Table 7. Mapping of other Diacritical Marks or Symbols
Fig. 1. Architecture of Transliteration and Pre-Processing The Output Text Generator has to pack these tokens well with other input text which may include punctuation marks and embedded Roman text. Finally, this will generate a Unicode formatted Gurmukhi text as shown in Figure 2.
Fig. 1. Architecture of Transliteration and Pre-Processing The Output Text Generator has to pack these tokens well with other input text which may include punctuation marks and embedded Roman text. Finally, this will generate a Unicode formatted Gurmukhi text as shown in Figure 2.
Table 8. Transliteration Results
Table 8. Transliteration Results
Table 10. Comparison of Output-A & B Clearly, our system is more practical in nature than PMT and we got good transliteration with different inputs having missing diacritical marks. But we are still having erroneous transliterations by the system. The main source of error is the Output-B of proposed system (left to right) fer ae fed ae mH ugs Undhat o Sud ot ot fyHtus sa fHw e Sy fes first aha ash Heehat dd et hes avs HoHH debut Jol At eH o fling age eT eet ade oud mus HES eA Ti fH o HS SF est HYS fea J fa ToT S Ba ga BTS Ha wué »ué HeTus’ eA us Can 3 AAs aS Hae dal wuel ays, wus Afsurerg, wtué fueas 3 nue feah 3S He age dal wruet aft uete SHS age USI ud Hat wer aren di fogs dl Hdarat FS a oH Bat Sa Una wa € ues fea ud udt sai mens dt afte Jal
Table 10. Comparison of Output-A & B Clearly, our system is more practical in nature than PMT and we got good transliteration with different inputs having missing diacritical marks. But we are still having erroneous transliterations by the system. The main source of error is the Output-B of proposed system (left to right) fer ae fed ae mH ugs Undhat o Sud ot ot fyHtus sa fHw e Sy fes first aha ash Heehat dd et hes avs HoHH debut Jol At eH o fling age eT eet ade oud mus HES eA Ti fH o HS SF est HYS fea J fa ToT S Ba ga BTS Ha wué »ué HeTus’ eA us Can 3 AAs aS Hae dal wuel ays, wus Afsurerg, wtué fueas 3 nue feah 3S He age dal wruet aft uete SHS age USI ud Hat wer aren di fogs dl Hdarat FS a oH Bat Sa Una wa € ues fea ud udt sai mens dt afte Jal
Table 1. Words with highest dynamic range (left column) and lowest dynamic range (right column) For illustration purposes, we now show a couple of synthetic examples in which BLEU and semantically weighted BLEU scores are compared depending on two sentences. The phrases used are made up and are not part of the corpus (nor representative of the corpus). Table 2 below shows the original sentence and the result of a translation and a roundtrip translation. The output of the back-translation is compared against the original and the scores are computed. We can see that when
Table 1. Words with highest dynamic range (left column) and lowest dynamic range (right column) For illustration purposes, we now show a couple of synthetic examples in which BLEU and semantically weighted BLEU scores are compared depending on two sentences. The phrases used are made up and are not part of the corpus (nor representative of the corpus). Table 2 below shows the original sentence and the result of a translation and a roundtrip translation. The output of the back-translation is compared against the original and the scores are computed. We can see that when
Table 2. Sample phrases, original and perturbed versions, with their BLEU and semantically weighted BLEU scores for two cases. 5.2 Rephrasing Experiments
Table 2. Sample phrases, original and perturbed versions, with their BLEU and semantically weighted BLEU scores for two cases. 5.2 Rephrasing Experiments
Fig. 1. Cross similarity (left) and self similarity (right) Cosine distance measure distribution. utterance in the other corpus we count one correct classification (and an incorrect otherwise). We did this both ways (classifying version A first and then classifying version B).The results are shown below in table 3. As we can see, about 80% accuracy is achieved with cosine distance. Table 3. Classification results using the Cosine Distance
Fig. 1. Cross similarity (left) and self similarity (right) Cosine distance measure distribution. utterance in the other corpus we count one correct classification (and an incorrect otherwise). We did this both ways (classifying version A first and then classifying version B).The results are shown below in table 3. As we can see, about 80% accuracy is achieved with cosine distance. Table 3. Classification results using the Cosine Distance
Fig. 2. Classification accuracy in the rephrase task as a function of rank k.
Fig. 2. Classification accuracy in the rephrase task as a function of rank k.
Table 1. Comparing different cooccurrence methods. The Baseline row corresponds to the results of the query without expansion. P@5 stands for precision after the first five documents retrieved, P@10 after the first ten, and R-PREC stands for R-precision. Best results appear in boldface.
Table 1. Comparing different cooccurrence methods. The Baseline row corresponds to the results of the query without expansion. P@5 stands for precision after the first five documents retrieved, P@10 after the first ten, and R-PREC stands for R-precision. Best results appear in boldface.
Fig. 1. Study of the best number of candidate terms to expand the original query with the different considered methods. R-PREC stands for R-Precision. Comparing and Combining Methods for Automatic Query Expansion 185
Fig. 1. Study of the best number of candidate terms to expand the original query with the different considered methods. R-PREC stands for R-Precision. Comparing and Combining Methods for Automatic Query Expansion 185
Fig. 2. Study of the best number of documents used to extract the set of candidate query terms. R-PREC stands for R-Precision.
Fig. 2. Study of the best number of documents used to extract the set of candidate query terms. R-PREC stands for R-Precision.
Table 5. Comparing MAP and GMAP for different methods considered for query expansion. Table 6. Comparing R-Precision (R-PREC), precision after 5 documents retrieved (P@5) and after 10 documents retrieved (P@10) for different methods considered for query expansion.
Table 5. Comparing MAP and GMAP for different methods considered for query expansion. Table 6. Comparing R-Precision (R-PREC), precision after 5 documents retrieved (P@5) and after 10 documents retrieved (P@10) for different methods considered for query expansion.
Table 8. Results of the MAP measure for the queries 41, 49, 53, 55, 58, 122. BoCo stands for the model combining Bol and cooccurrence. The best value appears in boldface.
Table 8. Results of the MAP measure for the queries 41, 49, 53, 55, 58, 122. BoCo stands for the model combining Bol and cooccurrence. The best value appears in boldface.
Fig. 1 Average relevancy and expertness learning results. Sample search queries were executed on WebFusion. The results are supposed to be classified to relevant, undecided, and irrelevant documents. Each exper- iment participant was asked to evaluate the result items in the corresponding classes.
Fig. 1 Average relevancy and expertness learning results. Sample search queries were executed on WebFusion. The results are supposed to be classified to relevant, undecided, and irrelevant documents. Each exper- iment participant was asked to evaluate the result items in the corresponding classes.
Fig. 2 Query optimization improvement for random initial population (left) and seeded initial population (right). We considered user models containing Boolean search queries representing ones individual search needs, long and short term context and document rele- vance estimation [18]. Several experiments were executed using data extracted from the LISA collection . The collection was indexed for Boolean IR and Ex- tended Bolean IR systems, using Saltons indexing function based on normalized term frequency and normalized inverse document term frequency [19] in the lat- ter case. Indexed collection contained 5999 documents and 18442 unique indexed terms. Due to stochastic character of GP process, all experiments were executed multiple times and mean experimental results evaluated.
Fig. 2 Query optimization improvement for random initial population (left) and seeded initial population (right). We considered user models containing Boolean search queries representing ones individual search needs, long and short term context and document rele- vance estimation [18]. Several experiments were executed using data extracted from the LISA collection . The collection was indexed for Boolean IR and Ex- tended Bolean IR systems, using Saltons indexing function based on normalized term frequency and normalized inverse document term frequency [19] in the lat- ter case. Indexed collection contained 5999 documents and 18442 unique indexed terms. Due to stochastic character of GP process, all experiments were executed multiple times and mean experimental results evaluated.
Fig. 3 Proposed search architecture. Implicit observation of user click-through behavior is used to provide data ba- sis for building both, SEEPs and IUPs. The simultaneous exploitation of search ngine expertness profiles and individual user profiles will be deployed to improve she result sets retrieved in response to user queries. Results will be re-rankec with respect to detected search engine expertness, preferring for particular cate- rory results retrieved by search engines with higher expertness. User profiles will »e employed to be a basis for personalized query optimization, performed on the op of retrieved and according to SEEP re-ranked result set. Neither the SEEPs 1or IUPs are needed to be specified explicitly. On the contrary, meta-information orovided by ordinary user tasks performed during web search ac tivities and hence ulmost omnipresent (and in majority of search environments omitted), will be *xtracted and reasonably exploited to discover SEEPs and UIPs. Hereby created und maintained search engine expertness uny additional feedback from the users, a profiles and user profi es do not require though an explicit feedback should be onsidered as a mean to speed up the expertness and user learning process and nake the profiles more literal.
Fig. 3 Proposed search architecture. Implicit observation of user click-through behavior is used to provide data ba- sis for building both, SEEPs and IUPs. The simultaneous exploitation of search ngine expertness profiles and individual user profiles will be deployed to improve she result sets retrieved in response to user queries. Results will be re-rankec with respect to detected search engine expertness, preferring for particular cate- rory results retrieved by search engines with higher expertness. User profiles will »e employed to be a basis for personalized query optimization, performed on the op of retrieved and according to SEEP re-ranked result set. Neither the SEEPs 1or IUPs are needed to be specified explicitly. On the contrary, meta-information orovided by ordinary user tasks performed during web search ac tivities and hence ulmost omnipresent (and in majority of search environments omitted), will be *xtracted and reasonably exploited to discover SEEPs and UIPs. Hereby created und maintained search engine expertness uny additional feedback from the users, a profiles and user profi es do not require though an explicit feedback should be onsidered as a mean to speed up the expertness and user learning process and nake the profiles more literal.
Fig. 1. Overview of the system architecture The first step to provide transparent access to announced job openings by an integrated platform was the automatic classification of Web pages into a database of organization home pages. The HTML-structure of these pages was then scanned for anchor elements that lead to job advertisements.
Fig. 1. Overview of the system architecture The first step to provide transparent access to announced job openings by an integrated platform was the automatic classification of Web pages into a database of organization home pages. The HTML-structure of these pages was then scanned for anchor elements that lead to job advertisements.
Fig. 2. Local grammar that models the context of the word leader. Though these five phrases are highly different on the morpho-syntactic level, they describe a highly similar context of the word “leader” that refers to the same semantic concept. By a local grammar, as shown in Figure 2, it is possible to efficiently represent all five phrases.
Fig. 2. Local grammar that models the context of the word leader. Though these five phrases are highly different on the morpho-syntactic level, they describe a highly similar context of the word “leader” that refers to the same semantic concept. By a local grammar, as shown in Figure 2, it is possible to efficiently represent all five phrases.
Table 2. Form of a structured job offer For the transformation of the initial HTML-documents into the form schema, we need different operations. The chronological application of the five preprocessing steps is shown in Figure 3 (see next page).
Table 2. Form of a structured job offer For the transformation of the initial HTML-documents into the form schema, we need different operations. The chronological application of the five preprocessing steps is shown in Figure 3 (see next page).
Fig. 3. Steps of the classification and transformation of a possible job annoucement
Fig. 3. Steps of the classification and transformation of a possible job annoucement
Table 3. Evaluation results gained on test corpora Table 3 shows promising results of precision (95.9% on average) and recall (93.0% on average) considering the recognition of entities typically found in job offers. The experimental evaluation presented in this paper was limited to the five highest weighted of the 20 information bits. The majority of difficulties could be observed for the explicit identification of company names.
Table 3. Evaluation results gained on test corpora Table 3 shows promising results of precision (95.9% on average) and recall (93.0% on average) considering the recognition of entities typically found in job offers. The experimental evaluation presented in this paper was limited to the five highest weighted of the 20 information bits. The majority of difficulties could be observed for the explicit identification of company names.
‘ig. 1. The Asium measure between p and q verbs.
‘ig. 1. The Asium measure between p and q verbs.
Fig. 2. The Asium measure between the verbs écouter (to listen) and convaincre (to convince) The example of the figure 2 gives an application of the Asium measure for the écouter (to listen) and convaincre (to convince) verbs. Next, we consider differ- ent similarity threshold values of the Asium measure. Over this one, verbs are considered as close by the Asium measure. The expansion method is described in the next section.
Fig. 2. The Asium measure between the verbs écouter (to listen) and convaincre (to convince) The example of the figure 2 gives an application of the Asium measure for the écouter (to listen) and convaincre (to convince) verbs. Next, we consider differ- ent similarity threshold values of the Asium measure. Over this one, verbs are considered as close by the Asium measure. The expansion method is described in the next section.
Table 1. Precision of the 100 first couples. Method M1 = LSA, Method M2 = Ex- pLSA with the intersections method, M3 = ExpLSA with the complementaries method. Asium threshold = 0.6. 220 Béchet N., Roche M. and Chauché J.
Table 1. Precision of the 100 first couples. Method M1 = LSA, Method M2 = Ex- pLSA with the intersections method, M3 = ExpLSA with the complementaries method. Asium threshold = 0.6. 220 Béchet N., Roche M. and Chauché J.
Table 2. Precision of the 100 first couples. Method M1 = LSA, Method M2 = Ex- pLSA with the intersections method, M3 = ExpLSA with the complementaries method. Asium threshold = 0.8.
Table 2. Precision of the 100 first couples. Method M1 = LSA, Method M2 = Ex- pLSA with the intersections method, M3 = ExpLSA with the complementaries method. Asium threshold = 0.8.
Fig. 3. Precision of the 200 first couples comparing the EzpLSA methods between concepts behavior and environment ExpLSA: An Approach Based on Syntactic Knowledge in Order to Improve LSA... 221
Fig. 3. Precision of the 200 first couples comparing the EzpLSA methods between concepts behavior and environment ExpLSA: An Approach Based on Syntactic Knowledge in Order to Improve LSA... 221
Fig. 4. Precision of the 200 first couples comparing the ExpLSA and LSA + Tree-tagger methods between concepts behavior and environment
Fig. 4. Precision of the 200 first couples comparing the ExpLSA and LSA + Tree-tagger methods between concepts behavior and environment
Table 3. Precision of the 100 first couples comparing the LSA (M1), the ExpLSA with Asium threshold = 0.8 (M2) and LSA + Tree-tagger approach (M4)
Table 3. Precision of the 100 first couples comparing the LSA (M1), the ExpLSA with Asium threshold = 0.8 (M2) and LSA + Tree-tagger approach (M4)
2.290 nickname for Venezuelan born Ilich Ramirez Sanchez Fig. 1. Comparison of the F-measures obtained with two different retrieval systems JIRS and Lucene.
2.290 nickname for Venezuelan born Ilich Ramirez Sanchez Fig. 1. Comparison of the F-measures obtained with two different retrieval systems JIRS and Lucene.
We performed experiments with three sets of definition question, the questions from the pilot evaluation of definition question 2002 [9], definition questions from TREC 2003 [10], and other questions from TREC 2004 [11]. (We did not compare our re- sults with the collections of the TREC 2005 and 2006 since in these years the list of nuggets was not readily available). To obtain passages, first we test the output of two retrieval systems JIRS and Lucene, since the overall performance of definition ques- tion system depends on the resources and tools used for answer finding [7,8]. Figure 1 shows the comparisons of the F-measure, the best results are obtained with the general propose system (Lucene) instead of JIRS, which is intended to retrieve pas- sages for factoid questions. These results led to the decision to keep using Lucene in further experiments (instead of JIRS), since it provides an improved set of para- graphs. Fig. 2. Comparison of the F-measures obtained with balanced sets LR20 and non-balanced sets L10_R30.
We performed experiments with three sets of definition question, the questions from the pilot evaluation of definition question 2002 [9], definition questions from TREC 2003 [10], and other questions from TREC 2004 [11]. (We did not compare our re- sults with the collections of the TREC 2005 and 2006 since in these years the list of nuggets was not readily available). To obtain passages, first we test the output of two retrieval systems JIRS and Lucene, since the overall performance of definition ques- tion system depends on the resources and tools used for answer finding [7,8]. Figure 1 shows the comparisons of the F-measure, the best results are obtained with the general propose system (Lucene) instead of JIRS, which is intended to retrieve pas- sages for factoid questions. These results led to the decision to keep using Lucene in further experiments (instead of JIRS), since it provides an improved set of para- graphs. Fig. 2. Comparison of the F-measures obtained with balanced sets LR20 and non-balanced sets L10_R30.
tained the four kinds of sentence fragments detailed above, in section 2. Finally, th sentence fragments were separated in two kinds of fragments, the fragment to the /e and right of the target term. The approach of information gain is then applied to thes sentence fragments to obtain the most relevant fragments. Also, we used non balanced sets of sentence fragments, as the results of the second experiment sug gested. Figure 3 displays the F-measure obtained with our system (DefQuestion_IG compared against the systems of the pilot version of definition questions proposed ii 2002. Figure 4 shows the comparison of the F-measures of our system with the sys tems that participated in the TREC 2003. Finally, figure 5 presents the comparison o the F-measures of the systems in the TREC 2004 and our system. From the figures : to 4, we can observe that our system DefQuestions IG showed very competitiv results. Fig. 3. Comparison of F-measure values of pilot evaluation of definition questions using the AUTHOR list of nuggets.
tained the four kinds of sentence fragments detailed above, in section 2. Finally, th sentence fragments were separated in two kinds of fragments, the fragment to the /e and right of the target term. The approach of information gain is then applied to thes sentence fragments to obtain the most relevant fragments. Also, we used non balanced sets of sentence fragments, as the results of the second experiment sug gested. Figure 3 displays the F-measure obtained with our system (DefQuestion_IG compared against the systems of the pilot version of definition questions proposed ii 2002. Figure 4 shows the comparison of the F-measures of our system with the sys tems that participated in the TREC 2003. Finally, figure 5 presents the comparison o the F-measures of the systems in the TREC 2004 and our system. From the figures : to 4, we can observe that our system DefQuestions IG showed very competitiv results. Fig. 3. Comparison of F-measure values of pilot evaluation of definition questions using the AUTHOR list of nuggets.
Fig. 4. Comparison of F-measure values of TREC 2003.
Fig. 4. Comparison of F-measure values of TREC 2003.
Fig. 5. Comparison of F-measure values of TREC 2004.
Fig. 5. Comparison of F-measure values of TREC 2004.
Fig. 1. A mobile phone keypad following the ITU-T E.161 mapping of numeric keys to alphabetic symbols. Non-standard mappings exist for key 1 (symbols such as punctuation), key 0 (spaces), and the * and # keys — one of which is usually assigned to “cycle”
Fig. 1. A mobile phone keypad following the ITU-T E.161 mapping of numeric keys to alphabetic symbols. Non-standard mappings exist for key 1 (symbols such as punctuation), key 0 (spaces), and the * and # keys — one of which is usually assigned to “cycle”
Fig. 2. The training environment for a Transformation-Based Learning system
Fig. 2. The training environment for a Transformation-Based Learning system
Table 2. Average accuracy (including standard-deviation) of three predictive text systems MFF; a Bigram model with back-off; and Transformation Based Learning
Table 2. Average accuracy (including standard-deviation) of three predictive text systems MFF; a Bigram model with back-off; and Transformation Based Learning
Fig. 3. 30 most-frequent errors made by MFF, with proportion remaining after TBL We should also examine the errors made by our original classifier, and the proportion of those errors that are corrected by our system. Figure 3 shows the 30 most costly words in our chosen test set. The bars show the original error
Fig. 3. 30 most-frequent errors made by MFF, with proportion remaining after TBL We should also examine the errors made by our original classifier, and the proportion of those errors that are corrected by our system. Figure 3 shows the 30 most costly words in our chosen test set. The bars show the original error
Table 3. The top 10 transformations for a randomly chosen experiment, ordered by number of corrections in training corpus. For each transformation, the most frequent triggers are shown. Each trigger is an ordered pair of preceding words, where * represents any word training corpus — an issue we will discuss below. Finally, there are examples of ambiguity in the triggering environments: in the context (* the) “no” is an abbreviation of “number”, and “us” is an abbreviation for “United States”.
Table 3. The top 10 transformations for a randomly chosen experiment, ordered by number of corrections in training corpus. For each transformation, the most frequent triggers are shown. Each trigger is an ordered pair of preceding words, where * represents any word training corpus — an issue we will discuss below. Finally, there are examples of ambiguity in the triggering environments: in the context (* the) “no” is an abbreviation of “number”, and “us” is an abbreviation for “United States”.

Related papers

DLSITE-1: Lexical analysis for solving textual entailment recognition
O. Ferrández

2007

This paper discusses the recognition of textual entailment in a text-hypothesis pair by applying a wide variety of lexical measures. We consider that the entailment phenomenon can be tackled from three general levels: lexical, syntactic and semantic. The main goals of this research are to deal with this phenomenon from a lexical point of view, and achieve high results considering only such kind of knowledge. To accomplish this, the information provided by the lexical measures is used as a set of features for a Support Vector Machine which will decide if the entailment relation is produced. A study of the most relevant features and a comparison with the best state-of-the-art textual entailment systems is exposed throughout the paper. Finally, the system has been evaluated using the Second PASCAL Recognising Textual Entailment Challenge data and evaluation methodology, obtaining an accuracy rate of 61.88%.

View PDFchevron_right
Fourth Recognising Textual Entailment
Alicia Ageno

2012

This paper describes our experiments on Textual Entailment in the context of the Fourth Recognising Textual Entailment (RTE-4) Evaluation Challenge at TAC 2008 contest. Our system uses a Machine Learning approach with AdaBoost to deal with the RTE challenge. We perform a lexical, syntactic, and semantic analysis of the entailment pairs. From this information we compute a set of semantic-based distances between sentences. We improved our baseline system for the RTE-3 challenge with more Language Processing techniques, an hypothesis classifier, and new semantic features. The results show no general improvement with respect to the baseline.

View PDFchevron_right
A Knowledge-Based Textual Entailment Approach Applied to the AVE Task
Rafael Pérez muñoz

Lecture Notes in Computer Science, 2007

The Answer Validation Exercise (AVE) is a pilot track within the Cross-Language Evaluation Forum (CLEF) 2006. The AVE competition provides an evaluation framework for answer validations in Question Answering (QA). In our participation in AVE, we propose a system that has been initially used for other task as Recognising Textual Entailment (RTE). The aim of our participation is to evaluate the improvement our system brings to QA. Moreover, due to the fact that these two task (AVE and RTE) have the same main idea, which is to find semantic implications between two fragments of text, our system has been able to be directly applied to the AVE competition. Our system is based on the representation of the texts by means of logic forms and the computation of semantic comparison between them. This comparison is carried out using two different approaches. The first one managed by a deeper study of the Word-Net relations, and the second uses the measure defined by Lin in order to compute the semantic similarity between the logic form predicates. Moreover, we have also designed a voting strategy between our system and the MLEnt system, also presented by the University of Alicante, with the aim of obtaining a joint execution of the two systems developed at the University of Alicante. Although the results obtained have not been very high, we consider that they are quite promising and this supports the fact that there is still a lot of work on researching in any kind of textual entailment.

View PDFchevron_right
A machine learning approach to textual entailment recognition
Alessandro Moschitti, F. Zanzotto

Natural Language Engineering, 2009

Designing models for learning textual entailment recognizers from annotated examples is not an easy task, as it requires modeling the semantic relations and interactions involved between two pairs of text fragments. In this paper, we approach the problem by first introducing the class of pair feature spaces, which allow supervised machine learning algorithms to derive first-order rewrite rules from annotated examples. In particular, we propose syntactic and shallow semantic feature spaces, and compare them to standard ones. Extensive experiments demonstrate that our proposed spaces learn first-order derivations, while standard ones are not expressive enough to do so.

View PDFchevron_right
Recognizing textual entailment using a machine learning approach
Alexander Gelbukh

2010

We present our experiments on Recognizing Textual Entailment based on modeling the entailment relation as a classification problem. As features used to classify the entailment pairs we use a symmetric similarity measure and a non-symmetric similarity measure. Our system achieved an accuracy of 66% on the RTE-3 development dataset (with 10-fold cross validation) and accuracy of 63% on the RTE-3 test dataset.

View PDFchevron_right
Acquiring Data for Textual Entailment Recognition
Zuzana Nevěřilová

Language resources are hardly ever large enough. Building language resources that can be used as a gold standard for semantic analysis requires effort and investment. We present a prototype for acquiring language resources by means of a language game which is a cheap but long-term method. Games employed to acquire language resources are not new. For example games with a purpose are used for collecting common sense knowledge. The game presented in this paper is a work in progress. It collects annotated pairs text-hypothesis suitable for recognizing textual entailment in Czech. The game narrative is based on Sherlock Holmes and dr. Watson dialogues. For generating the dialogue line we use rule-based approaches such as syntactic analysis, anaphora resolution, synonym and hypernym replacement, word order rearrangement and verb frame based inference. To generate natural sounding sentences we added a language model score (based on n-gram frequencies in a corpus).

View PDFchevron_right
Recognizing Textual Entailment Using Lexical, Syntactical, and Semantic Information
Lun-Wei Ku

2014

paper describes our system for participating in the system validation subtask of NTCIR-11 RITE-VAL. We trained a SVM model with LibSVM using features extracted from labeled sentence pairs. Besides features based on lexical, syntactic and semantic analysis, we introduce a novel approach of extracting &quot;concepts&quot; from a sentence and generating features based on it. Unlabeled testing sentence pairs&#39; features are extracted through the same process, and the SVM model that we have trained predicts their labels.

View PDFchevron_right
A semantic approach to textual entailment: System evaluation and task analysis
Anette Frank

Proceedings of the ACL- …, 2007

Recognizing and generating textual entailment and paraphrases are regarded as important technologies in a broad range of NLP applications, including, information extraction, summarization, question answering, information retrieval, machine translation and text generation. Both textual entailment and paraphrasing address relevant aspects of natural language semantics. Entailment is a directional relation between two expressions in which one of them implies the other, whereas paraphrase is a relation in which two expressions convey essentially the same meaning. Indeed, paraphrase can be defined as bi-directional entailment. While it may be debatable how such semantic definitions can be made well-founded, in practice we have already seen evidence that such knowledge is essential for many applications.

View PDFchevron_right
Different Models and Approaches of Textual Entailment Recognition
Marwa M. A. Elfattah

International Journal of Computer Applications, 2016

Variability of semantic expression is a fundamental phenomenon of a natural language where same meaning can be expressed by different texts. The process of inferring a text from another is called textual entailment. Textual Entailment is useful in a wide range of applications, including question answering, summarization, text generation, and machine translation. The recognition of textual entailment is one of the recent challenges of the Natural Language Processing (NLP) domain. This paper summarizes key ideas from the area of textual entailment recognition by considering in turn the different recognition models. The paper points to prominent testing data, training data, resources and Performance Evaluation for each model. Also this paper compares between textual entailment models according to the method which used, the result of each method and the strong and weakness of each method.

View PDFchevron_right
Design and realization of a modular architecture for textual entailment
Sebastian Padó

2013

A key challenge at the core of many NLP tasks is the ability to determine which conclusions can be inferred from a given natural language text. This problem, called the Recognition of Textual Entailment (RTE), has initiated the development of a range of algorithms, methods and technologies. Unfortunately, research on TE (like semantics research more generally), is fragmented into studies focussing on various aspects of semantics such as world knowledge, lexical and syntactic relations, or more specialized kinds of inference. This fragmentation has problematic practical consequences. Notably, interoperability among existing RTE systems is poor, and reuse of resources and algorithms is mostly infeasible. This also makes systematic evaluations very difficult to carry out. Finally, TE presents a wide array of approaches to potential end users with little guidance on which to pick. Our contribution to this situation is the novel EXCITEMENT architecture, which was developed to enable and encourage the consolidation of methods and resources in the TE area. It decomposes RTE into components with strongly typed interfaces. We specify (a) a modular linguistic analysis pipeline and (b) a decomposition of the "core" RTE methods

View PDFchevron_right

References (284)

  1. Bar-Haim, R., Dagan, I., Dolan, B., Ferro, L., Giampiccolo, D., Magnini B., Szpektor, I.: The Second PASCAL Recognising Textual Entailment Challenge. In Proceedings of the Second PASCAL Challenges Workshop on Recognizing Textual Entailment. Venice. Italy. (2006)
  2. Cunningham, H., Maynard, D., Bontcheva, K., Tablan, V.: GATE: A Framework and Graphical Development Environment for Robust NLP Tools and Applications. Proceedings of the 40th Anniversary Meeting of the Association for Computational Linguistics (ACL'02). Philadelphia, July (2002)
  3. Dagan I., Glickman O., Magnini, B.: The PASCAL Recognising Textual Entailment Challenge. In Quiñonero-Candela et al., editors, MLCW 2005, LNAI Volume 3944. Springer-Verlag (2006) 177-190
  4. Fellbaum, C.: WordNet: An Electronic Lexical Database. MIT Press, Cambridge, Mass. (1998)
  5. Platt, J.: Fast Training of Support Vector Machines using Sequential Minimal Optimization. Advances in Kernel Methods -Support Vector Learning, B. Schoelkopf, C. Burges, and A. Smola, eds., MIT Press (1998)
  6. Quinlan, J. R.: C4.5: Programs for machine learning. San Francisco: Morgan Kaufmann (1993) 1
  7. Seg.9 25 -29 Seg.9 21 -22 Seg.9 29 -32 Seg.9 14 -16
  8. References
  9. C. Corley, R. Mihalcea: "Measuring the semantic similarity of texts", Proceed- ings of the ACL Workshop on Empirical Modeling of Semantic Equivalence and Entailment, Ann Arbor, June, 2005, pages 13-18.
  10. I. Dagan, O. Glickman and B. Magnini: The PASCAL Recognising Textual En- tailment Challenge, Proceedings of the PASCAL Work-shop, 2005.
  11. M. Hearst: "TextTiling: A Quantitative Approach to Discourse Segmentation", Technical Report 93/24, University of California, Berkeley, 1993. 1997
  12. R. Mitkov (editor): "The Oxford Handbook of Computational Linguistics", Ox- ford University Press, 2003.
  13. Lewis, D. D.; Yang, Y.; Rose, T.; and Li, F. " RCV1: A New Benchmark Collec- tion for Text Categorization Research", Journal of Machine Learning Research, 5:361-397, 2004.
  14. D. Marcu: "From discourse structure to text summaries" in Proceedings of the ACL/EACL '97 Workshop on Intelligent Scalable Text Summarization, pp 82-88, Madrid,Spain.
  15. C. Orasan: "Comparative evaluation of modular automatic summarization sys- tems using CAST", PhD Thesis, University of Wolverhampton, UK, 2006.
  16. D. Radev, E. Hovy, K. McKeown : "Introduction to the Special Issues on Sum- marization", Computational Linguistics, Vol.28, 2002, pp 399-408.
  17. Donald Richie: "The Koan", in Zen Inklings. Some Stories, Fables, Parables and Sermons, New York and Tokyo: Weatherhill, 1991, pp.25-27
  18. D. Tatar, G. Serban, A. Mihis and R. Mihalcea: "Text Entailment as direc- tional relation", Proceedings of CALP07 Workshop at RANLP2007, pp 53-58, Borovets, Bulgaria, September 21-23
  19. D. Tȃtar, M. Frent ¸iu: "Textual inference by theorem proving and linguistic ap- proach", Studia Universitatis "Babeş-Bolyai", Seria Informatics, 2006, nr 2, pages 31-41.
  20. Tatar D., Tamaianu-Morita E., Mihis A. and Lupsa D. References
  21. Geach, P.: Reference and Generality. Ithaca, NY: Cornell University Press (1962)
  22. Kamp, H. and Reyle, U.: From Discourse to Logic. Introduction to Model-Theoretic Semantics of Natural Language, Formal Logic and Discourse Representation Theory. Springer (1993)
  23. Jespersen, B.: 'Predication and extensionalization'. Journal of Philosophical Logic, January 2008; retrievable at: http://www.springerlink.com/content/q31731r311428343/?p=60281b936ca249278691b0f 25135cc0a&pi=2
  24. Křetínský, J.: Use-mention Distinction in Transparent Intensional Logic. Bachelor thesis, Masaryk University Brno (2007); retrievable at: http://is.muni.cz/th/139914/fi_b/bachelor.pdf
  25. Materna, P.: Conceptual Systems. Logos Verlag, Berlin (2004)
  26. Neale, S.: Descriptions. The MIT Press, Cambridge (1990)
  27. Rescher, N.: Epistemic Logic. Pittsburgh: University of Pittsburgh Press (2005)
  28. Sandu, G.: 'On the theory of anaphora: dynamic predicate logic vs. game-theoretical semantics'. Linguistic and Philosophy 20 (1997), 147-174
  29. Sandu, G., Hintikka, J.: 'Aspects of compositionality'. Journal of Logic, Language, Information 10 (2001) 49-61
  30. Tichý, P.: The Foundations of Frege's Logic, Berlin, New York: De Gruyter (1988)
  31. Tichý, P.: Collected Papers in Logic and Philosophy, V. Svoboda, B. Jespersen, C. Cheyne (eds.), Prague: Filosofia, Czech Academy of Sciences, and Dunedin: University of Otago Press (2004) References
  32. Gildea, D.: Corpus variation and parser performance. In: EMNLP '01. (2001), 167-172
  33. Bikel, D.M.: Intricacies of collins' parsing model. Computational Linguististics 30(4) (2004) 479-511
  34. Klein, D., Manning, C.D.: Accurate unlexicalized parsing. In: ACL '03. (2003) 423-430
  35. Johnson, M.: PCFG models of linguistic tree representations. Computational Linguistics 24(4) (1998) 613-632
  36. Matsuzaki, T., Miyao, Y., Tsujii, J.: Probabilistic CFG with latent annotations. In: ACL '05. (2005) 75-82
  37. Petrov, S., Barrett, L., Thibaus, R., Klein, D.: Learning accurate, compact, and interpret- able tree annotation. In: COLING-ACL '06. (2006) 433-440
  38. Charniak, E., Johnson, M.: Coarse-to-fine n-best parsing and maxent discriminative reranking. In: ACL 2005. (2005) 173-180
  39. Collins, M.: Head-driven statistical models for natural language parsing. Computational Linguististics 29(4) (2003) 589-637
  40. Prescher, D.: Head-driven PCFGs with latent-head statistics. In: IWPT '05. (2005) 115- 124
  41. Charniak, E.: Statistical parsing with a context-free grammar and word statistics. In: AAAI/IAAI 1997. (2005) 598-603
  42. Seung-Hoon N., Meixun J., In-Su K. and Jong-Hyeok L. References
  43. Daniel M. Bikel and David Chiang. 2000. Two Statistical Parsing Models Applied to Chinese Treebank. In Proceedings of the 2nd Chinese language processing workshop.
  44. Daniel M. Bikel. 2004(a). Intricacies of Collins' Parsing Model. In Computational Lin- guistics, 30(4): 479-511.
  45. Daniel M. Bikel. 2004(b). On the Parameter Space of Generative Lexicalized Statistical Parsing Models. Ph.D. thesis, University of Pennsylvania.
  46. David Chiang and Daniel Bikel. 2002. Recovering Latent Information in Treebanks. In Proceedings of the 19th International Conference on Computational Linguistics.
  47. Deyi Xiong, Shuanglong Li, Qun Liu et al. 2005. Parsing the Penn Chinese Treebank with Semantic Knowledge. In Proceedings of the Second International Joint Conference Natu- ral language processing.
  48. Daniel Jurafsky, James H. Martin. 2000. Speech and Language Processing: An Introduc- tion to Natural Language Processing, Computational Linguistics and Speech Recognition. Prentice-Hall.
  49. Eugene Charniak, Glenn Carroll, John Adcock et al. 1996. Taggers for Parsers. Artificial Intelligence, 85(1-2):45-57.
  50. Fei Xia. 1999. Automatic Grammar Generation from Two Different Perspectives. PhD thesis, University of Pennsylvania.
  51. Kazuhiro Yoshida, Yoshimasa Tsuruoka, Yusuke Miyao, et al. 2007. Ambiguous Part-of- Speech Tagging for Improving Accuracy and Domain Portability of Syntactic Parsers. In Proceedings of the Twentieth International Joint Conference on Artificial Intelligence.
  52. Mary Hearne and Andy Way. 2004. Data-Oriented Parsing and the Penn Chinese Tree- bank. In Proceedings of the First International Joint Conference Natural language process- ing.
  53. Cao H., Zhang Y. and Isahara H.
  54. Mengqiu Wang, Kenji Sagae and Teruko Mitamura. 2006. A Fast, Accurate Deterministic Parser for Chinese. In Proceedings of COLING/ACL.
  55. Michael Collins. 1999. Head-driven Statistical Models for Natural Language Parsing. Ph.D. thesis, University of Pennsylvania.
  56. Nianwen Xue, Fei Xia, Fu-Dong Chiou et al. 2005. The Penn Chinese Treebank: Phrase Structure Annotation of a Large Corpus. Natural Language Engineering, 11(2):207-238.
  57. Pascale Fung, Grace Ngai, Yongsheng Yang et al. 2004. A Maximum-Entropy Chinese Parser Augmented by Transformation-Based Learning. ACM Transactions on Asian Lan- guage Processing, 3(2):159-168.
  58. Rebecca Watson. 2006. Part-of-speech Tagging Models for Parsing. In Proc. of CLUK.
  59. Roger Levy and Christopher Manning. 2003. Is it Harder to Parse Chinese, or the Chinese Treebank? In Proceedings of ACL.
  60. Xiaoqiang Luo. 2003. A Maximum Entropy Chinese Character-Based Parser. In Proceed- ings of the conference on Empirical methods in Natural Language Processing.
  61. Zhengping Jiang. 2004. Statistical Chinese parsing. Honours thesis, National University of Singapore. Study on Architectures for Chinese POS Tagging and Parsing 65
  62. Cutting, D., Kupiec, J., Pedersen, J., Sibun, P.: A practical part-of-speech tagger. In: Proceedings of the Third Conference on Applied Natural Language Processing. (1992) 133-140
  63. Merialdo, B.: Tagging english text with a probabilistic model. Comput. Linguist. 20(2) (1994) 155-171
  64. Brants, T.: TnT a Statistical Parts-of-Speech Tagger. In: Proceedings of the Sixth Conference on Applied Natural Language Processing ANLP-2000. (2000) 224-231
  65. Ratnaparkhi, A.: A maximum entropy part-of -speech tagger. In: Proc. of EMNLP'96. (1996)
  66. Lafferty, J.D., McCallum, A., Pereira, F.C.N.: Conditional Random Fields: Prob- abilistic Models for Segmenting and Labeling Sequence Data. In: ICML. (2001) 282-289
  67. Singh, S., Gupta, K., Shrivastava, M., Bhattacharyya, P.: Morphological richness offsets resource demand-experiences in constructing a pos tagger for hindi. In: Proceedings of the COLING/ACL 2006. (2006) 779-786
  68. Avinesh, P., Karthik, G.: Part Of Speech Tagging and Chunking using Conditional Random Fields and Transformation Based Learning. In: Proceedings of IJCAI Workshop on Shallow Parsing for South Asian Languages, India (2007) 21-24
  69. Dandapat, S.: Part Of Specch Tagging and Chunking with Maximum Entropy Model. In: Proceedings of the IJCAI Workshop on Shallow Parsing for South Asian Languages, Hyderabad, India (2007) 29-32
  70. Ekbal, A., Naskar, S., Bandyopadhyay, S.: Named Entity Recognition and Translit- eration in Bengali. Named Entities: Recognition, Classification and Use, Special Issue of Lingvisticae Investigationes Journal 30(1) (2007) 95-114
  71. Ekbal, A., Bandyopadhyay, S.: Lexicon Development and POS Tagging using a Tagged Bengali News Corpus. In: Proceedings of the 20th International Florida AI Research Society Conference (FLAIRS-2007), Florida (2007) 261-263
  72. Malouf, R.: A comparison of algorithms for maximum entropy parameter esti- mation. In: Proceedings of Sixth Conf. on Natural Language Learning. (2002) 49-55
  73. Ekbal, A., Bandyopadhyay, S.: A Web-based Bengali News Corpus for Named Entity Recognition. Language Resources and Evaluation Journal (accepted) (2007)
  74. Ekbal, A., Mondal, S., Bandyopadhyay, S.: POS Tagging using HMM and Rule- based Chunking. In: Proceedings of the IJCAI Workshop on Shallow Parsing for South Asian Languages, Hyderabad, India (2007) 31-34 Baseline 65.80( 46780 71097 ) 66.56( 46780
  75. Results of word segmentation and POS tagging Algorithm Recall Precision F1 Recall known Recall unknown Baseline 57.58( 40940 71097 ) 58.25( 40940
  76. Aroonmanakun, W.: Collocation and Thai Word Segmentation. Proc. of the 5th SNLP & 5th Oriental COCOSDA Workshop. (2002) 68-75
  77. Charoenpornsawat, P.: Feature-based Thai Word Segmentation. Master's The- sis, Computer Engineering, Chulalongkorn University. (1999)
  78. Jaruskulchai, C.: An Automatic Thai Lexical Acquisition from Text. Proc. of PRICAI. (1998) 289-296
  79. Jurafsky, D., Martin, J. H.: Speech and Language Processing: An Introduction to Natural Language Processing, Computational Prentice-Hall, Inc. (2000)
  80. Kruengkrai, C., Charoenporn, T., Sornlertlamvanich, V., Isahara, H.: Acquiring Selectional Preferences in a Thai Lexical Database. Proc. of IJCNLP. (2004)
  81. Kruengkrai, C., Sornlertlamvanich, V., Isahara, H.: A Conditional Random Field Framework for Thai Morphological Analysis. Proceedings of the Fifth International Conference on Language Resources and Evaluation. (2006)
  82. Kudo, T., Yamamoto, K., and Matsumoto, Y.: Applying Conditional Random Fields to Japanese Morphological Analysis. Proc. of EMNLP. (2004)
  83. Liu, D. C., Nocedal, J.: On the limited memory BFGS method for large scale optimization. Math. Programming. (1989) 45:503-528
  84. Sornlertlamvanich, V., Charoenporn, T., Isahara, H.: ORCHID: Thai Part-Of- Speech Tagged Corpus. Technical Report TR-NECTEC-1997-001, NECTEC. (1997)
  85. Uchimoto, K., Nobata, C., Yamada, A., Sekine, S., Isahara, H.: Morphological Analysis of a Large Spontaneous Speech Corpus in Japanese. Proc. of ACL. (2003) A Two-Pass Search Algorithm for Thai Morphological Analysis 91
  86. Xue, N.: Chinese word segmentation as character tagging. Computational Linguistics and Chinese Language Processing 8 (2003) 29-48
  87. Sproat, R., Emerson, T.: The first international Chinese word segmentation bakeoff. In: The Second SIGHAN Workshop on Chinese Language Processing, Sapporo, Japan (2003) 133-143
  88. Peng, F., Feng, F., McCallum, A.: Chinese segmentation and new word detection using conditional random fields. In: COLING 2004, Geneva, Switzerland (2004) 562-568
  89. Low, J.K., Ng, H.T., Guo, W.: A maximum entropy approach to Chinese word segmentation. In: Proceedings of the Fourth SIGHAN Workshop on Chinese Language Processing, Jeju Island, Korea (2005) 161-164
  90. Tseng, H., Chang, P., Andrew, G., Jurafsky, D., Manning, C.: A conditional random field word segmenter for SIGHAN bakeoff 2005. In: Proceedings of the Fourth SIGHAN Work- shop on Chinese Language Processing, Jeju Island, Korea (2005) 168-171
  91. Zhao, H., Huang, C.N., Li, M.: An improved Chinese word segmentation system with condi- tional random field. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 162-165
  92. Tsai, R.T.H., Hung, H.C., Sung, C.L., Dai, H.J., Hsu, W.L.: On closed task of Chinese word segmentation: An improved CRF model coupled with character clustering and automatically generated template matching. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 108-117
  93. Zhu, M.H., Wang, Y.L., Wang, Z.X., Wang, H.Z., Zhu, J.B.: Designing special post- processing rules for SVM-based Chinese word segmentation. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 217-220
  94. Lafferty, J.D., McCallum, A., Pereira, F.C.N.: Conditional random fields: Probabilistic mod- els for segmenting and labeling sequence data. In: ICML '01: Proceedings of the Eighteenth International Conference on Machine Learning, San Francisco, CA, USA, Morgan Kauf- mann Publishers Inc. (2001) 282-289
  95. Rosenfeld, B., Feldman, R., Fresko, M.: A systematic cross-comparison of sequence classi- fiers. In: SDM 2006, Bethesda, Maryland (2006) 563-567
  96. Zhao, H., Huang, C.N., Li, M., Lu, B.L.: Effective tag set selection in Chinese word seg- mentation via conditional random field modeling. In: Proceedings of the 20th Asian Pacific Conference on Language, Information and Computation, Wuhan, China (2006) 87-94
  97. Lü, X., Zhang, L., Hu, J.: Statistical substring reduction in linear time. In et al., K.Y.S., ed.: Proceeding of the 1st International Joint Conference on Natural Language Processing (IJCNLP-2004). Volume 3248 of Lecture Notes in Computer Science., Sanya City, Hainan Island, China, Springer (2004) 320-327
  98. Kit, C., Wilks, Y.: Unsupervised learning of word boundary with description length gain. In Osborne, M., Sang, E.T.K., eds.: CoNLL-99, Bergen, Norway (1999) 1-6
  99. Shannon, C.E.: A mathematical theory of communication. The Bell System Technical Jour- nal 27 (1948) 379-423, 623-656
  100. Cover, T.M., Thomas, J.A.: Elements of Information Theory. John Wiley and Sons, Inc., New York (1991)
  101. Feng, H., Chen, K., Deng, X., Zheng, W.: Accessor variety criteria for Chinese word extrac- tion. Computational Linguistics 30 (2004) 75-93
  102. Tung, C.H., Lee, H.J.: Identification of unknown words from corpus. Computational Pro- ceedings of Chinese and Oriental Languages 8 (1994) 131-145
  103. Chang, J.S., Su, K.Y.: An unsupervised iterative method for Chinese new lexicon extraction. Computational Linguistics and Chinese Language Processing 2 (1997) 97-148
  104. Huang, J.H., Powers, D.: Chinese word segmentation based on contextual entropy. In Ji, D.H., Lua, K.T., eds.: Proceedings of the 17th Asian Pacific Conference on Language, Infor- mation and Computation, Sentosa, Singapore, COLIPS Publication (2003) 152-158
  105. Jin, Z., Tanaka-Ishii, K.: Unsupervised segmentation of Chinese text by use of branching entropy. In: COLING/ACL 2006, Sidney, Australia (2006) 428-435
  106. Harris, Z.S.: Morpheme boundaries within words. In: Papers in Structural and Transforma- tional Linguistics. (1970) 68 77
  107. Levow, G.A.: The third international Chinese language processing bakeoff: Word segmen- tation and named entity recognition. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 108-117
  108. Zhao, H., Kit, C.: Unsupervised segmentation helps supervised learning of character tagging for word segmentation and named entity recognition. In: The Sixth SIGHAN Workshop on Chinese Language Processing, Hyderabad, India (2008) Exploiting Unlabeled Text with Different Unsupervised Segmentation Criteria... 103
  109. Zhao, H., Kit, C.: An empirical comparison of goodness measures for unsupervised Chinese word segmentation with a unified framework. In: The Third International Joint Conference on Natural Language Processing (IJCNLP-2008), Hyderabad, India (2008)
  110. Carpenter, B.: Character language models for Chinese word segmentation and named entity recognition. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Process- ing, Sydney, Australia, Association for Computational Linguistics (2006) 169-172
  111. Wang, X., Lin, X., Yu, D., Tian, H., Wu, X.: Chinese word segmentation with maximum entropy and N-gram language model. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 138-141
  112. Zhang, M., Zhou, G.D., Yang, L.P., Ji, D.H.: Chinese word segmentation and named en- tity recognition based on a context-dependent mutual information independence model. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Aus- tralia (2006) 154-157
  113. Grinstead, C., Snell, J.L.: Introduction to Probability. American Mathematical Society, Providence, RI (1997)
  114. Jacobs, A.J., Wong, Y.W.: Maximum entropy word segmentation of Chinese text. In: Pro- ceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Aus- tralia (2006) 108-117
  115. Liu, W., Li, H., Dong, Y., He, N., Luo, H., Wang, H.: France Telecom R&D Beijing word segmenter for SIGHAN bakeoff 2006. In: Proceedings of the Fifth SIGHAN Workshop on Chinese Language Processing, Sydney, Australia (2006) 108-117
  116. Jiao, F., Wang, S., Lee, C.H., Greiner, R., Schuurmans, D.: Semi-supervised conditional random fields for improved sequence segmentation and labeling. In: Proc. of COLING/ACL- 2006, Sydney, Australia (2006) 209 216
  117. Suzuki, J., Fujino, A., Isozaki, H.: Semi-supervised structured output learning based on a hybrid generative and discriminative approach. In: Proceedings of the 45rd Annual Meeting of the Association for Computational Linguistics (ACL'07), Prague, Czech, Association for Computational Linguistics (2007) 791
  118. References
  119. Aha, D., Kibler, D. Instance-based learning algorithms. Machine Learning. 6:37-66. (1991)
  120. Breiman, L. (1996). Bagging Predictors. Machine Learning, 24/2, 123-140.
  121. Breiman, L,, Friedman, J., Olshen, R., and Stone, C. (1983). CART: Classification and Regression Trees. Wadsworth, NY.
  122. Forman, G., and Cohen, I. Learning from Little: Comparison of Classifiers Given Little Training. In 15th European Conference on Machine Learning and the 8th European Conference on Principles and Practice of Knowledge Discovery in Databases (2004)
  123. Freund, Y. & R.E. Schapire (1996). Experiments with a New Boosting Algorithm, in Proceedings of the Thirteenth International Conference on Machine Learning.
  124. Legrand, S., Pulido JGR. A Hybrid Approach to Word Sense Disambiguation: Neural Clustering with Class Labeling. Knowledge Discovery and Ontologies workshop at 15th European Conference on Machine Learning (ECML) (2004).
  125. Mierswa, I., Wurst, M., Klinkenberg, R., Scholz, M. and Euler, T. YALE: Rapid Prototyping for Complex Data Mining Tasks. Proceedings of 12th ACM SIGKDD (2006)
  126. Quinlan, R. C4.5: Programs for Machine Learning. Morgan Kaufmann Publishers. (1993)
  127. Saarikoski, H., Legrand, S., Gelbukh, A. 2006. Case-Sensitivity of Classifiers for WSD: Complex Systems Disambiguate Tough Words Better. In proceedings of CicLING 2006. 10. 10.Vapnik, V. N. The Nature of Statistical Learning Theory. Springer (1995) 11.
  128. Witten, I., Frank, E. Data Mining: Practical Machine Learning Tools and Techniques (Second Edition). Morgan Kaufmann (2005).
  129. Yarowsky, D. and Florian, R. Evaluating sense disambiguation across diverse parameter spaces. Journal of Natural Language Engineering, 8(4) (2002) 293-311.
  130. Comrie, B. Language Universals and Linguistic Typology. Basil Blackwell, Oxford. Second edition. 1989.
  131. Nirenburg, S., Carbonell, J., Tomita, M., and Goodman, K. Machine Translation: A Knowledge-based Approach. Morgan Kaufmann. 1992.
  132. Pollard, C. and Sag, I.A. Head-Driven Phrase Structure Grammar. University of Chicago Press, Chicago, 1994.
  133. Mustajoki, A. Functional Syntax as the Basis for Contrastive Languae Study. Studia Slavica Finlandesia. Finnish Contributions to the 13th International Congress of Slavists, Ljubljana, August, 15-21, 2003, pp.100-127 (In Russian).
  134. Briscoe, E.J. "An introduction to tag sequence grammars and the RASP system parser". Technical Report N662, University of Cambridge, March, 2006.
  135. Shaumyan, S. A Semiotic Theory of Language. Indiana University Press, 1987.
  136. Bondarko A.V. Printsipy Funktsional'noi Grammatiki I Voprosy Aspektologhii. Moskwa, URSS, 2001 /Functional Grammar Principles and Aspectology Questions. Moscow, URSS, 2001 (In Russian).
  137. Kozerenko, E.B. Cognitive Approach to Language Structure Segmentation for Machine Translation Algorithms // Proceedings of the International Conference on Machine Learning, Models, Technologies and Applications, June, 23-26, 2003, Las Vegas, USA.// CSREA Press, pp. 49-55, 2003.
  138. Shaumyan, S. Categorial Grammar and Semiotic Universal Grammar. In Proceedings of The International Conference on Artificial Intelligence, IC-AI'03, Las Vegas, Nevada, CSREA Press, 2003.
  139. Kozerenko, E.B., Shaumyan, S. Discourse Projections of Semiotic Universal Grammar // Proceedings of the International Conference on Machine Learning, Models, Technologies and Applications, June, 27-30, 2005, Las Vegas, USA.// CSREA Press, pp. 3-9, 2005.
  140. Shaumyan, S. Intrinsic Anomalies in Meaning and Sound and Their Implications for Linguistic Theory and Techniques of Linguistic Analysis. // Proceedings of the International Conference on Machine Learning, Models, Technologies and Applications, June, 27-30, 2005, Las Vegas, USA.// CSREA Press, pp. 10-17, 2005.
  141. Chater, N. Reconciling simplicity and likelihood principles in perceptual organization Psychological Review 103 (3), pp. 566-581, 1996.
  142. Chater, N. The search for simplicity: a fundamental cognitive principle? The Quarterly Journal of Experimental Psychology. 52 (2), 273-302, 1999.
  143. Johnson, M. PCFG Models of Linguistic Tree Representations. Computational Linguistics, 24(4), pp. 613-632, 1998.
  144. Osborne, M. and Briscoe, T. Learning Stochastic Categorial Grammars. In T.M. Ellison, editor, Proceedings of CoNLL97: Computational Natural Language Learning, pp. 80-87, Somerset, NJ, 1998.
  145. Hockenmaier, Julia. Data and Models for Statistical Parsing with Combinatory Categorial Grammar. Ph.D. thesis, University of Edinburgh , 2003.
  146. Brown, P.F., J. Cocke, S.A. Della Pietra, V.J. Della Pietra, F. Jelinek, J.D. Lafferty, R.L. Mercer & P.S. Roossin. A statistical approach to machine translation. Computational Linguistics 16, pp. 79-85, 1990.
  147. Habash, Nizar and Bonnie Dorr. Handling Translation Divergences: Combining Statistical and Symbolic Techniques in Generation-Heavy Machine Translation. AMTA-2002.
  148. Dorr, Bonnie and Nizar Habash. Interlingua Approximation: A Generation-Heavy Approach. AMTA-2002 Interlingua Reliability Workshop. Tiburon, California, USA, 2002.
  149. Apresjan, Jurij, Igor Boguslavsky, Leonid Iomdin, Alexandre Lazursku, Vladimir Sannikov, and Leonid Tsinman. "ETAP-2: the linguistics of a machine translation system".META, 37(1):97-112, 1992.
  150. Boguslavsky, I., Chardin, I., Grigorieva, S., Grigoriev, N., Iomdin, L., Kreidlin, L. and Frid, N. "Development of a dependency treebank for Russian and its possible applications in NLP". In Proceedings of the 3rd International Conference on Language Resources and Evaluation, Las Palmas, Gran Canaria, Spain, 2002.
  151. Kakkonen T. "Dependency treebanks: methods, annotation schemes and tools". Proceedings of the 15th NODALIDA conference, Joensuu, 94-104, 2005.
  152. Copestake, A., Lascarides, A., & Flickinger, D. "An algebra for semantic construction in constraint-based grammars". In Proceedings of the 39th Meeting of the Association for ComputationalLinguistics. Toulouse, France, 2001.
  153. Flickinger, D. "On building a more efficient grammar by exploiting types". Natural Language Engineering, 6 (1) (Special Issue on Efficient Processing with HPSG), 15 -28, 2000.
  154. Marcus, M. P., Santorini, B., & Marcinkiewicz, M. A. "Building a large annotated corpus of English. The Penn Treebank". Computational Linguistics, 19, 313 -330, 1993.
  155. Oepen, S., Callahan, E., Flickinger, D., & Manning, C. D. "LinGO Redwoods. A rich and dynamic treebank for HPSG". In LREC workshop on parser evaluation. Las Palmas, Spain, 2002.
  156. McCarthy, D., and Carroll J. "Disambiguating nouns, verbs, and adjectives using automatically acquired selectional preferences" Computational Linguistics, vol. 29.4, 639- 654, 2003.
  157. Watson, R., J. Carroll and E. Briscoe "Efficient extraction of grammatical relations", Proceedings of the 9-th International Workshop on Parsing Technologies (IWPT'05), Vancouver, Ca., 2005. Features and Categories Design for the English-Russian Transfer Model 137
  158. Otto Jespersen. A Modern English Grammar on Historical Principles.London: George Allen & Unwin and Copenhagen: Ejnar Munksgaard(1954)
  159. R. Cattell. Composite Predicates in English. Academic Press, New York(1984)
  160. Butt, M. and Geuder, W.. On the (semi)lexical Status of Light Verbs. In Corver, N. and van Riemsdijk, H., editors, Semi-lexical Categories: On the content of function words and the function of content words. Mouton de Gruyter, Berlin. (2001) 323-370
  161. Suzanne Stevenson and Afsaneh Fazly and Ryan North Statistical Measures of the Semi-Productivity of Light Verb Constructions Second ACL Workshop on Multiword Expressions: Integrating Processing ( 2004) 1-8
  162. Gholamhossein Karimi-Doostan. Light verbs and Structural Case. Lingua 115 (2005) 1737-1756
  163. Miriam Butt. The Light Verb Jungle. http://ling.uni-konstanz.de/pages/home/butt/harvard-work.pdf (2003).
  164. Jane Grimshaw and Armin Mester. Light Verbs and 1 -Marking. Linguistic Inquiry. 19 (1988) 205-232,
  165. Atsuko Shimada and Valia Kordoni. Japanese "Verbal Noun and Suru" Constructions. Proceedings of the workshop on Multi-Verb constructions Trondheim Summer School 2003
  166. Atsushi Fujita Kentaro Furihata Kentaro Inui Yuji Matsumoto Koichi Takeuchi Paraphrasing of Japanese Light-verb Constructions Based on Lexical Conceptual Structure Second ACL Workshopon Multiword Expressions:Integrating Processing July (2004)
  167. Yo Matsumoto. A Syntactic Account of Light Verb Phenomena in Japanese. Journal of East Asian Linguistics 5 (1996) 107-149.
  168. Tadao Miyamoto. The Light Verb Construction in Japanese: the role of the verbal noun. John Benjamins Publishing Company. (2000)
  169. Yokota, K. The structure and meaning of Japanese light verbs. Language sciences 27,3 (2005) 247-280
  170. NTT Communication Science Basic Research. Goi-Taikei--A Japanese Lexicon. Iwanami Publisher. (1997)
  171. Malik, M. G. A.: Punjabi Machine Transliteration. In Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL (2006) 1137-1144.
  172. Afzal, M., Hussain S.: Urdu Computing Standards: Urdu Zabta Takhti (UZT) 1.01. In proceedings of the IEEE INMIC, Lahore (2001).
  173. Haizhou, L., Min, Z., and Jian S.: A Joint Source-Channel Model for Machine Transliteration. In Proceedings of the 42nd Annual Meeting of the Association for Computational Linguistics (2004) 159-166.
  174. Youngim, J., Donghun, L., Aesun, Y., Hyuk-Chul, K.: Transliteration System for Arabic-Numeral Expressions using Decision Tree for Intelligent Korean TTS, Vol. 1. 30th Annual Conference of IEEE (2004) 657-662.
  175. Nasreen Abdululjaleel, leah S. Larkey: Statistical Transliteration for English- Arabic Cross Language Information Retrieval. Proceedings of the 12th international conference on information and knowledge management (2003) 139- 146.
  176. Yan, Q., Gregory, G., David A. Evans: Automatic Transliteration for Japanese-to- English Text Retrieval. In Proceedings of the 26th annual international ACM SIGIR conference on Research and development in information retrieval (2003) 353-360.
  177. Arbabi, M., Fischthal, S. M., Cheng, V. C., and Bart E.: Algorithms for Arabic Name Transliteration. IBM Journal of research and Development (1994) 183-193.
  178. Knight, K., and Graehl, J.: Machine Transliteration. In Proceedings of the 35th Annual Meeting of the Association for Computational Linguistics (1997) 128-135.
  179. Stalls, B. G. and Kevin K.: Translating Names and Technical Terms in Arabic Text. COLING ACL Workshop on Computational Approaches to Semitic Languages (1998) 34-41.
  180. Papineni, K., Roukos, S., Ward, T., and Zhu, W. 2001. BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting on Association For Computational Linguistics (Philadelphia, Pennsylvania, July 07 -12, 2002). Annual Meeting of the ACL. Association for Computational Linguistics, Morristown, NJ, 311-318.
  181. Chu-Carroll, J, and R. Carpenter (1999), Vector-Based Natural Language Call Routing. Journal of Computational Linguistics, 25(30), pp. 361-388, 1999
  182. H.-K.J. Kuo,; Chin-Hui Lee Discriminative training of natural language call routers Speech and Audio Processing, IEEE Transactions on Volume 11, Issue 1, Jan 2003 Page(s): 24 -35.
  183. Thomas Landauer, P. W. Foltz, & D. Laham (1998). "Introduction to Latent Semantic Analysis". Discourse Processes 25: 259-284.
  184. Berger, A. L., Pietra, V. J., and Pietra, S. A. 1996. A maximum entropy approach to natural language processing. Comput. Linguist. 22, 1 (Mar. 1996), 39-71.
  185. J.M. Schultz and M. Liberman, "Topic Detection and Tracking using idfWeighted Cosine Coefficient," Proceedings of the DARPA Broadcast News Workshop, 189-192, 1999.
  186. Dasgupta, A., Kumar, R., Raghavan, P., and Tomkins, A. 2005. Variable latent semantic indexing. In Proceeding of the Eleventh ACM SIGKDD international Conference on Knowledge Discovery in Data Mining (Chicago, Illinois, USA, August 21 -24, 2005). KDD '05. ACM Press, New York, NY, 13-21.
  187. Gene Golub, and Charles Van Loan, Matrix Computations, Third Edition, Johns Hopkins University Press, Baltimore, 1996,
  188. Gregory Marton and Boris Katz Using Semantic Overlap Scoring in Answering TREC Relationship Questions, Proceedings of LREC 2006, Genoa, Italy; May, 2006
  189. Evgeniy Gabrilovich and Shaul Markovitch Computing Semantic Relatedness using Wikipedia-based Explicit Semantic Analysis Proceedings of The 20th International Joint Conference on Artificial Intelligence (IJCAI), Hyderabad, India, January 2007 Vector based Approaches to Semantic Similarity Measures 173
  190. Gianni Amati, Claudio Carpineto, and Giovanni Romano. Comparing weighting models for monolingual information retrieval. In CLEF, pages 310-318, 2003.
  191. Gianni Amati and Cornelis Joost Van Rijsbergen. Probabilistic models of informa- tion retrieval based on measuring the divergence from randomness. ACM Trans. Inf. Syst., 20(4):357-389, 2002.
  192. Claudio Carpineto, Renato de Mori, Giovanni Romano, and Brigitte Bigi. An information-theoretic approach to automatic query expansion. ACM Trans. Inf. Syst., 19(1):1-27, 2001.
  193. Thomas M. Cover and Joy A. Thomas. Elements of information theory. Wiley- Interscience, New York, NY, USA, 1991.
  194. E. N. Efthimiadis. Query expansion. Annual Review of Information Systems and Technology, 31:121-187, 1996.
  195. Y. Jing and W. Bruce Croft. An association thesaurus for information retrieval. In Proceedings of RIAO-94, 4th International Conference "Recherche d'Information Assistee par Ordinateur", pages 146-160, New York, US, 1994.
  196. C. Macdonald, B. He, V. Plachouras, and I. Ounis. University of Glasgow at TREC 2005: Experiments in Terabyte and Enterprise Tracks with Terrier. In Proceeddings of the 14th Text REtrieval Conference (TREC 2005), 2005.
  197. Helen J. Peat and Peter Willett. The limitations of term co-occurrence data for query expansion in document retrieval systems. JASIS, 42(5):378-383, 1991.
  198. Carol Peters and Martin Braschler. European research letter: Cross-language sys- tem evaluation: The clef campaigns. JASIST, 52(12):1067-1072, 2001.
  199. V. Plachouras, B. He, and I. Ounis. University of Glasgow at TREC2004: Exper- iments in Web, Robust and Terabyte tracks with Terrier. In Proceeddings of the 13th Text REtrieval Conference (TREC 2004), 2004.
  200. Yonggang Qiu and Hans-Peter Frei. Concept based query expansion. In SIGIR, pages 160-169, 1993.
  201. C. J. Van Rijsbergen. A theoretical basis for the use of cooccurrence data in information retrieval. Journal of Documentation, (33):106-119, 1977.
  202. J. J. Rocchio. Relevance feedback in information retrieval. In G Salton, editor, The SMART retrieval system, pages 313-323. Prentice Hall, 1971.
  203. Hinrich Schütze and Jan O. Pedersen. A cooccurrence-based thesaurus and two applications to information retrieval. Inf. Process. Manage., 33(3):307-318, 1997.
  204. Ellen M. Voorhees. Overview of the trec 2003 robust retrieval track. In TREC, pages 69-77, 2003.
  205. Ellen M. Voorhees. The trec robust retrieval track. SIGIR Forum, 39(1):11-20, 2005.
  206. Ellen M. Voorhees. The trec 2005 robust track. SIGIR Forum, 40(1):41-48, 2006.
  207. Jinxi Xu and W. Bruce Croft. Improving the effectiveness of information retrieval with local context analysis. ACM Trans. Inf. Syst., 18(1):79-112, 2000.
  208. Angel Zazo, Carlos G. Figuerola, and José Luis Alonso Berrocal. REINA at CLEF 2006 robust task: Local query expansion using term windows for robust retrieval. In A. Nardi, C. Peters, and J.L. Vicedo, editors, ABSTRACTS CLEF 2006 Work- shop, 20-22 September, Alicante, Spain. Results of the CLEF 2006 Cross-Language System Evaluation Campaign, 2006.
  209. Gulli, A., Signorini, A.: The indexable web is more than 11.5 billion pages. In: WWW '05: Special interest tracks and posters of the 14th international conference on World Wide Web, New York, NY, USA, ACM Press (2005) 902-903
  210. Losee, R.M.: When information retrieval measures agree about the relative quality of document rankings. Journal of the American Society of Information Science 51 (2000) pp. 834-840
  211. Gordon, M., Pathak, P.: Finding information on the world wide web: the retrieval effectiveness of search engines. Inf. Process. Manage. 35 (1999) 141-180
  212. Kraft, D.H., Petry, F.E., Buckles, B.P., Sadasivan, T.: Genetic Algorithms for Query Optimization in Information Retrieval: Relevance Feedback. In Sanchez, E., Shibata, T., Zadeh, L., eds.: Genetic Algorithms and Fuzzy Logic Systems, Singapore, World Scientific (1997)
  213. Brin, S., Page, L.: The anatomy of a large-scale hypertextual Web search engine. Computer Networks and ISDN Systems 30 (1998) 107-117
  214. Amento, B., Terveen, L., Hill, W.: Does authority mean quality? predicting expert quality ratings of web documents. In: SIGIR '00: Proceedings of the 23rd annual international ACM SIGIR conference on Research and development in information retrieval, New York, NY, USA, ACM Press (2000) 296-303
  215. Joachims, T.: Optimizing search engines using clickthrough data. In: Proceedings of the ACM Conference on Knowledge Discovery and Data Mining (KDD), ACM (2002)
  216. Xue, G.R., Zeng, H.J., Chen, Z., Yu, Y., Ma, W.Y., Xi, W., Fan, W.: Optimizing web search using web click-through data. In: CIKM '04: Proceedings of the thir- teenth ACM international conference on Information and knowledge management, New York, NY, USA, ACM Press (2004) 118-126
  217. Gomez, M., Abasolo, J.C.: Improving meta-search by using queryweighting and numerical aggregation operators. In: Proceedings of the Information Processing and Management of Uncertainty conference, IPMU 2002. (2002)
  218. Yager, R.R.: On ordered weighted averaging aggregation operators in multicriteria decisionmaking. In Dubois, D., Prade, H., Yager, R.R., eds.: Readings in Fuzzy Sets for Intelligent Systems. Kaufmann, San Mateo, CA (1993) 80-87
  219. Zhang, D., Dong, Y.: An efficient algorithm to rank Web resources. Computer Networks (Amsterdam, Netherlands: 1999) 33 (2000) 449-455
  220. Wang, P.H., Wang, J.Y., Lee, H.M.: Queryfind: Search ranking based on users' feedback and expert's agreement. In: EEE '04: Proceedings of the 2004 IEEE International Conference on e-Technology, e-Commerce and e-Service (EEE'04), Washington, DC, USA, IEEE Computer Society (2004) 299-304
  221. Keyhanipour, A.H., Moshiri, B., Piroozmand, M., Lucas, C.: Webfusion: Funda- mentals and principals of a novel meta search engine. In: Proceedings of Interna- tional Joint Conference on Neural Networks. (2006)
  222. Gauch, S., Wang, G.: Information Fusion with ProFusion. In: Proceedings of WebNet '96: The First World Conference of the Web Society, San Francisco, CA (1996) 174-179
  223. Cordn, O., de Moya, F., Zarco, C.: Fuzzy logic and multiobjective evolutionary algorithms as soft computing tools for persistent query learning in text retrieval en- vironments. In: IEEE International Conference on Fuzzy Systems 2004, Budapest, Hungary (2004) 571-576
  224. Mitchell, M.: An Introduction to Genetic Algorithms. MIT Press, Cambridge, MA (1996)
  225. Ramos, V., Abraham, A.: Evolving a stigmergic self-organized data-mining. CoRR cs.AI/0403001 (2004)
  226. Húsek, D., Snášel, V., Neruda, R., Owais, S.S.J., Krömer, P.: Boolean queries op- timization by genetic programming. WSEAS Transactions on Information Science and Applications 3 (2006) 15-20
  227. Salton, G., Buckley, C.: Term-weighting approaches in automatic text retrieval. Information Processing and Management 24 (1988) pp. 513-523
  228. Fondeur, Y., Tuchszirer, C.: Internet et les intermdiaires du march du travail. In: La lettre de l'IRES. Number 67. IRES -Institut de Recherches Economiques et Sociales, Noisy-le-Grand, France (2005)
  229. Focus RH: Le guide des 500 meilleurs sites emploi. Jeunes Editions, Levallois- Perret, France (2006)
  230. Salton, G., McGill, M.J.: Introduction to Modern Information Retrieval. McGraw- Hill, Tokyo (1983)
  231. Gross, M.: Local grammars and their representation by finite automata. In Hoey, M., ed.: Data, Description, Discourse, Papers on the English Language in honour of John McH Sinclair. Harper-Collins, London (1993) 26-38
  232. Gross, M.: The Construction of Local Grammars. In Roche, E., Schabs, Y., eds.: Finite-State Language Processing. Language, Speech, and Communication, Cambridge, Mass.: MIT Press (1997) 329-354
  233. Courtois, B., Silberztein, M.: Dictionnaires lectroniques du franais. Langues franaise 87 (1990) 11-22
  234. Courtois, B., Garrigues, M., Gross, G., Gross, M., Jung, R., Mathieu-Colas, M., Silberztein, M., Vivs, R.: Dictionnaire lectronique des noms composs DELAC : les composants NA et NN. Rapport Technique du LADL 55 (1997)
  235. Gross, M.: A bootstrap method for constructing local grammars. In: Contempo- rary Mathematics: Proceedings of the Symposium, University of Belgrad, Belgrad (1999) 229-250
  236. Senellart, J.: Locating noun phrases with finite state transducers. In: Proceedings of the 17th International Conference on Computational Linguistics, Montral (1998) 1212-1219
  237. Mallchok, F.: Automatic Recognition of Organization Names in English Business News. PhD thesis, Ludwig-Maximillians-Universitt Munich, Munich, Germany (2004)
  238. Harris, Z.S.: Mathematical Structures of Language. Interscience Tracts in Pure and Applied Mathematics 21 (1968) 230238
  239. Harris, Z.S.: Language and Information. Bampton Lectures in America 28 (1988) 120128
  240. Friburger, N., Maurel, D.: Elaboration d'une cascade de transducteurs pour l'extraction des noms personnes dans les textes. In: TALN 2001, Tours (2001)
  241. Paumier, S.: Manuel d'utilisation d'Unitex. (2004)
  242. Bsiri S., Geierhos M. and Ringlstetter C. References
  243. Landauer, T., Dumais, S.: A Solution to Plato's Problem: The Latent Semantic Analysis Theory of Acquisition, Induction and Representation of Knowledge. Psy- chological Review, Vol. 104 (1997) 211-240
  244. Turney P.D.: Mining the Web for Synonyms: PMI-IR versus LSA on TOEFL. Proceedings of ECML'01, Lecture Notes in Computer Science (2001) 491-502
  245. Rehder B., Schreiner M., Wolfe M., Laham D., Landauer T., Kintsch W.: Using Latent Semantic Analysis to assess knowledge: Some technical considerations. Dis- course Processes, Vol. 25 (1998) 337-354
  246. Landauer T., Laham D., Rehder B., Schreiner M. E.: How well can passage meaning be derived without using word order? A comparison of Latent Semantic Analysis and humans. Proceedings of the 19th annual meeting of the Cognitive Science Society (1997)
  247. Wiemer-Hastings P., Zipitria I.: Rules for syntax, vectors for semantics. Proceed- ings of the Twenty-third Annual Conference of the Cognitive Science Society (2001)
  248. Brill E.: Some Advances in Transformation-Based Part of Speech Tagging AAAI, Vol. 1 (1994) 722-727
  249. ExpLSA: An Approach Based on Syntactic Knowledge in Order to Improve LSA... 223
  250. Kanejiya D., Kumar A., Prasad S.: Automatic Evaluation of Students' Answers us- ing Syntactically Enhanced LSA. Proceedings of the Human Language Technology Conference (HLT-NAACL 2003) Workshop on Building Educational Applications using NLP (2003)
  251. Daille, B.: Study and Implementation of Combined Techniques for Automatic Ex- traction of Terminology. In: Resnik, P.; Klavans, J. (eds.): The Balancing Act: Combining Symbolic and Statistical Approaches to Language, MIT Press, Cam- bridge, MA, USA, (1996) 49-66.
  252. Bourigault D., Fabre C.: Approche linguistique pour l'analyse syntaxique de cor- pus.
  253. Cahiers de Grammaires, Vol. 25 (2000) 131-151
  254. Roche M., Heitz T., O. Matte-Tailliez O., Kodratoff Y.: Exit: Un système itératif pour l'extraction de la terminologie du domaine à partir de corpus spécialisés. Proceedings of JADT'04, Vol. 2 (2004) 946-956
  255. Bourigault D.: UPERY : un outil d'analyse distributionnelle étendue pour la con- struction d'ontologies à partir de corpus. Actes de TALN, Nancy (2002) 75-84
  256. Chauché J.: Un outil multidimensionnel de l'analyse du discours. Proceedings of Coling, Standford University, California (1984) 11-15
  257. Faure D.: de textes : le système ASIUM. Phd, Université de Paris Sud (2000)
  258. Ferri C., Flach P., Hernandez-Orallo J.: Learning decision trees using the area under the ROC curve Proceedings of ICML'02 (2002) 139-146
  259. Mary J.: Étude de l'Apprentissage Actif : Application à la conduite d'expériences. Phd, Université Paris Sud (2005)
  260. Béchet N., Roche M., Chauché J.: ExpLSA : utilisation d'informations syntaxico- sémantiques associées à LSA pour améliorer les méthodes de classification con- ceptuelle. Proceedings of EGC'08.
  261. Faure D., Nedellec C.: Knowledge acquisition of predicate argument structures from technical texts using machine learning: The system ASIUM. Proceedings of the 11th European Workshop, Knowledge Acquisition, Modelling and Management, number 1937 in LNAI (1999) 329-334
  262. Schmid H. Improvements in part-of-speech tagging with an application to Ger- man. Technical report, Universitat Stuttgart. Institut fur maschinelle Sprachver- arbeitung. (Revised version of a paper presented at EACL SIGDAT, Dublin 1995).
  263. Y. Yang An Evaluation of Statistical Approaches to Text Categorization. Informa- tion Retrieval, number 1-2, volume 1 (1999) 69-90
  264. Carmel, D., Farchi, E., Petruschka, Y., and Soffer, A.: Automatic Query Refinement using Lexical Affinities with Maximal Information Gain. SIGIR (2002): 283-290.
  265. Cui, H., Kan, M. Chua, T. and Xiao, J.: A Comparative Study on Sentence Retrieval for Definitional Questions Answering. SIGIR Workshop on Information Retrieval for Question Answering (IR4QA), (2004) 90-99.
  266. Denicia-Carral, C., Montes-y-Gómez, M., and Villaseñor-Pineda, L.: A Text Mining Ap- proach for Definition Question Answering. 5 th International Conference on Natural Lan- guage Processing, Fin Tal. Lecture Notes in Artificial Intelligence, Springer (2006).
  267. Grinberg, D., Lafferty, J., and Sleator, D.: A Robust Parsing Algorithm for Link Gram- mars. Carnegie Mellon University Computer Science technical report CMU-CS-95-125, and Proceedings of the Fourth International Workshop on Parsing Technologies, Prague, September, (1995).
  268. Harabagiu, S., and Lacatusu, F., Strategies for Advanced Question Answering. In Proceed- ing of the Workshop on Pragmatics of Questions Answering at HLT-NAACL. (2004): 1-9.
  269. Hildebranddt, W., Katz, B. and Lin, J.: Answering Definition Question Using Multiple Knowledge Sources. In Proceeding of HLT/NAACL, Boston (2004): 49-56.
  270. Martínez-Gil, C., and López-López, A.: Answering Definition Questions using a Statistical Method Combined with a Classifier Ensemble, in Proceedings of the Seventh International Symposium on Natural Language Processing, SNLP2007, Thailand, December, (2007): 189-194.
  271. Moldovan, D., Pasca, M., Harabagiu,S., and Surdeanu, M.: Performance Issues and Error Analysis in an Open-Domain Question Answering System. Proceeding of the 40 th Annual Meeting of the Association for Computational Linguistics (ACL). Philadelphia, July, (2002): 33-40.
  272. Voorhees, E.: Evaluating Answers to Definition Questions. NIST (2003) 1-3.
  273. Voorhees, E.: Overview of the TREC 2003 Question Answering Track. NIST (2003): 54- 68.
  274. Voorhees, E.: Overview of the TREC 2004 Question Answering Track. NIST (2004): 12- 20.
  275. Xu, J., Licuanan A., and Weischedel R.: TREC 2003 QA at BBN: Answering Definitional Questions. In the Twelfth Text Retrieval Conference (2003): 28-35.
  276. International Telecommunication Union: E.161: Arrangement of digits, letters and symbols on telephones and other devices that can be used for gaining access to a telephone network. http://www.itu.int/rec/T-REC-E.161-200102-I/en (2001)
  277. Gong, J., Tarasewich, P.: Alphabetically constrained keypad designs for text entry on mobile devices. In: CHI '05: Proceedings of the SIGCHI conference on Human factors in computing systems, New York, NY, USA, ACM Press (2005) 211-220
  278. MacKenzie, I.S.: KSPC (keystrokes per character) as a characteristic of text entry techniques. In: Mobile HCI '02: Proceedings of the 4th International Symposium on Mobile Human-Computer Interaction, London, Springer-Verlag (2002) 195-210
  279. Brill, E.: Transformation-based error-driven learning and natural language process- ing: A case study in part-of-speech tagging. Computational Linguistics 21 (1995) 543-565
  280. MacKenzie, I.S., Kober, H., Smith, D., Jones, T., Skepner, E.: Letterwise: Prefix- based disambiguation for mobile text input. Technical report, Eatoni (2001)
  281. Davies, M.: The View/BNC interface. http://corpus.byu.edu/bnc/ (2004) Accessed on 23rd October 2007.
  282. Hasselgren, J., Montnemery, E., Nugues, P., Svensson, M.: HMS: A predictive text entry method using bigrams. In: Proceedings of the Workshop on Language Modeling for Text Entry Methods, 10th Conference of the European Chapter of the Association of Computational Linguistics, Budapest, Hungary, Association for Computational Linguistics (2003) 43-49
  283. Hawes, N., Kelleher, J.: Context-sensitive word selection for single-tap text entry. In: STAIRS 2004: Proceedings of the Second Starting AI Researchers' Symposium, IOS Press (2004) 217-222
  284. Charniak, E., Hendrickson, C., Jacobson, N., Perkowitz, M.: Equations for part-of- speech tagging. In: AAAI. (1993) 784-789

Related papers

An open-source package for recognizing textual entailment
Matteo Negri

2010

View PDFchevron_right
An approach using named entities for recognizing textual entailment
julio castillo

Proceedings of the Fourth PASCAL Challenges …, 2008

This paper describes the Sagan system in the context of the Fourth Pascal Recognizing Textual Entailment (RTE-4) Evaluation Challenge. Sagan applies a Support Vector Machine classifier to examples characterized by four features based on: edit distance, distance in WordNet and Longest Common Substring between text and hypothesis. Additionally, we created a filter applying hand-crafted rules based on Named Entities to detect cases where no entailment was found. Despite this simple approach, results are promising. Applying the Named Entity filter yields a small improvement in precision.

View PDFchevron_right
A Statistics-Based Semantic Textual Entailment System
Dr. Partha Pakray

Advances in Artificial …, 2011

View PDFchevron_right
2007b. DLSITE-1: Lexical Analysis for Solving Textual Entailment Recognition
Daniel Micol

2015

Abstract. This paper discusses the recognition of textual entailment in a text-hypothesis pair by applying a wide variety of lexical measures. We consider that the entailment phenomenon can be tackled from three general levels: lexical, syntactic and semantic. The main goals of this research are to deal with this phenomenon from a lexical point of view, and achieve high results considering only such kind of knowledge. To accomplish this, the information provided by the lexical measures is used as a set of features for a Support Vector Machine which will decide if the entailment relation is produced. A study of the most relevant features and a comparison with the best state-of-the-art textual entailment systems is exposed throughout the paper. Finally, the system has been evaluated using the Second PASCAL Recognising Textual Entailment Challenge data and evaluation methodology, obtaining an accuracy rate of 61.88%. 1

View PDFchevron_right
1 Design and Realization of a Modular Architecture for Textual Entailment
Sebastian Padó

2016

A key challenge at the core of many NLP tasks is the ability to determine which conclusions can be inferred from a given natural language text. This problem, called the Recognition of Textual Entailment (RTE), has initiated the development of a range of algorithms, methods and technologies. Unfortunately, research on TE (like semantics research more generally), is fragmented into studies focussing on various aspects of semantics such as world knowledge, lexical and syntactic relations, or more specialized kinds of inference. This fragmentation has problematic practical consequences. Notably, interoperability among existing RTE systems is poor, and reuse of resources and algorithms is mostly infeasible. This also makes systematic evaluations very difficult to carry out. Finally, TE presents a wide array of approaches to potential end users with little guidance on which to pick. Our contribution to this situation is the novel EXCITEMENT architecture, which was developed to enable and ...

View PDFchevron_right
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved