Academia.eduAcademia.edu

Figure 4 – uploaded by Daniel Bekele

See full PDFdownloadDownload figure
Table 3.5 Minimum edit distance of strings In the following table the edit distance for the selected strings is presented. As shown in table 3.5 if the edit distance is greater, the strings are more different (i.e. they are not morphologically related). The edit distance is 0 (zero) if the strings are identical. The smaller value of edit distance indicates that the strings are morphologically related or likely variants of each others.

Table 3 5 Minimum edit distance of strings In the following table the edit distance for the selected strings is presented. As shown in table 3.5 if the edit distance is greater, the strings are more different (i.e. they are not morphologically related). The edit distance is 0 (zero) if the strings are identical. The smaller value of edit distance indicates that the strings are morphologically related or likely variants of each others.

Related Figures (9)

Table 3.1 Vocabulary file for the given Afaan Oromo sentence daniel knew how to worship A word alignment for a parallel sentence pair represents the correspondence between words in a source language and their translations in a target language (Brown et al., 1993). In this study, word alignment represents the mapping between Afaan Oromo (source language) and English (target language). Nowadays, word aligned bilingual corpora are being used as an important source of the knowledge. Word alignment model was first introduced in SMT by Brown et al. (1993). GIZA++ uses a statistical alignment model which computes a translation probability for each co-occurring word pair. A given word from the source language may appear as being aligned with several translation candidates of target words, each one with a given probability value. For example, for the following Afaan Oromo-English parallel sentence pairs selected from the collected corpora, the vocabulary files are given in table 3.1 and table 3.2 and the bitext file generated for the sentence pairs is given in table 3.3. daani’eel akkamitti akka waaqeffachuu qabu beeka ture
Table 3.1 Vocabulary file for the given Afaan Oromo sentence daniel knew how to worship A word alignment for a parallel sentence pair represents the correspondence between words in a source language and their translations in a target language (Brown et al., 1993). In this study, word alignment represents the mapping between Afaan Oromo (source language) and English (target language). Nowadays, word aligned bilingual corpora are being used as an important source of the knowledge. Word alignment model was first introduced in SMT by Brown et al. (1993). GIZA++ uses a statistical alignment model which computes a translation probability for each co-occurring word pair. A given word from the source language may appear as being aligned with several translation candidates of target words, each one with a given probability value. For example, for the following Afaan Oromo-English parallel sentence pairs selected from the collected corpora, the vocabulary files are given in table 3.1 and table 3.2 and the bitext file generated for the sentence pairs is given in table 3.3. daani’eel akkamitti akka waaqeffachuu qabu beeka ture
Figure 3.1 Architecture of the Afaan Oromo-English CLIR System The architecture of the Afaan Oromo-English CLIR system is shown diagrammatically in figure 3.1 (adopted from Aynalem, 2009). As illustrated in the figure, the proposed CLIR system uses a number of phases to translate a given Afaan Oromo query into an English query. The major components involved in the Afaan Oromo-English cross- language information retrieval system are explained in the following sections. Corpus-based CLIR method is based on multilingual text collections, from which translation knowledge is derived using various statistical methods (Talvensaari et al., 2007). It is one of the query translation ap proaches of CLIR that uses either parallel or comparable corpora (Kishida, 2005), to establish a link between the query and the documents. However, Talvensaari (2008) proved that more accurate translation knowledge is extracted than comparable corpus. This from parallel corpus rather research, therefore, uses parallel documents of Afaan Oromo and English to study the application of corpus-based query translation approach of CLIR.
Figure 3.1 Architecture of the Afaan Oromo-English CLIR System The architecture of the Afaan Oromo-English CLIR system is shown diagrammatically in figure 3.1 (adopted from Aynalem, 2009). As illustrated in the figure, the proposed CLIR system uses a number of phases to translate a given Afaan Oromo query into an English query. The major components involved in the Afaan Oromo-English cross- language information retrieval system are explained in the following sections. Corpus-based CLIR method is based on multilingual text collections, from which translation knowledge is derived using various statistical methods (Talvensaari et al., 2007). It is one of the query translation ap proaches of CLIR that uses either parallel or comparable corpora (Kishida, 2005), to establish a link between the query and the documents. However, Talvensaari (2008) proved that more accurate translation knowledge is extracted than comparable corpus. This from parallel corpus rather research, therefore, uses parallel documents of Afaan Oromo and English to study the application of corpus-based query translation approach of CLIR.
Ihe translation of Afaan Oromo queries into English was based on the Afaan Oromo- English bilingual dictionary which has been constructed automatically from the Afaan Oromo-English parallel corpora collected. The bilingual dictionary that is constructed stores both source words and their corresponding translation of the target words. The word alignment in the dictionary contains all possible translation of a word from the source text into a target word together with its probability of alignment. This probability value assigned for each possible translation of a word shows the degree to which Afaan Oromo word is most likely translated into its equivalent English word. The highest the probability value indicates the best translation among the candidates translations exist. The bilingual dictionary was, therefore, constructed by the help of this probability value assigned to each translation. Python script was developed to select the one that has the highest probability of alignment, if there is more than one alignment for the given source word. Table 3.4 shows sample of the constructed Afaan Oromo-English bilingual dictionary. Table 3.4 Sample Afaan Oromo-English bilingual dictionary constructed D. Translation This component is responsible for taking query in one language and translating it into another language, i.e. it is the query translation phase. Query translation is required to achieve CLIR by the help of a bilingual dictionary built using parallel corpora collected. The translation of the given query into another language is needed to retrieve documents in the translated (target) language. For this research, a given Afaan Oromo query was translated into its equivalent English query
Ihe translation of Afaan Oromo queries into English was based on the Afaan Oromo- English bilingual dictionary which has been constructed automatically from the Afaan Oromo-English parallel corpora collected. The bilingual dictionary that is constructed stores both source words and their corresponding translation of the target words. The word alignment in the dictionary contains all possible translation of a word from the source text into a target word together with its probability of alignment. This probability value assigned for each possible translation of a word shows the degree to which Afaan Oromo word is most likely translated into its equivalent English word. The highest the probability value indicates the best translation among the candidates translations exist. The bilingual dictionary was, therefore, constructed by the help of this probability value assigned to each translation. Python script was developed to select the one that has the highest probability of alignment, if there is more than one alignment for the given source word. Table 3.4 shows sample of the constructed Afaan Oromo-English bilingual dictionary. Table 3.4 Sample Afaan Oromo-English bilingual dictionary constructed D. Translation This component is responsible for taking query in one language and translating it into another language, i.e. it is the query translation phase. Query translation is required to achieve CLIR by the help of a bilingual dictionary built using parallel corpora collected. The translation of the given query into another language is needed to retrieve documents in the translated (target) language. For this research, a given Afaan Oromo query was translated into its equivalent English query
The normalized Levenshtein distance returns the value between 0 and 1. If the value is | there is a strong similarity between the strings, but if it is 0 there is no similarity between the strings. The closer a value is to 1, the more certain the character strings are the same; the closer to 0, the less certain. By using this normalized edit distance the difficulties indicated in the above situation can be minimized. The normalized value of edit distance of the strings in the table 3.5 is given in the table 3.6.
The normalized Levenshtein distance returns the value between 0 and 1. If the value is | there is a strong similarity between the strings, but if it is 0 there is no similarity between the strings. The closer a value is to 1, the more certain the character strings are the same; the closer to 0, the less certain. By using this normalized edit distance the difficulties indicated in the above situation can be minimized. The normalized value of edit distance of the strings in the table 3.5 is given in the table 3.6.
Figure 4.2 Average Recall-Precision graph of experimentation phase one for English documents Figure 4.1 Average Recall-Precision graph of experimentation phase one for Afaan Oromo documents
Figure 4.2 Average Recall-Precision graph of experimentation phase one for English documents Figure 4.1 Average Recall-Precision graph of experimentation phase one for Afaan Oromo documents
Figure 4.4 Average Recall-Precision graph of experimentation phase two for Figure 4.3 Average Recall-Precision graph of experimentation phase two for Afaan Oromo documents
Figure 4.4 Average Recall-Precision graph of experimentation phase two for Figure 4.3 Average Recall-Precision graph of experimentation phase two for Afaan Oromo documents
In both experiments, the number of queries for which no documents retrieved was larger for the bilingual run than that of monolingual run. This low performance of the bilingual run was caused because of the source words were not really aligned with the Corresponding target words. This wrong alignment was caused because of the limited data size of the parallel corpora used for building bilingual dictionary.
In both experiments, the number of queries for which no documents retrieved was larger for the bilingual run than that of monolingual run. This low performance of the bilingual run was caused because of the source words were not really aligned with the Corresponding target words. This wrong alignment was caused because of the limited data size of the parallel corpora used for building bilingual dictionary.
Related topics:
EngineeringComputer ScienceInformation RetrievalNatural Language ProcessingCross Language Information RetrievalGeo technical EngineeringTechnical Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved