Academia.eduAcademia.edu

Figure 11 – uploaded by John McCrae

See full PDFdownloadDownload figure
Table 1. Hypernyms and Linked Hypernyms datasets - size statistics. Table 2. Hypernyms and Linked Hypernyms datasets - accuracy .

Table 1 Hypernyms and Linked Hypernyms datasets - size statistics. Table 2. Hypernyms and Linked Hypernyms datasets - accuracy .

Related Figures (44)

Figure 1: Linguistic Linked Open Data cloud as of April 2014.
Figure 1: Linguistic Linked Open Data cloud as of April 2014.
Table 2: Results comparing the performance in terms of F_1 score for French, Finnish and Portuguese (highest in bold).
Table 2: Results comparing the performance in terms of F_1 score for French, Finnish and Portuguese (highest in bold).
Figure 1: Fl score for Finnish, French and Portuguese de- pending on the value of a and /.
Figure 1: Fl score for Finnish, French and Portuguese de- pending on the value of a and /.
Figure 2: Graphical representation of the F1 score against delta for our measure and other Level 2 Measures.
Figure 2: Graphical representation of the F1 score against delta for our measure and other Level 2 Measures.
Figure 1: Component Metadata Model (ISO/DIS 24622-1, 2013)
Figure 1: Component Metadata Model (ISO/DIS 24622-1, 2013)
Table 1: Extracted genres and number of descriptions tagged with them It should be noted that we excluded descriptions for ’short film’ genre ("Koporkomerpaxka" in Russian), 3103 in total. The reason was that ’short film’ is hardly a movie genre per se, it is more of a movie format. Thus, we instead added to our list the set of descrip- tions for "anime" genre. It was the next candidate by frequency and is certainly closer to the definition of n This collection possessed 84 different unique genres most of which occurred not more than 20 or 30 times. Trying to employ this "genre long tail" to get keywords seemed useless. Thus, we chose 20 most frequent gen- res, which together comprise 198375 (86%) of our de- scriptions. The genres we omit are mostly exotic ones like "movie about movie" and "vampires". Notable exceptions are adult movies (2300 descriptions), bio- graphic movies (1806 descriptions) and movies about animals (1500 descriptions). For now they fall out from top 20 genres, however in the future we plan to inte- grate them into our data set as well. The table 1 gives the list of selected genres (their English translations, to be more precise) with corresponding number of de- scriptions.
Table 1: Extracted genres and number of descriptions tagged with them It should be noted that we excluded descriptions for ’short film’ genre ("Koporkomerpaxka" in Russian), 3103 in total. The reason was that ’short film’ is hardly a movie genre per se, it is more of a movie format. Thus, we instead added to our list the set of descrip- tions for "anime" genre. It was the next candidate by frequency and is certainly closer to the definition of n This collection possessed 84 different unique genres most of which occurred not more than 20 or 30 times. Trying to employ this "genre long tail" to get keywords seemed useless. Thus, we chose 20 most frequent gen- res, which together comprise 198375 (86%) of our de- scriptions. The genres we omit are mostly exotic ones like "movie about movie" and "vampires". Notable exceptions are adult movies (2300 descriptions), bio- graphic movies (1806 descriptions) and movies about animals (1500 descriptions). For now they fall out from top 20 genres, however in the future we plan to inte- grate them into our data set as well. The table 1 gives the list of selected genres (their English translations, to be more precise) with corresponding number of de- scriptions.
Figure 1: Component Metadata Model (ISO/DIS 24622-1, 2013)
Figure 1: Component Metadata Model (ISO/DIS 24622-1, 2013)
Figure 1 - Annotating objects in an image
Figure 1 - Annotating objects in an image
because the target language of an annotation cannot be guaranteed without the intervention of a translation. This applies equally to intra- and inter-platform environments. Gracia et al. (2012) suggest that users expect access to the Web of Data to be intuitive, appealing, and effortless. We believe this is also true for Linguistic Linked Open Data. Linking semantically related resources across the Web of Data can improve the situation (Sasaki, 2013) of, what would otherwise be, disparate linguistic platforms. The alternative is to replicate and directly adapt every annota-
because the target language of an annotation cannot be guaranteed without the intervention of a translation. This applies equally to intra- and inter-platform environments. Gracia et al. (2012) suggest that users expect access to the Web of Data to be intuitive, appealing, and effortless. We believe this is also true for Linguistic Linked Open Data. Linking semantically related resources across the Web of Data can improve the situation (Sasaki, 2013) of, what would otherwise be, disparate linguistic platforms. The alternative is to replicate and directly adapt every annota-
Figure 1: Logical structure of a HyperGramm entry.
Figure 1: Logical structure of a HyperGramm entry.
Figure 2: Conceptual representation of HyperGramm terms. The model we describe in this section is depicted in Fig- ure 2. For reasons of readability, we did not include in this representation the DataTypeProperty “hasID”, which is obligatory for each class. IDs will be represented as URIs compliant with the LOD representation principles de- scribed above.
Figure 2: Conceptual representation of HyperGramm terms. The model we describe in this section is depicted in Fig- ure 2. For reasons of readability, we did not include in this representation the DataTypeProperty “hasID”, which is obligatory for each class. IDs will be represented as URIs compliant with the LOD representation principles de- scribed above.
Table 1: Examples of vocabularies used for the representa- tion of HyperGramm. With respect to the first issue, at this stage of the represen- tation, we have selected DC?, RDF-Schema*, OWL>, and SKOS®, as shown in Table 1.
Table 1: Examples of vocabularies used for the representa- tion of HyperGramm. With respect to the first issue, at this stage of the represen- tation, we have selected DC?, RDF-Schema*, OWL>, and SKOS®, as shown in Table 1.
Table 2: Sample entry of term Jntensitdtspartikel ‘intensity particle’ in the original list (with En- glish translations).
Table 2: Sample entry of term Jntensitdtspartikel ‘intensity particle’ in the original list (with En- glish translations).
Table 1: Statistics on the etymological dictionaries, see Sect. 1. for abbreviations
Table 1: Statistics on the etymological dictionaries, see Sect. 1. for abbreviations
Table 2: Verse-aligned texts in the Germanic parallel Bible corpus (parentheses indicate marginal fragments) Table 3: Selected translational dictionaries extracted from parallel Bibles
Table 2: Verse-aligned texts in the Germanic parallel Bible corpus (parentheses indicate marginal fragments) Table 3: Selected translational dictionaries extracted from parallel Bibles
older Germanic languages are fragmentary attested, only). To eliminate possible outliers, we eliminated all hapax legomena from the alignment table (filter >1 in Tab. 3).
older Germanic languages are fragmentary attested, only). To eliminate possible outliers, we eliminated all hapax legomena from the alignment table (filter >1 in Tab. 3).
Figure 6: Visualization and fine-tuning of parallel segment detection metrics for OS and OHG Tatian
Figure 6: Visualization and fine-tuning of parallel segment detection metrics for OS and OHG Tatian
Figure 1: The meaning of a lexical entry / over time. Taken together, at each time point ¢;, the meanings cj can be thought of as constituting a time slice of a perdurant ob- ject, d such that d represents the process of meaning shift of the lexical entry / with original meaning co. | a: a a ee) we ae a
Figure 1: The meaning of a lexical entry / over time. Taken together, at each time point ¢;, the meanings cj can be thought of as constituting a time slice of a perdurant ob- ject, d such that d represents the process of meaning shift of the lexical entry / with original meaning co. | a: a a ee) we ae a
Figure 2: The meaning of a lexical entry / over time repre- sented as a perdurant.
Figure 2: The meaning of a lexical entry / over time repre- sented as a perdurant.
In Figure 4 we present a diagram of the JemonDIA model.
In Figure 4 we present a diagram of the JemonDIA model.
5. Conclusions
5. Conclusions
Table 1: Table data Table 1 illustrates what we mean by table data; it provides a set of data, here observations about the last symbol in several words, where each word’s class is also given.
Table 1: Table data Table 1 illustrates what we mean by table data; it provides a set of data, here observations about the last symbol in several words, where each word’s class is also given.
Table 2: Matrix data
Table 2: Matrix data
Figure 1: Matrix transformed into a graph Table 2 can algorithmically be transformed into a graph by assigning the row and column labels as vertices and con- necting them via edges for cells that have a “1”. The result of this transformation is illustrated in Figure 1.
Figure 1: Matrix transformed into a graph Table 2 can algorithmically be transformed into a graph by assigning the row and column labels as vertices and con- necting them via edges for cells that have a “1”. The result of this transformation is illustrated in Figure 1.
Figure 2: Linked Data example
Figure 2: Linked Data example
We have demonstrated how to extract table data from the Linked Data graph and explained how table data can be transformed into numerical matrices. An illustration is given in Figure 3, which contrasts the graph, table and ma- trix formats. Figure 3: Toy example of a graph with the equivalent ta- ble and matrix formats. The graph has two kinds of links: between nodes of kind ‘F’ and ‘L’, and between nodes of kind ‘L’ and ‘S’. So, there are two tables with links, and two corresponding matrices.
We have demonstrated how to extract table data from the Linked Data graph and explained how table data can be transformed into numerical matrices. An illustration is given in Figure 3, which contrasts the graph, table and ma- trix formats. Figure 3: Toy example of a graph with the equivalent ta- ble and matrix formats. The graph has two kinds of links: between nodes of kind ‘F’ and ‘L’, and between nodes of kind ‘L’ and ‘S’. So, there are two tables with links, and two corresponding matrices.
Figure 4: Dot product number of paths connecting S-nodes to F-nodes.
Figure 4: Dot product number of paths connecting S-nodes to F-nodes.
We decided to try and identify major clusters of associa- tion between WALS and PHOIBLE. Using a weighted co- sine similarity, we identify levels of high association be- tween clusters of features in WALS chapters and PHOIBLE phonological inventory data. These are visualized as heat maps in Figures 5, 6 and 7.'? The point of these figures is mainly to illustrate the possible observation of clusters. The detailed extraction of linguistically sensible clusters and their interpretation will have to wait for another paper. We will here only comment on a few possible interpreta- tions. Figure 5: Heat map for all characteristics with frequency more than 10 (~1000 characteristics)
We decided to try and identify major clusters of associa- tion between WALS and PHOIBLE. Using a weighted co- sine similarity, we identify levels of high association be- tween clusters of features in WALS chapters and PHOIBLE phonological inventory data. These are visualized as heat maps in Figures 5, 6 and 7.'? The point of these figures is mainly to illustrate the possible observation of clusters. The detailed extraction of linguistically sensible clusters and their interpretation will have to wait for another paper. We will here only comment on a few possible interpreta- tions. Figure 5: Heat map for all characteristics with frequency more than 10 (~1000 characteristics)
Figure 7: Heat map for genera with most data in WALS only Figure 6: Heat map for languages with most data in WALS only
Figure 7: Heat map for genera with most data in WALS only Figure 6: Heat map for languages with most data in WALS only
Figure 2: Entity-relationship diagram of the WOLD data model; SynSets are sets of synonyms, a Counterpart is an instance of the many-to-many relation between Words and Meanings in the sense of Haspelmath and Tadmor (2009). tt at 2 ae BARD BARE POSUREIUEEUQULUEE CLLD applications are implemented using the clld framework.'? This framework in turn is based on the python packages pyramid and SQLA1chemy and allows
Figure 2: Entity-relationship diagram of the WOLD data model; SynSets are sets of synonyms, a Counterpart is an instance of the many-to-many relation between Words and Meanings in the sense of Haspelmath and Tadmor (2009). tt at 2 ae BARD BARE POSUREIUEEUQULUEE CLLD applications are implemented using the clld framework.'? This framework in turn is based on the python packages pyramid and SQLA1chemy and allows
In a concrete incarnation this core data model can be inter- preted as shown in Figure 2. Note the additional relation between Word and Counterpart which is not present in the core model. The clld framework uses the joined table inheritance feature of the SQLA1chemy package to trans- parently add attributes to entities of the core data model. (see section 2.2.2.).!8
In a concrete incarnation this core data model can be inter- preted as shown in Figure 2. Note the additional relation between Word and Counterpart which is not present in the core model. The clld framework uses the joined table inheritance feature of the SQLA1chemy package to trans- parently add attributes to entities of the core data model. (see section 2.2.2.).!8
Table 1: Number of lexical elements in the graphs. Figure 1: Some of the statistics available about the French extracted data.
Table 1: Number of lexical elements in the graphs. Figure 1: Some of the statistics available about the French extracted data.
Table 2: Number of lexico-semantic relations in the graphs.
Table 2: Number of lexico-semantic relations in the graphs.
Figure 1: The core of the /emon model.
Figure 1: The core of the /emon model.
Table 1: Statistics concerning the /emon-BabelNet 2.0 RDF dataset.
Table 1: Statistics concerning the /emon-BabelNet 2.0 RDF dataset.
Figure 1: Methodology for Legacy Language Resources Adaptation for Sentiment Analysis. Then we used a searching-chunking process to achieve the most relevant collocations of the corpora. This task con- sisted of identification of collocated context entities anc sentiment words using a 3-word sliding window. The cal- culated collocations were reviewed again by linguists. A simple web application helped the linguists to: the context of the entity room which is a noun with the synset id 04105893.
Figure 1: Methodology for Legacy Language Resources Adaptation for Sentiment Analysis. Then we used a searching-chunking process to achieve the most relevant collocations of the corpora. This task con- sisted of identification of collocated context entities anc sentiment words using a 3-word sliding window. The cal- culated collocations were reviewed again by linguists. A simple web application helped the linguists to: the context of the entity room which is a noun with the synset id 04105893.
Figure 2: Snapshot of the Web application that allows linguists to specify the sentiment scores. Listing 5: Namespaces used in the RDF lexicons..
Figure 2: Snapshot of the Web application that allows linguists to specify the sentiment scores. Listing 5: Namespaces used in the RDF lexicons..
Figure 3: Example lexicon for the domain *hotel’ in English.
Figure 3: Example lexicon for the domain *hotel’ in English.
Table 1: Most frequent properties used in PDEV-Lemon
Table 1: Most frequent properties used in PDEV-Lemon
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved