Academia.eduAcademia.edu

Figure 34 – uploaded by Amba Kulkarni

See full PDFdownloadDownload figure
Let’s consider the sentence, ‘Jug is on the table’. Using standard NLP tech- niques, one can identify jug and table to be some objects and that table is used in the locative sense. This is all the meaning that could be drawn when this sentence is considered as such. But when we replace ‘S,,,’ in the above structure with ‘jug’, and another ‘S',,’ (which is in contact with the earlier S,,) with ‘table’, we can draw the following inferences:

Figure 34 Let’s consider the sentence, ‘Jug is on the table’. Using standard NLP tech- niques, one can identify jug and table to be some objects and that table is used in the locative sense. This is all the meaning that could be drawn when this sentence is considered as such. But when we replace ‘S,,,’ in the above structure with ‘jug’, and another ‘S',,’ (which is in contact with the earlier S,,) with ‘table’, we can draw the following inferences:

Related Figures (104)

Table 2. The sUP affixes
Table 2. The sUP affixes
‘igure [4] (right) shows a Hasse diagram of the concept lattice, i.e., a Hasse liagram of the set of the eight intersection sets ordered by the inverted subset elation. In the diagram the nodes (circles) correspond to the intersection sets und an edge between two nodes indicates that the set associated with the upper 1ode is a subset of the one associated with the lower node and that no other ntersection set is a superset of the set associated with the upper and a subset of she set associated with the lower node. The diagram is labeled as follows: each egment is written above the node corresponding to the smallest set to which it »elongs, e.g., c labels the node for the set {c}, b the node for the set {b}, and 1 the one for {a,b}. The labeled Hasse diagram can be read as an inheritance uierarchy: each node corresponds to the set of segments by which the node or me of its supernodes is labeled. E.g., the node for ‘class 4’ corresponds to the et {a, b,c}, the bottom node corresponds to {a, b,c, d}, and the top node to the mpty set.
‘igure [4] (right) shows a Hasse diagram of the concept lattice, i.e., a Hasse liagram of the set of the eight intersection sets ordered by the inverted subset elation. In the diagram the nodes (circles) correspond to the intersection sets und an edge between two nodes indicates that the set associated with the upper 1ode is a subset of the one associated with the lower node and that no other ntersection set is a superset of the set associated with the upper and a subset of she set associated with the lower node. The diagram is labeled as follows: each egment is written above the node corresponding to the smallest set to which it »elongs, e.g., c labels the node for the set {c}, b the node for the set {b}, and 1 the one for {a,b}. The labeled Hasse diagram can be read as an inheritance uierarchy: each node corresponds to the set of segments by which the node or me of its supernodes is labeled. E.g., the node for ‘class 4’ corresponds to the et {a, b,c}, the bottom node corresponds to {a, b,c, d}, and the top node to the mpty set.
Figure [shows a Hasse diagram with intersecting edges of the concept lattice of the formal context constituted by the four classes ‘devoicing output’, ‘devoic- ing input’, ‘nasal assimilation labial’, and ‘nasal assimilation velar’. According to [II] and [16], a lattice has no plane diagram if it is possible to gain the graph Y (i.e. the complete graph K°) from the graph of the diagram enlarged by an additional edge connecting the lowest and the top most node by removing some of the edges and contracting others. It is obvious that the graph in Figure (top) can be constructed from the one in Figure [6] by leaving out some of the edges. The sequence of drawings at the bottom of Figure[f|starts with the graph from the top of the figure enlarged by an additional edge connecting the top and the bottom node. Each remaining graph is gained from its left neighbor by con- tracting the emphasized edge (thick grey edge). As the final graph is isomorphic to the graph ®, the sequence proves that it is impossible to draw a diagram of the concept lattice in Figure [6] without intersecting edges. Hence, the concept lattice of the formal context of our collection of phonological classes is not pla- nar. It follows that it is impossible to construct pratyahara sutras for the formal context in which each sound segment occurs only once. Hence, analogously to the Sivasiitras for Sanskrit we are forced to repeat at least one sound segment for our pratyahara sutras for German. This. the next step towards nratriiaghara sitras for German is the identification
Figure [shows a Hasse diagram with intersecting edges of the concept lattice of the formal context constituted by the four classes ‘devoicing output’, ‘devoic- ing input’, ‘nasal assimilation labial’, and ‘nasal assimilation velar’. According to [II] and [16], a lattice has no plane diagram if it is possible to gain the graph Y (i.e. the complete graph K°) from the graph of the diagram enlarged by an additional edge connecting the lowest and the top most node by removing some of the edges and contracting others. It is obvious that the graph in Figure (top) can be constructed from the one in Figure [6] by leaving out some of the edges. The sequence of drawings at the bottom of Figure[f|starts with the graph from the top of the figure enlarged by an additional edge connecting the top and the bottom node. Each remaining graph is gained from its left neighbor by con- tracting the emphasized edge (thick grey edge). As the final graph is isomorphic to the graph ®, the sequence proves that it is impossible to draw a diagram of the concept lattice in Figure [6] without intersecting edges. Hence, the concept lattice of the formal context of our collection of phonological classes is not pla- nar. It follows that it is impossible to construct pratyahara sutras for the formal context in which each sound segment occurs only once. Hence, analogously to the Sivasiitras for Sanskrit we are forced to repeat at least one sound segment for our pratyahara sutras for German. This. the next step towards nratriiaghara sitras for German is the identification
features yield a total of 2'© = 65536 classes which can be described by feature sets, thus far more than by pratyaharas. Even if one drops the requirement that features should have acoustic, articulatory or auditory definitions and allows for unnatural features, a featural description of the 10 phonological classes cannot perform better than our pratyahara description. As none of our 10 classes can be described in terms of an intersection of some of the other classes, every featural description which is able to distinguish those 10 classes has to make use of at least 10 different features. A minimal featural description of the 10 classes would be to use the class identifiers (e.g., ‘devoicing input’, ‘devoicing output’) as fea- tures. Such a description would be minimal since none of the features would be reducible to other features. 10 features still yield a total of 21° = 1024 classes which can be described by feature sets; this is nearly four times more than by the pratyaharas we employed. One objection against our approach could be that we are onlv considering
features yield a total of 2'© = 65536 classes which can be described by feature sets, thus far more than by pratyaharas. Even if one drops the requirement that features should have acoustic, articulatory or auditory definitions and allows for unnatural features, a featural description of the 10 phonological classes cannot perform better than our pratyahara description. As none of our 10 classes can be described in terms of an intersection of some of the other classes, every featural description which is able to distinguish those 10 classes has to make use of at least 10 different features. A minimal featural description of the 10 classes would be to use the class identifiers (e.g., ‘devoicing input’, ‘devoicing output’) as fea- tures. Such a description would be minimal since none of the features would be reducible to other features. 10 features still yield a total of 21° = 1024 classes which can be described by feature sets; this is nearly four times more than by the pratyaharas we employed. One objection against our approach could be that we are onlv considering
Fig. 2. Sample speech used for training with its labels, spectrogram and wave form
Fig. 2. Sample speech used for training with its labels, spectrogram and wave form
Table 1. MOS Test Results
Table 1. MOS Test Results
Fig. 1. Constituency representation of qaderecdcaay
Fig. 1. Constituency representation of qaderecdcaay
Fig. 3. Compound Splitter: System Data Flow
Fig. 3. Compound Splitter: System Data Flow
Table 1. Rank-wise Distribution The complete rank wise distribution is given in Table[I]
Table 1. Rank-wise Distribution The complete rank wise distribution is given in Table[I]
Appendix-A : Sanskrit Compound Type and Sub-type Classification
Appendix-A : Sanskrit Compound Type and Sub-type Classification
Indian languages is described in Bharati, et. al. (1995; 85-100). This parser is modeled as a bipartite graph matching problem. A statistical parser for analysing Sanskrit is described in Hellwig (2008). The shallow parser of Huet (2006, 2009) uses bare minimum information of transitivity of a verb as a sub-categorisation frame and models it as a graph-matching algorithm. The main purpose of this shallow parser is to filter out non-sensical interpretations. It is therefore natural for Huet to develop small tools such as ‘ca’ handler with more priority to rule out non-grammatical solutions (rather than to develop a full-fledged parser). While designing a grammar based parser, two major design issued one has to address are: a) what should be the level of semantic analysis, and b) which relations to represent in the parsed output. In order to decide on these issues, in what follows, we first look at the Sanskrit grammar to see what kind of semantic relations can be extracted from a language string, precisely where is the information about these relations coded, and whether the extracted relations are from primary sources or secondary. Later we discuss the issues the mechanical processing throws up, and the possible ways to handle them. Based on these observations, we decide various design parameters. The next section discusses mathematical formulation of the problem, its implementation and finally its performance analysis.
Indian languages is described in Bharati, et. al. (1995; 85-100). This parser is modeled as a bipartite graph matching problem. A statistical parser for analysing Sanskrit is described in Hellwig (2008). The shallow parser of Huet (2006, 2009) uses bare minimum information of transitivity of a verb as a sub-categorisation frame and models it as a graph-matching algorithm. The main purpose of this shallow parser is to filter out non-sensical interpretations. It is therefore natural for Huet to develop small tools such as ‘ca’ handler with more priority to rule out non-grammatical solutions (rather than to develop a full-fledged parser). While designing a grammar based parser, two major design issued one has to address are: a) what should be the level of semantic analysis, and b) which relations to represent in the parsed output. In order to decide on these issues, in what follows, we first look at the Sanskrit grammar to see what kind of semantic relations can be extracted from a language string, precisely where is the information about these relations coded, and whether the extracted relations are from primary sources or secondary. Later we discuss the issues the mechanical processing throws up, and the possible ways to handle them. Based on these observations, we decide various design parameters. The next section discusses mathematical formulation of the problem, its implementation and finally its performance analysis.
Table 1. Relations The relations used, along with their ranks are given in Table 1
Table 1. Relations The relations used, along with their ranks are given in Table 1
Stapll WiC Is tlle WiCallils OL seiucice, In fact, our goal is much bigger than just solving the problem of sabdabodha where one only tries to analyze the meaning of a sentence. Our goal is to con- struct a computational ontology which is foundational as well as generative in nature. We are aiming at generative graph grammar on which comprehensive ontology interpreter can be constructed. Once that task is accomplished, one could apply it to solve domain-specific problems like sabdabodha . Hence the main focus of the paper is on the building of graph grammar suitable for on- tology applications. Idea is to construct computational environment such that applications like sabdabodha can be built. Most of the existing ontologies follow top-down approach they start with the names of the entities (real as well as imaginary) existing in the universe, they try to classify these names and in turn try to classify the actual objects in the universe. But this approach has a problem since in the process of classifying names, one cannot help placing a name under two classes. In other words one name/class might have to have two parent classes. This is called crossing. Fot instance if one has the classes ‘flying objects’ and ‘horses’ in one’s classificatior and now if one encounters a new object called ‘flying horse’, in which of the above classes should he place this object? Obviously, both classes should be its parents since it has the features of both the classes, and hence there is crossing. A good ontological classification is one in which there are no crossings. Having saic this, we now elucidate an alternative bottom-up approach which we take to build not just any ontology but a formal, foundational and generative ontology. Fo1 this we utilized formalization of Vaisestka ontology which is a comprehensive foundational ontology. While formalizing it, there was some difference noted witl classical Vaisesika traditions, hence our approach has been called Neo-Vaisesike Formal Ontology (NVFO).
Stapll WiC Is tlle WiCallils OL seiucice, In fact, our goal is much bigger than just solving the problem of sabdabodha where one only tries to analyze the meaning of a sentence. Our goal is to con- struct a computational ontology which is foundational as well as generative in nature. We are aiming at generative graph grammar on which comprehensive ontology interpreter can be constructed. Once that task is accomplished, one could apply it to solve domain-specific problems like sabdabodha . Hence the main focus of the paper is on the building of graph grammar suitable for on- tology applications. Idea is to construct computational environment such that applications like sabdabodha can be built. Most of the existing ontologies follow top-down approach they start with the names of the entities (real as well as imaginary) existing in the universe, they try to classify these names and in turn try to classify the actual objects in the universe. But this approach has a problem since in the process of classifying names, one cannot help placing a name under two classes. In other words one name/class might have to have two parent classes. This is called crossing. Fot instance if one has the classes ‘flying objects’ and ‘horses’ in one’s classificatior and now if one encounters a new object called ‘flying horse’, in which of the above classes should he place this object? Obviously, both classes should be its parents since it has the features of both the classes, and hence there is crossing. A good ontological classification is one in which there are no crossings. Having saic this, we now elucidate an alternative bottom-up approach which we take to build not just any ontology but a formal, foundational and generative ontology. Fo1 this we utilized formalization of Vaisestka ontology which is a comprehensive foundational ontology. While formalizing it, there was some difference noted witl classical Vaisesika traditions, hence our approach has been called Neo-Vaisesike Formal Ontology (NVFO).
The above details can be concisely captured in the following diagram. If the match m fixes an occurrence of L in a given graph G, then G 24H denotes the direct derivation where p is applied to G leading to a derived graph HZ. Intuitively, H is obtained by replacing the occurrence of L in G by R. Negative Application Conditions (NACs). Rules can have exceptions - it may not be likely to apply a rule in some particular cases. Those cases/conditions can also be depicted in the form of graphs. Hence each rule may (or may not) have one or more NACs. So a rule will be applied only when its LHS matches with some subgraph of the host graph, and none of its NACs matches with any subgraph of the host graph.
The above details can be concisely captured in the following diagram. If the match m fixes an occurrence of L in a given graph G, then G 24H denotes the direct derivation where p is applied to G leading to a derived graph HZ. Intuitively, H is obtained by replacing the occurrence of L in G by R. Negative Application Conditions (NACs). Rules can have exceptions - it may not be likely to apply a rule in some particular cases. Those cases/conditions can also be depicted in the form of graphs. Hence each rule may (or may not) have one or more NACs. So a rule will be applied only when its LHS matches with some subgraph of the host graph, and none of its NACs matches with any subgraph of the host graph.
Rule 1 (Start). We start by replacing the start node, ‘G’, with anonymous entity. ‘BE’, and variable node, ‘X’, which are self-linked.
Rule 1 (Start). We start by replacing the start node, ‘G’, with anonymous entity. ‘BE’, and variable node, ‘X’, which are self-linked.
Rule 7 (MereologyConjunct). Here the middle graph represents LHS of this rule and the rightmost one, its RHS. It says that when some entity inheres in two othe entities, a conjunct can be created in between them on the condition (NAC) that there is already no conjunct between them, which is represented by the leftmost graph in the above image. This rule is to model the conjunct between parts of a whole. Rule 6 (Disjunct). LHS is replaced by separable punctuator where the new entity (£) and the old entity (1:£) are in disjunct with each other. In the process ‘D’ is generated which inheres in both the entities.
Rule 7 (MereologyConjunct). Here the middle graph represents LHS of this rule and the rightmost one, its RHS. It says that when some entity inheres in two othe entities, a conjunct can be created in between them on the condition (NAC) that there is already no conjunct between them, which is represented by the leftmost graph in the above image. This rule is to model the conjunct between parts of a whole. Rule 6 (Disjunct). LHS is replaced by separable punctuator where the new entity (£) and the old entity (1:£) are in disjunct with each other. In the process ‘D’ is generated which inheres in both the entities.
Rule 8 (MereologyDisjunct). It says that when some entity inheres in two other entities, a disjunct can be created in between them on the condition (NAC) that there is already no disjunct between them.
Rule 8 (MereologyDisjunct). It says that when some entity inheres in two other entities, a disjunct can be created in between them on the condition (NAC) that there is already no disjunct between them.
Rule 1 (DefineSubstantialWhole). A substantial whole is one in which something inheres, and also it itself inheres in some other entities. Hence ‘E” is replaced by ‘S,,’ if at least one entity inheres in it, and it itself inheres in minimum two other entities (It has to inhere in minimum two entities for it to be differentiated from Qualities and Actions which can inhere only in one entity).
Rule 1 (DefineSubstantialWhole). A substantial whole is one in which something inheres, and also it itself inheres in some other entities. Hence ‘E” is replaced by ‘S,,’ if at least one entity inheres in it, and it itself inheres in minimum two other entities (It has to inhere in minimum two entities for it to be differentiated from Qualities and Actions which can inhere only in one entity).
Rule 9 (DefineQuality(a)). In a Quality at least one Universal inheres, and it itself inheres in one and only one whole (its inherence in two or more wholes is handled in NACs).
Rule 9 (DefineQuality(a)). In a Quality at least one Universal inheres, and it itself inheres in one and only one whole (its inherence in two or more wholes is handled in NACs).
inheres in one and only one whole.
inheres in one and only one whole.
Prakarata-vigesyata analysis within tinanta word: In the case of verbs, root meaning is “head” and suffix meanings are modifiers. The tinanta suffix has four meanings (1) ka@raka, (2) purusa, (3) vacana and (4) kala. Tense or mood is related with the action and person and number are related with karakas expressed by a tin suffix. Overall two karakas can be expressed by a tin suffix and in the
Prakarata-vigesyata analysis within tinanta word: In the case of verbs, root meaning is “head” and suffix meanings are modifiers. The tinanta suffix has four meanings (1) ka@raka, (2) purusa, (3) vacana and (4) kala. Tense or mood is related with the action and person and number are related with karakas expressed by a tin suffix. Overall two karakas can be expressed by a tin suffix and in the
The tinianta suffix has four meanings (1) kdraka, (2) purusa, (3) vacana and (4) kala. The subanta ending in the first case-ending is identified with the meaning karaka of the tinanta suffix. Therefore, it is placed or moved in the karaka meaning. Other two meanings of the tinianta suffix namely, purusa, and vacana become the prakara of the subanta in the nominative case, linked with the karaka meaning of the tinanta suffix. The complete sentence meaning after prakarata and vigesyatad analysis will be as follows.(Table.4) If the sentence is transformed to or is in 1 the passive or impersonal | form then ee ee | lem ry rH ¥ 7— _ “- q
The tinianta suffix has four meanings (1) kdraka, (2) purusa, (3) vacana and (4) kala. The subanta ending in the first case-ending is identified with the meaning karaka of the tinanta suffix. Therefore, it is placed or moved in the karaka meaning. Other two meanings of the tinianta suffix namely, purusa, and vacana become the prakara of the subanta in the nominative case, linked with the karaka meaning of the tinanta suffix. The complete sentence meaning after prakarata and vigesyatad analysis will be as follows.(Table.4) If the sentence is transformed to or is in 1 the passive or impersonal | form then ee ee | lem ry rH ¥ 7— _ “- q
prakrtyartha-ramanistha-prakarata-nirt pita-pratvayartha-kartr-nistha-visesvata- nistha-prakarata-nirt pita-prathamapurusa-ekavacana-nistha-prakarata-niru pita- prakrtvartha-gramanistha-prakarata-nirt pita-pratvavartha-gramanistha-visesvata- abhinna-karmanistha-visesvata-nist ha-prakarata-niru pita-vartamana-nistha-prakarata- OR yD Se) Ne ey Eee Te eee Seen. ee ee, ee ew |
prakrtyartha-ramanistha-prakarata-nirt pita-pratvayartha-kartr-nistha-visesvata- nistha-prakarata-nirt pita-prathamapurusa-ekavacana-nistha-prakarata-niru pita- prakrtvartha-gramanistha-prakarata-nirt pita-pratvavartha-gramanistha-visesvata- abhinna-karmanistha-visesvata-nist ha-prakarata-niru pita-vartamana-nistha-prakarata- OR yD Se) Ne ey Eee Te eee Seen. ee ee, ee ew |
Even though the suffix in the impersonal verb conveys two meanings (1) kala, tense and (2) bhava, the later meaning does not figure in the head-modifier analysis because the meaning bhava actually means the action in general and therefore is connected to the meaning of the verbal root with non-difference (abheda). The head-modifier analysis of the sentence of the impersonal voice will be as shown below. (Table.6)
Even though the suffix in the impersonal verb conveys two meanings (1) kala, tense and (2) bhava, the later meaning does not figure in the head-modifier analysis because the meaning bhava actually means the action in general and therefore is connected to the meaning of the verbal root with non-difference (abheda). The head-modifier analysis of the sentence of the impersonal voice will be as shown below. (Table.6)
Table 8. Hierarchical bracket parsing of the active voice sentence ramah gramam gacchati Hierarchical Bracket Parsing
Table 8. Hierarchical bracket parsing of the active voice sentence ramah gramam gacchati Hierarchical Bracket Parsing
Table 6. Groups and Characters Present in the Groups
Table 6. Groups and Characters Present in the Groups
Table 7. Partial Result Set for Approximate Matching
Table 7. Partial Result Set for Approximate Matching
Table 1. Structure of Mahabharata (According to the BORI’s critical edition) this point. MBh is neither a history in the modern sense of the term, nor chronicle. But it stands in incomparable isolation, defying all definitions. It is’ veritable encyclopaedia comprising heterogeneous material from all branches c knowledge. Taking the core-story of the feud between two branches of a roye family and the circumstances leading to a catastrophic war, several branche of knowledge including philosophy, law, ethics, statecraft, warfare, history an ethnology are embodied in its structure. The MBh is a comment on the huma condition with all its richness, complexity and subtlety. The MBh is the source c many compositions like- abhijndanasakuntalam of Kalidasa, naisadhiya-caritar of sriharsa etc. It is the text that is most sought for in order to enrich cultura social and any other type of knowledge about Indian civilization. While on the one hand, the text is very important, on the other, it is so hug
Table 1. Structure of Mahabharata (According to the BORI’s critical edition) this point. MBh is neither a history in the modern sense of the term, nor chronicle. But it stands in incomparable isolation, defying all definitions. It is’ veritable encyclopaedia comprising heterogeneous material from all branches c knowledge. Taking the core-story of the feud between two branches of a roye family and the circumstances leading to a catastrophic war, several branche of knowledge including philosophy, law, ethics, statecraft, warfare, history an ethnology are embodied in its structure. The MBh is a comment on the huma condition with all its richness, complexity and subtlety. The MBh is the source c many compositions like- abhijndanasakuntalam of Kalidasa, naisadhiya-caritar of sriharsa etc. It is the text that is most sought for in order to enrich cultura social and any other type of knowledge about Indian civilization. While on the one hand, the text is very important, on the other, it is so hug
Table 2. The ‘shloka’ table
Table 2. The ‘shloka’ table
Table 5. The ‘upaparva’ table Table 4. The ‘akhyana’ table
Table 5. The ‘upaparva’ table Table 4. The ‘akhyana’ table
Development of the MBh Indexer
Development of the MBh Indexer
Step II: MBh Indexer and Database. At this step, the indexer makes an indexed list of exact and partially matching words. Getting the query as an input, the indexer, after a light preprocessing, sends it to the database. If the word has its occurrence in the database, the system gives the output.
Step II: MBh Indexer and Database. At this step, the indexer makes an indexed list of exact and partially matching words. Getting the query as an input, the indexer, after a light preprocessing, sends it to the database. If the word has its occurrence in the database, the system gives the output.
Fig.1. Multi-tiered architecture of the MBh indexing system
Fig.1. Multi-tiered architecture of the MBh indexing system
Fig. 4. Search page of the MBh indexing system Fig. 3. Homepage of the MBh Indexing System
Fig. 4. Search page of the MBh indexing system Fig. 3. Homepage of the MBh Indexing System
The front-end of the system is developed in utf-8 enabled Java Server Pages (JSP) and HTML. The front-end of the software enables the user to interact with the indexing system with the help of Apache Tomcat web-server. The JSP technology helps to create web based applications by combining Java code with HTML. The web server runs the Java code and displays the results as HTML. For this system, there are two JSP pages, one is the main search page and the other is the cross-referential search page which searches the searched query in different online lexical and linguistic resources. The snapshots of the indexing system are as follow:
The front-end of the system is developed in utf-8 enabled Java Server Pages (JSP) and HTML. The front-end of the software enables the user to interact with the indexing system with the help of Apache Tomcat web-server. The JSP technology helps to create web based applications by combining Java code with HTML. The web server runs the Java code and displays the results as HTML. For this system, there are two JSP pages, one is the main search page and the other is the cross-referential search page which searches the searched query in different online lexical and linguistic resources. The snapshots of the indexing system are as follow:
Fig. 8. Diagram of the interconnection between the database tables A design of the indexing system of MBh database is given below: 5.4 Database Connectivity
Fig. 8. Diagram of the interconnection between the database tables A design of the indexing system of MBh database is given below: 5.4 Database Connectivity
to lose important information. This tagset does not include attributes of lexical categories such as gender, number, person etc. Apart from this the tags for incorporated adjective and adverb are not applicable in Sanskrit. In Sanskrit, like noun and verb, Particles also require more tags as they have different and valuable function in the language, so in order to suit Sanskrit this tagset will have to be modified significantly. Keeping in mind the complex morphology of Sanskrit this flat tagset will hardly be able to do WSD.
to lose important information. This tagset does not include attributes of lexical categories such as gender, number, person etc. Apart from this the tags for incorporated adjective and adverb are not applicable in Sanskrit. In Sanskrit, like noun and verb, Particles also require more tags as they have different and valuable function in the language, so in order to suit Sanskrit this tagset will have to be modified significantly. Keeping in mind the complex morphology of Sanskrit this flat tagset will hardly be able to do WSD.
Table 2. JNU-Sanskrit Tagset This is a flat and fine-grained tagset capturing linguistic features accordi (o the traditional Sanskrit grammar. Its coverage is very high. Almost eve <ind of linguistic item was treated discretely (especially taddhita and kridan constructions). This tagset is very informative but we assume that its specifici will suffer while tagging the data. For instance, among noun tags there are 8 ta and the basis of counting them separately seems to be the special morpholo; ‘or these items like Patronymic Noun (e.g. raghavah, janaki), Metronymic Nov e.g. parthah, saumitrih), (both tags could be treated as Proper Noun), Age Noun (e.g. kartuh, adhyapakah) (could be treated as Common Noun). Likewi ander Compound category there are 14 types. Many of which could be clubbe
Table 2. JNU-Sanskrit Tagset This is a flat and fine-grained tagset capturing linguistic features accordi (o the traditional Sanskrit grammar. Its coverage is very high. Almost eve <ind of linguistic item was treated discretely (especially taddhita and kridan constructions). This tagset is very informative but we assume that its specifici will suffer while tagging the data. For instance, among noun tags there are 8 ta and the basis of counting them separately seems to be the special morpholo; ‘or these items like Patronymic Noun (e.g. raghavah, janaki), Metronymic Nov e.g. parthah, saumitrih), (both tags could be treated as Proper Noun), Age Noun (e.g. kartuh, adhyapakah) (could be treated as Common Noun). Likewi ander Compound category there are 14 types. Many of which could be clubbe
Table 3. LDCIL Tagset under other categories as well. For example, the words like rajapurusah and devendrah have been assigned Determinative Compound (genitive) which could be very well given Common noun and Proper noun respectively. The division of participle tags is also very lengthy; kridanta vartamana has two separate tags which does not seem serving any purpose and the same is with kridanta vidhyarthaka which has 3 tags.
Table 3. LDCIL Tagset under other categories as well. For example, the words like rajapurusah and devendrah have been assigned Determinative Compound (genitive) which could be very well given Common noun and Proper noun respectively. The division of participle tags is also very lengthy; kridanta vartamana has two separate tags which does not seem serving any purpose and the same is with kridanta vidhyarthaka which has 3 tags.
Table 4. IL-POSTS Sanskrit tagset categories and their Types Table 5. IL-POSTS Sanskrit tagset attributes and their values
Table 4. IL-POSTS Sanskrit tagset categories and their Types Table 5. IL-POSTS Sanskrit tagset attributes and their values
In the introduction, we mentioned the theory that the amount of training data influences the analysis accuracy. To assess this theory, the analysis of the RASAMANJARI was repeated using statistical key values that were calculated from randomly selected subsets of the text corpus. Figure[3] displays graphically the influence between the amount of training data and the analysis quality. Although the analysis accuracy obviously increases with the amount of training data, we achieve comparatively high rates of accuracy even when only 20% of the corpus are used for training. Therefore, it seems that substantial increases in the accuracy of the tagger cannot be achieved by increasing the size of the corpus, but only by improving the analysis procedures. Figure [J] displays the boxplots of H, Hpzx and Epog. Since the analysis of Sanskrit differs strongly from that of other languages, H and Hrpx are hardly comparable with rates recorded in previous research. Obviously, however, both A and Hyex need significant improvements in future versions of the program. If we take into account the complex considerations that the linguistic phenomena of composite creation and bahuvrihi make necessary for POS tagging of Sanskrit, the POS accuracy of SanskritTagger compares acceptably with POS error rates reported for less complex languages (J4e Pos = 0.090, o¢ pos = 0.035). To go deeper into the details of table [] we should try to determine the influence of literary features on the analysis quality. If lexical richness and the average nhrace lencth are used to vet an estimation of the complexity of a literarv
In the introduction, we mentioned the theory that the amount of training data influences the analysis accuracy. To assess this theory, the analysis of the RASAMANJARI was repeated using statistical key values that were calculated from randomly selected subsets of the text corpus. Figure[3] displays graphically the influence between the amount of training data and the analysis quality. Although the analysis accuracy obviously increases with the amount of training data, we achieve comparatively high rates of accuracy even when only 20% of the corpus are used for training. Therefore, it seems that substantial increases in the accuracy of the tagger cannot be achieved by increasing the size of the corpus, but only by improving the analysis procedures. Figure [J] displays the boxplots of H, Hpzx and Epog. Since the analysis of Sanskrit differs strongly from that of other languages, H and Hrpx are hardly comparable with rates recorded in previous research. Obviously, however, both A and Hyex need significant improvements in future versions of the program. If we take into account the complex considerations that the linguistic phenomena of composite creation and bahuvrihi make necessary for POS tagging of Sanskrit, the POS accuracy of SanskritTagger compares acceptably with POS error rates reported for less complex languages (J4e Pos = 0.090, o¢ pos = 0.035). To go deeper into the details of table [] we should try to determine the influence of literary features on the analysis quality. If lexical richness and the average nhrace lencth are used to vet an estimation of the complexity of a literarv
Table 1. Tokens and types in each kandas Polysemy. Amarakosga has 4,017 synsets. Some of the words fall under mor than one synset, and thus are ambiguous. Most of these polysemous words belon; to the nanarthavarga of the third kanda which lists the polysemous words alpha betically according to their endings. The polysemy distribution in the Amarakos. is summarised in Table 2. There is only one word hari in Amarakosga which ha as many as 14 senses, the word antara belongs to 13 synsets, and the word g has 12 synsets. We note that almost 65% words (7459 words) belong to a singl synset and thus are not ambiguous.
Table 1. Tokens and types in each kandas Polysemy. Amarakosga has 4,017 synsets. Some of the words fall under mor than one synset, and thus are ambiguous. Most of these polysemous words belon; to the nanarthavarga of the third kanda which lists the polysemous words alpha betically according to their endings. The polysemy distribution in the Amarakos. is summarised in Table 2. There is only one word hari in Amarakosga which ha as many as 14 senses, the word antara belongs to 13 synsets, and the word g has 12 synsets. We note that almost 65% words (7459 words) belong to a singl synset and thus are not ambiguous.
Table 2. Polysemy distribution from the meta-language indicating the gender of the following word siva to be feminine. In addition to these general guidelines, in the lingadisangrahavarga Amarasimha gives certain grammatical and phonological clues for deciding the gender of a word. In the event of absence of any rule, the gender of the remaining words in 2.5.5, constituting two compound words” bhurimayagomayumrgadhurta kah” and ” srgalavancakakrostupherupheravajambukah” is inferred to be mascu- line from their compounding-forms. The avvavavarea, lists svnsets consistine of indeclinables.
Table 2. Polysemy distribution from the meta-language indicating the gender of the following word siva to be feminine. In addition to these general guidelines, in the lingadisangrahavarga Amarasimha gives certain grammatical and phonological clues for deciding the gender of a word. In the event of absence of any rule, the gender of the remaining words in 2.5.5, constituting two compound words” bhurimayagomayumrgadhurta kah” and ” srgalavancakakrostupherupheravajambukah” is inferred to be mascu- line from their compounding-forms. The avvavavarea, lists svnsets consistine of indeclinables.
Table 3. Words and references of the svarga-synset
Table 3. Words and references of the svarga-synset
Table 4. Example of Head-Word The relations are among various Head Words and are marked as records. Each record corresponds to one synset ID. The first field of each record consists of the synset ID, and remaining six fields correspond to the Head Words that bear a relation of is_a_part_of (avayavavayavi), is.a_kind_of(pardparajati), janya- janaka-bhava, pati-patni-bhava, sva-svami-bhava, ajivika with the synset ID in the first field.
Table 4. Example of Head-Word The relations are among various Head Words and are marked as records. Each record corresponds to one synset ID. The first field of each record consists of the synset ID, and remaining six fields correspond to the Head Words that bear a relation of is_a_part_of (avayavavayavi), is.a_kind_of(pardparajati), janya- janaka-bhava, pati-patni-bhava, sva-svami-bhava, ajivika with the synset ID in the first field.
Table 6. Example of is_a_kind_of relations
Table 6. Example of is_a_kind_of relations
Table 8. Example of Pati-patni relation
Table 8. Example of Pati-patni relation
Table 10. Example of Ajzvika relation
Table 10. Example of Ajzvika relation
Table 11. Relational statistics
Table 11. Relational statistics
Fig. 2. Histogram showing the WSD accuracy across domains for Hindi Words Table 2. WSD accuracy across domains for Hindi Words
Fig. 2. Histogram showing the WSD accuracy across domains for Hindi Words Table 2. WSD accuracy across domains for Hindi Words
Table 5. Possible synsets for the word hari in the present Sanskrit Wordnet
Table 5. Possible synsets for the word hari in the present Sanskrit Wordnet
Table 3. Differential Divergence
Table 3. Differential Divergence
Table 4. Alternative Divergence 2.5 Non-karaka Divergence
Table 4. Alternative Divergence 2.5 Non-karaka Divergence
Table 5. Non-karaka Divergence 2.6 Hindi Specific Divergences
Table 5. Non-karaka Divergence 2.6 Hindi Specific Divergences
Table 1. Sanskrit anaphora Pronominal Anaphors. The major classification in pronominal is the first, second, and third person pronoun. First and second person are commonly used as deictic, though they are used in discourse. The third person pronoun is the most commonly used anaphora in a language. In Sanskrit tat is the third person pronoun which has the following separate forms in different cases for masculine and feminine gender|] rT 4 row . r << whee wih 7 a | as 4 =a os AMsyye 4Y
Table 1. Sanskrit anaphora Pronominal Anaphors. The major classification in pronominal is the first, second, and third person pronoun. First and second person are commonly used as deictic, though they are used in discourse. The third person pronoun is the most commonly used anaphora in a language. In Sanskrit tat is the third person pronoun which has the following separate forms in different cases for masculine and feminine gender|] rT 4 row . r << whee wih 7 a | as 4 =a os AMsyye 4Y
Table 2. Forms of pronominal tat in different cases
Table 2. Forms of pronominal tat in different cases
Fig. 1. Computational Structure of Astadhyayt with Asiddhatva Principle
Fig. 1. Computational Structure of Astadhyayt with Asiddhatva Principle
Fig. 6. Step 3: Selecting one rule from the set of applicable rules
Fig. 6. Step 3: Selecting one rule from the set of applicable rules
Related topics:
Optimality Theory
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved