Paninian Grammar Based Hindi Dialogue Anaphora Resolution

Transcription

1 Paninian Grammar Based Hindi Dialogue Anaphora Resolution by darshan.agarwal, vandan.mujadia, Radhika Mamidi, Dipti Misra Sharma in The 19th International Conference on Asian Language Processing (IALP) Soochow University, China Report No: IIIT/TR/2015/-1 Centre for Language Technologies Research Centre International Institute of Information Technology Hyderabad , INDIA October 2015

2 Paninian Grammar Based Hindi Dialogue Anaphora Resolution Vandan Mujadia, Darshan Agarwal, Radhika Mamidi, Dipti Misra Sharma Language Technology Research Centre, IIIT - Hyderabad {vandan.mujadia, {radhika.mamidi, Abstract In this paper, we present a Paninian grammar based heuristic model 1 to resolve entitypronoun references in Hindi dialogue. We explore the use of Paninian based dependency structures as a source of syntactico-semantic information. Our experiments illustrate that the use of dependency and dialogue structures help to resolve specific types of references. We also show that named entity, discourse information like subtopic boundary and animacy features increase the overall resolution accuracy to 64% for user-user interaction data and 59% for playstory corpora. I. Introduction Any automated NLP system is evaluated upon its ability to relate one piece of text to another piece of text in a given discourse. In order to achieve such coherence, system needs to identify and resolve cohesive constructions like anaphora occurring in the given discourse. Anaphora resolution plays one of the important roles in any NLP system. According to Hist (1981) in [1] : Anaphora in discourse, is a device for making an abbreviated reference (containing fewer bits of disambiguating information, rather than being lexically or phonetically shorter) to some entity (or entities) in exceptions that the receiver of the discourse will be able to dis-abbreviate the reference and thereby, determine the identity of the entity. For Hindi language, anaphoric reference type can be classified into abstract (event) and concrete (entity) references where an abstract (event) anaphora refers to an event or a proposition and while concrete (entity) anaphora can refer to a concrete entity like noun phrase (person, place etc), quantifiers, etc. We focused on resolving only entity pronouns for Hindi dialogue. In this paper, we show how the structure of dialogue helps to resolve concrete anaphora with Computational Paninian Grammar (CPG) framework as a main parameter. Other important features which help in anaphora resolution include named entity, animacy and subtopic boundary information. 1 AnaphoraResolution The corpus detail is discussed in section III. Section IV describes Hindi dialogue anaphora resolution algorithm for different pronoun types and we will conclude our claim in section V. II. Background Discourse structure with syntactic information has been used as an important feature in various rule based and data driven approaches for anaphora resolution like Hobbs algorithm [2] and Brennans Centering Theory [3]. However for dialogue, there are few dialogue structure (discourse-oriented approaches) based approaches like Groszs [4], Eckerts [5] and Martinezs [6]. Niraula [7] and Eckert [8] showed anaphora resolution based on syntactic and semantic features. These approaches are primarily demonstrated for English. Like Hindi, most Indian languages are relatively free word ordered. Paninian grammar based dependency structures are suitable representations for relatively free word order languages as described by Bharti [9] and Melcuk [10]. For English, Jorkelund [11] explores the possibility of using dependency relations as a feature for coreference resolution in a learning based approach. Most of the approaches are limited in their exploration of dependency for anaphora resolution as they only use dependency relations either as a salience for ranking the candidate referents or as an additional feature in a learning approach. Prasad [12] applied a discourse salience ranking to pronoun resolution algorithms, the BFP and the S-List algorithm. For Hindi, Uppalapu [13] extends the S-List algorithm by using two different lists in place of a single list. Dakwale [14] reported the best performance in Anaphora Resolution on news corpus by proposing a hybrid approach with limited linguistic knowledge such as Named Entity(NE) categories and verb similarity. However for Hindi natural dialogue, we are the first to tackle this anaphora resolution problem. III. Corpus details and representation We evaluated our experiments on three different kinds of anaphora annotated dialogue corpora. They are user-user

3 Table I: Corpus Details Type Conversations Utterances Sentences Pronouns Play - Story[PS] User - User[UU] User - System[US] interaction(chat), user-system interaction and Hindi playstory corpus. Table I shows the number of conversations, sentences and pronouns for each corpus. Table II shows the pronoun type distribution. We observe that personal pronouns have relatively high presence in dialogue corpora. All corpora have rich linguistic information from shallow features to deep features like dependency labels, animacy, speaker and subtopic boundary. The corpus is annotated in Shakti Standard Format (SSF) which is a highly readable and easily accessible textual representation for language analysis as showed by Bharti [15]. Table II: Anaphora distribution Type FP-SP TP Loc Ref Rel Play-Story[PS] User-User[UU] User-System[US] Label CPG relation Grammatical/thematic equivalent k1 karta Subject k2 karma Object k4 sampradan Experiencer/reciever k7p(or k2p) adhikaran Location r6 sambandh Genitive Table III: CPG relations and equivalents The dependency annotation in corpora were based on the Computational Paninian Grammar (CPG-henceforth) framework, as explained in [16] and [17]. This framework is based on the karaka notations which are syntacticosemantic relations representing the participant elements in the action specified by the verb. It emphasizes the role of case endings or markers such as post-positions and verbal inflections. Table III describes the karaka based dependency label with CPG relation and its thematic role equivalent label. The anaphora annotation in these corpora are as per the annotation guidelines described in Dakwale [18]. The speaker and subtopic boundary are marked accordingly in the corpus. IV. Anaphora resolution In this section we describe the set up for Hindi dialogue anaphora resolution from various experiments. Anaphora resolver requires the output of full parser/analyze 2 and subtopic boundary identifier 3 in SSF format. The mention detection or anaphora determination step is not done by anaphora resolution module but it is triggered, when shallow parser gives pronoun as POS tag for particular word in given dialogue. For the identified mention/pronoun, anaphora resolution module tries to find the antecedent. For the Hindi language, as concluded in [14] certain pronominal forms are based on their different syntactic behavior and they can be resolved quite successfully with some specific heuristic rules using the dependency information. Therefore we categorize pronominal forms into reflexive, relative, locative and personal. The anaphora resolver tries to resolve a pronoun differently based on their pronominal types. A. First and Second person pronoun resolution (FP-SP) It is obvious that for any natural dialogue, there are high occurrences of first and second person pronouns (Table II). As the speaker marking module 4 marks the speaker per utterance of the given dialogue, the resolution of first and second person pronoun becomes relatively easier. First person pronoun can be resolved directly by assigning current speaker of utterance as its referent. Similarly for second person pronoun resolution, one can assign previous speaker or current listener as referent of pronoun. Only with this two consideration we are able to achieve relatively high accuracy for first and second person pronoun resolution. B. Third-person pronoun resolution (TP) Mostly Third-person pronouns are either inter-clausal or inter-sentential and for their resolution, similar to [14], we used salience of dependency relation for reordering of the candidate mentions with some modifications. Such as instead of salience ordering (k1 > k2 > r6 > k4 > k3 > others) based on relative frequency of dependency relation for candidate mention ranking, we modified ordering of mentions selection to (k1 > r6 > k2 >k4 > others) considering speaker possessiveness (sentences are relatively small as compare to other text and they are more inclined towards speaker) in dialogue. These assumptions guide us to modify salience order by giving less preference to object compare to possessive relation. The subtopic boundary with gender information helps to prune the unlikely mention list for this pronoun category. Example 4.1: र म : [k1स दर ] i [k7tकब] ज रह ह? Ram : Sundar when go+present-cont. Ram: when is Sundar i going? म हन : [k1वह] i [k7pइ डय ] [k7tश नव र क ] ज एग Mohan : he India Saturday.LOC go.future. Mohan: He i will go to India on Saturday. In the above example (4.1) the pronoun वह (he) is gender neutral. For this pronoun, According to our algorithm, mention स दर (sundar) with k1 (doer) dependency label is in same subtopic boundary with the pronoun and has 4 integral part of subtopic identification module

4 (1)move up to verb घर (home) r6 अपन (own) k7p आय थ (came) k2 आज ह (today only) Figure 1 (2)select k1 k1 गण श (Ganesh) agreement on number and animacy, so it is selected as the referent. C. Reflexive pronoun resolution (Ref) In Hindi dialogue, reflexive pronouns are categorized into possessive reflexive which include अपन, अपन (inflected forms of own ) and other types of reflexives for describing objects like various forms of oneself [सवय, ख द ]. As mentioned in [14] the referent of the reflexive pronoun is the accessible subject in its own governing category. Also, the k1 relation of Paninian based framework roughly corresponds to SUBJECT of the traditional framework. Sentences of dialogues are mostly restricted by its length, so less dependency confusions, therefore in most of the cases the referent of the reflexive pronoun is the k1 of the same clause. Example 4.2: गण श i आज ह आय थ अपन i घर स Ganesh.ERG today only come+past his.poss.ref home from Ganesh i come today only from (his own) i home. Figure (1) shows the dependency structure of example (4.2). The verb of the clause containing possessive reflexive अपन (his/her own) is आय थ (came). The verb has a descendant node गण श (Ganesh) with a dependency relation k1 with the verb. Thus according to the rule गण श (Ganesh) should be selected as the referent of the pronoun. D. Locative pronoun resolution (Loc) In Hindi, locative pronouns include वह ( there ), यह ( here ) and their inflected forms. The referent for locative pronoun can be identified based on Location as NE type within same subtopic boundary. Mentions with k2p or k7p as dependency relation with corresponding verb represent location in theta role and it can act as a referent of given locative pronoun. Better results were observed when we considered animacy feature with karaka labels to prune the candidate mention list. Another pattern we observed that if any locative pronoun other than current pronoun occurred in same subtopic boundary then there are high chances that first occurred locative pronoun is referent of the current locative pronoun. Example 4.3: [ NER=LC ग ल ड क स ट ] i गय थ Goldcoast.LOC go+past. Went to Goldcoast i. वह i क तस व र ह there.loc (the photos) be+pres The photos are from there i. select k7p गय थ (had gone) k7p ग ल ड क स ट i (Goldcoast) Figure 2 k1 - (NONE) Figure (2) shows the dependency structure of example (4.3). In the example the locative pronoun वह ( there ) is the descendant of the verb ह (is). The previous sentence has a descendant node ग ल ड क स ट (Gold Coast) with a dependency relation k7p and location as NE type with the verb (गय थ - went). Thus it should be selected as the referent of the pronoun. E. Relative pronoun resolution (Rel) In Hindi, relative pronouns include ज (which) and its case forms like जस (to which), जनम (in which), जसस (from which) etc. According to Paninian framework, relative clauses are marked with dependency label nmod-relc within same clause. The referent of relative pronoun should be selected as the noun-phrase to which the relative pronoun is attached. We used salience of dependency relation for reordering of the candidate mentions ( nmod-relc > k1 > r6 > k2 > k4 > others) based on relative frequency of dependency relation. We incorporated subtopic boundary with animacy and gender to prune the mentions which are likely to be candidate. Example 4.4: र म : [adj स दर ] [k2 तस व र ] i [k5 भ ई] Ram : beautiful photos brother. Ram: Beautiful photos i brother. म हन : [k7 जनम ] i [k1गण श] ह [व ]? Mohan : (in which) Ganesh is there that? Mohan: That (in which) i Ganesh is there? In the above example (4.4) the pronoun जनम (in which) is resolved based on dependency wise salience ordering of the possible referents for this pronouns is [(तस व र - k2), (भ ई - k5)].(तस व र ) agrees with the pronoun in number,animacy and in same subtopic boundary, so it is selected as the referent for जनम (in which). F. Experiments We developed rule based anaphora resolver by incrementally adding different syntactic, semantic and discourse features with optimal combination. All above pronoun wise algorithms are used from second experiment and in third-fourth experiments we added other semantic and structure features. In the first experiment, for a given pronoun irrespective of pronoun type, we considered only shallow features like POS, morph and vibhakti. In the second experiment, we have added Paninian dependency grammar based karaka labels and speaker [SPR] as feature rules with existing shallow features and got comparatively better accuracy than the first experiment. In the third experiment, as feature rules we added animacy and named

5 entity information. Which further improved the overall result. For further improvement, in the fourth experiment we used subtopic boundary as false positive mention pruner. During those experiments we observed the most effective advantage of dialogue structure in dependency based anaphora resolution is due to relatively smaller length of utterance as compared to monologue text. Hence, there is less ambiguity in karaka role assignment and because of this advantage there is very less confliction in dependency based rule application for anaphora resolution. Table IV: Results : Anaphora Resolution Type Features #Pro #Res. UU Shallow features[sf] UU SF+Speaker[SPR]+Panian Grammar[CPG] UU SF+SPR+CPG+Animicy[AM]+Named Entity[NE] UU SF+SPR+CPG+AM+NE+SPR+Subtopic Boundary PS Shallow features[sf] PS SF+Speaker[SPR]+Panian Grammar[CPG] PS SF+SPR+CPG+Animicy[AM]+Named Entity[NE] PS SF+SPR+CPG+AM+NE+Subtopic Boundary Table V: Results : Type-wise Anaphora Resolution for [SF+SPR+CPG+AM+NE+LCG] Type Corpus #Total #Resolved Accuracy FP and SP User - User[UU] TP User - User[UU] Reflexive User - User[UU] Relative User - User[UU] Locative User - User[UU] FP and SP Play - Story[PS] TP Play - Story[PS] Reflexive Play - Story[PS] Relative Play - Story[PS] Locative Play - Story[PS] V. Results and Conclusion From the results we observed that the use of subtopic boundary within Paninian dependency framework, reduces the search space for antecedent identification and also gives a chance to highly possible referents of same subtopic. But upto some extent, the clausal anaphora especially relative pronoun in complex sentences cause difficulty in resolution because the resolution module thoroughly depends on karaka based dependency parser. Table - IV shows the feature wise accuracy for two corpora whereas table - V describes the pronoun type wise accuracy(acc) for best performing feature set obtained from table - IV. There are also many difficulties in the identification of the correct subtopic boundary which was because of the incompleteness of existing resources like Hindi WordNet. We found that subtopic boundary evaluation is also an ambiguous task for evaluators. From this work, we claim that factors like Paninian dependency grammar, subtopic boundary, animacy and speaker identification contribute considerably to anaphora resolution in dialogue. In the current work, we experimented only on user-user interactions and playstory corpora. We were unable to perform this experiment on user-system interaction data due to sparsity of such an anaphora annotated corpus in Hindi and other Indian languages. In future, this work can be used as a baseline for further improvement of dependency based anaphora resolution in Hindi and other Indian languages. In future we also aim to identify the event and noun-to-noun coreference for Hindi natural dialogue using several language specific features. References [1] M. Palomar and P. Martínez-Barco, Computational approach to anaphora resolution in Spanish dialogues, J. Artif. Intell. Res. (JAIR), vol. 15, pp , [2] J. Hobbs, Resolving pronoun references, in Readings in natural language processing. Morgan Kaufmann Publishers Inc., [3] S. E. Brennan, M. W. Friedman, and C. J. Pollard, A centering approach to pronouns, in Proceedings of 25th ACL, [4] B. J. Grosz et al., The representation and use of focus in a system for understanding dialogs, in IJCAI, vol. 67, 1977, p. 76. [5] M. Eckert and M. Strube, Dialogue acts, synchronizing units, and anaphora resolution, Journal of Semantics, vol. 17, no. 1, pp , [6] P. Martínez Barco, R. Muñoz Guillena, S. Azzam, M. Palomar Sanz, A. Ferrández Rodríguez et al., Evaluation of pronoun resolution algorithm for spanish dialogues, [7] N. B. Niraula, V. Rus, and D. Stefanescu, Dare: Deep anaphora resolution in dialogue based intelligent tutoring systems, in Proceedings of the 6th International Conference on Educational Data Mining (EDM 2013), 2013, pp [8] M. Eckert and M. Strube, Resolving discourse deictic anaphora in dialogues, in Proceedings of the ninth conference on European chapter of the Association for Computational Linguistics. Association for Computational Linguistics, 1999, pp [9] A. Bharati, V. Chaitanya, R. Sangal, and K. Ramakrishnamacharyulu, Natural language processing: a Paninian perspective. Prentice-Hall of India, [10] I. A. Melcuk, Dependency syntax: theory and practice. State University of New York Press, [11] A. B. Jörkelund and J. Kuhn, Phrase structures and dependencies for end-to-end coreference resolution, [12] R. Prasad and M. Strube, Discourse salience and pronoun resolution in Hindi, U. Penn Working Papers in Linguistics, [13] B. Uppalapu and D. M. Sharma, Pronoun resolution for Hindi. in DAARC-7, [14] P. Dakwale, V. Mujadia, and D. M. Sharma, A hybrid approach for anaphora resolution in Hindi, in Proceedings of the Sixth International Joint Conference on Natural Language Processing, 2013, pp [15] A. Bharati, R. Sangal, and D. M. Sharma, Ssf: Shakti standard format guide, Language Technologies Research Centre, International Institute of Information Technology, Hyderabad, India, pp. 1 25, [16] R. Begum, S. Husain, A. Dhwaj, D. M. Sharma, L. Bai, and R. Sangal, Dependency annotation scheme for indian languages, in Proceedings of IJCNLP, [17] A. Bharati, S. Husain, B. Ambati, S. Jain, D. Sharma, and R. Sangal, Two semantic features make all the difference in parsing accuracy, Proc. of ICON, [18] P. Dakwale, H. Sharma, and D. M. Sharma, Anaphora annotation in Hindi dependency treebank, in Proceedings of PACLIC- 26, 2012.