Discourse Level Tagger for Mah¯Abh¯As.Ya - a Sanskrit Commentary on P¯An

Discourse Level Tagger for Mah¯abh¯as.ya - a Sanskrit Commentary on P¯an. ini’s Grammar

Monali Das Amba Kulkarni

University of Hyderabad University of Hyderabad [email protected] [email protected]

Abstract (Mann and Thompson, 1988) proposed a Rhetorical Structure Theory (RST) for dis- Mah¯abh¯as.ya is an important commen- course analysis. The intentional and informa- tary on P¯an. ini’s grammar for Sanskrit tional relations between adjacent texts are rep- and is highly structured. The tradi- resented in the form of a binary tree structure. tional scholars have tagged it manually (Wolf and Gibson, 2005) suggested a Depen- showing its underlying discourse struc- dency Graph Structure (DG) to show the re- ture. The traditional grammar also dis- lations between the topic and sub-topic seg- cusses clues for discourse level annota- ments. The segments here unlike RST need tions. Taking into account these clues not be adjacent. (Polanyi, 1988) suggested a we have developed an automatic tag- Linguistic Discourse Model (LDM) which re- ger for tagging the Mah¯abh¯as.ya. This sembles RST. In this model each node inherits tagger is described in this paper, along the properties of its parent node and a parent with its performance evaluation. We node is an interpretation of its children nodes have also extended this tag-set to on and the relations between them. another important text S¯abarabh¯as´ ya. . There have been several efforts in the recent 1 Introduction past in computational linguistics in the field of discourse analysis. The major efforts is the Discourse level analysis is an important mod- proposal of discourse tag-set by (Joshi et al., ule in NLP which takes us beyond sentence 2007) in the form of Penn Dependency Tree- level analysis. While grammar is basically Bank (PDTB). The annotation scheme orig- about how words combine to form sentences, inally developed for English was further ex- text and discourse analysis is about how sen- tended to Arabic (Alsaif, 2012), Chinese (Xue, tences combine to form texts (Salike, 1995). 2005), Turkish (Zeyrek and Webber, 2008) and Four types of discourse structures are dis- Hindi (Prasad et al., 2009). (Mladova et al., cussed in the modern linguistics. 2008) suggested a scheme for Czech and En- 1. Topic Structure: The topic structure glish (Praguian Discourse TreeBank) which is gives a broad outline of the topics in a consistent with the dependency tree bank of given text. Prague. The automatic annotation and evaluation 2. Functional Structure: It identifies various help us in understanding the discourse struc- sections within a topic serving different tures further. (Jurafsky et al., 2000) presented functions. a framework for modelling and automatic clas- 3. Event Structure: This identifies various sification of dialogue acts with an accuracy of events in the discourse and show them on 74.7%. (Jinova et al., 2012) experimented with the time-line. semi-automatic annotation of intra-sentential discourse relations in PDT. The automatic an- 4. Coherence Relations: Based on the lin- notation of discourse structure of various user guistic clues and from the functional and forums for thread solvedness classification ac- event structure various coherence rela- curacy is an important application of the au- tions are identified. tomatic discourse structure analysis being ex- plored by (Baldwin et al., 2012), (Xi et al., we glance the Indian literature, we notice dis- 2004), (Fortuna et al., 2007), (Kim et al., cussions on various aspects of textual analy- 2010). sis which basically deal with the coherence of All these discourse theories have been cen- texts. The coherence is judged at different lev- tered around English and other European lan- els right from the relations between two ad- guages, though PDT and RST have been re- joining sentences to the coherence at the level cently being applied to Arabic, Hindi and of texts with respect to the discipline. The co- Tamil as well. These theories may be tried herence levels discussed in the literature are as on Sanskrit corpus as well. But there are two follows - important considerations. The first and fore- most is that Sanskrit comes with a rich her- 1. S¯astrasa´ ˙ngati(Bhat.t.¯ac¯arya, 1989): This itage of linguistic theory of its own. The rich is a subject level coherence. grammatical tradition of India has a vast literature on almost all aspects of language analy- 2. Adhy¯aya sa ˙ngati (Bhat.t.¯ac¯arya, 1989): sis ranging from phonology to discourse anal- This is a chapter or a book level coher- ysis. Thus it is natural to look at these theo- ence. Among the scientific literature in ries which have been discussed over centuries, Sanskrit, we find three different trends with a focus on Sanskrit literature. The sec- under the adhy¯aya sa ˙ngati.One is known ond reason is Sanskrit literature comes with as Bh¯as.ya parampar¯a. Here the origi- annotations, though not by original authors, nal text is in the form of s¯utras(com- but by the later commentators. These annota- pact aphorisms). This is followed by a tions are done following the grammatical the- commentary explaining the s¯utras, op- ories. Hence it is natural to use these theories tionally followed by an explanation (t.¯ık¯a), for computational models related to Sanskrit. a note (t.ippan. i) etc. The commen- In this paper in the next section we give taries may be nested, i.e. there is a a brief survey of Indian theories of discourse commentary on the original s¯utrasand analysis. In section 3 we take up one particu- then sub-commentary on this commen- lar Sanskrit text viz. Mah¯abh¯as.ya of Patañjali tary, and further sub-commentary on the which is available as a tagged text and discuss sub-commentaries and so on. At each its internal structure providing tagging guide- stage the number of commentaries may lines for the tag-set. In section 4 we look at be more than one. The s¯utras(com- the clues suggested in the Sanskrit literature pact aphorism) as well as the commen- and built a FSA for automatic tagging of the taries and sub-commentaries follow a cer- discourse relations. In section 5 we discuss the tain discourse structure. extension of this tagger to handle another im- Another trend is where the original text ´ portant text - S¯abarabh¯as.ya - a commentary establishes a theory, and the later scholars on M¯ım¯a˙ms¯as¯utra. write criticisms on it attacking the original view and proposing a new view. This 2 Indian Theories of Discourse trend is known as Khan. d. ana-man. d. ana Analysis parampar¯a. And there can be a series of such texts criticizing the previous theory Indian Grammatical Theories evolved in order in the series and proposing a new theory. to understand the Vedic literature and protect The structure of these texts then leads to them. The theories developed in three direc- a tree structure, where the siblings indi- tions. The first one, dealing with word and cate different criticisms of the same text sentence formation and analysis. This lead leading to different view points. to the establishment of Vy¯akaran. a (grammar) school. The second one dealing with the se- The third trend is to write Prakaran. a mantics, logic and inference lead to the for- granthas (books dealing with a specific mation of Ny¯aya (logic) school of thought and important topic among several topics dis- the third one dealing with discourse analysis cussed in the texts in s¯utraform). These resulted in M¯ım¯a˙ms¯a (exegesis) school. As books are thus related to the original s¯utratexts, but also have their own nested (a) Hetuhetumadbh¯avah. (cause effect re- commentaries. lationship): yatah. (because), tatah. (hence), yasm¯at (because), tasm¯at 3. P¯adasa ˙ngati(Bhat.t.¯ac¯arya, 1989): This (hence), atah. (therefore). is a section level coherence. (b) As¯aphalyam (failure): kintu (but). (c) Anantarak¯al¯ınatvam (following ac- 4. Adhikaran. a sa ˙ngati (Bhat.t.¯ac¯arya, 1989): This is a topic level coherence. tion): atha (then). M¯ım¯a˙msakas (exegesists) discussed about (d) K¯aran. asatve’api k¯ary¯abh¯avah. / this sa ˙ngati. Further Naiy¯ayikas (logi- k¯aran. ¯abh¯ave’api k¯aryotpattih. (non- cians) discussed about six topic level productive effort or product without relations (S¯astri,1916).´ They are cause): yadyapi (even-though), tath¯api(still), ath¯api(hence). (a) Prasa ˙nga- Corollary. (e) Pratibandhah. (conditional): yadi (b) Upodgh¯ata- Pre-requisite. (if), tarhi (then), cet (if/provided), (c) Hetut¯a- Causal dependence. tarhyeva (then only). (d) Avasara - Provide an opportunity for (f) Samuccayah. (conjunction): ca, apica further inquiry. (and also). (e) Nirv¯ahakaikya - The adjacent sec- (g) P¯urvak¯al¯ıkatvam: The non-finite tions have a common end. verb form ending with suffix ktv¯a ‘ad- (f) K¯aryaikya - The adjacent sections verbial participial’. are joint causal factors of a common (h) Prayojanam (Purpose of the main ac- effect. tivity): The non-finite verb form ending with suffix tumun ‘to-infinitive’. 5. Sub-topic level analysis: Under this level (i) Sam¯anak¯al¯ıkatvam (Simultaneity): we see two different tag sets one found in The non-finite verb form ending exegesist’s analysis and the other in gram- with suffix Satr´ . and S¯anac´ ‘present marian’s analysis. participle’. A. Exegesist’s (M¯ım¯a˙msakas) tag-set Thus we notice that the Indian grammati- (Bhatt¯ac¯arya, 1989): .. cal theories have, to a large extent, guidelines ¯ (a) Aks.epa - Objection for analysis of a text at different levels of dis- (b) Dr.s.t.¯anta - Example course. (c) Pratyud¯aharan. a - Counter- example 3 Structure of Mah¯abh¯as.ya (d) Pr¯asa˙ngika - Corollary P¯an. ini’s As..t¯adhy¯ayi (circa 500 BC) is an im- (e) Upodgh¯ata- Pre-requisite portant milestone in the developmental his- (f) Apav¯ada- Exception tory of Indian theories of language analysis. B. Grammarians (Vaiy¯akaran. as) tag-set As..t¯adhy¯ayi is in s¯utra1 (compact aphorism) (Joshi, 1968): form and hence to understand it one needs an (a) Pra´sna- Question explanation. Patañjaliwrote a detailed commentary on Ast¯adhy¯ayi known as Mah¯abh¯asya (b) Aks¯ .epa - Objection .. . (c) Sam¯adh¯ana- Justification (the great/large commentary). It provides explanation of these s¯utrasand also throws (d) Uttara - Answer light on important aspects of linguistic anal- (e) Vy¯akhy¯a- Elaboration ysis. It is the second important milestone of 6. Inter-sentential relations: In (Ramakrish- the Indian grammatical tradition. The text 1 namacaryulu, 2009), inter-sentential rela- alp¯aks.ara ˙masandigdha ˙ms¯aravat vi´svato mukha ˙m. tions are classified into 9 sub-headings. astobham anavadya ˙mca s¯utrams¯utravido viduh. . A s¯utra contains minimum number of words, is unam- These are given below along with the lex- biguous, contains the essence of the topic, has universal ical clues which mark these relations. validity and is devoid of any faults. of Mah¯abh¯as.ya follows a very well defined dis- and at least an example from the Mah¯abh¯as.ya course structure. illustrating the tag. Mah¯abh¯asya contains commentaries on vari- . 3.1 Mah¯abh¯asya’s Sub-Topic Level ous s¯utrasand the supplementary s¯utrascalled . Tag-set as V¯arttika-s. The relevant level of discourse analysis for Mah¯abh¯as.ya is then the The annotated Mah¯abh¯as.ya has 9 major tags adhikaran. a (topic) level analysis and all the and several sub-tags under each major tag. lower level analysis viz. sub-topic level and These sub-tags are different in different vol- inter-sentential. The s¯utra/v¯arttika sets up umes. But the major tags are the same in all the new topic and all the discussions under the books. We give below a brief description of this follow a well defined structure. The topic these 9 tags, with an example each from the level analysis of a commentary on one s¯utra actual tagged Mah¯abh¯as.ya. The example in was taken up ealier (Kulkarni and Das, 2012). most cases consists of a pair of Sanskrit sen- Figure 1 shows the structure of a commentary tences/paragraphs. The first one sets the con- on the s¯utraP2.1.1 (Samarthah. Padavidhih. ). text under which the next set of sentence(s) is uttered. The label is attached to the second one.

1. Pra´sna (Question) An independent question about some topic or an argument is marked with a question tag. For example, • Skt: atha vidhi´sabd¯arthanirupan. ¯adhikaran. am. Eng: Now starts the section in which the meaning of the word Vidhi is ex- amined.2 • Skt: (pra´snabh¯as.yam): vidhih. iti 3 Figure 1: Structure of a commentary kah. ayam ´sabdah. . [1.1] Eng: What is this word Vidhi? ‘samarthah. padavidhih. ’ 2. Uttara (Answer) An answer to a ques- There are 14 topics under the main headings tion is tagged with this tag. For example, of the s¯utra P2.1.1 (samarthah. padavidhih. ). These 14 topics are related to each other by • Skt: (pra´snabh¯as.yam): vidhih. iti a set of relations, which show the coherence kah. ayam ´sabdah. . of the discussion under this s¯utra.These rela- Eng: What is this word vidhi? tions are the topic level or adhikarana sa˙ngati- . • Skt: (uttarabh¯as.yam): s. The tagging at this level involves semantic vip¯urv¯adghajñah. karmas¯adhana analysis of the text. ik¯arah. . vidh¯ıyate vidhiriti. ki ˙m The original Mah¯abh¯as.ya does not have any punarvidh¯ıyate. sam¯aso vibhak- of the mark-ups. The first version of marked tividh¯anam par¯a˙ngavadbh¯ava´sca. up texts is published by Nirn. aya S¯agaraPress, [1.2] Bombay in early 1900 in 6 volumes. Eng: The letter i denoting the pas- We find the text marked up at sub-topic sive sense (is added) after (the root) level analysis. No text contains the description ghajñ preceded by the pre-verb vi. or explanation of the tags used, nor is there 2English translations are taken from Patañjali’s any prologue mentioning the purpose of this Vy¯akaran. a Mah¯abh¯as.ya Samarth¯ahnika (P 2.1.1) tagging, as is the case with any typical San- Edited with Translation and Explanatory Notes by S skrit texts centuries old. So the first task we D Joshi. 3The first number denotes to the topic number and took up is to provide a manual listing the tags the second number denotes to the sub-topic number of used, the semantics associated with these tags the s¯utra samarthah. padavidhih. P2.1.1. What is prescribed by P¯an. ini’s rules r¯ajñah. purus.ah. and dar´san¯ıyah. is vidhi: ‘operation’. But what r¯ajñah. purus.ah. . could that be which is prescribed? • Skt: (sam¯adh¯anabh¯as.yam): na ‘compounding’, ‘prescription of case- es.ah. dos.ah. . pradh¯anam atra ending’ and ‘treatment as a part of s¯apeks.am bhavati ca pradh¯anasya the following word’. s¯apeks.asya api sam¯asah. . [3.27-28] ¯ Eng: Nothing wrong here. Because 3. Aks.epa (Objection) An objection to an answer or resolution is marked as an it is here the main member which requires an outside word. And com- ¯aks.epa. For example, pounding does take place, even if • Skt: (uttarabh¯as.yam): v¯akye the main member requires an outside pr.thagarth¯ani. r¯ajñah. purus.ah. iti. word. sam¯asepunah. ek¯arth¯anir¯ajapurus.ah. iti. 5. B¯adhaka (Rejection): This tag is used Eng: In the compounded word- to mark the rejection of the arguments group words have separate meanings such as, objection, answer of an objection, of their own, like r¯ajñah. purus.ah. : refutation, criticism etc. For example, king’s man. But in a compound, • Skt: (sam¯adh¯anabh¯as.yam): na words have a single meaning, like in es.ah. dos.ah. . samud¯ay¯apeks.¯a atra r¯ajapurusah: king-man. . . s.as.t.h¯ısarva ˙mgurukulam apeks.ate. • Skt: (¯aks.epabh¯as.yam): kimucy- Eng: Nothing wrong here. Here the ate pr.thagarth¯aniiti y¯avat¯ar¯ajñah. genitive qualifies the whole word gu- purus.a ¯an¯ıyat¯amityukte r¯ajapurus.a rukulam. ¯an¯ıyate r¯ajapurusa iti ca sa eva. . • Skt: (sam¯adh¯anab¯adhakabh¯asyam): [4.44-45] . yatra tarhi na samud¯aya apeks¯a Eng: Why do you say ‘words have . sasth¯ı tatra vrttih na pr¯apnoti. separate meanings of their own’? Be- . .. . . [3.30-31] cause when we say ‘let the king’s Eng: Then when a word in genitive man be brought’, the king-man is does not qualify the whole, it should brought. And when we say ‘let the not result in a compound formation. king-man be brought’, the same man is brought. 6. S¯adhaka (Reaffirmation): This tag is 4. Sam¯adh¯ana(Resolution) This tag is used to mark the reaffirmation of an argu- used to mark an answer to an objection. ment which has been earlier rejected. For For example, example, • Skt: (¯aksepab¯adhakabh¯asyam): • Skt: (¯aks.epabh¯as.yam): yadi . . nanu ca gamyate tatra s¯amarthyam. s¯apeks.amasamartham bhavati kumbhak¯arah nagarak¯arah iti. iti ucyate r¯ajapurus.o’bhir¯upah. . . r¯ajapurus.o dar´san¯ıyah. atra vr.ttirna Eng: But is it not so, that, when we pr¯apnoti. say kumbhak¯arah. : ‘pot-maker’ or na- Eng: If we accept the statement, garak¯arah. : ‘city-maker’, we do appre- ‘what requires an outside word hend semantic connection between is treated as semantically uncon- pot and maker. nected’ then the word-composition • Skt: (¯aks.epas¯adhakabh¯as.yam): r¯ajapurus.a: king-man in the ex- satya ˙m gamyate utpanne pressions r¯ajapurus.ah. abhir¯upah. : tu pratyaye. sa eva t¯avat handsome king-man, r¯ajapurus.ah. samarth¯adutp¯adyah. . [2.12-13] dar´san¯ıyah. : goodlooking king-man Eng: Yes, that is true. It is appre- would not result from the uncom- hended once a suffix has been added. pounded word-groups abhir¯upah. But that same suffix must first be generated after the semantically con- Tags Frequency nected word. Pra´sna 10 (Question) 7. Ud¯aharana (Example): This tags the . Pratipra´sna 1 example. (Counter question) • Skt: (subalo- Pratipra´snottara 1 pod¯aharan. abh¯as.yam): supah. (Answer to a counter question) alopah. bhavati v¯akye. r¯ajñah. Uttara 6 purus.ah. iti. sam¯ase punah. na (Answer) bhavati. r¯ajapurus.a iti. [5.49] Aks¯ .epa 47 Eng: Non-disappearance of case- (Objection) ending occurs in an un-compounded Praty¯aks.epa 6 word-group, like r¯ajñah. purus.ah. : (Counter objection) king’s man. But in a compound it Praty¯aks.epasam¯adh¯ana 2 does not occur, as in r¯ajapurus.ah. : (Answer to a counter objection) king-man. Sam¯adh¯ana 40 (Answer to an objection) 8. D¯usana (Criticism): This tag is to . . Vy¯akhy¯a 34 mark criticism. For example, (Explanation) • Skt: (vy¯akhy¯abh¯as.yam): Ud¯aharan. a 8 sam¯anav¯akya iti prakr.tya (Example) nigh¯atayus.madasmad¯ade´s¯a vak- tavy¯ah. . Table 1: Tags with their frequencies of Eng: Under the heading of ‘within Samarth¯ahnikam the same sentence’ the accents and substitutions for yus.mad and asmad pra´sna, uttara, ¯aks.epa and sam¯adh¯ana also are to be stated. have sub-tags viz. pratipra´sna (a question • Skt: (d¯us.an. av¯arttikam): yoge to a question), pratipra´snottara (answer to a pratis.edha´sc¯adibhih. . [9.115] question to a question), praty¯aks.epa (counter Eng: When there is connection with objection) and praty¯aks.epasam¯adh¯ana(answer and the prohibition should also be to a counter objection), which are more fre- stated. quent. The distribution of these 10 tags - 6 main tags and 4 sub-tags, in one 9. Vy¯akhy¯a(Explanation): Explanation book (samarth¯ahnikam) among 87 books of of either an objection, answer or alterna- Mah¯abh¯asya is shown in Table 1. tive view is marked with this tag. For . example, 4 Mah¯abh¯as.ya Tagger • Skt: (sam¯adh¯anav¯arttikam): The tags used in Mah¯abh¯asya are general pr.thagarth¯an¯amek¯arth¯ıbh¯avah. . samarthavacanam. Eng: The word and are found to be used in the texts samartha means single integrated from other disciplines such as philosophy meaning of the separate meanings. etc. Traditional texts also discuss clues to mark these tags. For example, Bhoja R¯aj¯a’s • Skt: (vy¯akhy¯abh¯asyam): . Sr´ ˙ng¯araPrak¯asa (Dvived¯ıand Dvived¯ı,2007), prthagarth¯an¯a˙m . . . S¯abdabodham¯ım¯a˙ms¯a(T¯at¯ach¯arya, 2005) and pad¯an¯amek¯arth¯ıbh¯avah samarthami- . Avyaya Ko´sa(Srivats¯a˙nk¯ac¯arya, 2004) do pro- tyucyate. [4.42] vide such clues for some tags. The clues for Eng: The single integrated mean- two tags pra´sna (question) and ¯aksepa (ob- ing of the words which have separate . jection), and their sub-tags viz. pratipra´sna meaning is called samartha. (counter question) and praty¯aks.epa (counter Out of these 9 tags, 3 tags viz. s¯adhaka, objection) are in the form of possible Lexical b¯adhakaand d¯us.an. a are rare. The four tags word(s). A. The list of words that classify a paragraph as a pra´sna (question) are the wh-words viz. kim, kah. , kimartham, katara, kutah. , kva, k¯ani,katham, kay¯a,kena etc.

B. The list of words that classify a paragraph as a ¯aks.epa (objection) are evam api, katham, kvacit, yadi, kasm¯atna, nanu ca, yatra tarhi, kim punah. etc.

We looked at one book (¯ahnika) of the tagged text for the clues for other two tags viz. pratipra´snaand praty¯aks.epa. The lexical clues are, Pratipra´sna: kah. punah. , kah. ca, kim ca. Praty¯aks.epa: na v¯a, kasm¯at na, kasya punah. , kah. v¯a,kva ca, kiñca. Figure 2: Finite State Automata of Tags From the semantics of the tags, it is clear that an uttara (answer) tag should follow a 5 Extending the tagger pra´sna (question). The pratipra´sna (counter question) follows a pra´sna (question) and The tagger performed satisfactorily on a part pratipra´snottara (answer to a counter ques- of Mah¯abh¯as.ya text. Till the digital copies of tion) follows a pratipra´sna (counter ques- other parts of the Mah¯abh¯as.ya text become tion). An ¯aks.epa (objection) arises only af- available, we thought of evaluating the tag- ter an answer (uttara). The sam¯adh¯ana (an- ger on other commentary. Since the com- swer to an objection) is for an ¯aks.epa (ob- mentaries are to establish the necessity of a jection) and praty¯aks.epa (counter objection) s¯utra, the assumption is that the basic struc- is to a ¯aks.epa (objection). The praty¯aks.epa ture of commentary should be the same. We sam¯adh¯ana (answer to the counter objection) chose a commentary on the M¯ım¯a˙ms¯as¯utra-s. will follow the praty¯aks.epa (counter objec- The M¯ım¯a˙ms¯as¯utra-s are written by Jaimin¯ı tion). Any sam¯adh¯ana (answer to an objec- around the end of 2nd century AD. It consists tion) can be followed by a ¯aks.epa (objection) of 12 adhy¯aya-s (chapters) and 60 p¯ada-s(sec- only. Varttika-s are the supplementary rules, tions). This text provides rules for the inter- which can occur at the very beginning or it pretation of the Veda-s. Earlier scholars wrote can upon one pra´sna (question) or ¯aks.epa (ob- commentaries on M¯ım¯a˙ms¯as¯utra but unfor- jection). This is represented as a finite state tunately they are lost. A major commentary automata in Figure 2. is composed by Sabarasv¯am¯ı´ around 5th cen- The digital version of the tagged text of tury and is well known as S¯abarabh¯as´ .ya. This complete Mah¯abh¯as.ya is not available. Only is the only extant and authoritative commen- 4 books out of 87 books were available at tary on full 12 chapters of the M¯ım¯a˙ms¯as¯utra the time of testing. Of these we used one of Jaimin¯ı. book for framing the rules and gathering cues The commentary used for our study is and tested our automata on the rest of the the Hindi translation of S¯abarabh¯as´ .yam by three books. Once digitized versions of other Yuddhis..thira M¯ım¯a˙msaka. This Hindi trans- books are available the automata tagger will lation is marked up with only two tags viz. be tested on them. The precision and recall Aks¯ .epa and Sam¯adh¯ana. From the tags in for the three books is given in Table 2. Ta- Hindi translation, we constructed the digital- ble 3 gives the precision and recall for each ized version of the tagged Sanskrit commen- tag in all the three books together. In these tary. Since the tagged text had only two three books, we did not find any instance of tags, we modified our automaton to suit this praty¯aks.epa and praty¯aks.epasam¯adh¯ana. structure removing the nodes corresponding to Books (Ahnika-s)¯ Precision Recall F Score K¯arak¯ahnikam 90.04% 86.45% 88.20% Paspa´s¯ahnikam 76.47% 79.13% 77.77% Pr¯atipadik¯artha´ses.¯ahnikam 85.37% 77.21% 81.08% Table 2: Precision Recall Table

Tags Total Tags Precision Recall F Score P 78 80.00% 76.92% 78.42% PP 4 100.00% 75.00% 85.71% PPU 1 50.00% 100.00% 66.66% U 64 82.76% 75.00 78.68% A 65 70.49% 66.15% 58.25% S 70 67.24% 55.71% 60.93% VA 111 98.23% 100.00% 99.10% VYA 109 95.58% 99.08% 97.29%

Table 3: Precision Recall Table for Each Tags other tags. This led us to modify the lexical References ¯ cues for the tags as well. So the cues for Aks.epa Amal AlSaif and Katja Markert. 2010. The leeds in S¯abarabh¯as´ .ya included the clues from both arabic discourse treebank: Annotating discourse Pra´sna as well as Aks¯ .epa of the Mah¯abh¯as.ya. connectives for arabic. In Nicoletta Calzolari, In addition, we found some stylistic variation Khalid Choukri, Bente Maegaard, Joseph Mar- in the cues. While many of the cues from iani, Jan Odijk, Stelios Piperidis, Mike Rosner, and Daniel Tapias, editors, Proceedings of the Mah¯abh¯as.ya did not ﬁnd any place in the Seventh International Conference on Language S¯abarabh¯as´ .ya it included one new phrase na Resources and Evaluation (LREC’10), pages br¯umah. (literally do not say this), as a marker 2046–2053, Valletta, Malta. European Language Resources Association (ELRA). for Aks¯ .epa. Modifying the tagger accordingly, rd st we tested it on the 3 chapter 1 section Amal Alsaif. 2012. Human and Automatic Anno- 6th adhikaran. a’s 12th s¯utraof S¯abarabh¯as´ .ya tation of Discourse Relations for Arabic. Ph.D. viz. Arun. ¯adhik¯ara and the precision and recall thesis, School of Computing, The University of were found to be 96.00% and 82.76% respec- Leeds. tively. Timothy Baldwin, Li Wang, and Su Nam Kim. 2012. The utility of discourse structure in iden- tifying resolved threads in technical user forums. In Proceedings of COLING 2012: Technical Pa- 6 Conclusion pers, pages 2739–2756, Mumbai. COLING 2012 Organizing Commitee.

´ Sanskrit has a rich grammatical tradition and Sri J¯ıv¯ananda Vidy¯as¯agar Bhat.t.¯ac¯arya. ´ offers theoretical insights for discourse analy- 1989. Sr¯ı M¯adhav¯ac¯aryaviracita Jaimin¯ıya Ny¯ayam¯al¯avistarah. . Kr.s.n. ad¯as Academy, sis as well. The principles for analysis being Varanasi. language independent, these insights should be applicable to other languages as well. In Rev¯apras¯ad Dvived¯ı and Sad¯a´sivakum¯ar this paper we have shown the applicability of Dvived¯ı. 2007. Sarasvat¯ı Kan. .th¯abharan. asya Bhojar¯ajasya S¯ahityaprak¯a´sa N¯ama this analysis in the context of scientific text Sr´. ˙ng¯araprak¯a´sah. , Prathamabh¯agah. 1 - 14 in Sanskrit. The implementation as a FSA is Prak¯a´s¯ah. . Indiragandhi Kala Kendram, New language independent with the clue set as a Delhi and Kalidasa Samsthan, Varanasi. language dependent component. It will be in- B. Fortuna, E.M. Rodrigues, and N. Milic- teresting to use this further on various forums Frayling. 2007. Improving the classifica- on internet. tion of newsgroup messages through social net- work analysis. In In Proceedings of the 16th ACM Conference on Information and Knowl- William C Mann and Sandra A Thompson. 1988. edge Management (CIKM 2007), page 877880, Rhetorical structure theory: Toward a func- Lisbon, Portugal. tional theory of text organization. Text - Inter- disciplinary Journal for the Study of Discourse, Pavlina Jinova, Jiri Mirovsky, and Lucie Po- 8(3):243–281. lakova. 2012. Semi-automatic annotation of intra-sentential discourse relations in pdt. In Yudhis.t.ira M¯ım¯am. saka. 1990. Proceedings of the Workshop on Advances in M¯ım¯am. s¯a´s¯abarabh¯as.yam, volume 2. Ram- Discourse Analysis and its Computational As- lal Kapur Trust, Sonipat, Haryana. pects (ADACA), COLING, pages 43–58, Mum- bai. COLING 2012 Organizing Commitee. Lucie Mladova, Sarka Zikanova, and Eva Hajicova. 2008. From sentence to discourse: Building an S D Joshi and J.A.F. Roodbergen. 1975. annotation scheme for discourse based on prague Patañjali’s Vy¯akarana Mah¯abh¯asya . . dependency treebank. In Nicoletta Calzolari, K¯arak¯ahnikam (P 1.4.23–1.4.55). Center Khalid Choukri, Bente Maegaard, Joseph Mar- of Advanced Study in Sanskrit, Pune. iani, Jan Odijk, Stelios Piperidis, and Daniel S D Joshi and J.A.F. Roodbergen. 1981. Tapias, editors, In Proceedings of the 6th Inter- Patañjali’s Vy¯akaran. a Mah¯abh¯as.ya national Conference on Language Resources and Pr¯atipadik¯artha´ses.¯ahnikam ( P2.3.46–2.3.71). Evaluation (LREC), pages 2564–2570, Marakes, Center of Advanced Study in Sanskrit, Pune. Maroko. European Language Resources Associ- ation (ELRA). S D Joshi and J.A.F. Roodbergen. 1986. Patañjali’s Vy¯akaran. a Mah¯abh¯as.ya Livia Polanyi. 1988. A formal model of discourse Paspa´s¯ahnikam. Center of Advanced Study in structure. Journal of Pragmatics, 12(5–6):601– Sanskrit, Pune, first edition. 638.

Aravind Joshi, Rashmi Prasad, Eleni Miltsakaki, Rashmi Prasad, Umangi Oza, Sudheer Kolachina, Nikhil Dinesh, Alan Lee, Livio Robaldo, and Dipti Misra Sharma, and Aravind Joshi. 2009. Bonnie Webber, 2007. The Penn Discourse The hindi discourse relation bank. In Proceed- Treebank 2.0 Annotation Manual. The PDTB ings of the Third Linguistic Annotation Work- Research Group. shop, ACL-IJCNLP, pages 158–161, Suntec, Singapore. Association for Computational Lin- S D Joshi. 1968. Patañjali’s Vy¯akaran. a guistics. Mah¯abh¯as.ya Samarth¯ahnika (P 2.1.1) Edited with Translation and Explanatory Notes. Cen- ter of Advanced Study in Sanskrit, Poona, first K V Ramakrishnamacaryulu. 2009. Annotating ´ edition. sanskrit texts based on S¯abdabodha systems. In Amba Kulkarni and GérardHuet, editors, Pro- Daniel Jurafsky, Andreas Stolcke, Klaus Ries, ceedings Third International Sanskrit Computa- Noah Coccaro, Elizabeth Shriberg, Rebecca tional Linguistics Symposium, pages 26–39, Hy- Bates, Paul Taylor, Rachel Martin, Carol Van derabad, India. Springer-Verlag, LNAI 5406. Ess-Dykema, and Marie Meteer. 2000. Dia- logue act modeling for automatic tagging and Raphael Salike. 1995. Text and Discourse Anal- recognition of conversational speech. In Asso- ysis, Language Workbooks series, series editor ciation for Computational Linguistics, volume Richard Hudson. Routledge, London. 26:3, pages 339–373. Association for Computa- tional Linguistic. Ananta S¯astri.´ 1916. K¯arik¯aval¯ı with the commentaries Mukt¯aval¯ı,Dinakar¯ı,R¯amarudr¯ı. Su Nam Kim, Li Wang, and Timothy Baldwin. Nirnaya S¯agaraPress, Bombay. 2010. Tagging and linking web forum posts. In . In Proceedings of the 14th Conference on Com- Sri V Srivats¯a˙nk¯ac¯arya. 2004. Avyaya Ko´sa. Sam- putational Natural Language Learning (CoNLL- skrit Education Society, Chennai 4. 2010), pages 192–202, Uppsala, Sweden. Sivadatta D. Kudala. 1912. Patañjali’s N S R T¯at¯ach¯arya. 2005. S¯abdabodham¯ım¯am´ . s¯a: Vy¯akaran. a Mah¯abh¯as.ya with Kaiyat.a’s Prad¯ıpa The sentence and its significance Part-I to V. and N¯age´sa’s Uddyota, volume 2. Nirn. aya Institute of Pondicherry and Rashtriya Sanskrit s¯agarapress, Bombay. Sansthan, New Delhi. Amba Kulkarni and Monali Das. 2012. Dis- Florian Wolf and Edward Gibson. 2005. Rep- course analysis of sanskrit texts. In Lucie Po- resenting discourse coherence: A corpus-based lakova Eva Hajicova and Jiri Mirovsky, editors, study. Journal of Computational Linguistics, Proceedings of the Workshop on Advances in 31(2):249–287. Discourse Analysis and its Computational As- pects (ADACA), COLING, pages 1–16, Mum- W. Xi, J. Lind, and E. Brill. 2004. Learning ef- bai. COLING 2012 Organizing Commitee. fective ranking functions for newsgroup search. In In Proceedings of 27th International ACM- SIGIR Conference on Research and Develop- ment in Information Retrieval (SIGIR 2004), pages 394–401, Sheffield, UK. Nianwen Xue. 2005. Annotating discourse connectives in the chinese treebank. In Proceedings of the Workshop on Frontiers in Corpus Annota- tion II: Pie in the Sky, pages 84–91, Ann Arbor, Michigan. Association for Computational Lin- guistics. Deniz Zeyrek and Bonnie Webber. 2008. A discourse resource for turkish: Annotating discourse connectives in the metu corpus. In The 6th Workshop on Asian Language Resources (ALR 6), IJCNLP, pages 65–71, Hyderabad, In- dia. Asian Federation of Natural Language Pro- cessing. Yuping Zhou and Nianwen Xue. 2012. Pdtb-style discourse annotation of chinese text. In Pro- ceedings of the 50th Annual Meeting of the As- sociation for Computational Linguistics, pages 69–77, Jeju Island, Korea. The Association for Computational Linguistics.