Treatment of Multiword Expressions and Compounds in Bulgarian

Home , Gapping, Pseudogapping, Scrambling (linguistics)

Petya Osenova and Kiril Simov Linguistic Modelling Deparment, IICT-BAS Acad. G. Bonchev 25A, 1113 Soﬁa, Bulgaria [email protected] and [email protected]

Abstract 2012)), we will adopt the Multiword Expressions The paper shows that catena represen- classification, presented in (Sag et al., 2001). They tation together with valence information divide them into two groups: lexicalized phrases can provide a good way of encoding and institutionalized phrases. The former are fur- Multiword Expressions (beyond idioms). ther subdivided into fixed-expressions, semi-fixed It also discusses a strategy for mapping expressions and syntactically-flexible expressions. noun/verb compounds with their counter- Fixed expressions are said to be fully lexicalized part syntactic phrases. The data on Mul- and undergoing neither morphosyntactic variation tiword Expression comes from BulTree- nor internal modification. Semi-fixed expressions Bank, while the data on compounds comes have a fixed word order, but “undergo some degree from a morphological dictionary of Bul- of lexical variation, e.g. in the form of inflection, garian. variation in reflexive form, and determiner selection” (non-decomposable idioms, proper names). 1 Introduction Syntactically-flexible expressions show more vari- Our work is based on the annotation of Multi- ation in their word order (light verb constructions, word Expressions (MWE) in the Bulgarian tree- decomposable idioms). We follow the understand- bank — BulTreeBank (Simov et al., 2004). We ing of (O’Grady, 1998) that MWEs have their in- use this representation for parsing and analysis of ternal syntactic structure which needs to be rep- compounds. BulTreeBank exists in two formats: resented in the lexicon as well as in the sentence HPSG-based (original - constituent-based with analysis. Such a mapping would provide a mech- head annotation and grammatical relations) and anism for accessing the literal meaning of MWE Dependency-based (converted from the HPSG- when necessary. The inclusion of the compounds based format). In both of them the representations into the MWE classification raises additional chal- of the various kinds of Multiword Expressions is lenges. As it was mentioned, an important ques- an important problem. We need a mechanism for tion is the prediction of the compound semantics connecting the MWE in the lexicon with their ac- formed on the basis of the related phrases contain- tual usages within the sentences. As an interesting ing verb + dependents. case of MWE at the interface of morphology and In this paper we discuss the usage of the same syntax we consider Compounds. They are usually formal instrument - catena - for their representa- derived from several lexical units, have an internal tion and analysis. Catena is a path in the syntac- structure with a specified derivation model and se- tic or morphemic analysis that is continuous in the mantics. In the paper we are especially interested vertical dimension. Its potential is discussed fur- in the mapping among deverbal compounds and ther in the text. The paper is structured as follows: their counterpart syntactic phrases. In the next section a brief review of previous works Since there is no broadly accepted standard on catena is presented. In Section 3 a typology of for Multiword Expressions (see about the vari- the Miltiword Expressions in BulTreeBank is out- ous classifications in (Villavicencio and Kordoni, lined. Section 4 considers possible approaches for consistent analyses of MWE. Section 5 introduces In this paper we consider catena as a unit of the relation of syntax with compound morphology. syntax and morphology1. In a syntactic or mor- Section 6 concludes the paper. phological tree (constituent or dependency) catena is: Any element (word) or any combination of ele- 2 Related works on catena ments that are continuous in the vertical dimension (y-axis). In syntax it is applied to the id- The notion of catena (chain) was introduced in iosyncratic meaning of all sorts, to the syntax of (O’Grady, 1998) as a mechanism for represent- ellipsis mechanisms (e.g. gapping, stripping, VP- ing the syntactic structure of idioms. He showed ellipsis, pseudogapping, sluicing, answer ellipsis, that for this task there is a need for a definition of comparative deletion), to the syntax of predicate- syntactic patterns that do not coincide with con- argument structures, and to the syntax of discon- stituents. He defines the catena in the following tinuities (topicalization, wh-fronting, scrambling, way: The words A, B, and C (order irrelevant) extraposition, etc.). In morphology it is applied form a chain if and only if A immediately dom- to the bracketing paradox problem. It provides inates B and C, or if and only if A immediately a mechanism for a (partial) set of interconnected dominates B and B immediately dominates C. In syntactic or morphological relations. The set is recent years the notion of catena revived again and partial in cases when the elements of the catena it was applied also to dependency representations. can be extended with additional syntactic or mor- Catena is used successfully for modelling of prob- phological relations to elements outside of the lematic language phenomena. catena. The relations within the catena cannot be (Gross, 2010) presents the problems in syntax changed. and morphology that have led to the introduction These characteristics of catena make it a good of the subconstituent catena level. Constituency- candidate for representing the various types of based analysis faces non-constituent structures in Multiword Expressions in lexicons and treebanks. ellipsis, idioms, verb complexes. In morphol- In the lexicons each MWE represented as a catena ogy the constituent-oriented bracketing paradoxes might specify the potential extension of each ele- have been also introduced ([moral] [philosoph - ment of the catena. As part of the morphemic anal- er] vs. [moral philosoph]-er). Catena is viewed ysis of compounds, catena is also a good candidate as a dependency grammar unit. At the mor- for mapping the elements of the syntactic para- phological level morphemes (affixes) receive their phrase of the compound to its morphemic analy- own nodes forming chains with the roots (such sis. as tenses: has...-(be)en; be...(be)ing, etc.). In (Gross, 2011) the author again advocated his ap- 3 Multiword Expressions in proach on providing a surface-based account of BulTreeBank the non-constituent phenomena via the contribu- In its inception and development phase, the tion of catena. Here the author introduces a no- HPSG-based Treebank adopted the following tion at the morphological level — morph catena. principles: When the MWE is fixed, which Also, he presents the morphological analysis in the is inseparable, with fixed order and can be Meaning-Text Theory framework, where (due to viewed as a part-of-speech, it receives lexical its strata) there is no problem like the one present treatment. This group concerns the multiword in constituency. closed class parts-of-speech: multiword prepo- Apart from linguistic modeling of language sitions, conjunctions, pronouns, adverbs. There phenomena, catena was used in a number of NLP are 1081 occurrences of such multiword closed applications. (Maxwell et al., 2013) presents an class parts-of-speech in the treebank, which makes approach to Information retrieval based on cate- around 1.9% of the token occurrences in the text. nae. The authors consider catena as a mecha- Thus, this group is not problematic. Of course, nism for semantic encoding which overcomes the there are also exceptions. For example, one problems of long-distance paths and elliptical sen- of the multiword indefinite pronouns in Bulgar- tences. The employment of catena in NLP appli- ian shows variation in its ending part: êàêâè- cations is additional motivation for us to use it in òî è äà å/ñà/áèëî (whatever). The varying the modeling of an interface between the valence lexicon, treebank and syntax. 1http://en.wikipedia.org/wiki/Catena (linguistics) part is a 3-person-singular-present-tense-auxiliary, as ïîðòôåéë (wallet) or ðîäíèíà (relative), the 3-person-plural-present-tense-auxiliary or its 3- verb takes canonical complements. In the latter person-neuter-singular-past-participle. The semi- cases, verbs like - îáðúùàì ‘to turn’ pay - take fixed expressions (mainly proper names) have only noun - âíèìàíèå ‘attention’ attention - for been interpreted as Multiword Expressions. How- making an idiom - îáðúùàì âíèìàíèå íà íÿ- ever, all the idioms, light verb constructions, etc. êîãî ‘to turn attention to somebody’ pay attention have been treated syntactically. This means that in to somebody. Another example is the verb - âçå- the annotations there is no difference between the ìàì ‘to take’, which combines in such cases with literal and idiomatic meaning of the expression: äóìà ‘word’ - âçåìàì äóìàòà ‘to take the word’ kick the bucket (= kick some object) and kick the take the floor. However, light expressions with bucket (= die). In both cases we indicated that the desemantisized verbs, such as èìàì have or ñòà- verb kick takes its nominal complement. âà happens (èìàì äóìàòà ‘I have the word’ to After some exploration of the treebank, such have the floor or ñòàâà äóìà çà íåùî ‘it happens as the extraction of the valency frames and train- word’ something refers to something)can take nu- ing of statistical parsers, we discovered that the merous semantic classes as dependants. In this present annotations of Multiword Expressions are case we mark the information only on the head of not the most useful ones. In both applications the these Multiword Expressions. In this approach the corresponding generalizations are overloaded with assumption is that the verb posits its requirements specific cases which are not easy to incorporate on its dependants. However, a very detailed va- in more consistent classifications. The group of lency lexicon is required. One problem with this lexically treated POS remained stable. However, approach is when the dependant elements allow the other two groups were reconsidered. Proper for modifications. names, as semi-fixed, are treated separately, i.e. as The second approach is construction-based. In non-Multiword Expressions, since we need coref- this case there is no head. Multiword Expres- erencing the single occurrence of the name with sions are with fixed order and inseparable parts. the occurrence of two or more parts of the name. They are annotated via brackets at the lexical level. Light verb constructions have to be marked as One example is the idiom îò èãëà äî êîíåö such explicitly in order to differentiate its specific ‘from needle to thread’ from the beginning to the semantics from the semantics of the verbal phrases end. This approach is problematic for syntacti- with semantically non-vacuous verbs. The same cally flexible Multiword Expressions. holds for the idioms. The third approach marks all the parts of the Multiword Expressions. It is based on the notion 4 Possible Approaches for Encoding of catena as introduced above. Here is an example Multiword Expressions in Treebanks of this annotation: (VPS Òîé (VPC-C (V-C ðèòíà) (N-C êàìáàíàòà))) There are a number of possible approaches for (VPS He (VPC-C (V-C kicked) (N-C bell.DEF))) handling idioms, light verb constructions and col- He kicked the bucket locations. The approaches are not necessarily con- where the suffix ”-C” marked the catena. This flicting with each other. However, we also seek for approach maybe adds some spurious composition- an approach that would give us the mapping be- ality to the idioms, but it would be indispensable tween compounds and their syntactic paraphrases. for handling idiosyncratic cases, such as separa- The first approach is selection-based. This ble MWE. However, in order to model the vari- approach is appropriate for Multiword Expres- ous MWE and to ensure mappings among com- sions in which there is a word that can play the role pounds and related syntactic phrases, the combi- of a head. For example, a verb subcategorizes for nation of catena with selection-based approach is only one lexical item or a very constrained set of needed. In case the MWE does not allow for any lexical items. When combined with nouns, such as modifications, for each element of the catena it is âðåìå (time), ôîðìà (shape), íàäåæäà (hope), specified that the element does not allow any mod- the verb forms idioms - ãóáÿ âðåìå ‘lose time’ ifications. Thus, catena plus selection-based ap- waste one’s time; ãóáÿ ôîðìà ‘lose shape’ to be proach is a powerful means for challenging anal- unfit; ãóáÿ íàäåæäà ‘lose hope’ lose one’s hope). yses. Construction-based approach does not make However, when combined with other nouns, such any difference for the strict idioms, since there is no lexical variation envisaged there. the lexical unit in the lexicon needs to specify the In Fig. 1 and Fig. 2 we present two sentences role of these modifiers. For example, the canon- from BulTreeBank in which the same verb çàòâà- ical form of the MWE in Fig. 2 is çàòâàðÿì ñè ðÿì ‘to close’ is used in its literal meaning and as î÷èòå. Its representation in the lexicon could be a part of idiomatic expression. The catena is high- as follows: lighted. [ form: < çàòâàðÿì ñè î÷èòå > catena: (VPC-C (V-C (V-C çàòâàðÿì) (Pron-C ñè)) (N-C î÷èòå) ) semantics: íå-îáðúùàì-âíèìàíèå-íà-ôàêòèòå_rel(e,[1]ôàêò) valency: < indobj; (PP (P x) (N [1]y)) : x ∈ { ïðåä, çà } > ] The specification above shows that the catena includes the elements ’shut my eyes’ in the sense of ’run away from facts’, which is presented in the semantics part as a relation. In this part the noun ’fact’ is indicated via a structure-sharing mechanism - [1]. This is necessary, because in the va- Figure 1: HPSG-based tree for the sentence “Ãå- lency part of the lexical unit the noun within the íèèòå ñè çàòâàðÿò î÷èòå ïðè ñâèðåíå” (’Ge- subcategorized PP by the catena ’shut my eyes’ re- niuses REFL.POSS.SHORT close eyes at playing’ produces some fact from the world. Also, if more Geniuses close their eyes when playing some in- than one preposition is possible, they are presented strument. ). as a set of x-values. 5 From Syntax to Compound Morphology The catena approach is also very appropriate for modeling the connection among compounds and their syntactic counterparts in Bulgarian. In (Gross, 2011) the notion of ‘morph catena’ has been explicitly introduced. By granting a node to each morpheme2, the author makes the problematic morpheme a dominant element over the other depending morphemes. Thus, all these morphemes are under its scope. The catena set includes also the intended meaning. Here we have in mind examples like the following: a) compound deverbal noun whose counterpart can be expressed only through a free syntactic Ãå- Figure 2: HPSG-based tree for the sentence “ phrase (áèëêîëå÷åíèå (‘herbcuring’, curing by íèèòå ñè çàòâàðÿò î÷èòå ïðåä äðåáíèòå íå- herbs), *áèëêîëåêóâàì (*‘herbcure.1PERS.SG’, ùà ” (’Geniuses REFL.POSS.SHORT close eyes to cure with herbs) and ëåêóâàì ñ áèëêè before minor things’ Geniuses run away from the (‘cure.1PERS.SG with herbs’, to cure with herbs); minor things. ). and b) compound deverbal noun whose verbal counterpart can be either a compound too, In the lexicon each MWE is represented in its but verbal, or a free syntactic phrase (ðúêîìà- canonical (lemmatized) form. The catena is stored õàíå (‘handwaving’, gesticulating), ðúêîìàõàì in the lexical unit. Additionally, the valency of MWE is expressed for the whole catena or for it 2Such as, histor-ic-al novel-ist where the morpheme ’ist’ dominates the rest of the morphemes, thus resolving the parts. When the MWE allows for some modifi- bracketing paradoxes of the type [historical [novel-ist]] and cation of its elements - i.e. modifiers of a noun, [[historical novel]-ist] (‘handwave.1PERS.SG’, gesticulate) and ìàõàì the deverbal nominal compound connects directly ñ ðúêà (‘wave with hand’, gesticulate). to a syntactic phrase (with no grammatical verb A previously done survey in (Osenova, 2012) compound counterpart). The morph catena will performed over an extracted data from a mor- straightforwardly present the tree of: áèëê-î-ëå÷- phological dictionary (Popov et al., 2003) shows åí-èå. However, in the syntactic catena a prepo- that in Bulgarian head-dependant compounds are sition is inserted according to the valence frame more typical for the nominal domain (with a head- of the verb ëåêóâàì (cure): ëåêóâàì ñ áèëêè final structure), while the free syntactic phras- (‘cure.1PERS.SG with herbs’, to cure with herbs). ing is predominant in the verbal domain. Also, Using catena, we can safely connect the non- regarding the occurrence of dependants in the constituent phrase ëåêóâàì ñ (cure with) with the compounds, subject is rarely present in the ver- root morpheme of the head in the compound - ëå÷. bal domain, while complements and adjuncts are Also, all the possible modifiers of áèëêè (herbs) frequent - ãëàñîïîäàâàì ‘votegive.1PERS.SG’, in the syntactic phrase would be connected to the vote - where ‘vote’ is a complement of ‘give’. On head morpheme áèëê. the contrary, in the nominal domain also subjects The second case is the one where the nominal are frequently present, since they are transformed compound has mappings to both - verb compound into oblique arguments - ñíåãîâàëåæ ‘snowrain’, and syntactic phrase. The connection among snowing. the nominal and verb compounds is rather triv- ðúêîìà- Irrespectively of the blocking on some com- ial, since only the inflections differ. ( õàíå ðúêîìàõàì pound verbs, there is a need to establish a map- (‘handwaving’, gesticulating), ìàõàì ping between the nominal compound and its free (‘handwave.1PERS.SG’, gesticulate) and ñ ðúêà ðúê-î- syntactic phrase counterpart. Both expressions (’wave with hand’, gesticulate): ìàõ-à-íå ðúê-î-ìàõ-à-ì are governed by the selection-based rules. Thus, vs. . The connection the realization of the dependants in the syntactic with the syntactic phrase follows the same rules phrases relies on the valency information of the as in the previous case. head verb only, while the realization of the depen- Here is the representation of the lexical unit for áèëêîëå÷åíèå dants in the nominal or verb compounds respects compound nouns: ( (‘herbcuring’, also the compound-building constraints. curing by herbs): [ form: < áèëêîëå÷åíèå > A mechanism is needed which relates the exter- catena: nal syntactic representations with the internal syn- (MorphVerbObj-C (MorphVerb-C [1]áèëê-) (MorphoObj-C [2]ëå÷-) tax of the counterpart morphological compounds. ) Moreover, some external arguments which are derivational catena: missing in the compound structures may well ap- (VPC-C [2]ëåêóâàì ñ [1]áèëêè pear in the free syntactic phrases, such as: ðú- (V-C (PP-C (P-C ) (N-C ))) êîìàõàì ñ ëÿâàòà ðúêà ) ‘handwave.1PERS.SG semantics: with left.DEF hand’, I am gesticulating with my ëåêóâàì_rel(e,x,y,[4]áèëêè) & íîìèíàë_rel(e) left hand, where the complement ðúêà (hand) is valency: ñ áèëêè further specified and for that reason is explicitly < mod; (PP (P ) [4](NP ModP* (N ) ModS*) ) : ModP* or ModS* is not empty > present. Thus, we can imagine that in the lex- ] icon we have the deverbal noun compounds as In this example we present two relations. First, well as verb compounds, presented via morpho- the morph catena is presented with its roots (the logical catena. These words are then connected to role of affixes omitted for simplicity). Then, the the heads of the corresponding syntactic phrases catena reflecting the derivational syntactic phrase (again in the lexicon), but this time the relations is shown. The correspondences are marked with are presented via a syntactic catena tree. We can tags [1] and [2]. The second relation is at the se- think of the morphological catena as a rather fixed mantic level, where the semantics of the syntac- one, while of the syntactic catena as a rather flexi- tic phrase (ëåêóâàì_rel(e,x,y,[3]áèëêè)) is rep- ble one, since it would allow also additional argu- resented fully, and additionally the event is nomi- ments or modifiers in specific contexts. nalized by the second predicate íîìèíàë_rel(e). Let us see in more detail how this mapping will In the valency list we might have a PP modifier be established. The first case is the one where (corresponding to the indirect object in the verb phrase) of the compound only if the preposition is number 610516: “QTLeap: Quality Translation ñ (by), the head noun of the preposition comple- by Deep Language Engineering Approaches” and ment is the same as the noun in the verbal phrase by European COST Action IC1207: “PARSEME: áèëêè (herbs) and there is at least one modifier PARSing and Multi-word Expressions. Towards of the noun. Thus phrases like: áèëêîëå÷åíèå linguistic precision and computational efficiency ñ áúëãàðñêè áèëêè (‘herbcuring with Bulgar- in natural language processing.” ian herbs’, curing with Bulgarian herbs) and áèë- We are grateful to the two anonymous review- êîëå÷åíèå ñ áèëêè, êîèòî ñà ñúáðàíè ïðåç ers, whose remarks, comments, suggestions and íîùòà (‘herbcuring with Bulgarian herbs that are encouragement helped us to improve the initial collected during the night’, curing with Bulgarian variant of the paper. All errors remain our own herbs that were collected at night) are allowed. But responsibility. phrases with dublicate internal and external arguments like áèëêîëå÷åíèå ñ áèëêè (‘herbcuring with herbs’, curing with herbs) are not allowed. References Many of the other details are left out here in order Thomas Gross. 2010. Chains in syntax and morphol- to put the focus on the important relations. Among ogy. In Otoguro, Ishikawa, Umemoto, Yoshimoto, the omitted phenomena are the representation of and Harada, editors, PACLIC, pages 143–152. Insti- tute for Digital Enhancement of Cognitive Develop- the subject and patient information as well as the ment, Waseda University. inflection of the compounds. As a result, we propose a richer valence lexi- Thomas Gross. 2011. Transformational grammari- ans and other paradoxes. In Igor Boguslavsky and con, extended with information on mappings be- Leo Wanner, editors, 5th International Conference tween compounds and their counterpart syntactic on Meaning-Text Theory, pages 88–97. phrases. The morph catena remains steady, while K. Tamsin Maxwell, Jon Oberlander, and W. Bruce the syntactic one is flexible in the sense that it en- Croft. 2013. Feature-based selection of dependency codes the predictive power of adding new material. paths in ad hoc information retrieval. In Proceed- When connectors (such as prepositions) are added, ings of the 51st Annual Meeting of the ACL, pages the prediction is easy due to the information in the 507–516, Sofia, Bulgaria. valence lexicon. However, when some modifiers William O’Grady. 1998. The syntax of idioms. Natu- come into play, the prediction might become non- ral Language and Linguistic Theory, 16:279–312. trivial and difficult for realization. Petya Osenova. 2012. The syntax of words (in Bulgar- ian). In Diana Blagoeva and Sia Kolkovska, editors, 6 Conclusions and Future Work “The Magic of Words”, Linguistic Surveys in honour of prof. Lilia Krumova-Tsvetkova, Sofia, Bulgaria. The paper confirms the conclusions from previous works that catena is indispensable means for en- Dimitar Popov, Kiril Simov, Svetlomira Vidinska, and coding idioms. Especially for cases where the lit- Petya Osenova. 2003. Spelling Dictionary of Bul- eral meaning also remained a possible interpreta- garian (in Bulgarian). Nauka i izkustvo, Sofia, Bul- garia. tion in addition to the figurative meaning in the re- spective contexts. We also extend this observation Ivan A. Sag, Timothy Baldwin, Francis Bond, Ann to other types of MWE. Copestake, and Dan Flickinger. 2001. Multiword expressions: A pain in the neck for nlp. In In Proc. Apart from that, we show that catena is a tool of the CICLing-2002, pages 1–15. that together with the selection-based approach can ensure mappings between the expressions Kiril Simov, Petya Osenova, Alexander Simov, and which have paraphrases on the level of morphol- Milen Kouylekov. 2004. Design and implemen- tation of the Bulgarian HPSG-based treebank. In ogy as well as syntax. While at the morphological Journal of Research on Language and Computation, level the catena is stable, in syntax domain it han- pages 495–522, Kluwer Academic Publishers. dles also additional material on prediction from Aline Villavicencio and Valia Kordoni. 2012. There’s valence lexicons and beyond them. light at the end of the tunnel: Multiword Expressions in Theory and Practice, course materials. Techni- 7 Acknowledgements cal report, Erasmus Mundus European Masters Pro- gram in Language and Communication Technolo- This research has received support by the EC’s gies (LCT). FP7 (FP7/2007-2013) under grant agreement