View metadata, citation and similar papers at core.ac.uk brought to you by CORE
provided by Repositorio Institucional de la Universidad de Alicante Procesamiento del Lenguaje Natural, Revista nº 46 marzo de 2011, pp 11-18 recibido 17-12-10 revisado 15-02-11 aceptado 24-02-11
AnCora-Nom: A Spanish Lexicon of Deverbal Nominalizations
AnCora-Nom: Un Léxico de Nominalizaciones Deverbales del Español
Aina Peris, Mariona Taulé CLiC-Centre de Llenguatge i Computació, Universitat de Barcelona. Gran Via de les Corts Catalanes, 585 08007, Barcelona {aina.peris, mtaule}@ub.edu
Resumen: En este artículo se describe un nuevo recurso: AnCora-Nom, un léxico de nominalizaciones deverbales del español. Actualmente, contiene 1.655 entradas léxicas y 3.094 sentidos, donde cada sentido tiene asociado el tipo denotativo y la estructura argumental con los papeles temáticos correspondientes. Este léxico se ha extraído automáticamente a partir de la información anotada en el corpus AnCora-Es. AnCora-Nom se derivó teniendo en cuenta no sólo la información estrictamente relacionada con las nominalizaciones deverbales sino también con información morfológica y sintáctico-semántica previamente anotada en el corpus. Palabras clave: nominalización deverbal, denotación, estructura argumental, recurso léxico, anotación de corpus.
Abstract: This paper describes a new lexical resource: Ancora-Nom, a Spanish lexicon of deverbal nominalizations. At present, it contains 1,655 lexical entries and 3,094 senses. Each sense has a denotation type associated, and the mapping of nominal complements with arguments and the corresponding theta roles is also annotated. A particular interest of this lexicon is that it has been automatically extracted from the annotated AnCora-Es corpus. AnCora-Nom was derived taking into account the information directly related to nominalizations, but also the morphological and syntactic-semantic information annotated in the corpus. Keywords: deverbal nominalization, denotation, argument structure, lexical resource, corpus annotation.
of the verb from which the nominalization is 1 Introduction derived. We use the AnCora-Verb lexicon (Aparicio et al., 2008) for obtaining the This paper presents AnCora-Nom, a Spanish semantic verbal information related to the lexicon of deverbal nominalizations. At present, argument structure. Having a lexical resource it contains 1,655 lexical entries corresponding that represents the semantics of nominalizations to the different deverbal nominalizations ─denotation type and argument structure─ and appearing in the annotated corpus AnCora-Es links these deverbal nouns with the (Taulé et al., 2008). In AnCora-Nom, each corresponding verbs can be very helpful for sense of a deverbal noun has a denotation type information extraction tasks. It can also give (i.e., event, result, and underspecified) more insight into the linguistic relation between associated to it, and the mapping of nominal verbs and their nominalizations, since it complements with arguments and the captures the similarity between sentences such corresponding theta roles are also annotated. as those in (1). The mapping between syntactic constituents (1a) La solución pasaba por [abaratar el and arguments is established taking into despido]. account the syntactic and semantic information
ISSN 1135-5948 © 2011 Sociedad Española para el Procesamiento del Lenguaje Natural Aina Peris, Mariona Toulé
‘The solution was to make dismissal cheaper’. for each predicate and describing the network (1b) La solución consistía en [el of relations between the semantic frames. abaratamiento del despido]. Concretely, deverbal nominalizations are ‘The solution consisted of the cost reduction of annotated as events or entities, in a similar way dismissal’. to our event-result distinction (see Section 4.1). Both sentences in (1) refer to the same In the case of events, they are included in the event, expressed either by a verb (1a) or by a semantic frame of the base verb while in the deverbal nominalization (1b). In fact, they are case of results they are considered to be a paraphrases. different semantic frame. The main goal of this paper is to present the Although English is the language that has content of the AnCora-Nom lexicon. received the most attention, there are also The remainder of the paper is organized as resources devoted to nominalizations in other follows. Section 2 briefly discusses the related languages. In the line of the Berkeley FrameNet work. Section 3 focuses on the methodology Project, there are proposals aiming at creating developed to obtain the nominal lexicon. FrameNets for other languages such as German Section 4 describes the information coded in the (Burchardt et al., 2009), Japanese (Ohara, 2009) nominal entries. Finally, main conclusions are and Spanish (Subirats, 2009). The Spanish drawn in Section 5. FrameNet, for instance, contains 600 lexical units spread over 100 different semantic frames. 2 Related work Besides the FrameNet framework, there are There exist similar lexical resources for other a number of proposals for languages other than 1 English. An example is “The Essex Database of languages. NOMLEX (Macleod et al., 1998) is 3 a lexicon of English deverbal morphologically Russian Verbs and Their Nominalizations” derived nominalizations consisting of 1,025 (Spencer and Zaretskaya, 1999). This database entries. This resource seeks not only to describe contains information about 800 Russian verbs the allowed complements for a nominalization, and their corresponding nominalizations, but also to relate the nominal complements to distinguishing those verbal entries that the arguments of the corresponding verb. The nominalize the whole event and preserving the NomBank project (Meyer, 2007), which carried verbal argument structure from those that out the semantic annotation of nominal denote a concrete or abstract result of the verb. argument structure in the PennTreeBank corpus This database includes morphological and (Palmer, 2005), used the same kind of semantic information of this kind of nouns. representation for nominal complements as Each nominal sense has one of the three NOMLEX. In fact, NOMLEX-PLUS (Meyers semantic categories assigned to it (i.e., complex et al., 2004), is an extension of NOMLEX, that event, simple event and result) based on automatically increases the number of lexical Grimshaw (1990) proposals, which are entries to 7,050. It also includes deverbal- comparable to our denotation-type attribute (see cousin and deadjectival nominalizations, and Section 4.1). argument-taking nouns such as relational There is also a French lexicon of nominalizations (Balvet et al., 2010) developed nouns. 4 Another proposal to represent deverbal by the Nomage research group . They aim at nominalizations is the Berkeley FrameNet describing the aspectual properties of deverbal Project (Ruppenhoffer et al., 2006). This project nouns. This group focuses on the development has created an on-line lexical resource2 for of two resources: a semantically annotated English with 11,600 lexical units based on corpus of deverbal nouns, and an electronic frame semantics and supported by corpus lexicon. At present, the lexicon contains 815 evidence. The aim is to document the range of lexical entries. For each one, they provide a semantic and syntactic combinatory definition, information about the source verb, possibilities -valences- of each predicative word, including verbs, nouns and adjectives, in each of its senses. The project produces 3http://privatewww.essex.ac.uk/~spena/res_intere semantic frames, annotating a set of examples sts.htm 4http://stl.recherche.univ- 1 http://nlp.cs.nyu.edu/nomlex/index.html lille3.fr/programmesetcontrats/NOMAGE/NOMAG 2 http://framenet.icsi.berkeley.edu/ Eenglish.html
12 AnCora-Nom: A Spanish Lexicon of Deverbal Nominalizations
the argument structure and the aspectual class extraction of nominalization senses in the of the nominalization. lexicon (see Section 3.1). The second step was Most of these lexicons, with the exception of to check both types of information manually NOMLEX-PLUS, were built manually. This is and to measure inter-annotator agreement to in contrast to the AnCora-Nom lexicon, which ensure the quality of the final corpus annotation is automatically obtained from the data (Peris et al., 2010) and thus, of the extracted annotated in AnCora-Es. lexicon presented here. The lexicon was derived taking into account 3 Methodology the information directly related to nominalizations, as well as the morphological As mentioned above, the methodology followed and syntactic-semantic information annotated in for building the AnCora-Nom lexicon takes the corpus. advantage of the information annotated in
AnCora-Es. It is a 500,000-word Spanish corpus consisting of newspaper texts annotated 3.1 Extraction process at different linguistic levels: morphology (PoS The 23,000 tokens annotated correspond to and lemmas), syntax (constituents and 1,655 different lemmas. For each of these functions), semantics (verbal argument lemmas we created a lexical entry using the structure, thematic roles, semantic verb classes, information annotated in the corpus. The named entities, and WordNet nominal senses), extraction process consists of consulting all and pragmatics (coreference (Recasens and occurrences of each lemma and determining the Martí, 2010))5. Recently, 23,000 deverbal different nominal senses a nominalization can nominalization tokens were also semantically have. In this process, we also extract the annotated in the corpus (Peris et al., 2010). By different features associated with each deverbal noun we mean a noun nominalization sense: argument structure, type morphologically derived from a verb or a so- of specifier, number, the verb from which it is called cousin noun (Meyers, 2007). Cousin derived, the verbal frame, etc. nouns are nouns that give rise to verbs (e.g., To establish the senses of the relación ‘relation’ > relacionar ‘to relate’), or nominalizations we take into account (1) the nouns semantically related to verbs (e.g., denotation type, (2) the sense of the base verb, escarnio ‘mocking’ is related to mofarse ‘to and (3) whether or not the nominalization is make fun’). Specifically, we annotated the part of a lexicalized construction. The criteria argument structure of these deverbal followed are: nominalizations, their semantic interpretation - if a nominalization is part of a (i.e., result, event, and underspecified)6, and lexicalized construction, it constitutes a whether or not they appear in a lexicalized sense in itself; and construction. To this end, the first step was to - those nominalizations with the same run two independent automated processes: one denotation and that are derived from the for the annotation of denotation types and same verb sense are grouped under the lexicalizations, using the ADN-classifier (Peris same nominal sense. et al., 2010), and another one for the annotation Once the senses are established, the different of argument structures (Peris and Taule, features extracted from the AnCora corpus are forthcoming). In the latter case, we designed a associated to them (detailed in Section 4). set of heuristic rules in a decision-list format Each sense can also contain one or more that take into account the argument structures nominal frames, depending on the verbal frame specified in the AnCora-Verb lexicon. This is from which the nominalization is derived. because we assume that deverbal There are a total of 3,204 different nominal nominalizations inherit the argument structure frames. At the frame level, three important of the base verb. In fact, in the corpus each features are coded: the argument structure with deverbal noun is linked to its corresponding the corresponding theta roles, the type of verbal lexical entry sense via the attribute specifier, and a feature indicating whether or
13 Aina Peris, Mariona Toulé
specifier occurs in the corpus. The plurality resources that are completely linked. Both feature is not coded in the lexicon until all the lexicons are in XML format. nominalization occurrences in the same frame AnCora-Nom Lexicon are explored; if one appears in plural, the value In this section we detail the structure and the is positive. This extraction process has not information specified in the nominal entries. produced errors in the generation of the lexicon. Figure 1 shows the full information associated In the manual revision process, a 5% of errors with the lexical entry aceptación ‘acceptance’. was found, which corresponds to errors from We can observe that the lemma aceptación the corpus annotation. In the next section we has two nominal senses. In this case, they are
Figure 1: aceptación lexical entry detail all the information contained in the established because of the different denotation lexical entry. ─the first one denotes an event and, the second Once the 1,655 entries were created, we a result─ the verbal sense being the same for erased from the corpus the information related both (originlink="verb.aceptar.1"). to the denotation type and whether or not the Furthermore, each sense contains one or nominalization is lexicalized. We leave only a more nominal frames. By nominal frame, we pointer to the corresponding nominal entry, mean the link to the corresponding verbal frame where this information is declared. (which was established depending on diathesis Furthermore, each nominal lexical entry is alternations). It is important to preserve this related to its corresponding verbal one, so link because the argument structure of the AnCora-Nom and AnCora-Verb are two nominalization is inherited from the
14 AnCora-Nom: A Spanish Lexicon of Deverbal Nominalizations
corresponding verb, and this information is with Spanish nominalizations, but in the future coded in the verbal frames. In Figure 1, each Catalan ones will also be taken into account. sense contains only a frame with the “default” c) The attribute origin indicates the type of type. This is because the occurrences of word from which the nouns are derived. In aceptación in the corpus link correspond to the Figure 1, the value for this attribute is default verbal frame, that is, the less marked "deverbal", meaning that this lexical entry frame of aceptar ‘to accept’ (‘A21.transitive- concerns a noun derived from a verb. At agent-patient’). The different nominal senses present, AnCora-Nom only contains deverbal and frames contain a series of attributes nouns but in the future it will include other type extracted from the morphosyntactic and of nominalizations such as deadjectivals. semantic information specified in the AnCora- d) The attribute type refers to the word Es corpus. We present them in two different class, "noun" in Figure 1. groups ─sense attributes and frame attributes─ in the following subsections. In addition to the
Figure 2: lexicalized sense golpe de estado features extracted from the corpus, the 3.2 Sense attributes following features are also included: The features related to sense attributes are the a) Lemma. In Figure 1, the value for this following: attribute is the noun aceptación a) The attribute cousin marks whether the (lemma="aceptación"). b) The attribute language ("lng") codifies nominalization is morphologically derived from a verb ("cousin=no", in Figure 1) or is a cousin the language represented in the lexical entry. noun ("cousin=yes"). AnCora resources work with Spanish and b) The denotation attribute indicates the Catalan, so the values of this attribute are "es" semantic interpretation of the deverbal noun. for Spanish (lng="es") and "ca" for Catalan The possible values are: "event", if the (lng="ca"). At present, AnCora-Nom only deals nominalization expresses an action; "result", if
15 Aina Peris, Mariona Toulé
the nominalization denotes the result of an aceptación corresponds to two synsets action, and "underspecified" if the contextual (wordnetsynset="1600117820+16:10039397"), features of the nominalization do not permit the while the second has only one distinction between the two previous readings. (wordnetsynset="16:00117820"). In Figure 1, there are two senses, the first one being "result" (denotation="result") and the 3.3 Frame attributes second one "event" (denotation="event"). As mentioned above, in each sense we c) Each sense contains an identifier ("id") distinguish different nominal frames, as well as to indicate the sense number. In Figure 1, the the arguments and thematic roles of the first sense is "id=1" and the second "id =2". d) The lexicalized attribute indicates nominalization, the constituents that realize these arguments and how frequently they whether the nominalization is part of a appear. Other attributes contained at the frame lexicalized construction ─lexicalized="yes" level are the specifier type of the (Figure 2) or not ‘lexicalized="no" (Figure 1). nominalization, and if the deverbal noun In the former case, two additional attributes are nominalization appears in plural. The latter two added: (a) the alternativelemma, specifying attributes are important for establishing the the whole lexicalized construction of which the nominalization denotation. Next, we detail the nominalization is part ("golpe_de_estado" in information at the frame level: Figure 2), and (b) lexicalization_type, to a) The attribute type indicates the verbal distinguish between six types of lexicalizations according to their similarity to different word frame from which the nominalization is derived. In the AnCora-Verb, each verbal sense classes: "nominal" (e.g., golpe_de_estado ‘coup can be realized in more than one frame: default, d’état’), "verbal" (e.g., estar de acuerdo ‘to passive, anticausative, locative, etc. In the agree’), "adjectival" (e.g., al alza ‘rising’), nominal entries, we mark the corresponding "adverbial" (e.g., con cuidado ‘with care’), verbal frames, which are the possible values for "prepositional" (e.g., en busca de ‘in search this attribute. In most cases, its value is of’), or "conjunctive" (e.g., en la medida que ‘as far as’). In the case of nominal "default" as in Figure 1 ("type=default"). b) Argument (Structure). In this lexicalizations one of the three above- complex attribute, the different arguments mentioned denotation values is assigned to the ("argument") and the corresponding thematic whole lexicalized construction. In Figure 2, the roles ("thematicrole") are specified. To lexicalized construction "golpe_de_estado" is a represent the arguments we follow the same nominal lexicalization (lexcalization_type= annotation scheme used in AnCora-Verb "nominal"), therefore, it has a denotation value (Aparicio et al., 2008). In Figure 1, the result ("denotation= result"). sense has one argument ("arg0") with the agent e) The attribute originlemma specifies the thematic role ("agt"). This argument is realized verb lemma from which the noun is derived. In Figure 1, the value for this attribute is "aceptar" once ("frequency=1") by a prepositional phrase ("constituent type =sp") introduced by de ‘of’ in both senses (originlemma="aceptar"). ("preposition=de"), and once by an adjective f) Since verbs can have different senses, phrase ("type=s.a"). The event sense has one the attribute originlink points to the concrete argument ("arg1") with the patient thematic role verbal sense of the base verb. Recall that ("pat"). This argument is realized twice different verbal senses linked to a ("frequency=2") by a prepositional phrase nominalization would imply different nominal ("constituent type =sp") introduced by the senses. In Figure 1, however, the "originlink" preposition de ‘of’ ("preposition=de") and once attribute takes the same value in both senses by a possessive determiner ("type=determiner", "verb.aceptar.1" ("oringinlink=verb.aceptar.1") the one used for establish the relation between "postype=possessive"). c) The attribute referencemodifier the nominal and the verbal entries. represents the nominal complements that are g) Since nouns in the AnCora corpus are 6 not arguments but which modify the reference annotated with WordNet synsets , we incorporate this information into the attribute of the nominalization. Frequency is also taken "wordnetsynset". In Figure 1, the first sense of into account. In Figure 2, the lexicalized construction golpe de estado has an adjectival
6 We used WordNet 1.6 for Spanish.
16 AnCora-Nom: A Spanish Lexicon of Deverbal Nominalizations
phrase as reference modifier ("constituent type fact, these lexicalized constructions represent a =s.a") appearing once ("frequency=1"). 42% of the total number of lexicalized d) The type of specifier is a widely constructions. The remaining 68% correspond accepted criterion to distinguish between result to nominal lexicalizations, which mostly belong and event readings, so we include this to the result class. information in the attribute specifier. The Table 2 presents the distribution of senses possible values are: "article", "indefinite", taking into account the denotation type and the "demonstrative", "exclamative", "numeral", number of arguments. It shows that result "interrogative" "possessive", "ordinal" and senses are the most likely to appear without any "void", when there is no specifier. In this argument (26%) in comparison to event (8%) attribute, we also take into account the and underspecified (4%) senses. In fact, most of frequency the specifiers are realized. In Figure the times event (53%) and underspecified 1, the result sense is specified once by an article (63%) senses appear with at least one argument determiner ("type=determiner", in contrast to result senses (32%). When "postype=article"). nominalizations have two arguments, event and e) The attribute appearsinplural indicates underspecified senses are also more common if an occurrence of a nominalization in a (27% and 23% respectively) compared to the particular frame appears in plural. It is a result senses (18%). However, result senses boolean attribute. In Figure 1, neither of the appear more frequently with more than two senses appear in plural, thus, the value is "no". arguments (24%) than event (12%) and Finally, each lexical entry also contains all underspecified (10%) senses. This is explained the examples from which the information has by the fact that result nominalizations appear been extracted, specifying the corpus file, the with more adjunct arguments (argMs), which node path and the sentence where it is. can be associated with several theta roles (manner, time, location, etc.) in the same NP. 4 Quantitative Analysis of Data However, event and underspecified nominalizations appear with real arguments AnCora-Nom consists of 1,655 deverbal lexical (arg0, arg1…) that can only appear once in the entries corresponding to 3,094 senses and 3,204 same NP. frames. Table 1 shows the distribution of nominal senses in denotation types and Arguments Event Result Underspecified lexicalized (or non-lexicalized). 0 48 499 22 Denotation Lexicalized Non- Total 1 336 603 312 lexicalized 2 168 340 111 Event 0 631 631 More than 2 79 444 47 Result 115 1,771 1,886 Total 631 1,886 492 Underspecified 2 490 492 None 85 0 85 Table 2: Distribution of senses taking into Total 202 2,892 3,094 account argument realization
Table 1: Distribution of nominal senses 5 Conclusions
The most frequent denotation among In this paper, we present a new lexical resource: nominalizations is the result class (61%) AnCora-Nom, a Spanish lexicon containing followed by those nominalizations belonging to 1,655 lexical entries of deverbal the event (20%) and the underspecified class nominalizations. This lexicon was developed (16%). It is not surprising that result from the information encoded in the AnCora-Es nominalizations are the most frequent since corpus. It includes all the nominalizations found events tend to be realized mostly by verbal in the corpus with their possible denotations clauses and nominalizations are more and argument structure combinations. AnCora- frequently used for the result-concrete meaning, Nom is linked to the AnCora-Es corpus and to more typical for nouns. The remaining 3% the AnCora-Verb Spanish lexicon, constituting correspond to nominalizations without a an excellent resource for studying the argument denotation value because they are part of a realization of both nouns and verbs. lexicalized construction that is not nominal. In
17 Aina Peris, Mariona Toulé
In the future, we intend to enlarge this lexicon case of risk and kakeru’ In Boas, Hans, C. with deadjectival nominalizations and relational (ed.), Mulitilingual FrameNets in nouns since we consider that they can also have Computational Lexicography: Methods and an argument structure. We also intend to build a Applications. Mouton de Gruyter. similar lexicon for Catalan nominalizations. Palmer, M., Kingsbury, P. and Gildea, D. (2005): The Proposition Bank: An Akcnowledgments Annotated Corpus of Semantic Roles, This research has received support from the Computational Linguistics, 21 (1). USA: projects Text-Knowledge 2.0 (TIN2009-13391- MIT Press. C04-04) and AnCora-Net (FFI2009-06497- Peris. A., Taulé, M., Boleda, G. & Rodríguez, E/FILO) from the Spanish Ministry of Science H. (2010). ‘AND-Classifier: Automatically and Innovation, and a FPU grant (AP2007- assigning denotation types to 01028) from the Spanish Ministry of Education. nominalizations’. Proceedings of the 7th We also thank three anonymous reviewers for International Conference on Language their comments and suggestions. Resources and Evaluation, Valleta, Malta. References Peris A., Taulé, M. & Rodríguez, H.(2010). ‘Semantic Annotation of Deverbal Aparicio, J., Taulé, M. & Martí, M.A. (2008). Nominalizations in the Spanish AnCora ‘AnCora-Verb: A Lexical Resource for the corpus'. Proceedings of The Ninth Semantic Annotation of Corpora’. International Workshop on Treebanks and Proceedings of Language, Resources and Linguistic Theories, University of Tartu, Evaluation. Marrakech, Morocco. Estonia. Balvet, A., Barque, L. & Marín, R. (2010). Peris, A and Taulé, M. (forthcoming). ‘Building a lexicon of French deverbal ‘Annotating the Argument Structure of nouns from a semantically annotated Deverbal Nominalizations in Spanish’. corpus’. Proceedings of the 7th International Conference on Language Recasens, M., Martí, M.A. (2010). ‘AnCora- Resources and Evaluation, Valleta, Malta. CO: Coreferentially annotated corpora for Spanish and Catalan’. Language Resources Burchardt, A. et al. (2009). ‘FrameNet for the and Evaluation, Springer Science. semantic analysis of German: Annotation, representation and automation’. In Hans C. Ruppenhofer, J., Ellsworth, M., Petruck, Boas (ed.): Multilingual FrameNets in M.L.R., Johnson, C.R. & Scheffczyk, J. Computational Lexicography: Methods and (2006). ‘FrameNet II: Extended Theory and Applications. Mouton de Guyter, 2009. Practice’. Online publication. Grimshaw, J. (1990). Argument Structure. Spencer, A. and Zaretskaya, M. (1999), ‘The Cambridge, Massachussets: The MIT Press. Essex Database of Russian Verbs and their Nominalizations’. Essex Research Reports Macleod C., Grishman, R., Meyers, A., Barrett, in Linguistics. University of Essex, L. & Reeves, R. (1998). ‘NOMLEX: A Colchester, UK, 25. Lexicon of Nominalizations’. Proceedings of EURALEX’98, Liege, Belgium. Subirats, C. 2009 ‘Spanish FrameNet: A frame- semantic analysis of the Spanish lexicon’. In Meyers, A. (2007). ‘Anotation Guidelines for Hans Boas, ed. Multilingual FrameNets in NomBank-Noun Argument Structure for Computational Lexicography. Methods and PropBank’. Online Publication. Applications. Berlin/New York: Mouton de Meyers, A., Reeves, R., Macleod, C., Szekely, Gruyter. R., Zielinska V., Young, B., & Grishman, R. Taulé, M., Martí, M.A. & Recasens, M. (2008). (2004). ‘The Cross-Breeding of ‘Ancora: Multilevel Annotated Corpora for Dictionaries’. Proceedings of the 4th Catalan and Spanish’. Proceedings of 6th International Conference on Language International Conference on Language Resources and Evaluation, Lisbon, Portugal. Resources and Evaluation. Marrakesh Ohara, K.H, (2009). ‘Frame-based contrastive (Morocco). lexical semantics in Japanese FrameNet: The
18