Treelex Meets Adjectival Tables

International Conference RANLP 2009 - Borovets, Bulgaria, pages 203–207 Treelex Meets Adjectival Tables Anna Kup´s´c Universit´e de Bordeaux / CLLE-ERSS, Signes, IPIPAN Abstract The paper presents Treelex, a valence lexicon of French adjectives automatically extracted from a treebank. The corpus contains morphological and syntactic annotations but no subcategorisa- tion information is present for adjectives. Due to rich corpus annotations, our extraction method is guided by linguistic knowledge. The obtained lexicon (about 2000 adjectives and 40 frames) has been evaluated against hand-crafted adjec- tival tables described in [13] and achieved 0.46 F-measure. Keywords adjectives, valence, treebank, syntactic lexicon, lexicon- grammar tables, French 1 Introduction The importance of subcategorisation information is unquestionable with respect to performance of vari- ous NLP applications [4, 15, 5, 9]. So far, creating valence lexicons has been mostly devoted to obtaining such resources for verbs whereas valence lexicons for other predicates, e.g., nouns, adjectives or adverbs, are scarce. For French, the only available resources for adjectives (lexicon-grammar tables in [8] and [13]) exist only on paper and have not been adapted yet to automatic processing. In this paper, we present a method for obtaining an electronic valence lexicon of French adjectives which can be used in various NLP applications. The adopted technique consists in automatically ex- tracting the lexicon from a treebank. We exploit a journalistic corpus, richly annotated with both mor- phological and syntactic information (constituents and functions), cf. [1]. The corpus is relatively small as it contains about 1 million words. The treebank has been automatically pre-tagged and then manually verified by human experts following annotation guidelines in [2]. We rely on linguistic knowledge and corpus anno- tations in order to obtain subcategorisation patterns for adjectives. Syntactic annotations in the corpus provide infor- mation about major constituents (including adjectival phrases) but grammatical functions are indicated only for phrases related to verbs. Therefore, no distinction between argumental and non-argumental dependents of adjectives is made in the treebank. Specifying va- lence (i.e., arguments) of adjectives is more difficult than for verbs. First, in most cases, the syntactic realization of adjective’s arguments is optional. For instance, [12] mentions just a few adjectives, such as enclin ‘inclined’, exempt ‘exempted’ or d´esireux ‘de- sirous’, among those for which a complement is oblig- atory. This makes the strongest ‘obligatoriness’ cri- terion, often used to identify complements of verbs, practically inapplicable to adjectives. Also results of other linguistic tests, e.g., topicalisation or pronomi- nalisation, are in general less reliable than for verbs, cf. [13]. Second, syntactic realization of adjective’s argu- ments is more variable than with verbs: several equiv- alent syntactic realizations of one semantic argument are possible. For example, affable ‘affable’ may appear with two semantically equivalent PP complements: af- fable envers/avec les clients ‘affable towards/with cus- tomers’. Such variability makes the specification of re- quired components even more challenging. Finally, ad- jectives may appear in many syntactic constructions, e.g., comparative or impersonal phrases. Hence, argu- ments specific to individual adjectives should be dis- tinguished from elements regularly appearing in a …

Content truncated to 500 words