Aligning the Romanian Reference Treebank and the Valence Lexicon of Romanian Verbs

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 6626–6634 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 6626 Aligning the Romanian Reference Treebank and the Valence Lexicon of Romanian Verbs Ana-Maria Barbu1,2, Verginica Barbu Mititelu3, Cătălin Mititelu1 1“Iorgu Iordan – Al. Rosetti” Institute of Linguistics, 2University of Bucharest, 3Romanian Academy Research Institute for Artificial Intelligence 13 Calea 13 Septembrie, Bucharest, Romania anamaria.barbu@g.unibuc.ro, vergi@racai.ro, catalin.mititelu@lingv.ro Abstract We present here the efforts of aligning two language resources for Romanian: the Romanian Reference Treebank and the Valence Lexicon of Romanian Verbs: for each occurrence of those verbs in the treebank that were included as entries in the lexicon, a set of valence frames is automatically assigned, then manually validated by two linguists and, when necessary, corrected. Validating a valence frame also means semantically disambiguating the verb in the respective context. The validation is done by two linguists, on complementary datasets. However, a subset of verbs were validated by both annotators and Cohen’s κ is 0.87 for this subset. The alignment we have made also serves as a method of enhancing the quality of the two resources, as in the process we identify morpho-syntactic annotation mistakes, incomplete valence frames or missing ones. Information from each resource complements the information from the other, thus their value increases. The treebank and the lexicon are freely available, while the links discovered between them are also made available on GitHub. Keywords: treebank, verbal valence frames, alignment, Romanian, semantic disambiguation 1. Introduction Among the language resources for Romanian, there are ones that describe the syntactic and semantic aspects of the language. Two of them are the Romanian Reference Treebank (henceforth RRT) (Barbu Mititelu, 2018) and 1 the Valence Lexicon of Romanian verbs (henceforth 2 DCV) (Barbu, 2017). The former is a source of possible syntactic structures in the language, complemented by morphological and lexical information, as well as genre specifications, but no semantics can be found here. The latter resource is an inventory of verbal subcategorization frames designed for different verbal senses (clustered together when possible), complemented with as much as necessary morphologic information. We present here the work for coupling these two resources, thus increasing their value: information from each resource complements the information from the other: a verb in DCV, where its semantics and subcategorization frames with examples are presented, is now presented “in action”: its occurrences in RRT show how the frames are lexicalized, how contextual elements can sometimes prevent some valences from being lexicalized. Given that DCV was created starting from examples in a journalistic corpus, they are now tested against occurrences in other genres, those represented in RRT (see below). Vice versa, verbs in RRT are associated with possible senses as defined in DCV, are grouped together under the same frame irrespective of the variations of their syntactic structures favoured by the context of occurrence. After presenting related work on aligning a treebank and a valence lexicon in Section 2, we describe the two …

Content truncated to 500 words