Senseval-3: The Italian all-words task

Senseval-3: The Italian All-words Task Marisa ULIVIERI ILC-CNR Via Moruzzi 1 Pisa, Italy, 56124 marisa.ulivieri@ilc.cnr.it Francesca BERTAGNA ILC-CNR Via Moruzzi 1 Pisa, Italy, 56124 francesca.bertagna@ilc.cnr.it Elisabetta GUAZZINI ILC-CNR Via Moruzzi 1 Pisa, Italy, 56124 elisabetta.guazzini@ilc.cnr.it Nicoletta CALZOLARI ILC-CNR Via Moruzzi 1 Pisa, Italy, 56124 glottolo@ilc.cnr.it Abstract This paper describes the Italian all-words sense disambiguation task for Senseval-3. The annotation procedure and criteria together with the encoding of multiwords are presented. 1 Introduction This paper describes the Italian all-words sense disambiguation task for Senseval-3: about 5000 words were manually disambiguated according to the ItalWordNet (IWN) word senses. The first section briefly describes of the corpus and the lexical reference resource. The second section contains some general criteria adopted for the annotation of the corpus and illustrated by a series of examples. Issues connected to the treatment of phenomena typically found in corpora, e.g. abbreviations, foreign words, jargon, locutions are discussed. Furthermore, the encoding of compounds, metaphorical usages, and multiword units is described. Problems connected with i) the high granularity of sense distinctions in the lexical resource and ii) unsolvable ambiguities of the contexts are dealt with. Finally, it is evidenced how the annotation exercise can be of help in updating or tuning IWN, by adding missing senses and/or entries. 2 The Corpus and the Lexical Resource The Italian all-words corpus consists of about 13600 word tokens, extracted from the SI-TAL 1, Italian Syntactic Semantic Treebank (ISST). The 1 SI-TAL (Integrated System for the Automatic treatment of Language) was a National Project devoted to the creation of large linguistic resources and software tools for Italian written and spoken language processing. ISST (Montemagni et al. 2003) consists of i) a generic corpus of about 215,000 tokens, extracted from different periodicals and newspaper articles (La Repubblica, Il Corriere della Sera) and ii) a specialised corpus of about 90,000 tokens, with texts belonging to the financial domain (Il Sole- 24Ore). The annotated corpus consists of about 5000 words and comprises a selection of Italian newspaper articles about various topics: politics, sport, news, etc. The common data format is XML. The reference lexical resource used for the Senseval-3 sense tagging task is the lexical- semantic database IWN, developed within the framework of two different research projects: EuroWordNet (Vossen 1999) and SI-TAL, during which IWN was extended by the insertion of adjectives, adverbs and a subset of proper nouns. The IWN database contains about 64,000 word senses corresponding to about 50,000 synsets. It has inherited the EWN linguistic model (Alonge et al., 1998) which provides a rich set of semantic relations, and the first nucleus of verbs and nouns. IWN was structured around the notion of synset, a set of synonymous word meanings, and the information is encoded in the form of lexical- semantic relations between pairs of synsets. The IWN database comprises also an Interlingual Index (ILI), based on the Princeton WordNet 1.5 used to link wordnets of different languages so that it is possible to go from the Italian words to similar words in English …

Content truncated to 500 words