Contribution of Complex Lexical Information to Solve Syntactic Ambiguity in Basque
Aitziber Atutxa, Eneko Agirre, Kepa Sarasola · International Conference on Computational Linguistics · 2012
In this study, we explore the impact of complex lexical information to solve syntactic ambiguity, including verbal subcategorization in the form of verbal transitivity and verb-noun-case or verb-noun-case-auxiliary relations. The information was obtained from different sources, including a subcategorization dictionary extracted from a Basque corpus, the web as a corpus, an English corpus and a Basque dictionary. Functional ambiguity between subject and object is a widespread problem in Basque, where 22% of subjects and objects are ambiguous, and this ambiguity surfaces in 33% of the sentences. This problem is comparable to PP attachment ambiguities in other languages. Our results show that, using complex lexical information, our results are better than a state-of-the-art statistical parser, obtaining a statistically significant error reduction of 20%. The disambiguation system is independent on the actual parsing algorithm used. The analysis revealed that the most relevant information are the case carried by the noun and the transitivity of the verb. TITLE AND ABSTRACT IN BASQUE Informazio lexikal konplexuaren ekarpena euskarazko anbiguotasun sintaktikoen ebazpenean Lan honetan informazio lexikal konplexua erabiltzearen garrantzia aztertzen dugu euskarazko anbiguotasun sintaktikoen ebazpenean. Aditzen iragankortasuna erakusten duen azpikategorizazioaren ekarpena aztertu dugu, baita aditz-izen-kasu eta aditz-izen-kasu-laguntzaile erlazioena ere. Informazio horiek hainbat iturritatik jaso ditugu: euskarazko corpus batetik, webetik berau corpus gisa hartuta, ingelesezko corpus batetik eta euskarazko hiztegi batetik. Subjektu eta objektuaren arteko anbiguotasun funtzionala maiz aurkitzen dugu euskarazko testuetan; subjektua edo objektua bereiztea kasuen %22an ambiguoa da, eta hori gertatzen da perpausen %33an. Horrela, arazo horren garrantzi handia konparagarria da beste hizkuntza batzuek duten PP attachment arazoarenarekin. Gure sistemaren emaitzak hobeak dira artearen egoerako analizatzaile sintaktiko estatistiko batenak baino, estatistikoki esanguratsua den %20ko errore-murrizketa lortzen baitu. Analisi sintaktikoa egiteko edozein algoritmorekin erabil daiteke desanbiguazio-sistema hau.