Exploring the Vector Space Model for Finding Verb Synonyms in Portuguese

International Conference RANLP 2009 - Borovets, Bulgaria, pages 393–398 Exploring the Vector Space Model for Finding Verb Synonyms in Portuguese Lu´ıs Sarmento1, Paula Carvalho2 and Eug´enio Oliveira1 1Faculdade de Engenharia da Universidade do Porto - DEI - LIACC Rua Dr. Roberto Frias, s/n, 4200-465 Porto, Portugal 2Faculdade de Ciˆencias da Universidade de Lisboa - DI - XLDB Campo Grande, 1749-016 Lisboa, Portugal las@fe.up.pt, pcc@xldb.di.fc.ul.pt and eco@fe.up.pt Abstract We explore the performance of the Vector Space Model (VSM) in finding verb synonyms in Por- tuguese by analyzing the impact of three oper- ating parameters: (i) the weighting function, (ii) the context window used for automatically ex- tracting features, and (iii) the minimum num- ber of vector features. We rely on distributional statistics taken from a large n-gram database to build feature vectors, using minimal linguis- tic pre-processing. Automatic evaluation of syn- onym candidates using gold-standard informa- tion from the OpenOffice and Wiktionary the- saurus shows that low frequency features carry most information regarding verb similarity, and that a [0, +2] window carries more information than a [-2, 0] window. We show that satis- factory precision levels require vectors with 50 or more non-nil components. Manual evalua- tion over a set of declarative verbs and psycho- logical verbs show that VSM-based approaches achieve good precision in finding verb synonyms for Portuguese, even when using minimal linguis- tic knowledge. This lead us to proposing a per- formance baseline for this task. Keywords Vector Space Model, Semantics, Relation Extraction, Statistical Methods, Language Resources, Evaluation 1 Introduction Large-coverage and fine-grained linguistic resources are crucial for the majority of the applications in nat- ural language processing, but they are still scarce and, in most cases, they do not satisfy every particular information need. Manual creation of linguistic re- sources is time-consuming and requires linguistic ex- pertise. Therefore, there is a rising interest in de- veloping automatic or semi-automatic methods and techniques for building language resources with min- imal human intervention. However, automatic meth- ods usually involve a large set of parameters, whose impact on final results is difficult to assess, and thus to optimize. In this paper, we address the task of au- tomatically creating a lexicon of verb synonyms for Portuguese using the Vector Space Model (VSM), and we explore the impact of three of its core parameters: (i) the context used for extracting vector features, (ii) the function used for weighting features, and (iii) the cut-off threshold for removing vectors with insufficient feature information. We rely on n-gram information collected from a large dump of the Portuguese web, in order to obtain distributional statistics for verb lemmas. For performing parameter exploration, we evaluate results automatically using gold-standard in- formation extracted from the OpenOffice thesaurus and from Wiktionary. Fine-grained evaluation was achieved by manually assessing the synonym candi- dates obtained for a sample of two syntactic-semantic classes of verbs: psychological verbs and declarative verbs. We chose these two specific verb classes for two reasons. First, they exhibit different syntactic and se- mantic behavior, and thus …

Content truncated to 500 words