UBC-ALM: Combining k-NN with SVD for WSD

Proceedings of the 4th International Workshop on Semantic Evaluations (SemEval-2007), pages 342–345, Prague, June 2007. c⃝2007 Association for Computational Linguistics UBC-ALM: Combining k-NN with SVD for WSD Eneko Agirre and Oier Lopez de Lacalle IXA NLP Group University of the Basque Country Donostia, Basque Country {e.agirre,jibloleo}@ehu.es Abstract This work describes the University of the Basque Country system (UBC-ALM) for lexical sample and all-words WSD subtasks of SemEval-2007 task 17, where it per- formed in the second and fifth positions re- spectively. The system is based on a com- bination of k-Nearest Neighbor classifiers, with each classifier learning from a distinct set of features: local features (syntactic, col- locations features), topical features (bag-of- words, domain information) and latent fea- tures learned from a reduced space using Singular Value Decomposition. 1 Introduction Our group (UBC-ALM) participated in the lexical sample and all-words WSD subtasks of SemEval- 2007 task 17. We applied a combination of different k-Nearest Neighbor (k-NN) classifiers. Each clas- sifier manages different information sources (fea- tures), making the combination a powerful solution. This algorithm was previously tested on the datasets from previous editions of Senseval (Agirre et al., 2005; Agirre et al., 2006). Before submission, the performance of the system was tested on the Se- mEval lexical sample training data. For learning we use a rich set of features, including latent features obtained from a reduced space using Singular Value Decomposition (SVD). This paper is organized as follows. The learning features are presented in section 2, and the learning algorithm and the combinations of single k-NNs are given in section 3. Section 4 focuses on the tuning experiments. Finally, section 5 summarizes the offi- cial results and some conclusions. 2 Feature set We relied on an extensive set of features of differ- ent types, obtained by means of different tools and resources. We defined two main groups: the origi- nal features extracted directly from the text, and the SVD features obtained after applying SVD decom- position and projecting the original features into the new semantic space (Agirre et al., 2005). 2.1 Original features Local collocations: bigrams and trigrams formed with the words around the target. These features are constituted by lemmas, word-forms, or PoS tags1. Other local features are those formed with the previ- ous/posterior lemma/word-form in the context. Syntactic dependencies: syntactic dependencies were extracted using heuristic patterns, and regular expressions defined with the PoS tags around the tar- get2. The following relations were used: object, sub- ject, noun-modifier, preposition, and sibling. Bag-of-words features: we extract the lemmas of the content words in the whole context, and in a ±4-word window around the target. We also obtain salient bigrams in the context, with the methods and the software described in (Pedersen, 2001). Domain features: The WordNet Domains re- source was used to identify the most relevant do- mains in the context. Following the relevance for- mula presented in (Magnini and Cavagli´a, 2000), we defined 2 feature types: (1) the most relevant do- main, and (2) a list of domains above …

Content truncated to 500 words