HLTDI: CL-WSD Using Markov Random Fields for SemEval-2013 Task 10

Second Joint Conference on Lexical and Computational Semantics (*SEM), Volume 2: Seventh International Workshop on Semantic Evaluation (SemEval 2013), pages 171–177, Atlanta, Georgia, June 14-15, 2013. c⃝2013 Association for Computational Linguistics HLTDI: CL-WSD Using Markov Random Fields for SemEval-2013 Task 10 Alex Rudnick, Can Liu and Michael Gasser Indiana University, School of Informatics and Computing {alexr,liucan,gasser}@indiana.edu Abstract We present our entries for the SemEval- 2013 cross-language word-sense disambigua- tion task (Lefever and Hoste, 2013). We submitted three systems based on classifiers trained on local context features, with some elaborations. Our three systems, in increasing order of complexity, were: maximum entropy classifiers trained to predict the desired target- language phrase using only monolingual fea- tures (we called this system L1); similar clas- sifiers, but with the desired target-language phrase for the other four languages as features (L2); and lastly, networks of five classifiers, over which we do loopy belief propagation to solve the classification tasks jointly (MRF). 1 Introduction In the cross-language word-sense disambiguation (CL-WSD) task, given an instance of an ambigu- ous word used in a context, we want to predict the appropriate translation into some target language. This setting for WSD has an immediate application in machine translation, since many words have mul- tiple possible translations. Framing the resolution of lexical ambiguities as an explicit classification task has a long history, and was considered in early SMT work at IBM (Brown et al., 1991). More recently, Carpuat and Wu have shown how to use CL-WSD techniques to improve modern phrase-based SMT systems (Carpuat and Wu, 2007), even though the language model and phrase-tables of these systems mitigate the problem of lexical ambiguities some- what. In the SemEval-2013 CL-WSD shared task (Lefever and Hoste, 2013), entrants are asked to build a system that can provide translations for twenty ambiguous English nouns, given appropri- ate contexts – here the particular usage of the am- biguous noun is called the target word. The five tar- get languages of the shared task are Spanish, Dutch, German, Italian and French. In the evaluation, for each of the twenty ambiguous nouns, systems are to provide translations for the target word in each of fifty sentences or short passages. The translations of each English word may be single words or short phrases in the target language, but in either case, they should be lemmatized. Following the work of Lefever and Hoste (2011), we wanted to make use of multiple bitext corpora for the CL-WSD task. ParaSense, the system of Lefever and Hoste, takes into account evidence from all of the available parallel corpora. Let S be the set of five target languages and t be the particular target language of interest at the moment; ParaSense cre- ates bag-of-words features from the translations of the target sentence into the languages S−{t}. Given corpora that are parallel over many languages, this is straightforward at training time. However, at test- ing time it requires a complete MT system for each of the four other languages, which is computation- ally …

Content truncated to 500 words