Enlarging Monolingual Dictionaries for Machine Translation with Active Learning and Non-Expert Users

Proceedings of Recent Advances in Natural Language Processing, pages 339–346, Hissar, Bulgaria, 12-14 September 2011. Enlarging Monolingual Dictionaries for Machine Translation with Active Learning and Non-Expert Users Miquel Espl`a-Gomis, V´ıctor M. S´anchez-Cartagena, Juan Antonio P´erez-Ortiz Transducens Research Group Departament de Llenguatges i Sistemes Inform`atics Universitat d’Alacant, Spain {mespla,vmsanchez,japerez}@dlsi.ua.es Abstract This paper explores a new approach to help non-expert users with no background in linguistics to add new words to a mono- lingual dictionary in a rule-based machine translation system. Our method aims at choosing the correct paradigm which ex- plains not only the particular surface form introduced by the user, but also the rest of inflected forms of the word. A large mono- lingual corpus is used to extract an initial set of potential paradigms, which are then interactively refined by the user through active machine learning. We show the re- sults of experiments performed on a Span- ish monolingual dictionary. 1 Introduction Rule-based machine translation (MT) systems heavily depend on explicit linguistic data such as morphological dictionaries, bilingual dictionaries, grammars, and structural transfer rules (Hutchins and Somers, 1992). Although some automatic ac- quisition is possible, collecting these data usually requires in the end the intervention of domain ex- perts (mainly, linguists) who master all the encod- ing and format details of the particular MT system. We should, however, open the door to a broader group of non-expert users who could collabora- tively enrich MT systems through the web. In this paper we present a novel method for en- larging the monolingual dictionaries in rule-based MT systems by non-expert users. An automatic process is first run to collect as much linguistic in- formation as possible about the new word to be added to the dictionary and, after that, the result- ing set of potential hypothesis is filtered by elicit- ing additional knowledge from non-experts with no linguistic background through active learn- ing (Olsson, 2009; Settles, 2010), that is, by in- teractively querying the user in order to efficiently reduce the search space. As these users do not possess the technical skills which are usually re- quired to fill in the dictionaries, this elicitation is performed via a series of simple and easy yes/no questions which only require speaker-level under- standing of the language. Our method does not only incorporate to the dictionary the particular surface form introduced by the user (for example, wants), but it also discovers a suitable paradigm for the new word so that all the word forms of the corresponding lexeme and their morphological in- formation (such as wanted, verb, past or wanting, verb, gerund) are also inserted. This work focuses on monolingual dictionar- ies. These dictionaries have basically two types of data: paradigms, that group regularities in in- flection, and word entries. The paradigm assigned to many common English verbs, for instance, indi- cates that by adding the ending -ing, the gerund is obtained. Paradigms make easier the management of dictionaries in two ways: 1. by reducing the quantity of information that needs to be stored, …

Content truncated to 500 words