Methods to Integrate a Language Model with Semantic Information for a Word Prediction Component

Proceedings of the 2007 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pp. 506–513, Prague, June 2007. c⃝2007 Association for Computational Linguistics Methods to integrate a language model with semantic information for a word prediction component Tonio Wandmacher Laboratoire d’Informatique (LI) Université François Rabelais de Tours 3 place Jean-Jaurès, 41000 Blois, France tonio.wandmacher@ univ-tours.fr Jean-Yves Antoine Laboratoire d’Informatique (LI) Université François Rabelais de Tours 3 place Jean-Jaurès, 41000 Blois, France jean-yves.antoine@ univ-tours.fr Abstract Most current word prediction systems make use of n-gram language models (LM) to es- timate the probability of the following word in a phrase. In the past years there have been many attempts to enrich such lan- guage models with further syntactic or se- mantic information. We want to explore the predictive powers of Latent Semantic Analysis (LSA), a method that has been shown to provide reliable information on long-distance semantic dependencies be- tween words in a context. We present and evaluate here several methods that integrate LSA-based information with a standard language model: a semantic cache, partial reranking, and different forms of interpola- tion. We found that all methods show sig- nificant improvements, compared to the 4- gram baseline, and most of them to a sim- ple cache model as well. 1 Introduction: NLP for AAC systems Augmented and Alternative Communication (AAC) is a field of research which concerns natural language processing as well as human-machine interaction, and which aims at restoring the com- municative abilities of disabled people with severe speech and motion impairments. These people can be for instance cerebrally and physically handi- capped persons or they suffer from a locked-in syndrome due to a cerebral apoplexy. Whatever the disease or impairment considered, oral communica- tion is impossible for these persons who have in addition serious difficulties to control physically their environment. In particular, they are not able to use standard input devices of a computer. Most of the time, they can only handle a single switch de- vice. As a result, communicating with an AAC sys- tem consists of typing messages by means of a vir- tual table of symbols (words, letters or icons) where the user successively selects the desired items. Basically, an AAC system, such as FASTY (Trost et al. 2005) or SIBYLLE (Schadle et al, 2004), consists of four components. At first, one finds a physical input interface connected to the computer. This device is adapted to the motion capacities of the user. When the latter must be restricted to a single switch (eye glimpse or breath detector, for instance), the control of the environment is reduced to a mere Yes/No command. Secondly, a virtual keyboard is displayed on screen. It allows the user to select successively the symbols that compose the intended message. In SIBYLLE, key selection is achieved by pointing let- ters through a linear scan procedure: a cursor suc- cessively highlights each key of the keyboard. The last two components are a text editor (to write e-mails or other documents) and …

Content truncated to 500 words