Predicting Linguistic Difficulty by Means of a Morpho-Syntactic Probabilistic Model

Predicting Linguistic Difficulty by Means of a Morpho-Syntactic Probabilistic Model Philippe Blache and St´ephane Rauzy Aix-Marseille Univ & CNRS LPL, 5 Avenue Pasteur, 13100 Aix-en-Provence, France {blache;rauzy}@lpl-aix.fr Abstract. We propose in this paper a new contribution to the evaluation of linguistic diffi- culty. At the opposite of classical approaches relying on syntax, we show that a probabilis- tic morpho-syntactic analysis provides information enough to calculate different parameters, including surprisal. This method constitutes an original and robust model to linguistic com- plexity. It constitutes a solution towards complexity experiments using spoken languages. Keywords: Linguistic complexity, difficulty model, computational psycholinguistics, sen- tence processing. 1 Introduction Linguistic complexity is an important problem both from linguistic, computational and cognitive perspectives: it is important to know for a given language what are the phenomena or the construc- tions that can lead to a processing problem for a human subject. Such knowledge is the basis of a better understanding of the cerebral mechanisms of language processing. It is also necessary in the persepctive of the development of a cognitively-plausible automatic parser. Today, several compu- tational models have been developed in this perspective. However, most of them are partial, only take into account local phenomena, and are experimented on highly controlled written text (cf. (Gibson, 2000), (Vasishth, 2003) or (Lewis and Vasishth, 2005) for example). More recent works propose to go one step forward in experimenting models on larger corpora (see in particular (Dem- berg and Keller, 2008), making it possible to validate complexity models. One of the conclusion of these works is that generic models have to bring together many different complexity parameters in order to take into account a large variety of phenomena. Moreover, it still remains necessary to develop and validate generic computational models dealing with unrestricted linguistic material, in particular spoken languages. Among the different parameters involved in its definition, the evaluation of the difficulty human subjects encounter in language production or perception plays a major role. In the literature, one of the most frequently used approaches relies on surprisal effects that can occur when integrating a new word to the structure (cf. (Hale, 2001)): roughly speaking, surprisal is high when the probability for a word to occur in a specific context is low. This parameter is evaluated as the negative logarithm of the probability of a word wt to appear in its context, taking into account the sequence w1...wt−1. Surprisal is evaluated as a function of the probabilities of the possible parses when integrating the word wt. When using a probabilistic grammar, evaluating surprisal consists in measuring the difference between the probability of all parses spanning the sequence w1...wt−1 and those spanning w1...wt: Copyright 2011 by Philippe Blache and St´ephane Rauzy 25th Pacific Asia Conference on Language, Information and Computation, pages 160–167 160 SIt = − log P(w1...wt) P(w1...wt−1) (1) Surprisal has been shown by different studies as being correlated with difficulty. This has been demonstrated experimentally, in particular by several works relying on eye-tracking in reading tasks (cf. (Boston et al., …

Content truncated to 500 words