Towards a Psycholinguistics-Based Parser.

Doan Nguyen Hai


Le 6 novembre 2002 à 11 h 30

Salle 3195, Pavillon André-Aisenstadt

This work consists of designing and implementing a natural language parsing prototype which is based on psycholinguistic knowledge of human sentence parsing. Traditional NLP parsing based on context-free grammars often gives all solutions for an input sentence. The number of solutions is exponential to the length of the input, and is therefore a great problem for any practical application which only needs the best solution. There are parsing strategies which enumerate the solutions one by one, but they hardly guarantee that the first output is actually the best solution. On the contrary, human sentence processing is generally very accurate and efficient. Simulating this process in a computer parser is thus an interesting approach. Unfortunately, it seems that very few influential prototypes have been implemented using this approach, maybe due to alack of understanding of the human parsing process. However, recent progress in psycholinguistics has provided us with better knowledge on this subject. I propose an algorithm based on the serial model of human parsing, in which the parser tries to produce only the best solution, i.e. the most plausible parse, taking into account all the linguistic knowledge the parser has during the analysis process. In general, the parser reads the input sentence word-by-word, and for each word, tries to attach it into the current (partial) parse tree, guided by psycholinguistic considerations, such as Late Closure, Minimal Attachment, and the Construal Hypothesis. If the attachment fails, reanalysis will be triggered, i.e. the parser will revise and modify the current parse tree to make it compatible with the current word. Attachment and reanalysis are carried out by mechanisms which exploit context information to achieve efficiency. The prototype implemented so far covers a relatively large fragment of present/past participles as adjectives), verb phrases (subcategorization, auxiliaries, tense, passive, infinitive), relative clauses, questions (yes-no and wh-questions), movement and empty elements, etc. In future work, I will focus on some remaining difficult parsing issues, such as lexical ambiguities, conjunctions, punctuation, etc. The prototype shows that a psycholinguistic-based parser is implementable, and may be very beneficial. The approach is promising in that it tries to combine the human's amazing performance with computer-specific techniques. An interesting research direction to pursue in the future is to integrate psycholinguistic studies on semantics and pragmatics into the computer parser. Applications of statistical methods to the model will also be worth studying.

Pour recevoir les annonces hebdomadaires par courriel, visitez

Liste des autres séminaires