A psychologically plausible and computationally effective approach to learning syntax

A Psychologically Plausible and Computationally Effective Approach to Learning Syntax Stephen Watkinson and Suresh Manandhar, Department of Computer Science, University of York, York YO10 5DD, UK. Abstract Computational learning of natural lan- guage is often attempted without using the knowledge available from other re- search areas such as psychology and linguistics. This can lead to systems that solve problems that are neither theoretically or practically useful. In this paper we present a system CLL which aims to learn natural language syntax in a way that is both compu- tationally effective and psychologically plausible. This theoretically plausible system can also perform the practically useful task of unsupervised learning of syntax. CLL has then been applied to a corpus of declarative sentences from the Penn Treebank (Marcus et al., 1993; Marcus et al., 1994) on which it has been shown to perform comparatively well with respect to much less psycho- logically plausible systems, which are significantly more supervised and are applied to somewhat simpler problems. 1 Introduction Computational learning of natural language can be considered from two common perspectives. Firstly, there is the psychological perspective, which leads to the investigation of learning prob- lems similar to those faced by people and the building of systems that seek to model human lan- guage learning faculties. Secondly, there is the computational perspective, which seeks to build systems that effectively solve language learning problems that are of practical importance. In principle, there is significant overlap be- tween these two perspectives. The most common language learning problems that we wish to solve computationally are frequently those that humans have to solve. For example when humans learn language, especially syntax, it seems to be in a mostly unsupervised setting i.e. there is no an- notation of training examples. From a computa- tional perspective, while there are some annotated resources available, in general we have very large amounts of unannotated text available from which we desire to be able to extract grammars, mean- ing etc. Given this overlap, it seems wise to in- vestigate what we know of the human approach, as humans are good at solving these problems. In this work we present a system for learning syntax that seeks to maintain both the psycho- logical and computational perspectives. We also show that this is an effective way to build natural language learning systems. We represent the syn- tactic knowledge using the Categorial Grammar (CG) formalism, so in Section 2 we introduce CG. In Section 3 we aim to define the problem that is to be solved in a way that is psychologically plau- sible. This is followed in Section 4 by the descrip- tion of CLL a computational effective solution to the problem, which we maintain is also reason- ably psychologically plausible. Related work is discussed in Section 5. The results of experiments using CLL on examples from the Penn Treebank are presented in Section 6 and we draw some con- clusions from this work in Section 7. 2 Categorial Grammar Categorial Grammar (CG) (Steedman, 1993; …

Content truncated to 500 words