Incorporating Linguistics Constraints into Inductive Logic Programming

In: Proceedings of CoNLL-2000 and LLL-2000, pages 184-193, Lisbon, Portugal, 2000. Incorporating Linguistics Constraints into Inductive Logic Programming James Cussens Stephen Pulman Dept. of Computer Science University of Cambridge Computer Laboratory University of York New Museums Site, Pembroke Street Heslington, York, YO10 5DD, UK Cambridge CB2 3QG, UK j c@cs. york. ac. uk Stephen. Pulman@c 1. cam. ac. uk Abstract We report work on effectively incorporating lin- guistic knowledge into grammar induction. We use a highly interactive bottom-up inductive logic programming (ILP) algorithm to learn 'missing' grammar rules from an :incomplete grammar. Using linguistic constraints on, for example, head features and gap threading, re- duces the search space to such an extent that, in the small-scale experiments reported here, we can generate and store all candidate gram- mar rules together with information about their coverage and linguistic properties. This allows an appealingly simple and controlled method for generating linguistically plausible grammar rules. Starting from a base of highly spe- cific rules, we apply least general generalisation and inverse resolution to generate more general rules. Induced rules are ordered, for example by coverage, for easy inspection by the user and at any point, the user can commit to a hypothe- sised rule and add it to the grammar. Related work in ILP and computational linguistics is discussed. 1 Introduction A major advantage of inductive logic program- ming is the ability to incorporate domain knowl- edge (background knowledge) into the inductive process. In ILP domain knowledge is usually encoded by (i) a set of definite clauses declar- ing rules and facts which are true (or assumed to be true) in the domain and (ii) extra-logical constraints on the hypothesis space. The ILP approach thus allows a very direct and flexible method of expressing domain knowledge. In this paper, we report on continuation of the work described in (Cussens and Pulman, 2000), which attempts to maximise the effectiveness of linguistic knowledge when inducing a grammar. We take an existing grammatical formalism (de- rived from the FraCaS Project (1996)) and ex- tend it with inductive capabilities, rather than shoe-horning a grammar learning problem into a form suitable for some particular ILP algo- rithm. This has major practical benefits, since the required linguistic knowledge can be en- coded in a linguistically natural manner. As in all real applications of ILP most effort is required in 'getting the background knowledge right'. Being able to express this knowledge in a representation specifically developed to enable linguists to write down a grammar makes this step easier and quicker. The paper is organised in a manner analo- gous to that of our algorithm. In Section 2, we describe how to generate naive grammar rules directly from the chart produced during a failed parse. The essentials of this approach have al- ready been described in (Cussens and Pulman, 2000), but we briefly describe it here for com- pleteness and also because we have altered its implementation. Section 3 describes the most important step of the algorithm--the represen- tation and use …

Content truncated to 500 words