Computational Semantics in the Natural Language Toolkit

Proceedings of the 2006 Australasian Language Technology Workshop (ALTW2006), pages 26–33. Computational Semantics in the Natural Language Toolkit Ewan Klein School of Informatics University of Edinburgh Scotland, UK ewan@inf.ed.ac.uk Abstract NLTK, the Natural Language Toolkit, is an open source project whose goals include providing students with software and lan- guage resources that will help them to learn basic NLP. Until now, the pro- gram modules in NLTK have covered such topics as tagging, chunking, and parsing, but have not incorporated any aspect of semantic interpretation. This paper de- scribes recent work on building a new se- mantics package for NLTK. This currently allows semantic representations to be built compositionally as a part of sentence pars- ing, and for the representations to be eval- uated by a model checker. We present the main components of this work, and con- sider comparisons between the Python im- plementation and the Prolog approach de- veloped by Blackburn and Bos (2005). 1 Introduction NLTK, the Natural Language Toolkit,1 is an open source project whose goals include providing stu- dents with software and language resources that will help them to learn basic NLP. NLTK is imple- mented in Python, and provides a set of modules (grouped into packages) which can be imported into the user’s Python programs. Up till now, the modules in NLTK have covered such topics as tagging, chunking, and parsing, but have not incorporated any aspect of semantic inter- pretation. Over the last year, I have been working on remedying this lack, and in this paper I will de- scribe progress to date. In combination with the 1http://nltk.sourceforge.net/ NLTK parse package, NLTK’s semantics pack- age currently allow semantic representations to be built compositionally within a feature-based chart parser, and allows the representations to be evalu- ated by a model checker. One source of inspiration for this work came from Blackburn and Bos’s (2005) landmark book Representation and Inference for Natural Lan- guage (henceforth referred to as B&B). The two primary goals set forth by B&B are (i) automating the association of semantic representations with expressions of natural language, and (ii) using log- ical representations of natural language to auto- mate the process of drawing inferences. I will be focussing on (i), and the related issue of defin- ing satisfaction in a model for the semantic rep- resentations. By contrast, the important topic of (ii) will not be covered—as yet, there are no the- orem provers in NLTK. That said, as pointed out by B&B, for many inference problems in NLP it is desirable to call external and highly sophisticated first-order theorem provers. One notable feature of B&B is the use of Pro- log as the language of implementation. It is not hard to defend the use of Prolog in defining logical representations, given the presence of first-order clauses in Prolog and the fundamental role of res- olution in Prolog’s model of computation. Never- theless, in some circumstances it may be helpful to offer students access to an alternative framework, such as the Python …

Content truncated to 500 words