A Modular Architecture for the Wide-Coverage Translation of Natural Language Texts into Predicate Logic Formulas

A Modular Architecture for the Wide-Coverage Translation of Natural Language Texts into Predicate Logic Formulas Yusuke Miyao* Alastair Butler†‡ Kei Yoshimoto‡ Jun’ichi Tsujii§ *National Institute of Informatics †Japan Science and Technology Agency, PRESTO ‡Center for the Advancement of Higher Education, Tohoku University §Graduate School of Information Science and Technology, Tokyo University Abstract. We present a new method for translating unrestricted natural language texts into predicate logic formulas. This relies on the semantic evaluation procedure of Scope Control Theory (SCT), a variant of Dynamic Semantic formalisms. The key benefit is that parsed syntactic structures are shown to form sufficient input for semantic evaluation, eliminating the need to build distinct semantic expressions to feed semantic evaluation. To have parsed syntactic structures for SCT to evaluate we apply an existing wide-coverage syntactic parser by converting the parser output into a form SCT can receive. This modularity has led to the rapid attainment of a broad coverage on real text. An experiment revealed our system achieved 82.7% coverage on real-world sentences, generating representations that make ex- plicit the scopes of quantifiers (e.g., ∃x), operators (e.g., negation), connectives (e.g., con- junction) and embedding predicates (e.g., thinks), while also capturing the inter and intra sentential dependencies and cross-sentential anaphoric dependencies that connect predicates. Keywords: semantic evaluation, parser output, predicate logic formulas, unrestricted natural language 1 Introduction Translating natural language expressions into logical formulas has been extensively studied for decades, attracting theoretical and practical interests. To date techniques have essentially involved adding methods of building a semantic expression piecewise to the process of identifying syntactic structure with a parser. In this paper we propose a departure from this tradition, with a system of semantic evaluation, Scope Control Theory or SCT (Butler, 2010), that accepts parsed syntactic structures as input. We employ semantic evaluation to generate explicitly semantic translations in the form of predicate logic formulas. It is also possible to have SCT evaluation generate the representations of alternative formalisms and offer different granularities for making semantic in- formation explicit (e.g., generate encodings with more fine-grained presuppositional information, modal information, or tense information). In using semantic evaluation to generate semantic translations a notable contribution is the elimination of the need for enriched guidance information that is extra to a conventional parsed form. Notably the co-indexing of syntactic constituents is rendered unnecessary. Moreover sen- tence information and discourse information are dealt with equally at the time of evaluation as providing binding information, which allows for their seamless integration. The language formalism of SCT is supplemented by a compact grammar to determine the contribution of morphosyntactic information. This offers flexibility in what is acceptable as input, so there is no need to develop a SCT-specific parser from scratch. Here we adopt an HPSG-based syntactic parser (Miyao and Tsujii, 2008), because it offers wide coverage and high accuracy, and provides detailed syntactic information that is sufficient for SCT. Although the output of the parser is not directly compatible with the input assumed by SCT, a small conversion program can fill this …

Content truncated to 500 words