Incremental chart parsing with predictive hints

Incremental Chart Parsing with Predictive Hints Abstract This paper describes an incremental chart parser that generates look-ahead catego- ries on the fly for a controlled natural lan- guage. These predictive hints tell the au- thor what kind of syntactic (or semantic) structure can follow the current input string and thereby aim at helping the au- thor to reduce the cognitive burden to learn and remember the rules of the con- trolled language. The parser can handle modifications (insertion, deletion, and re- placement) to the input string without the need to reparse the entire string. These modifications are a function of the size of the tokens changed rather than the size of the entire input. 1 Introduction Over the last decade two types of controlled natu- ral languages have been developed: human-ori- ented controlled languages and machine-oriented controlled languages (Huijsen, 1998; O’Brien, 2003). The main goal of human-oriented controlled natural languages is to improve the readability and understandability of technical documents for hu- man readers, especially non-native speakers (AECMA, 2001). Machine-oriented controlled nat- ural languages, on the other hand, try to ease the translation process (Kamprath, 1998) and to facili- tate the subsequent inference processes (Fuchs and Schwertel, 2003; Schwitter et al., 2003; Sukkarie, 2003). In general, a controlled natural language can be defined as a subset of a natural language that has been restricted with respect to its grammar and its lexicon. Grammatical restrictions usually result in less complex and less ambiguous texts. Lexical re- strictions reduce the size of the vocabulary and the meaning of the lexical entries for a particular ap- plication domain. To allow writing unambiguous and precise spe- cifications, we have developed a machine-oriented controlled natural language (PENG – Processable ENGlish) for knowledge representation (Schwitter, 2002). Specifications written in PENG can be translated unambiguously into first-order predicate logic via discourse representation structures and can be automatically checked for consistency and informativeness with the help of third-party reas- oning services (McCune, 2001; Bos, 2001; McCune, 2003; Bos, 2003). It is well known that writing documents in a controlled natural language is hard and time-con- suming without the support of intelligent writing assistance (Goyvaerts, 1996; Power et al., 2003). To ease the writing process and to guarantee well- formed syntactic structures, PENG uses a look- ahead editor that displays after each word form a set of syntactic categories that inform the author how the input string can be continued (Schwitter et al., 2003). The look-ahead editor of PENG communicates with a chart parser that processes a unification- based grammar. The chart parser generates these predictive hints dynamically while the text is writ- ten and thereby enforces the restrictions placed upon the language. Apart from these hints, the chart parser also generates for each input string a discourse representation structure (Kamp and Reyle, 1993) as well as a paraphrase (Schwitter Rolf Schwitter Centre for Language Technology Macquarie University Sydney, NSW 2109, Australia schwitt@ics.mq.edu.au and Ljungberg, 2002) that reflects the interpreta- tion of the machine in controlled natural language. So …

Content truncated to 500 words