Scalable Deep Linguistic Processing: Mind the Lexical Gap

Scalable Deep Linguistic Processing: Mind the Lexical Gap∗ Timothy Baldwin Department of Computer Science and Software Engineering and NICTA Victoria Laboratories, University of Melbourne, VIC 3010, Australia, tim@csse.unimelb.edu.au Abstract. Coverage has been a constant thorn in the side of deployed deep linguistic process- ing applications, largely because of the difficulty in constructing, maintaining and domain- tuning the complex lexicons that they rely on. This paper reviews various strands of research on deep lexical acquisition (DLA), i.e. the (semi-)automatic creation of linguistically-rich language resources, particularly from the viewpoint of DLA for precision grammars. 1. Introduction Over recent years, computational linguistics has benefitted considerably from advances in statis- tical modelling and machine learning, culminating in methods capable of deeper, more accurate automatic analysis, over a wider range of languages. Implicit in much of this work, however, has been the existence of deep language resources (DLR hereafter), that is resources which en- code precise symbolic linguistic knowledge. DLRs aim either to capture a particular feature of a language, such as verb argument structure (e.g. COMLEX (Grishman et al., 1994) or PropBank (Palmer et al., 2005)) or lexical semantics (e.g. WordNet (Fellbaum, 1998) or FrameNet (Baker et al., 1998)), or alternatively to model a language in its entirety, in the form of a precision grammar (e.g. the English Resource Grammar (Flickinger, 2002), various ParGram grammars (Butt et al., 2002) or CCGBank (Hockenmaier and Steedman, 2007)). In line with Baldwin (2005a), we consider the development of DLRs to be made up of two basic tasks: (1) design of a data representation to systematically capture the generalisations and idiosyncracies of the dataset of interest (system design); and (2) classification of data items ac- cording to the predefined data representation (data classification). In the case of a deep grammar, for example, system design encompasses the construction of the system of lexical types, templates, and/or phrase structure rules, and data classification corresponds to the determination of the lexi- cal type(s) each individual lexeme conforms to. Naturally, the two tasks are tightly integrated, and actual DLR construction involves successive iterations over the design of analyses (with the data classification informing the system design), and the application of those analyses over data (with the system design informing the data classification). One common problem pervading all DLRs is that of coverage, which we define as the pro- portion of relevant data points in a representative text that are adequately described in a given DLR. Based on our bifurcation of the DLR development process, lack of coverage is caused by either: (a) deficiencies in the system design, e.g. a precision grammar not containing an analysis of a given construction, or a lexical semantic DLR not capturing a particular verb frame type; or (b) deficiencies in data classification, that is certain words simply not having been classified, ∗Copyright 2007 by Timothy Baldwin 3 or being only partially described in the DLR. This arises through simple resource constraints on DLR development, domain effects, or the DLR not capturing generalisations effectively (e.g. not supporting conversion …

Content truncated to 500 words