Integrated Shallow and Deep Parsing: TopP Meets HPSG

Integrated Shallow and Deep Parsing: TopP meets HPSG Anette Frank, Markus Becker z, Berthold Crysmann, Bernd Kiefer and Ulrich Sch¨afer DFKI GmbH School of Informatics z 66123 Saarbr¨ucken, Germany University of Edinburgh, UK firstname.lastname@dfki.de M.Becker@ed.ac.uk Abstract We present a novel, data-driven method for integrated shallow and deep parsing. Mediated by an XML-based multi-layer annotation architecture, we interleave a robust, but accurate stochastic topological field parser of German with a constraint- based HPSG parser. Our annotation-based method for dovetailing shallow and deep phrasal constraints is highly flexible, al- lowing targeted and fine-grained guidance of constraint-based parsing. We conduct systematic experiments that demonstrate substantial performance gains.1 1 Introduction One of the strong points of deep processing (DNLP) technology such as HPSG or LFG parsers certainly lies with the high degree of precision as well as detailed linguistic analysis these systems are able to deliver. Although considerable progress has been made in the area of processing speed, DNLP systems still cannot rival shallow and medium depth tech- nologies in terms of throughput and robustness. As a net effect, the impact of deep parsing technology on application-oriented NLP is still fairly limited. With the advent of XML-based hybrid shallow- deep architectures as presented in (Grover and Las- carides, 2001; Crysmann et al., 2002; Uszkoreit, 2002) it has become possible to integrate the added value of deep processing with the performance and robustness of shallow processing. So far, integration has largely focused on the lexical level, to improve upon the most urgent needs in increasing the robust- ness and coverage of deep parsing systems, namely 1This work was in part supported by a BMBF grant to the DFKI project WHITEBOARD (FKZ 01 IW 002). lexical coverage. While integration in (Grover and Lascarides, 2001) was still restricted to morphologi- cal and PoS information, (Crysmann et al., 2002) ex- tended shallow-deep integration at the lexical level to lexico-semantic information, and named entity expressions, including multiword expressions. (Crysmann et al., 2002) assume a vertical, ‘pipeline’ scenario where shallow NLP tools provide XML annotations that are used by the DNLP system as a preprocessing and lexical interface. The per- spective opened up by a multi-layered, data-centric architecture is, however, much broader, in that it en- courages horizontal cross-fertilisation effects among complementary and/or competing components. One of the culprits for the relative inefficiency of DNLP parsers is the high degree of ambiguity found in large-scale grammars, which can often only be re- solved within a larger syntactic domain. Within a hy- brid shallow-deep platform one can take advantage of partial knowledge provided by shallow parsers to pre-structure the search space of the deep parser. In this paper, we will thus complement the efforts made on the lexical side by integration at the phrasal level. We will show that this may lead to considerable per- formance increase for the DNLP component. More specifically, we combine a probabilistic topological field parser for German (Becker and Frank, 2002) with the HPSG parser of (Callmeier, 2000). The HPSG grammar used is the one originally …

Content truncated to 500 words