Constraint Based Integration of Deep and Shallow Parsing Techniques

Constraint Based Integration of Deep and Shallow Parsing Techniques Michael Daum and Kilian A. Foth and Wolfgang Menzel Natural Language Systems Department of Computer Science University of Hamburg (michalfothlwolfgang) @nats . informatik .uni—hamburg . de Abstract To investigate the contributions of tag- gers or chunkers to the performance of a deep syntactic parser, Weighted Constraint Dependency Grammars have been extended to also take into con- sideration information from external sources. Using a weak information fu- sion scheme based on constraint opti- mization techniques, a parsing accuracy has been achieved which is comparable to other (stochastic) parsers. 1 Introduction To investigate the contributions of different shal- low processing components like taggers or chun- kers to the performance of a deep syntactic parser, Weighted Constraint Dependency Gram- mars (WCDG) have been extended to also take into consideration information from external sources. Constraints in the WCDG framework are (partially) defeasible conditions on admissible structural configurations in a dependency tree for a given natural language utterance. By defining conditions (1) on word forms and their positional relationships, (2) on links between word forms, and (3) on the labels attached to these links, they license certain dependency relations or combina- tions thereof and assign a score to them. These scores give an estimate of the appropriateness of a particular partial structure. Constraint optimiza- tion techniques are used to obtain the structure with the best overall score, which is chosen as the optimal representation of the utterance. Due to the use of weighted constraints throughout the gram- mar, WCDG comes with a number of advantages: • A WCDG grammar is able to accomodate conflicting requirements as they can often be found in natural language, e.g. in case of ordering preferences. WCDG shares this property with Optimality Theory (Prince and Smolensky, forthc). • Since constraints can be violated, the gram- mar also accepts unusual or even deviant in- put (usually with a lower global score) and assigns a structural interpretation to it. • To a large degree a WCDG parser does not depend crucially on a complete set of con- straints. Although performance degrades if some constraints are deactivated, it usually does so in a graceful manner. • Constraints offer an ideal interface to inte- grate external processing components as ad- ditional sources of evidence into the decision about the optimal structural interpretation. It is this last advantage which this paper is fo- cussing upon. Shallow processing components like taggers or chunkers can be integrated into the deep analysis simply by providing additional con- straints. If a scoring scheme is also available for such an external contribution (and it usually is), it can be combined with the internal scores as- signed by the constraints of the grammar. Sim- ilar to the arbitration of conflicting requirements within the grammar itself, the general optimiza- tion procedure allows the WCDG parser to han- dle possible conflicts between predictions of dif- ferent origin. This weak integration of hypotheses not only makes the approach an ideal platform for information fusion. Moreover the …

Content truncated to 500 words