Zurück zur Übersicht

Coreference Resolution by Global Optimization under Linguistic Constraints

Gesuchsteller/in Klenner Manfred
Nummer 118108
Förderungsinstrument Projekte
Forschungseinrichtung Institut für Computerlinguistik Universität Zürich
Hochschule Universität Zürich - ZH
Hauptdisziplin Schwerpunkt Germanistik und Anglistik
Beginn/Ende 01.11.2007 - 31.10.2010
Bewilligter Betrag 157'555.00
Alle Daten anzeigen

Alle Disziplinen (2)

Schwerpunkt Germanistik und Anglistik

Keywords (12)

coreference resolution; anaphora; corpus annotation; machine learing; anaphora resolution; bridging; statistical NLP; optimization; integer linear programming; binding constraints; machine learning; classifier

Lay Summary (Englisch)

Lay summary
The project aims at the conceptualization and implementation of a coreference resolution system for German. The most important new feature of our model is that it takes full advantage of linguistically motivated global constraints, e.g. the transitivity of the anaphoric relation. Moreover, we will try to bridge the gap between purely data-driven machine learning approaches to coreference resolution on the one hand, and traditional preference-based, theory-driven, non-statistical methods on the other.

Methodologically this is to be achieved through an architecture that proved successful in previous work on different NLP tasks. It couples one or more machine learning classifiers that can take only local decisions, with a constraint-based optimization tool that enforces global constraints on the output of the classifier and at the same time finding an optimal solution. In such a scenario, the probabilities of the classifier no longer provide the final solution, as they usually do, but count as probabilistic suggestions for a globally consistent solution. These probabilistic suggestions are used as weights in the optimization task. The constraints at this level are mostly of a linguistic nature (e.g. intra-sentential binding constraints), which means that our model brings linguistic theory back into the system, to act as prescriptive knowledge that guides empirically derived preferences towards a consistent solution. A feasibility study (Klenner, 2007) has shown that such models do not have to become brittle. There is also evidence that such an approach can boost performance (precision,recall) considerably without a substantial increase in computation time.

Manfred Klenner (2007). Enforcing consistency on coreference sets.
In Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP07).

Direktlink auf Lay Summary Letzte Aktualisierung: 21.02.2013

Verantw. Gesuchsteller/in und weitere Gesuchstellende