Jointly Combining Implicit Constraints Improves Temporal Ordering

Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing, pages 698–706, Honolulu, October 2008. c⃝2008 Association for Computational Linguistics Jointly Combining Implicit Constraints Improves Temporal Ordering Nathanael Chambers and Dan Jurafsky Department of Computer Science Stanford University Stanford, CA 94305 {natec,jurafsky}@stanford.edu Abstract Previous work on ordering events in text has typically focused on local pairwise decisions, ignoring globally inconsistent labels. How- ever, temporal ordering is the type of domain in which global constraints should be rela- tively easy to represent and reason over. This paper presents a framework that informs lo- cal decisions with two types of implicit global constraints: transitivity (A before B and B be- fore C implies A before C) and time expression normalization (e.g. last month is before yes- terday). We show how these constraints can be used to create a more densely-connected network of events, and how global consis- tency can be enforced by incorporating these constraints into an integer linear programming framework. We present results on two event ordering tasks, showing a 3.6% absolute in- crease in the accuracy of before/after classifi- cation over a pairwise model. 1 Introduction Being able to temporally order events is a neces- sary component for complete document understand- ing. Interest in machine learning approaches for this task has recently been encouraged through the cre- ation of the Timebank Corpus (Pustejovsky et al., 2003). However, most work on event-event order- ing has focused on improving classifiers for pair- wise decisions, ignoring obvious contradictions in the global space of events when misclassifications occur. A global framework to repair these event or- dering mistakes has not yet been explored. This paper addresses three main factors involved in a global framework: the global optimization al- gorithm, the constraints that are relevant to the task, and the level of connectedness across pairwise de- cisions. We employ Integer Linear Programming to address the first factor, drawing from related work in paragraph ordering (Bramsen et al., 2006). After finding minimal gain with the initial model, we ex- plore reasons for and solutions to the remaining two factors through temporal reasoning and transitivity rule expansion. We analyze the connectivity of the Timebank Cor- pus and show how textual events can be indirectly connected through a time normalization algorithm that automatically creates new relations between time expressions. We show how this increased con- nectivity is essential for a global model to improve performance. We present three progressive evaluations of our global model on the Timebank Corpus, showing a 3.6% gain in accuracy over its original set of re- lations, and an 81% increase in training data size from previous work. In addition, we present the first results on Timebank that include an unknown rela- tion, establishing a benchmark for performance on the full task of document ordering. 2 Previous Work Recent work on classifying temporal relations within the Timebank Corpus built 6-way relation classifiers over 6 of the corpus’ 13 relations (Mani et al., 2006; Mani et al., 2007; Chambers et al., 2007). A …

Content truncated to 500 words