Learning Constraints for Consistent Timeline Extraction

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 873–882, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Learning Constraints for Consistent Timeline Extraction David McClosky and Christopher D. Manning Natural Language Processing Group Computer Science Department Stanford University, Stanford, CA, USA {mcclosky,manning}@stanford.edu Abstract We present a distantly supervised system for extracting the temporal bounds of fluents (re- lations which only hold during certain times, such as attends school). Unlike previous pipelined approaches, our model does not as- sume independence between each fluent or even between named entities with known con- nections (parent, spouse, employer, etc.). In- stead, we model what makes timelines of flu- ents consistent by learning cross-fluent con- straints, potentially spanning entities as well. For example, our model learns that someone is unlikely to start a job at age two or to marry someone who hasn’t been born yet. Our sys- tem achieves a 36% error reduction over a pipelined baseline. 1 Introduction Many information extraction (IE) systems tradition- ally extracted just relations, but a great many real world relations such as attends school or has spouse vary over time. To capture this, some recent IE systems have extended their focus from relations to fluents (relations combined with temporal bounds). This can be seen in the temporal slot filling track in the TAC-KBP 2011 shared task (Ji et al., 2011). A direct application of this work is the automatic im- provement of online resources such as Freebase and Wikipedia infoboxes. Indirect applications include question answering systems. Fluents can be grouped together to form time- lines (see Figure 1 for an example) and provide eas- ily capturable consistency constraints. Our goal is Figure 1: A timeline of two named entities. Each time span represents a fluent (a relation with temporal bounds). Temporal bounds are denoted by spans on the timeline. Fluents can create links between entities (e.g., marriage). to learn these constraints and use them to produce more accurate timelines of significant events for people and organizations. For example, it is com- mon knowledge that someone cannot attend a school if they haven’t been born yet. Constraints on con- sistent timelines do not need to be hard constraints, though: it is rare, although possible, to become the CEO of a company at the age of 21. Despite the rich constraints on valid timelines, there is relatively little work on exploiting these con- straints for mutual disambiguation. Many existing systems extract different parts of a timeline sepa- rately and use heuristics to combine them. These heuristics tend to optimize only local consistency (within a single fluent) but ignore more global con- straints across fluents (e.g., attending a school be- fore being born) or across fluents of two linked entities (e.g., attending a school before the school was founded). In this work, we explore using joint inference to enforce these constraints. We show that these techniques can yield substantial improve- ments. Additionally, our general approach is not …

Content truncated to 500 words