Acquiring Topic Features to improve Event Extraction: in Pre-selected and Balanced Collections

Proceedings of Recent Advances in Natural Language Processing, pages 9–16, Hissar, Bulgaria, 12-14 September 2011. Acquiring Topic Features to Improve Event Extraction: in Pre-selected and Balanced Collections Shasha Liao Ralph Grishman Computer Science Department Computer Science Department New York University New York University liaoss@cs.nyu.edu grishman@cs.nyu.edu Abstract Event extraction is a particularly challenging type of information extraction (IE) that may require inferences from the whole article. However, most current event extraction systems rely on local information at the phrase or sentence level, and do not consider the article as a whole, thus limiting extraction performance. Moreover, most annotated corpora are artificially enriched to include enough positive samples of the events of interest; event identification on a more balanced collection, such as unfiltered newswire, may perform much worse. In this paper, we investigate the use of unsupervised topic models to extract topic features to improve event extraction both on test data similar to training data, and on more balanced collections. We compare this unsupervised approach to a supervised multi-label text classifier, and show that unsupervised topic modeling can get better results for both collections, and especially for a more balanced collection. We show that the unsupervised topic model can improve trigger, argument and role labeling by 3.5%, 6.9% and 6% respectively on a pre-selected corpus, and by 16.8%, 12.5% and 12.7% on a balanced corpus. 1 Introduction The goal of event extraction is to identify instances of a class of events in free text, along with their arguments. In this paper, we focus on the ACE 2005 event extraction task, which involved a set of 33 generic event types and subtypes appearing frequently in the news. It generally expresses the core arguments plus place and time information of a single event, like Attack, Marry or Arrest. In general, identifying an ACE event can be quite difficult. Given a narrow scope of information, even a human cannot make a confident decision. For example, for the sentence: (1) So he returned to combat … it is hard to tell whether it is an Attack event, which is defined as a violent physical act causing harm or damage, or whether it refers to a more innocent endeavor such as a tennis match. A broader field of view is often helpful to understand how facts tie together. If we read the whole article, and find it to be a terrorist story, it is easy to tag this as an Attack event; however, if it is in a tennis report, we probably won’t tag it as an Attack event. The problem of event identification is exacerbated if we shift to corpora with a topic distribution different from the training and official test corpus. In general, an effort is made to have the test corpora be representative of the sort of texts to which the NLP process is intended to be applied. In the case of the event extraction, this has generally been news sources such as newswires or broadcast news transcripts. However, a particular event type is …

Content truncated to 500 words