Cascading Use of Soft and Hard Matching Pattern Rules for Weakly Supervised Information Extraction

Cascading Use of Soft and Hard Matching Pattern Rules for Weakly Supervised Information Extraction Jing Xiao School of Computing, National University of Singapore, 117543 xiaojing@comp.nus.edu.sg Tat-Seng Chua School of Computing, National University of Singapore, 117543 chuats@comp.nus.edu.sg Hang Cui School of Computing, National University of Singapore, 117543 cuihang@comp.nus.edu.sg Abstract Current rule induction techniques based on hard matching (i.e., strict slot-by-slot matching) tend to fare poorly in extracting information from natural language texts, which often exhibit great variations. The reason is that hard matching techniques result in relatively high precision but low recall. To tackle this problem, we take advantage of the newly proposed soft pattern rules which offer high recall through the use of probabilistic matching. We propose bootstrapping framework in which soft and hard a matching pattern rules are combined in a cascading manner to realize a weakly supervised rule induction scheme. The system starts with a small set of hand-tagged instances. At each iteration, we first generate soft pattern rules and utilize them to tag new training instances automatically. We then apply hard pattern rule induction on the overall tagged data to generate more precise rules, which are used to tag the data again. The process can be repeated until satisfactory results are obtained. Our experimental results show that our bootstrapping scheme with two cascaded learners approaches the performance of a fully supervised information extraction system while using much fewer hand- tagged instances. 1 Introduction Information Extraction (IE) aims to extract specific information items of interest from free or semi- structured texts, and pattern rule induction is one of the most common techniques for IE tasks (Muslea, 1999). There has been much work in learning extraction pattern rules from tagged data, e.g., AutoSlog-TS (Riloff, 1996), WHISK (Soderland, 1999) and LP2 (Ciravegna, 2001). In a typical IE system, generalized pattern rules are usually represented as regular expressions and matched against test instances through exact matching for each slot, which we call hard matching. Utilizing hard matching pattern rules could obtain precise results from test instances. However, the approach is problematic in dealing with natural language text, such as news articles, which often exhibits great variations in both lexical and syntactic constructions. For instance, in the terrorism domain, given a common rule “<victim> be kidnapped by …”, hard matching pattern rules cannot pick up the instance “<victim> , kidnapped by …” due to the mismatch in only one token. Such hard matching techniques often result in low recall. To achieve flexibility in pattern matching for natural language texts, soft matching pattern rules have been proposed for question answering (Cui, et al., 2004). Soft pattern rules match test instances using a probabilistic model to better accommodate variations in expressions. However, differing from the question answering problem, the IE task needs to precisely locate the boundaries of the extracted slots. As such, soft pattern rules may not meet the precision requirement of the task. In this paper, we aim to minimize the number of hand-tagged training instances needed to start the learning process …

Content truncated to 500 words