Using Random Walks for Question-focused Sentence Retrieval

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 915–922, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics Using Random Walks for Question-focused Sentence Retrieval Jahna Otterbacher1, G¨unes¸ Erkan2, Dragomir R. Radev1,2 1School of Information, 2Department of EECS University of Michigan {jahna,gerkan,radev}@umich.edu Abstract We consider the problem of question- focused sentence retrieval from complex news articles describing multi-event sto- ries published over time. Annotators gen- erated a list of questions central to under- standing each story in our corpus. Be- cause of the dynamic nature of the stories, many questions are time-sensitive (e.g. “How many victims have been found?”) Judges found sentences providing an an- swer to each question. To address the sentence retrieval problem, we apply a stochastic, graph-based method for com- paring the relative importance of the tex- tual units, which was previously used suc- cessfully for generic summarization. Cur- rently, we present a topic-sensitive version of our method and hypothesize that it can outperform a competitive baseline, which compares the similarity of each sentence to the input question via IDF-weighted word overlap. In our experiments, the method achieves a TRDR score that is sig- nificantly higher than that of the baseline. 1 Introduction Recent work has motivated the need for systems that support “Information Synthesis” tasks, in which a user seeks a global understanding of a topic or story (Amigo et al., 2004). In contrast to the clas- sical question answering setting (e.g. TREC-style Q&A (Voorhees and Tice, 2000)), in which the user presents a single question and the system returns a corresponding answer (or a set of likely answers), in this case the user has a more complex information need. Similarly, when reading about a complex news story, such as an emergency situation, users might seek answers to a set of questions in order to un- derstand it better. For example, Figure 1 shows the interface to our Web-based news summarization system, which a user has queried for information about Hurricane Isabel. Understanding such stories is challenging for a number of reasons. In particular, complex stories contain many sub-events (e.g. the devastation of the hurricane, the relief effort, etc.) In addition, while some facts surrounding the situation do not change (such as “Which area did the hurri- cane first hit?”), others may change with time (“How many people have been left homeless?”). There- fore, we are working towards developing a system for question answering from clusters of complex sto- ries published over time. As can be seen at the bot- tom of Figure 1, we plan to add a component to our current system that allows users to ask questions as they read a story. They may then choose to receive either a precise answer or a question-focused sum- mary. Currently, we address the question-focused sen- tence retrieval task. While passage retrieval (PR) is clearly not a new problem (e.g. (Robertson et al., 1992; Salton et al., 1993)), it remains important and yet often overlooked. As noted by …

Content truncated to 500 words