Modeling Context in Scenario Template Creation

Modeling Context in Scenario Template Creation Long Qiu, Min-Yen Kan, Tat-Seng Chua Department of Computer Science National University of Singapore Singapore, 117590 {qiul,kanmy,chuats}@comp.nus.edu.sg Abstract We describe a graph-based approach to Sce- nario Template Creation, which is the task of creating a representation of multiple re- lated events, such as reports of different hur- ricane incidents. We argue that context is valuable to identify important, semantically similar text spans from which template slots could be generalized. To leverage context, we represent the input as a set of graphs where predicate-argument tuples are ver- tices and their contextual relations are edges. A context-sensitive clustering framework is then applied to obtain meaningful tuple clus- ters by examining their intrinsic and extrin- sic similarities. The clustering framework uses Expectation Maximization to guide the clustering process. Experiments show that: 1) our approach generates high quality clus- ters, and 2) information extracted from the clusters is adequate to build high coverage templates. 1 Introduction Scenario template creation (STC) is the problem of generating a common semantic representation from a set of input articles. For example, given multiple newswire articles on different hurricane incidents, an STC algorithm creates a template that may in- clude slots for the storm’s name, current location, di- rection of travel and magnitude. Slots in such a sce- nario template are often to be filled by salient entities in the scenario instance (e.g., “Hurricane Charley” or “the coast area”) but some can also be filled by prominent clauses, verbs or adjectives that describe these salient entities. Here, we use the term salient aspect (SA) to refer to any of such slot fillers that people would regard as important to describe a par- ticular scenario. Figure 1 shows such a manually- built scenario template in which details about im- portant actions, actors, time and locations are coded as slots. STC is an important task that has tangible bene- fits for many downstream applications. In the Mes- sage Understanding Conference (MUC), manually- generated STs were provided to guide Information Extraction (IE). An ST can also be viewed as reg- ularizing a set of similar articles as a set of at- tribute/value tuples, enabling multi-document sum- marization from filled templates. Despite these benefits, STC has not received much attention by the community. We believe this is because it is considered a difficult task that re- quires deep NL understanding of the source articles. A problem in applications requiring semantic simi- larity is that the same word in different contexts may have different senses and play different roles. Con- versely, different words in similar contexts may play similar roles. This problem makes approaches that rely on word similarity alone inadequate. We propose a new approach to STC that incor- porates the use of contextual information to address this challenge. Unlike previous approaches that con- centrate on the intrinsic similarity of candidate slot fillers, our approach explicitly models contextual ev- idence. And unlike approaches to word sense disam- biguation (WSD) and other semantic analyses that 157 use neighboring …

Content truncated to 500 words