Text Segmentation and Graph-based Method for Template Filling in Information Extraction

Proceedings of the 5th International Joint Conference on Natural Language Processing, pages 723–731, Chiang Mai, Thailand, November 8 – 13, 2011. c⃝2011 AFNLP Text Segmentation and Graph-based Method for Template Filling in Information Extraction Ludovic Jean-Louis, Romaric Besanc¸on, Olivier Ferret CEA, LIST, Vision and Content Engineering Laboratory Fontenay-aux-Roses, F-92265 France {ludovic.jean-louis,romaric.besancon,olivier.ferret}@cea.fr Abstract In event-based Information Extraction sys- tems, a major task is the automated fill- ing from unstructured texts of a template gathering information related to a partic- ular event. Such template filling may be a hard task when the information is scattered throughout the text and mixed with similar pieces of information relative to a differ- ent event. We propose in this paper a two- step approach for template filling: first, an event-based segmentation is performed to select the parts of the text related to the target event; then, a graph-based method is applied to choose the most relevant entities in these parts for characterizing the event. An evaluation of this model based on an annotated corpus for earthquake events shows that we achieve a 77% F1-measure for the template-filling task. 1 Introduction Information Extraction (IE) is a process that aims at extracting pieces of information from texts. Fol- lowing the paradigm defined in the Message Un- derstanding Conferences (MUC) (Grishman and Sundheim, 1996), IE systems focus on extract- ing structured information concerning events to fill predefined templates. These templates make it possible to highlight the information that is spe- cific to a type of events and to discard pieces of in- formation that are not relevant in this respect. Fig- ure 1 gives an example of the filling of a template by information extracted from a news article. Common issues addressed by IE systems for filling a template include identifying named enti- ties, finding relations between these entities, re- solving entity coreference, gathering scattered in- formation, etc. (Turmo et al., 2006). Currently, there is no standard approach for fill- ing templates. However, most IE systems have Text EV1There are no reports of damage or injuries after a small earthquake rattled the Chino Hills area Tuesday morning. EV1The 3.1-magnitude quake hit at 6:40 a.m. and was centered about two miles west of Chino Hills. EV1It was felt in several surrounding communities. EV2Last July, a 5.4-magnitude quake hit the same area. EV2That quake resulted in cracked walls and broken water and gas lines. Template EV1 ● EVENT: earthquake ● DATE: Tuesday morning ● TIME: 6:40 a.m ● MAGNITUDE: 3.1 ● LOCATION: Chino Hills EV2 ● EVENT: quake ● DATE: Last July ● TIME: ● MAGNITUDE: 5.4 ● LOCATION: Figure 1: Example of template filling been relying on a sentence-oriented approach: first, domain-specific patterns or classifiers are used to process sentences separately; then, ad-hoc strategies for merging these local results into tem- plates are applied. Even if such approach has been used widely, it does not take into account two im- portant problems: (i) events can be described in more than one sentence; (ii) patterns/classifiers mainly capture binary relations among entities …

Content truncated to 500 words