Improving Mention Detection Robustness to Noisy Input

Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 335–345, MIT, Massachusetts, USA, 9-11 October 2010. c⃝2010 Association for Computational Linguistics Improving Mention Detection Robustness to Noisy Input Radu Florian, John F. Pitrelli, Salim Roukos and Imed Zitouni IBM T.J. Watson Research Center Yorktown Heights, NY, U.S.A. {raduf,pitrelli,roukos,izitouni}us.ibm.com Abstract Information-extraction (IE) research typically focuses on clean-text inputs. However, an IE engine serving real applications yields many false alarms due to less-well-formed input. For example, IE in a multilingual broadcast processing system has to deal with inaccu- rate automatic transcription and translation. The resulting presence of non-target-language text in this case, and non-language mate- rial interspersed in data from other applica- tions, raise the research problem of making IE robust to such noisy input text. We ad- dress one such IE task: entity-mention de- tection. We describe augmenting a statistical mention-detection system in order to reduce false alarms from spurious passages. The di- verse nature of input noise leads us to pursue a multi-faceted approach to robustness. For our English-language system, at various miss rates we eliminate 97% of false alarms on in- puts from other Latin-alphabet languages. In another experiment, representing scenarios in which genre-specific training is infeasible, we process real financial-transactions text con- taining mixed languages and data-set codes. On these data, because we do not train on data like it, we achieve a smaller but significant im- provement. These gains come with virtually no loss in accuracy on clean English text. 1 Introduction Information-extraction (IE) research is typically per- formed on clean text in a predetermined language. Lately, IE has improved to the point of being usable for some real-world tasks whose accuracy require- ments are reachable with current technology. These uses include media monitoring, topic alerts, sum- marization, population of databases for advanced search, etc. These uses often combine IE with tech- nologies such as speech recognition, machine trans- lation, topic clustering, and information retrieval. The propagation of IE technology from isolated use to aggregates with such other technologies, from NLP experts to other types of computer scientists, and from researchers to users, feeds back to the IE research community the need for additional inves- tigation which we loosely refer to as “information- extraction robustness” research. For example: 1. Broadcast monitoring demands that IE handle as input not only clean text, but also the tran- scripts output by speech recognizers. 2. Multilingual applications, and the imperfection of translation technology, require IE to contend with non-target-language text input (Pitrelli et al., 2008). 3. Naive users at times input to IE other material which deviates from clean text, such as a PDF file that “looks” like plain text. 4. Search applications require IE to deal with databases which not only possess clean text but at times exhibit other complications like mark- up codes particular to narrow, application- specific data-format standards, for example, the excerpt from a financial-transactions data set shown in Figure 1. Legacy industry-specific standards, such as il- lustrated in this example, are part …

Content truncated to 500 words