Learning to Tag Multilingual Texts Through Observation

Learning to Tag Multilingual Texts Through Observation Scott W. Bennett and Chinatsu Aone and Craig Lovell SRA International 4300 Fair Lakes Court Fairfax, VA 22033 {bennett,aonec,lovellc} @sra.com Abstract This paper describes RoboTag, an ad- vanced prototype for a machine learning- based multilingual information extraction system. First, we describe a general client/server architecture used in learning from observation. Then we give a detailed description of our novel decision-tree tag- ging approach. RoboTag performance for the proper noun tagging task in English and Japanese is compared against human- tagged keys and to the best hand-coded pattern performance (as reported in the MUC and MET evaluation results). Re- lated work and future directions are pre- sented. 1 Introduction The ability to tag proper names such as organi- zation, person, and place names in multilingual texts has great value for tasks like information ex- traction, information retrieval, and machine trans- lation (Aone, Charocopos, and Gorlinsky, 1997). The most successful systems currently rely on hand- coded patterns to identify the desired names in texts (Adv, 1995; Def, 1996). This approach achieves its best performance using different hand- coded rule sets for each language/domain pair. Sev- eral of these systems have improved in ease of use, particularly in the speed of the write pat- tern/evaluate performance/refine pattern loop which plays the central role in the development process. One approach in name tagging is to assist in the creation of hand-coded rules by making it easier for the developer to mark parts of the name and its surrounding context to include in the pattern. This boosts productivity in hand-coding rules but still re- quires a significant amount of effort by the developer to identify key parts of the pattern. A step up from this is to determine how to generalize the rule so that it is more broadly applicable or to suggest to the developer which parts of the context have high- value for inclusion in the pattern. Nevertheless, a skilled developer with a thorough knowledge of the particular pattern language is still essential. Our goal in developing RoboTag was to make it possible for an end-user to build a tagging system simply by giving examples of what should be tagged, rather than requiring the user to understand a pat- tern language. RoboTag uses a machine learning algorithm to discover features that the training ex- amples have in common. This knowledge is used to construct a tagging procedure that can find addi- tional, previously unseen examples for extraction. It was important (for the confidence of our users) that the tagging procedure induced by the system be easily explained in terms of how it makes its de- cisions. This was one of the factors that led us to consider using decision trees (Quinlan, 1993) as a key component of the system. Other potential learning or statistical approaches for a problem like this (e.g., Neural Nets or Hidden Markov Models) did not offer this advantage. The RoboTag system is particularly well instrumented for exploration of different learn- ing system …

Content truncated to 500 words