QA@L2F, Second Steps at QA@CLEF

This paper describes the participation of QA@LF, the question-answering system from LF/INESC-ID, at the QA track of CLEF in 2008. Making intensive use of a Natural Language Processing chain (which includes, among others, a morphological analyzer, a disambiguation module, a multi-word recognizer, a chunker and a named entities recognizer), QA@LF is based on a three module approach to answer questions: corpora pre-processing (where the information sources are processed and potentially relevant information is extracted), question interpretation (where the question is converted into a frame) and answer extraction (where different strategies are used to retrieve the final answer to the input question). QA@LF system was created in 2007 and had its first participation at CLEF07, with results we considered auspicious. Nevertheless, with the objectives of correcting some detected failures, increasing the percentage of questions the system deals with and correctly answers, and also experiment new techniques using the same processing tools, the system suffered modifications during this year: the question interpretation step was improved to better profit from the results of the Natural Language Processing chain; an anaphora solver module was introduced, which allowed us to answer some questions containing backwards references; finally, some other small improvements were done on the system, especially in the answer extraction module. QA@LF had 20% of precision at the competition this year, which represents an increase in the number of correct answers returned by the system of 6%, as compared to the last year results. The system highest accuracy values are on definition questions, in which it achieved 60.714% of precision. However, much work is still to be done in order to improve the system’s results, like, for instance, the introduction of an answer validation module, in order to minimize the number of answers given with different type from the expected type, which was the case this year with 10 of our wrong answers.