Request Correction

Language Resources as by-Product of Evaluation: The MULTITAG Example

In this paper, we show how the paradigm of evaluation can function as language resource producer for high quality and low cost validated language resources. First the paradigm of evaluation is presented, the main points of its history are recalled, from the first deployment that took place in the USA during the DARPA/NIST evaluation campaigns, up to latest efforts in Europe (SENSEVAL2/ROMANSEVAL2, CLEF, CLASS etc.). Then the principle behind the method used to produce high-quality validated language at low cost from the by-products of an evaluation campaign is exposed. It was inspired by the experiments (Recognizer Output Voting Error Recognition) performed during speech recognition evaluation campaigns in the USA and consists of combining the outputs of the participating sys-tems with a simple voting strategy to obtain higher performance results. Here we make a link with the existing strategies for system combination studied in machine learning. As an illustration we describe how the MULTITAG project funded by CNRS has built from the by-products of the GRACE evaluation campaign (French Part-Of-Speech tagging system evaluation campaign) a corpus of around 1 million words, annotated with a fine grained tagset derived from the EAGLES and MULTEXT projects. A brief presentation of the state of the art in Part-Of-Speech (POS) tagging and of the problem posed by its evaluation is given at the beginning, then the corpus itself is presented along with the procedure used to produce and validate it. In particular, the cost reduction brought by using this method instead of more classical methods is presented and its generalization to other control task is discussed in the conclusion.