Do we still Need Gold Standards for Evaluation?

The availability of a huge mass of textual data in electronic format has increased the need for fast and accurate techniques for textual data processing. Machine learning and statistical approaches have been increasingly used in NLP since the 1990s, mainly because they are quick, versatile and efficient. However, despite this evolution of the field, evaluation still rely (most of the time) on a comparison between the output of a probabilistic or statistical system on the one hand, and a non-statistic, most of the time hand-crafted, gold standard on the other hand. In order to be able to compare these two sets of data, which are inherently of a different nature, it is first necessary to modify the statistical data so that they fit with the hand-crafted reference. For example, a statistical parser, instead of producing a score of grammaticality, will have to produce a binary value for each sentence (grammatical vs ungrammatical) or a tree similar to the one stored in the treebank used as a reference. In this paper, we take the example of the acquisition of subcategorization frames from corpora as a practical example. Our study is motivated by the fact that, even if a gold standard is an invaluable resource for evaluation, a gold standard is always partial and does not really show how accurate and useful results are. We describe the task (SCF acquisition) and show how it is a typical NLP task. We then very briefly describe our SCF acquisition system before discussing different issues related to the evaluation using a gold standard. Lastly, we adopt the classical distinction between intrinsic and extrinsic evaluation and show why this framework is relevant for SCF acquisition. We show that, even if intrinsic evaluation correlates with extrinsic evaluation, these two evaluation frameworks give a complementary insight on the results. In the conclusion, we quickly discuss the case of other NLP tasks.