Detection of synonymy links between terms
This paper presents a new approach for the evaluation of the detection of synonymy relations between terms. Our goal is to help the terminology structuring. This approach exploits the synonymy relationships which have been extracted from lexical resources to infer synonymy links between complex terms. The inferred links are then validated by an human expert in the context of a terminological application. In a previous evaluation on documents dealing with electric power plant, the expert has underlined that the most important point is to increase the recall even if the precision is low and if some links are mistyped. This paper reports new experiments which help to understand how this synonymy detection approach is to be used. Various lexical resources — from general language dictionary to very specialized semantic information — are exploited and compared as bootstrapping knowledge. Results show the complementary of the different sources. The first evaluation relied on traditional recall and precision measures. However, those scores do not reflect the usefulness of the inferred links for the terminology structuring. From the terminologist’s point of view, erroneous links are quick to eliminated. They may even suggest good ones. Above all, the system points out relations between terms which are generally not found manually. We thus aim at proposing a new evaluation criteria which better reflects the expert’s and terminologist’s point of view in the application context. This score points out the quality of the results and the validation cost rather than the proportion of validated links. We have designed an evaluation score which takes into account the productivity of the dictionary links. It can be viewed as a normalization of the precision.