Open Access

Annemarie Friedrich, Nikos Engonopoulos, Stefan Thater, Manfred Pinkal

• Standard word sense disambiguation (WSD) data sets annotate each word instance in context with exactly one sense of a predefined inventory, and WSD systems are traditionally evaluated with regard to how good they are at picking this sense. Recently, the notion of graded word sense assignment (GWSA) has gained attention as a more natural view of the contextual specification of word meaning; multiple senses may apply simultaneously to one instance of a word, and they may be applicable to different degrees. In this paper, we apply three different WSD algorithms to the task of GWSA. The three models belong to the class of knowledge-based models in the WSD terminology; they are unsupervised in the sense that they do not depend on annotated training material. We evaluate the models on two recently published GWSA data sets. We find positive correlations with the human judgments for all models, and develop a metric based on the notion of accuracy that highlights differences in the behaviorsStandard word sense disambiguation (WSD) data sets annotate each word instance in context with exactly one sense of a predefined inventory, and WSD systems are traditionally evaluated with regard to how good they are at picking this sense. Recently, the notion of graded word sense assignment (GWSA) has gained attention as a more natural view of the contextual specification of word meaning; multiple senses may apply simultaneously to one instance of a word, and they may be applicable to different degrees. In this paper, we apply three different WSD algorithms to the task of GWSA. The three models belong to the class of knowledge-based models in the WSD terminology; they are unsupervised in the sense that they do not depend on annotated training material. We evaluate the models on two recently published GWSA data sets. We find positive correlations with the human judgments for all models, and develop a metric based on the notion of accuracy that highlights differences in the behaviors of the models.…