Identifying Polysemous Words and Inferring Sense Glosses in a Semantic Network
Abstract
Introduction
The present paper aims at detecting polysemous words from their hypernyms. For instance, a native speaker knowing that the French word frégate (frigate) is a ship and a bird can easily guess that frégate is polysemous. Indeed, it is difficult to conceive something being both a ship and a bird at the same time. We can say that those two hypernyms are "incompatible". If one had a list of all incompatible hypernyms (which will be referred as incompatibility rules later in this paper), one could easily detect polysemous words. Is it possible to create such a list ? Can it be done automatically ? To answer these questions we experimented on the French lexical-semantic network JeuxDeMots, Lafourcade (2007), which a free and open resource. Identifying polysemous words is crucial in order to understand a text. It is usually done by detecting high density components in co-occurrence graphs created from large corpora, as in Véronis (2003). Similar methods have been used by Dorow and Widdows (2003) and Ferret (2004) to discover word senses also in corpora. To detect the different dense areas of their graphs, Dorow and Widdows (2003) used the Markov Cluster Algorithm, van Dongen (2000). These methods are very effective, but they highly depend on the corpora used to create the graphs which might induce many biases. To choose the proper glosses for naming the different word senses, Dorow and Widdows (2003) used the hypernyms present in the lexical network WordNet, Fellbaum (1998). WordNet is also used by Ferret (2004) to evaluate his results. We experimented our approach on the French lexical-semantic network JeuxDeMots, and there is no other complete enough french resources equivalent to WordNet to automatically compare our results to. Hence, we had to rely on some manual evaluation. In this paper, we will first present the JeuxDeMots network and some of its specificities. Then, we will detail the method we used (a) for generating list of incompatible hypernym and then (b) for inferring glosses for naming word senses, followed by some evaluations.
The present paper aims at detecting polysemous words from their hypernyms. For instance, a native speaker knowing that the French word frégate (frigate) is a ship and a bird can easily guess that frégate is polysemous. Indeed, it is difficult to conceive something being both a ship and a bird at the same time. We can say that those two hypernyms are "incompatible". If one had a list of all incompatible hypernyms (which will be referred as incompatibility rules later in this paper), one could easily detect polysemous words. Is it possible to create such a list ? Can it be done automatically ? To answer these questions we experimented on the French lexical-semantic network JeuxDeMots, Lafourcade (2007), which a free and open resource. Identifying polysemous words is crucial in order to understand a text. It is usually done by detecting high density components in co-occurrence graphs created from large corpora, as in Véronis (2003). Similar methods have been used by Dorow and Widdows (2003) and Ferret (2004) to discover word senses also in corpora. To detect the different dense areas of their graphs, Dorow and Widdows (2003) used the Markov Cluster Algorithm, van Dongen (2000). These methods are very effective, but they highly depend on the corpora used to create the graphs which might induce many biases. To choose the proper glosses for naming the different word senses, Dorow and Widdows (2003) used the hypernyms present in the lexical network WordNet, Fellbaum (1998). WordNet is also used by Ferret (2004) to evaluate his results. We experimented our approach on the French lexical-semantic network JeuxDeMots, and there is no other complete enough french resources equivalent to WordNet to automatically compare our results to. Hence, we had to rely on some manual evaluation. In this paper, we will first present the JeuxDeMots network and some of its specificities. Then, we will detail the method we used (a) for generating list of incompatible hypernym and then (b) for inferring glosses for naming word senses, followed by some evaluations.
Origin | Files produced by the author(s) |
---|
Loading...