Associating Collocations with WordNet Senses Using Hybrid Models

Associating Collocations with WordNet Senses Using Hybrid Models � � ��� Yi-Chun Chen ������������ Department of Computer Science National Tsing Hua University jordanchengno1@hotmail.com ��� Tzu-Xi Yen ������������ Department of Computer Science National Tsing Hua University joseph.yen.@gmail.com ��� Jason S. CHang ������������ Department of Computer Science National Tsing Hua University jason.jschang@gmail.com Abstract In this paper, we introduce a hybrid method to associate English collocations with sense class members chosen from WordNet. Our combinational approach includes a learning-based method, a paraphrase-based method and a sense frequency ranking method. At training time, a set of collocations with their tagged senses is prepared. We use the sentence information extracted from a large corpus and cross-lingual information to train a learning-based model. At run time, the corresponding senses of an input collocation will be decided via majority voting. The three outcomes participated in voting are as follows: 1. the result from a learning-based model; 2. the result from a paraphrase-based model; 3. the result from sense frequency ranking method. The sense with most votes will be associated with the input collocation. Evaluation shows that the hybrid model achieves significant improvement when comparing with the other method described in evaluation time. Our method provides more reliable result on associating collocations with senses that can help lexicographers in Proceedings of the Twenty-Fourth Conference on Computational Linguistics and Speech Processing (ROCLING 2012) 47 compilation of collocations dictionaries and assist learners to understand collocation usages. ����������������������������������� Keywords: supersense tagging, collocation classification, word sense disambiguation, WordNet, maximum entropy model, Paraphrase. 1 Introduction A collocation is a pair of words that co-occur with more frequency than random. A collocation usually contains a base word (e.g., “oil” in fuel oil) and a collocate (e.g., “fuel” in fuel oil). In a collocations dictionary, we can find many collocates of a base word (e.g., fuel oil, motor oil, peanut oil, salad oil). Some collocations dictionaries show the collocates for all senses, while other collocations dictionaries present the collocates by senses of a base word so learners can better grasp the usage of a collocation. Determining the set of broad senses to classify collocations is not an easy task. Researches have used thesaurus topics such as Roget’s (Yarowsky, 1992) or arbitrarily top-level WordNet senses as classes. There are 44 semantic classes called lexicographer-files and each synset in WordNet is assigned to one lexicographer-file. There are 26 lexicographer-files (or supersenses), which can be used to tag common nouns. Consider the word “oil” which can be used as fuel/to make machines work smoothly, or as belonging to the noun.substance supersense and used in cooking could be seen as belonging to the noun.food supersense. In this paper, we present a hybrid model that automatically associated a given collocation with the corresponding supersense. The hybrid model is composed of a learning-based method, a paraphrase-based method and a sense frequency ranking method. The output supersense of a collocation is decided via majority vote of the above three methods. At training time, we need some collocations tagged with supersenses as seeds. …

Content truncated to 500 words