(A) method for computing noun similarities using adjectives as semantic contexts

Abstract

Noun similarity measures the semantic likeness between two nouns, and it generally means semantic similarity. Measuring semantic similarity requires an information resource such as a corpus or knowledge base. In this thesis, we focus on methods for using corpus data. Previous research on computing semantic similarity using corpus data still has some critical limitations. First, the target nouns should directly or indirectly co-occur in the corpus. Also, the words that are semantically unrelated to the target words in the context can be incorrectly used as representing the meaning. To overcome these limitations, we propose a method of utilizing the modifying adjectives in the context of a target noun. By using adjectives for a target noun, we can extract contextual information regardless of whether or not it co-occur with the other noun being compared in the corpus. To effectively make use of adjective information, we adopt the adjective classification method from past research. With the method we form vectors, each representing attributes of each adjective. We evaluate the proposed method with existing benchmarks and compare the performance with past studies. The result shows that adjective information has a positive impact on measuring noun similarity.