Measuring Termhood in Automatic Terminology Extraction

Automatic terminology extraction can be divided into two tasks. The first task measures the unithood which is used to identify a string as a lexical unit. The second task measures the so called termhood, used to identify a lexical unit being a domain specific term. This paper proposes a method to measure termhood in Chinese ATE. It considers the domain specificity of both the components of a candidate term as well as statistical information and other contextual information across different domains and applied to a support vector machine model for terminology extraction. The experiments are based on the Chinese corpus in the IT domain with cross validation of data from outside of the IT domain. Results show that the precision of the open tests can reach over 80% for the top 2,000 candidates and around 50% for the top 20,000 candidate. Furthermore, experiments with different lexicon size shows that the algorithm does not require a comprehensive domain lexicon of a large size. A few thousand basic domain terms would be sufficient to achieve the above mentioned performance.