Towards more accurate content categorization of API discussions

• Main Authors: Zhou, Bo, Xia, Xin, LO, David, Tian, Cong, Wang, Xinyu
• Format: text
• Language: English
• Published: Institutional Knowledge at Singapore Management University 2014
• Subjects: API Discussion; Text Categorization; Composite Method; CacheBased Method; Software Engineering
• Online Access: https://ink.library.smu.edu.sg/sis_research/2420; https://ink.library.smu.edu.sg/context/sis_research/article/3420/viewcontent/p95_zhou.pdf
• Institution: Singapore Management University

Nowadays, software developers often discuss the usage of various APIs in online forums. Automatically assigning pre-defined semantic categorizes to API discussions in these forums could help manage the data in online forums, and assist developers to search for useful information. We refer to this process as content categorization of API discussions. To solve this problem, Hou and Mo proposed the usage of naive Bayes multinomial, which is an effective classification algorithm. In this paper, we propose a Cache-bAsed compoSitE algorithm, short formed as CASE, to automatically categorize API discussions. Considering that the content of an API discussion contains both textual description and source code, CASE has 3 components that analyze an API discussion in 3 different ways: text, code, and original. In the text component, CASE only considers the textual description; in the code component, CASE only considers the source code; in the original component, CASE considers the original content of an API discussion which might include textual description and source code. Next, for each component, since different terms (i.e., words) have different affinities to different categories, CASE caches a subset of terms which have the highest affinity scores to each category, and builds a classifier based on the cached terms. Finally, CASE combines all the 3 classifiers to achieve a better accuracy score. We evaluate the performance of CASE on 3 datasets which contain a total of 1,035 API discussions. The experiment results show that CASE achieves accuracy scores of 0.69, 0.77, and 0.96 for the 3 datasets respectively, which outperforms the state-of-the-art method proposed by Hou and Mo by 11%, 10%, and 2%, respectively.