Extracting Phrases as Software Features From Overlapping Sentence Clusters in Product Descriptions

Extracting software features from the public product descriptions in the natural language is beneficial for developing new products. Because software features are often expressed in phrases, many approaches currently propose to define phrase patterns and extract phrases as features from product descriptions accordingly. However, there are often lots of noisy phrases extracted because public product descriptions are described freely by different designers and it is difficult to obtain accurate phrase patterns in practice. It is also not suitable to filter those noisy phrases according to frequencies because some important features may be infrequent. To address such issues, this paper proposes a feature extraction approach by extracting phrases as features from the sentence clusters among product descriptions rather than directly from the product descriptions. Considering that more than one feature can be described in one sentence, a new algorithm is designed to detect the overlapping sentence clusters from public product descriptions. It can detect all potential sentence clusters and reduce the affection of noisy descriptions. By taking bigram collocations as the phrase pattern, the bigram collocations containing cluster keywords are elicited as features from each detected sentence cluster. The evaluations conducted on the public software product descriptions from the application market of Softpedia.com, have shown that the proposed approach has better performance than the competitive approaches in terms of precision and time consumption.