Purpose: This study aims to build an automatic survey generation tool, named CitationAS,based on citation content as represented by the set of citing sentences in the original articles.Design/methodology/approach: Firstly, we apply LDA to analyse topic distribution ofcitation content. Secondly, in CitationAS, we use bisecting K-means, Lingo and STC tocluster retrieved citation content. Then Word2Vec, WordNet and combination of them areapplied to generate cluster labels. Next, we employ TF-IDF, MMR, as well as consideringsentence location information, to extract important sentences, which are used to generatesurveys. Finally, we adopt manual evaluation for the generated surveys.Findings: In experiments, we choose 20 high-frequency phrases as search terms. Resultsshow that Lingo-Word2Vec, STC-WordNet and bisecting K-means-Word2Vec have betterclustering effects. In 5 points evaluation system, survey quality scores obtained by designingmethods are close to 3, indicating surveys are within acceptable limits. When consideringsentence location information, survey quality will be improved. Combination of Lingo,Word2Vec, TF-IDF or MMR can acquire higher survey quality.Research limitations: The manual evaluation method may have a certain subjectivity. We usea simple linear function to combine Word2Vec and WordNet that may not bring out theirstrengths. The generated surveys may not contain some newly created knowledge of somearticles which may concentrate on sentences with no citing.Practical implications: CitationAS tool can automatically generate a comprehensive,detailed and accurate survey according to user’s search terms. It can also help researcherslearn about research status in a certain field.
展开▼
