Automatic keyword extraction using Latent Dirichlet Allocation topic modeling: Similarity with golden standard and users' evaluation

This study investigates the automatic keyword extraction from the table of contents of Persian e-books in the field of science using LDA topic modeling, evaluating their similarity with the golden standard, and users' viewpoints of the model keywords.

This is mixed text-mining research in which LDA topic modeling is used to extract keywords from the table of contents of scientific e-books. The evaluation of the used approach has been done by two methods of cosine similarity computing and qualitative evaluation by users.

Table of contents are medium-length texts with a trimmed mean of 260.02 words, about 20% of which are stop-words. The cosine similarity between the golden standard keywords and the output keywords is 0.0932 thus very low. The full agreement of users showed that the extracted keywords with the LDA topic model represent the subject field of the whole corpus, but the golden standard keywords, the keywords extracted using the LDA topic model in sub-domains of the corpus, and the keywords extracted from the whole corpus were respectively successful in subject describing of each document.

The keywords extracted using the LDA topic model can be used in unspecified and unknown collections to extract hidden thematic content of the whole collection, but not to accurately relate each topic to each document in large and heterogeneous themes. In collections of texts in one subject field, such as mathematics or physics, etc., with less diversity and more uniformity in terms of the words used in them, more coherent and relevant keywords are obtained, but in these cases, the control of the relevance of keywords to each document is required. In formal subject analysis procedures and processes of individual documents, this approach can be used as a keyword suggestion system for indexing and analytical workforce.