Article Preview
Top1. Introduction
There are millions of academic documents in the current academic network. How to discover knowledge from the literature is a challenging problem. Keywords as a concise summary of the literature have been widely used in Natural Language Processing and information retrieval tasks. Such as scientific and technical literature summarization (Sarkar, Nasipuri, & Ghose, 2010), text clustering (Hammouda, Matute, & Kamel, 2005) recommendation (Pudota, Dattolo, Baruzzo, Ferrara, & Tasso, 2010), query(Jones & Staveley, 1999). The extraction methods of key words are mainly divided into two kinds: supervised learning and unsupervised learning (Hasan & Ng, 2010, 2011; Merrouni, Frikh, & Ouhbi, 2017). In the literature review, Hasan (Hasan & Ng, 2011) summarized the supervised and unsupervised learning methods in the extraction of keyword extraction.
In supervised learning, the core idea is to convert the task of extracting key words into a two-classification problem. Using positive and negative examples of artificial annotation as training corpus, we train models by acquiring the various feature values and then classify the correct keywords from the literature. i.e. decision tree algorithm (C4.5) (Turney, 2002), Bayesian method (Frank, Paynter, Witten, Gutwin, & Nevill-Manning, 1999), neural network(Sarkar et al., 2010), SVM (Jiang, Hu, & Li, 2009), maximum entropy (Yih, Goodman, & Carvalho, 2006). These algorithms use all kinds of features in the text and integrate the features into the classification model for solving. In early studies, Frank built a system called KEA (Frank et al., 1999). The two features: TF-IDF and phrase first appear in the relative position of the document are incorporated into naive Bayes formula to predict. The prediction quality of the system depends on whether the training set can match the document which need to handle. KEA make the researcher realize that word frequency information and position information is most important feature to judge the keyword. The GenEx (Turney, 2000) method developed by Turney extracts the three features TF-IDF, the position of the phrase which first appears in the document and the length of phrases, and integrates them into the decision tree generation algorithm based on heuristic rules, which improves the evaluation of F value. Nguyen (Nguyen & Kan, 2007) extends the KEA to integrate the phrase's position information in the different chapters of the article into the model. Medelyan (Medelyan, Frank, & Witten, 2009) extends KEA and integrates the information of Wikipedia access into the model. Chuang (Chuang, Manning, & Heer, 2012) presents a model incorporating statistical information and language information to identify descriptive keywords from text. Caragea (Caragea, Bulgarov, Godea, & Gollapalli, 2014) proposed a supervised method of merging citation information with traditional feature information in the paper to extract key words from the scientific literature. Compared with previous methods, this method has improved the quality of extraction. However, when important words do not appear in important positions, such as the beginning of a sentence, or are not quoted, the prediction effect of the model will be affected.
In the unsupervised learning, the core idea is to transform the keyword extraction task into a kind of sorting work based on probability statistics or based on graphs. Use the various types of features in the text or the associations between features for statistics or sorting.