基于主题N元语法模型的科技报告主题分析

作为科技情报的重要载体之一,科技报告可以反映科技发展的脉络,可以揭示科技前沿的动态,甚至可以洞察科技发展的趋势等。中国科技报告的开发利用研究目前主要集中在书本型科技报告或电子出版物的出版发行、数据库建设、服务方式和知识产权等方面,在深度数据挖掘方面的研究工作相对较少。笔者尝试利用主题N元语法模型对科技报告进行领域深层主题分析,为了确定特定领域科技报告的主题数目,笔者借助动态规划的思想针对主题N元语法模型提出了困惑度的有效计算方法。最后,以肿瘤领域1344条科技报告为实验数据,揭示了以“分子机制/肿瘤细胞”和“系统生物学/关键方法”为代表的70个主题,验证了利用主题N元语法模型揭示科技报告领域深层主题的可行性和有效性。

Topical Analysis of Scientific and Technical Reports based on Topical N-Grams Model

As one of the important carriers of scientific & technical (S&T) intelligence, S&T reports can reflect the line of S&T development, recover the latest news of S&T fronts, and even insight the trends of S&T development. Researches on developing and utilizing S&T reports in our country mainly focus on the following: publication and distribution of S&T reports in the form of book and electrical publication;database construction;service mode;intelligent property and so on. The deep data mining on S&T reports remains largely under-studied. This work tries to discover the domain latent topics of S&T reports with the topical n-grams model. In order to determine the number of topics of S&T reports for some specific domain, the calculation method of perplexity of the topic n-grams model is put forward with the dynamic programming in this study. Finally, 70 domain topics are discovered from 1 344 S&T reports in the tumor domain, such as "molecular mechanisms/tumor cells","system biology/key methods" and so on. Experimental results show that it is feasible and efficient to discover the latent topics from S&T reports with the topical n-grams model.