An Integrated Latent Dirichlet Allocation and Word2vec Method for Generating the Topic Evolution of Mental Models from Global to Local

Mental models play a crucial role in explaining and driving human innovation activities. To help researchers clarify the changes of mental models in various innovation situations, an exploration of its topic dynamic evolution changes is urgently needed. However, most existing works have discussed the topic-semantic distributions of collected documents along the overall timeline, which ignores the semantic details of fusion and evolution between topics in continuous time. This paper discovers and reveals the multi-level information evolving of topics, by integrating latent Dirichlet allocation (LDA) and Word2vec harmoniously to generate the topic evolution maps of the corpus from global to local perspectives. Which include topic distribution trends and their dynamic evolution under the overall time series, as well as the merging and splitting of semantic information between topics in the adjacent time span. These reveal the correlation between topics and the full life cycle of a topic emerging, developing, maturing, and fading. Then, the integrated method was used to perform an analysis of topic evolution with 3984 abstracts of mental model-related papers published between 1980 and 2020. Finally, the performance of the proposed method was compared to that of three traditional topic evolution generated methods based on the standard evaluation metrics. The experimental results demonstrated that our method outperforms other methods both in terms of the content and strength of topic evolution. The proposed method could mine the latent evolution information more clearly and comprehensively from a vast number of papers and is also suited to the various applications of expert systems related to information mining works.