Measuring the diffusion of innovations with paragraph vector topic models

Measuring the diffusion of innovations from textual data sources besides patent data has not been studied extensively. However, early and accurate indicators of innovation and the recognition of trends in innovation are mandatory to successfully promote economic growth through technological progress via evidence-based policy making. In this study, we propose Paragraph Vector Topic Model (PVTM) and apply it to technology-related news articles to analyze innovation-related topics over time and gain insights regarding their diffusion process. PVTM represents documents in a semantic space, which has been shown to capture latent variables of the underlying documents, e.g., the latent topics. Clusters of documents in the semantic space can then be interpreted and transformed into meaningful topics by means of Gaussian mixture modeling. In using PVTM, we identify innovation-related topics from 170, 000 technology news articles published over a span of 20 years and gather insights about their diffusion state by measuring the topic importance in the corpus over time. Our results suggest that PVTM is a credible alternative to widely used topic models for the discovery of latent topics in (technology-related) news articles. An examination of three exemplary topics shows that innovation diffusion could be assessed using topic importance measures derived from PVTM. Thereby, we find that PVTM diffusion indicators for certain topics are Granger causal to Google Trend indices with matching search terms.除专利数据外，利用其他文本数据源测度创新扩散的相关研究尚未得到充分探索。然而，若要通过技术进步、依托循证决策推动经济增长，精准的早期创新指标与创新趋势识别方法是不可或缺的。本研究提出段落向量主题模型（Paragraph Vector Topic Model，PVTM），并将其应用于科技领域新闻文本，以时序维度剖析创新相关主题，并探究其扩散进程。PVTM将文档映射至语义空间，该空间已被证实能够捕获文本蕴含的潜在变量，例如潜在主题。借助高斯混合模型，可对语义空间内的文档聚类进行解读，并将其转化为具备实际语义的主题。本研究依托PVTM，从20年间发布的17万篇科技新闻文本中识别出与创新相关的主题，并通过时序维度测度语料库中的主题重要性，以探析其扩散状态。研究结果表明，针对（科技领域）新闻文本的潜在主题发现任务，PVTM可作为当前主流主题模型的可靠替代方案。针对三个示例主题的验证分析表明，可借助PVTM导出的主题重要性指标评估创新扩散水平。研究发现，部分主题的PVTM扩散指标与对应搜索词的谷歌趋势（Google Trends）指数存在格兰杰因果关系。