Visualising changes in the construction of meaning with Word Vector Space

Natural language processing (NLP) techniques concerned with language and meaning (e.g., neural word embeddings) are mostly based on distributional semantics theory (Harris 1954; Firth 1957) and typically focus on mapping different senses expressed by single words. In recent years, although there has been an upsurge in such NLP studies, the investigation of language and meaning continues to stay for the majority on word level while its relation with discourse remains largely neglected. In this project, we propose an experimental method that aims to overcome such limitation. Specifically, we explore the value of merging neural word embeddings with the discourse-historical approach (DHA) (Reisigl and Wodak 2001) to investigate the historical changes in the semantic space of public discourse of migration in the United Kingdom. As data-set, we use the Times Digital Archive (TDA) from 1900 to 2000. For the computational part, we use publicly available TDA word2vec models (Kenter et al. 2015; Martinez-Ortiz et al. 2016); these models have been trained according to sliding time windows with the specific intention to map conceptual change. We then use DHA to triangulate the results generated by the word vector models with social and historical data to identify plausible explanations for the changes in the public debate. By bringing the focus of the analysis to the level of discourse, with this method, we aim to go beyond mapping different senses expressed by single words and to add the currently missing sociohistorical and sociolinguistic depth to the computational results. This repository shows the visualisation of the word2vec results. Please refer to Viola and Verheul (2020) for the discussion of the socio-historical triangulation.