An Artificial Intelligence Method for the Analysis of Marketing Scientific Literature: An Abstract

Abstract

We suggest a machine-based research literature reading method specific to the academic discipline of marketing, adopting artificial intelligence (AI) developments from the field of materials science. Keeping up with research publications is untenable due to exponential growth. Researchers have become much better at the generation of information than at its deployment. AI can help to simplify the use of such knowledge. In materials science, Tshitoyan et al. (2019) have made steps in trying to achieve ‘a generalized approach to the mining of scientific literature’ using text mining and natural language processing. Using AI, research can be extracted from documents, classified, tokenised in individual words, and encoded as information-dense word embeddings, which are vector representations of words, without human supervision (Tshitoyan et al. 2019). Building on these developments we suggest a methodology specific to marketing science. The first step is to compile consolidated bodies of offline marketing research on topics such as branding, retail or advertising following Tshitoyan et al. (2019) method of knowledge extraction and relationships for the handling of large bodies of scientific literature. For this we shall use CrossRef Application Programming Interface (API) for the retrieval of large lists of article Digital Object Identifiers (DOIs). This is used by a number of publisher APIs, such as Elsevier https://dev.elsevier.com and Springer Nature https://dev.springernature.com to download full-text journal articles. Secondly, the embeddings will be trained with the scientific abstracts from each of the topics. For this we shall use article abstracts from 1975 to 2021 from more than a thousand journals and also articles likely to contain marketing-related research directly retrieved from the aforementioned databases (i.e. Elsevier and Science Direct) combined with web scraping. The performance of the algorithm is deemed to improve when irrelevant abstracts are removed. The remaining abstracts will then be classified as relevant and tokenised using ChemDataExtractor (Swain and Cole 2016). Correct pre-processing, especially the choice of phrases to be included as individual tokens should improve the results. In the third and final step, we shall repeat the first two steps integrating offline topics with the equivalent online topics, e.g. online branding. As AI is also capable of predictive writing using bidirectional encoders BERT and ELMo used to produce contextual word embeddings (Devlin et al. 2018), our work in progress will consider developing automated hypotheses formulation in marketing science (Spangler et al. 2014). Simplification of knowledge could also facilitate its transfer to practice.