Knowing what you get when seeking semantic similarity: exploring classic NLP method biases
Johanne Saint-Charles, Pierre Mongeau, Louis Renaud-Desjardins · Edward Elgar Publishing eBooks · 2024
Various Natural Language Processing (NLP) methods are called upon to establish similarity between texts in the context of socio-semantic studies. This chapter addresses the methodological diversity in the field by asking to what extent classical NLP methods converge in their identification of similarity between various texts. We compare the results of well-known (and often used) NLP methods in social sciences and humanities: Jaccard, LDA, LSA and TF–IDF, on corpora with different characteristics. Results show that these methods have specific bias and cannot be substituted for one another. Our observations invite social sciences and humanities scholars to consider new criteria for the selection of an NLP method suited to their research objectives.