Implementing Eco’s Model Reader with Word Embeddings. An Experiment on Facebook Ideological Bots

In semiotics, the concept of model reader is used to describe the felicity conditions of a text, that is, the information and pragmatic competence needed to interpret the text with reference to a hypothesis on its producer's intention. The model reader permits to formulate inferences about the implicit content of sentences and of the entire text. In this paper we propose to formalize the model reader as a function that takes as inputs a text and a larger context and produces as output what is needed to complete the text's implicit information, filling up its “blank spaces”. One possible technique to implement this function is word embedding. We performed an experiment in this sense, using the data collected and analyzed by the Tracking Exposed group (TREX) during the Italian 2018 elections. For their study, six blank Facebook profiles were created, each characterized by a political orientation: all profiles followed the same common set of 30 pages, representative of the entire Italian political spectrum at the time, but each profile interacted only with content linked to its distinctive political orientation. TREX's study of the profiles' newsfeeds demonstrated that each profile was prompted with an uneven distribution of information sources, biased by its political orientation. For our study, we created six different word spaces, one for each profile. Then we identified a certain number of politically neutral terms and observed the semantic associations of these terms in each word space. To identify the terms, we performed a classification of the entire corpus with the software Iramuteq and selected the most significant terms associated with each cluster. Finally, by performing some operations within each word space, we observed some differences in semantic associations that are coherent with the political orientation of the corresponding profile. These results appear to show that word embedding is a valuable approach for computational text pragmatics, as they can help to model the inferences performed by a certain reader. Also, these results suggest the pertinence of such analyses for the study of filter bubbles resulting from algorithmic personalization.