DSAI seminar with Denitsa Saynova

Abstract

We address two understudied areas related to explainability for neural text models. First, class explanations. What features are descriptive across a class, rather than explaining single input instances? Second, the type of features that are used for providing explanations. Does the explanation involve the statistical pattern of word usage or the presence of domain-specific content words? Here, we present a method to extract both class explanations and strategies to differentiate between two types of explanations - domain-specific signals or statistical variations in frequencies of common words. We demonstrate our method using a case study in which we analyse transcripts of political debates in the Swedish Riksdag.

Our algorithm can identify statistical speech patterns of speakers, which we see in explanations where stop words appear, but can also separate them from explanations containing domain-specific words, hinting at policy. Additionally, we find indications that domain-specific explanations correlate with model performance. Patterns related to policy in our experiment may be more robust than learned speech patterns of stop words.

About the speaker

Denitsa Saynova is a PhD student in the DSAI division and works on NLP methods for studying political behaviour. She is particularly interested in ways to represent and measure political views from official parliamentary texts in the multiparty Swedish context.

This is a seminar from the DSAI seminars series usually held every Monday at 14:00 by the Data Science and AI division. The seminars are usually hybrid.