Cognitive Insights into Sentic Spaces Using Principal Paths

The availability of an effective embedding to represent textual information is important in commonsense reasoning. Assessing the quality of an embedding is challenging. In most approaches, embeddings are built using statistical properties of the data that are not directly interpretable by a human user. Numerical methods can be inconsistent with respect to the target problem from a cognitive view point. This paper addresses the issue by developing a protocol for evaluating the coherence between an embedding space and a given cognitive model. The protocol uses the recently introduced notion of principal path, which can support the exploration of a high-dimensional space. The protocol provides a qualitative measure of concept distributions in a graphical format, which allows the embedding properties to be analyzed. As a consequence, the tool mitigates the black-box effect that is typical of automatic inference processes. The experimental section involves the characterization of AffectiveSpace, demonstrating that the proposed approach can be used to describe embeddings. The reference cognitive model is the hourglass model of emotions.