Connotation frames are a new formalism for analyzing subjective roles and relationships implied by a given predicate. For example, with the verb "violate": the writer has spun the sentence in such a way that the agent seems malicious while the theme seems sympathetic. Their word choice seems to convey their positive perspective towards the theme and negative sentiment towards the agent. They have also suggested to the reader that the theme has been negatively impacted and will have a poor mental state afterwords. These relationships are unified by connotation frames which contains labels for relationships inferrable from the predicate:
For example, with the verb violate, if somebody writes a sentence where "X violates Y", then we can infer:
The development and investigation of connotation frames is part of a project with Sameer Singh and Yejin Choi started in 2014. There is a paper on arxiv that will be updated with our latest edition, recently accepted to ACL.
The labelled connotation frames were collected using AMT annotations as described in our latest paper, "Connotation Frames: A data driven investigation". We received annotations for about 950 of the 1000 most frequently used transitive verbs in the NY times Corpus. The labels were created by aggregating over 15 annotators' responses. In the annotated files, we have labels for 12 different relationship aspects connoted by the predicate:
Design by TEMPLATED.