Embedding Contextual Information in Seq2seq Models for Grounded Semantic Role Labeling

Natural Language interactions between humans and robots are meant to be situated, in the sense that both the user and the robot can access and make reference to the shared environment. Contextual knowledge plays thus a key role in the solution of inherent ambiguities in interpretation tasks, such as Grounded Semantic Role Labeling (GSRL). Explicit representations for the context (i.e. the map description of the surroundings) are crucial and the possibility of injecting such information in the training stages of semantic interpreters is very appealing. In this paper, we propose to make a sequence-to-sequence model for GSRL, thus eliminating the traditional cascade of tasks and effectively linking real-world entities with their identifiers, that is sensitive to map information in form of linguistic descriptions. The corresponding generation process, based on BART, achieves results competitive with the state-of-the-art on the GSRL task.