Recognizing textual entailment is a key task for many seman- tic applications, such as Question Answering, Text Summa- rization, and Information Extraction, among others. Entail- ment scenarios can range from a simple syntactic variation to more complex semantic relationships between pieces of text, but most approaches try a one-size-fits-all solution that usually favors some scenario to the detriment of another. We propose a composite approach for recognizing text entailment which analyzes the entailment pair to decide whether it must be resolved syntactically or semantically. We also make the answer interpretable: whenever an entailment is solved se- mantically, we explore a knowledge base composed of structured lexical definitions to generate natural language human- like justifications, explaining the semantic relationship hold- ing between the pieces of text. Besides outperforming well- established entailment algorithms, our composite approach gives an important step towards Explainable AI, using world knowledge to make the semantic reasoning process explicit and understandable.
Vivian Silva, Andre Freitas, Siegfried Handschuh
29 Nov 2018