EAGER: Learning a High-Fidelity Semantic Parser

University Of Rochester, Rochester NY

Investigators

Abstract

Communication with computers in ordinary language is a long-sought goal of AI researchers, educational, commercial, and government enterprises, and everyone who uses computers. The most impressive systems to date depend on coding of thousands of specialized "skills" by thousands of expert programmers. Ordinary comments such as "I'm afraid I won't make it to the meeting" and "She managed to get the insulin shot in time" are not understood well enough to draw obvious conclusions such as "I won't be at the meeting" and "She got the insulin shot in time". This exploratory EAGER project takes a step towards machine understanding of ordinary language, by providing a comprehensive way of representing the content of language in machines, and developing a machine learning technique that allows computers to translate language into that representation, and hence make the kinds of inferences mentioned. This in turn provides immediate tools for improving systems that require some degree of general understanding and inference, such as dialogue systems, sentiment analysis systems, and systems that extract desired knowledge from text. The high-fidelity representations of meaning produced by the semantic parser also provides a substrate for deriving deeper meanings, using what we know about the way discourse segments form coherent passages, and making use of general knowledge about word meanings and the world. The project team consists of a diverse group guided by the project principal investigators, several graduate-level and a dozen undergraduate-level researchers. This project focuses on deriving "unscoped logical forms" (ULFs) reflecting the semantic type structure of standard English sentences with unprecedented fidelity, covering not only predication but also quantification, tense, modality, reification, predicate and sentence modification, comparison structures, and other semantic phenomena. As such, it moves well beyond the expressive range of current mainstream approaches, such as Abstract Meaning Representation (AMR). Thanks to its type coherence, ULF supports forward discourse inferences from text in a more comprehensive way than Natural Logic, and without requiring knowledge of a target hypothesis to be confirmed or disconfirmed. Demonstrating inferences from clause-taking verbs, counterfactuals, questions, and requests provides an important proof of concept. The semantic ULF parser is produced by supervised learning of a cache transition parser, much like one previously applied successfully to AMR parsing, but enhanced by prioritizing type-consistent operator-operand combinations. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.