ABSTRACT

The use of intelligent agents that can communicate with us in human language has become an essential part of our daily lives. Today's intelligent agents can respond appropriately to many things we say or text to them, but they cannot yet communicate fully like humans. They lack our general ability to arrive quickly at accurate and relevant interpretations of what others communicate to us and to form appropriate responses, particularly in sustained interactions. The typical way we teach a machine to acquire such ability is to provide it with approximations of the meanings of utterances in the contexts in which they have occurred in the past. Over the years these approximations have become increasingly rich and detailed, enabling ever more sophisticated systems for interacting with computers using natural language, such as searching for information, getting up-to-date recommendations for products and services, and translating foreign languages. The goal of this project is to bring together linguists and computer scientists to jointly develop a practical meaning representation formalism based on these rich approximations that can be applied to a much more diverse set of languages. This will allow us to use machine learning to develop techniques to automatically translate human utterances into our meaning formalism. In turn, this will enable intelligent agents to acquire more advanced communication capabilities, and for a wider range of languages. The languages considered for the project include those spoken by large populations such as English, Chinese and Arabic, as well as native tongues of smaller groups such as Norwegian, and Arapaho and Kukama-Kukamira, two indigenous languages of the Americas. As such, this project will help bring modern technology to smaller groups so that all people can benefit equally from technological advancement. The project will also contribute to the development of the US workforce by training a new generation of researchers on cutting-edge technologies in artificial intelligence.

This project brings together an interdisciplinary team of linguists and computer scientists from three institutions to jointly develop a Uniform Meaning Representation (UMR). UMR is a practical, formal, computationally tractable, and cross-linguistically valid meaning representation of natural language that can impact a wide range of downstream applications requiring deep natural language understanding (NLU). UMR will extend existing meaning representations to include quantifier types and relations, modality, negation, tense and aspect, and be tested on a typologically diverse set of languages. Methods and techniques for UMR annotation, parsing and generation, and evaluation will be uniform across languages. The project will also develop novel algorithms and models for UMR-based broad-coverage and general-purpose multilingual semantic parsers. Students participating in the project will receive training in the full cycle of conceptualizing, producing, processing, and consuming meaning representations at the sites of participating institutions. This project will help to build a community of NLP researchers that will contribute to the development of UMR-based data and tools and advance the state of the art in Natural Language Processing (NLP) in particular, and Artificial Intelligence (AI) in general.

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.

Please report errors in award information by writing to: awardsearch@nsf.gov.