RI: Small: Cache transition systems for sentence understanding and generation

University Of Rochester, Rochester NY

Investigators

Abstract

Graph-based semantic representations allow computers to store and process information in natural language text. Such graphs contain nodes representing events and entities, and edges between nodes representing relations. This project will develop algorithms that operate directly on graphs, and will allow statistical natural processing techniques to better represent semantic structures. These advances can improve systems that extract information from text, translate between human languages such as English and Chinese, and interact with humans in natural language dialog. Improved language understanding can help in accessing the enormous amount of information available in unstructured text on the web as well as in databases of newspapers and scanned books. Improved translation between languages increases opportunities for trade as well as for dissemination of information generally between nations and cultures. Graph-based representations of the meaning of natural language sentences are being used to an increasing degree for reasoning tasks including question answering, merging information from disparate sources, and generating responses in dialog systems. However, automatic interpretation of sentences into such structures remains a very difficult task, despite recent progress in syntactic parsing. This project will develop algorithms for parsing into and generating text from semantic graphs, focusing on Abstract Meaning Representation or AMR, although the techniques generalize to other representations. Existing statistical systems for the AMR parsing task generally use ad-hoc algorithmic approaches; fundamentally new algorithms are necessary to advance the state of the art. This project is based on a new transition system, called a cache transition system, tailored to the task of parsing into graph structures. Preliminary experiments show that the system is a good match to real datasets of semantic graphs, in that it is able to produce the vast majority of graphs observed while at the same time simplifying the machine learning problem of predicting the next transition at each step. This project aims to advance the state of the art in semantic parsing and generation by developing and training a neural version of the transition system to predict semantic graphs from input strings. In its final year, the project will apply the parsing and generation methods to the task of machine translation, providing semantic graphs along with source language strings to a neural machine translation system. 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.