Deep language understanding involves mapping language to a formalism that captures its intended meaning in context, using concepts and relations in an ontology that supports reasoning. It is generally thought that deep understanding is too difficult to accomplish except in very limited domains. As a result, most of the field has shifted to studying so-called shallow methods. Work in shallow language processing has been greatly enhanced in recent years due to the availability of considerable annotated corpora and off-the-shelf software. If one had to build this infrastructure from scratch, then most current work in the field would not be feasible. Despite the successes of shallow language understanding, however, one cannot, for instance, build sophisticated dialogue systems or NL interfaces to intelligent systems, (e.g., humanrobot interaction) using off-the-shelf shallow components. For these and many other applications, we need deeper understanding. However, there are currently no resources or toolkits to support deep understanding.
This planning grant focuses on identifying the resources that the research community would find most useful and exploring what would be useful APIs to this resource (e.g., deep semantic parsers, semantic lexicons and ontologies, discourse processing capabilities such as ontology-based reference resolution, surface speech act interpretation). The interested communities fall into two broad camps. The first are what we might call technology users, people who want to use deep understanding capabilities but whose research interests are in some other area of intelligent systems (e.g., reasoning, learning, image interpretation, robotics). The second camp, technology developers, consist of researchers in NLP who are pushing research in NLP further. A critical subpart of this second group are students studying to become natural language researchers and technology developers. Through a series of case studies and a workshop, we explore the viability of different delivery mechanisms and needs for the different types of users and for different applications.
There are many potential applications for natural language interfaces, including systems that can improve internet search, dialogue-based interfaces to computers, phones and cars, and tutoring and coaching systems to improve education. Progress towards such applications, however, requires that systems perform a deeper level of understanding than possible with current techniques. Specifically, systems need to be able to identify the intended meaning underlying language in order to respond to it appropriately. This project was a preliminary planning grant to assess the possibility of creating resources for the reserach community that could accelerate work on deep language understanding. We considered the needs of two related communities: first, the natural language researchers who are actively pursuing deep language understanding, and second, the community of possible users for natural language technology in building applications and performing related work in fields such as artificial intelligence and linguistics. We identified a set of needs, but concluded that the state of the art is not ready yet for a focussed effort on large scale development of resources. Rather, more basic research is needd first. We identified some critical problems that need to be addressed, with a prime focus on the need to develop techniques that integrate language understanding with knowledge representation and reasoning. We are currently focusing our work on these problems. As a demonstration of the current state of the art, we also constructed a web site where researchers and others can experiment with the system we have built that performs deep parsing. It can be found at www.cs.rochester.edu/research/cisd/projects/trips/parser/cgi/web-parser-xml.cgi.
