Research
We contribute to research on a variety of topics in intelligent systems. Keeping up with the latest advances in the field is central to our commitment to providing the best solutions to you. Below are a few of the topics we have contributed to recently.
Relational Learning and Inference
This is the task of making simultaneous inferences about multiple related entities or events. For example, you may want to identify the type of a web page in a set of linked web pages. It has been shown that making simultaneous inferences about a group of related entities can be better than reasoning about the entities independently. However, the existing techniques can be difficult to use and too slow for many practical applications. We are working on basic research on this topic to develop faster techniques that perform as well as current techniques.
Text Categorization
Text categorization is a family of techniques involving the grouping and labeling of text records such as emails and news reports. We have developed several novel methods for text categorization that improve upon the accuracy of state-of-the-art techniques.
Linguistic Ontologies
The improvement of natural language understanding is a long-term goal for a large research community. Future systems that can make use of the knowledge in arbitrary texts will be orders of magnitude more capable than those of today because of the wealth of information that humans have recorded in a textual format. One facet of our research into natural language understanding systems involves a dictionary of concepts that assign meaning to words in the context of sentences. The way word meanings are encoded and stored, called lexical semantics, has a strong impact on the accuracy and robustness of language interpretation. We are developing a novel system of meaning representation called the Generative Sublanguage Ontology (GSO). GSO uses a well-principled approach for compact encoding of word meanings and allows the meaning of words to be extended by contextual information, which is key to resolving spatial and temporal references. This approach allows for much more accurate interpretation and greatly increases flexibility in approaching new texts.
Case-Based Reasoning
Case-Based Reasoning (CBR) is a technique based on solving new problems by reusing past solutions. We have pioneered the development of Taxonomic Conversational Case-Based Reasoning (TCCBR), allows users to converse with a computer to find relevant solutions for situations such as troubleshooting printer problems. By internally organizing conversational material into taxonomies, TCCBR guides the conversation toward relevant knowledge with significantly greater accuracy, providing a more efficient and less frustrating interactive experience.
Knowledge Extraction
Many complex software systems make use of highly structured and encoded data, such as cases, rules, plans, and taxonomies, to make useful inferences and solve problems. Knowledge engineering, the process of encoding this data, is currently conducted only by knowledgeable experts; it's notoriously repetitive, error-prone, and expensive. However, much of the needed data already exists in free text that cannot currently be processed by a computer. We have a long-term research agenda to automate this knowledge engineering process by extracting structured data from available texts using both shallow and deep natural language processes.