Robert J. Bobrow

A Fully Statistical Approach to Natural Language Interfaces

We present a natural language interface system which is based entirely on trained statistical models. The system consists of three stages of processing: parsing, semantic interpretation, and discourse. Each of these stages is modeled as a statistical process. The models are fully integrated, resulting in an end-to-end system that maps input utterances into meaning representation frames.

HIDDEN UNDERSTANDING MODELS OF NATURAL LANGUAGE

We describe and evaluate hidden understanding models, a statistical learning approach to natural language understanding. Given a string of words, hidden understanding models determine the most likely meaning for the string. We discuss 1) the problem of representing meaning in this framework, 2) the structure of the statistical model, 3) the process of training the model, and 4) the process of understanding using the model. Finally, we give experimental results, including results on an ARPA evaluation.

Motion to support rapid interactive queries on node--link diagrams

Many different problems can be represented as graphs displayed in the form of node--link diagrams. However, when a graph is large it becomes visually uninterpretable because of the tangle of links. We describe a set of techniques that use motion in an interactive interface to provide effective access to larger graphs. Touching a node with the mouse cursor causes that node and the subgraph of closely connected nodes to oscillate. We argue from perceptual principles that this should be a more effective way of interactively highlighting a subgraph than more conventional static methods. The MEGraph system was developed to gain experience with different forms of motion highlighting. Based on positive feedback, three experiments were carried out to evaluate the effectiveness of motion highlighting for specific tasks. All three showed motion to be more effective than static highlighting, both in increasing the speed of response for a variety of visual queries, and in reducing errors. We argue that motion highlighting can be a valuable technique in applications that require users to understand large graphs.

Supporting visual queries on medium-sized node-link diagrams

For reasons of clarity, a typical node–link diagram statically displayed on paper or a computer screen contains fewer than 30 nodes. However, many problems would benefit if far more complex information could be diagrammed. Following Munzner et al., we suggest that with interactive diagrams this may be possible. We describe an interactive technique whereby a subset of a larger network diagram is highlighted by being set into oscillatory motion when a node is selected with a mouse. The subset is determined by a breadth first search of the underlying graph starting from the selected node. This technique is designed to support visual queries on moderately large node-link diagrams containing up to a few thousand nodes. An experimental evaluation was carried out with networks having 32, 100, 320, 1000, and 3200 nodes respectively, and with four highlighting techniques: static highlighting, motion highlighting, static+motion highlighting, and none. The results show that the interactive highlighting methods support rapid visual queries of nodes in close topological proximity to one another, even for the largest diagrams tested. Without highlighting, error rates were high even for the smallest network that was evaluated. Motion highlighting and static highlighting were equally effective. A second experiment was carried out to evaluate methods for showing two subsets of a larger network simultaneously in such a way that both are clearly distinct. The specific task was to determine if the two subsets had nodes in common. The results showed that this task could be performed rapidly and with few errors if one subset was highlighted using motion and the other was highlighted using a static technique. We discuss the implications for information visualization.

Semantic Interpretation Using KL-ONE

This paper presents extensions to the work of Bobrow and Webber [Bobrow & Webber 80a, Bobrow & Webber 80b] on semantic interpretation using KL-ONE to represent knowledge. The approach is based on an extended case frame formalism applicable to all types of phrases, not just clauses. The frames are used to recognize semantically acceptable phrases, identify their structure, and, relate them to their meaning representation through translation rules. Approaches are presented for generating KL-ONE structures as the meaning of a sentence, for capturing semantic generalizations through abstract case frames, and for handling pronouns and relative clauses.

The BBN/HARC spoken language understanding system

The design and performance of a complete spoken language understanding system under development at BBN are described. The system, dubbed HARC (Hear And Respond to Continuous speech), successfully integrates state-of-the-art speech recognition and natural language understanding subsystems. The system has been tested extensively on a restricted airline travel information (ATIS) domain with a vocabulary of about 2000 words. HARC is implemented in portable, high-level software that runs in real time on todays workstations to support interactive online human-machine dialogs. No special-purpose hardware is required other than an A/D (analog-to-digital) converter to digitize the speech. The system works well for any native speaker of American English and does not require any enrollment data from the users. Results of formal DARPA tests in Feb. and Nov. 1992 are presented.<<ETX>>

Statistical language processing using hidden understanding models

This paper introduces a class of statistical mechanisms, called hidden understanding models, for natural language processing. Much of the framework for hidden understanding models derives from statistical models used in speech recognition, especially the use of hidden Markov models. These techniques are applied to the central problem of determining meaning directly from a sequence of spoken or written words. We present an overall description of the hidden understanding methodology, and discuss some of the critical implementation issues. Finally, we report on experimental results, including results of the December 1993 ARPA evaluation.

Gisting conversational speech

A novel system for extracting information from stereotyped voice traffic is described. Off-the-air recordings of commercial air traffic control communications are interpreted in order to identify the flights present and determine the scenario (e.g., takeoff, landing) that they are following. The system combines algorithms from signal segmentation, speaker segregation, speech recognition, natural language parsing, and topic classification into a single system. Initial evaluation of the algorithm on data recorded at Dallas-Fort Worth airport yields performance of 68% detection of flights with 98% precision at an operating point where 76% of the flight identifications are correctly recognized. In tower recording containing both takeoff and landing scenarios, flights are correctly classified as takeoff or landing 94% of the time.<<ETX>>

Fragment processing in the DELPHI system

This paper presents the fallback understanding component of BBNs DELPHI NL sysystem. This component is invoked when the core DELPHI system is unable to understand an input. It incorporates both syntax- and frame-based fragment combination sub-components, in an attempt to provide a smoother path from accurate but fragile conventional parsers on the one hand to the robust but less accurate schema-based methods on the other. The frame-based sub-component is fully integrated with the DELPHIs core grammar and parser, and represents an advance over previous proposals.The complete fallback understanding component, incorporating both sub-components, was used in the February 1992 NL and SLS evaluations of the DELPHI system and we report on its contribution to these results, and those of its two separate sub-components. For SLS, use of the frame-based sub-component alone resulted in a figure 39.2% Weighted Error---signifigantly lower than our lowest official score of 43.7% Weighted Error.

Statistical agenda parsing

This paper presents the results of converting a standard Graham/Harrison/Ruzzo (GHR) parser for a unification grammar into an agenda-driven parsing system. The agenda is controlled by statistical measures of grammar-rule likelihood obtained from a training set.The techniques in the agenda parser lead to substantial reductions in chart size and parse time, and can be applied to any chart-based parsing algorithm without hand-tuning.