Search

In a natural language database query, the existence of syntactic ambiguity within the input utterance results in more than one possible interpretation of the input. This paper describes a procedure to resolve this type of ambiguity by making use of domain knowledge and information in the database schema, and by examining the sentential structure of the input utterance.

In the process of generating text, competent speakers/writers take into consideration the effect their utterances are likely to have on listeners/readers†. In other words, speakers try to generate utterances which are best suited to attain their communicative goals with respect to a particular audience [Hovy 1987].

The workshop entitled "Gaps and Bridges: New Directions in Planning and Natural Language Generation" was held on 12 August 1996 in Budapest, Hungary. This article describes the four sessions of the workshop and summarizes the important themes that were revealed.

The DORIS project (Dialogue Oriented Roaming Interactive System) aims to develop a spoken dialogue module for an autonomous robotic agent. This paper examines the techniques used by Scusi?, the speech interpretation component of DORIS, to postulate and assess hypotheses regarding the meaning of a spoken utterance. The results of our evaluation are encouraging, yielding good interpretation performance for utterances of different types and lengths.

We describe a probabilistic reference disambiguation mechanism developed for a spoken dialogue system mounted on an autonomous robotic agent. Our mechanism performs probabilistic comparisons between features specified in referring expressions (e.g. size and colour) and features of objects in the domain. The results of these comparisons are combined using a function weighted on the basis of the specified features. Our evaluation shows high reference resolution accuracy across a range of spoken referring expressions.

This paper describes the first step in a project for topic identification in help-desk applications. In this step, we apply a clustering mechanism to identify the topics of newsgroup discussions. We have used newsgroup discussions as our testbed, as they provide a good approximation to our target application, while obviating the need for manual tagging of topics. We have found that the postings of individuals who contribute repeatedly to a newsgroup may lead the clustering process astray, in the sense that discussions may be grouped according to their author, rather than according to their topic. To address this problem, we introduce a filtering mechanism, and evaluate it by comparing clustering performance with and without filtering.

Multimodal presentation planning must solve the constraints imposed by the structure of the discourse. In this paper, we discuss a multi-agent implementation of the design and realization layers of the Standard Reference Model (SRM) for Intelligent Multimedia Presentation Systems (IMMPS), where design constraints for multimodal presentations are distributed over a number of agents. Thus, a constraint propagation mechanism is required so that the distributed constraints can be satisfied cooperatively by these agents.

Analysis of queries posed to open-domain question-answering systems indicates that particular types of queries are dominant, e.g., queries about the identity of people, and about the location or time of events We applied a rule-based mechanism and performed manual classification to classify queries into such commonly occurring types We then experimented with different adjustments to our basic document retrieval process for each query type The application of the best retrieval adjustment for each query type yielded improvements in retrieval performance Finally, we applied a machine learning technique to automatically learn the manually classified query types, and applied the best retrieval adjustments obtained for the manual classification to the automatically learned query classes The learning algorithm exhibited high accuracy, and the retrieval performance obtained for the learned classes was consistent with the performance obtained for the rule-based and manual classifications.

This paper describes a probabilistic mechanism for the interpretation of sentence sequences developed for a spoken dialogue system mounted on a robotic agent. The mechanism receives as input a sequence of sentences, and produces an interpretation which integrates the interpretations of individual sentences. For our evaluation, we collected a corpus of hypothetical requests to a robot. Our mechanism exhibits good performance for sentence pairs, but requires further improvements for sentence sequences.

We describe an information-theoretic argument-interpretation mechanism embedded in an interactive system. Our mechanism applies the Minimum Message Length principle to evaluate candidate interpretations and select the best candidate. It receives as input an argument entered through a web interface, and produces an interpretation in terms of its underlying knowledge representation -- a Bayesian network. The results of our preliminary evaluations are encouraging, with the system generally producing plausible interpretations of users' arguments.

The vast amounts of information presented in museums can be overwhelming to a visitor, whose receptivity and time are typically limited. Hence, s/he might have difficulties selecting interesting exhibits to view within the available time. Mobile, context-aware guides offer the opportunity to improve a visitor's experience by recommending exhibits of interest, and personalising the delivered content. The first step in this recommendation process is the accurate prediction of a visitor's activities and preferences. In this paper, we present two adaptive collaborative models for predicting a visitor's next locations in a museum, and an ensemble model that combines their predictions. Our experimental results from a study using a small dataset of museum visits are encouraging, with the ensemble model yielding the best performance overall.

The explicit consideration of an addressee’s inferences during discourse planning affects the content and coherence of the generated discourse. The content of the discourse is affected because the consideration of an addressee’s inferences may require the addition of information that addresses erroneous inferences or suggest the omission of easily inferred information. The coherence of the discourse is affected because the generated discourse should incorporate inferential relations that emerge from the consideration of an addressee’s inferences. In this article, we describe a discourse planner that takes into consideration a user’s inferences. In particular, we discuss its content planning and discourse structuring mechanisms. The content planning mechanism generates a set of rhetorical devices that achieves a given communicative goal. The discourse structuring mechanism takes into consideration schematic, inferential, and prescriptive relations between these rhetorical devices in order to organize them into a coherent sequence. Our discourse planner has been implemented in a system called WISHFUL that generates explanations about concepts in technical domains. We evaluated WISHFUL’s performance to test the effect of considering a user’s inferences during content planning and discourse organization and also to determine the suitability of the discourse produced by WISHFUL for different types of users. The results of our evaluation endorse the ideas embodied in the WISHFUL system.

We present a realistic simulation framework to examine the impact of sensor noise on the performance of user models in the museum domain. Our contributions are (1) models to simulate noisy visit trajectories as time-stamped sequences of (x,y) positional coordinates which reflect walking and hovering behaviour; (2) a discriminative inference model that distinguishes between hovering and walking on the basis of (simulated) noisy sensor observations; (3) a model that infers viewed exhibits from hovering coordinates; and (4) a model that predicts the next exhibit on the basis of inferred (rather than known) viewed exhibits. Our staged evaluation assesses the effect of these models (in combination with sensor noise) on inferential and predictive performance, thus shedding light on the reliability attributed to inferences drawn from sensor observations.

Sentiment analysis deals with inferring people's sentiments and opinions from texts An important aspect of sentiment analysis is polarity classification, which consists of inferring a document's polarity – the overall sentiment conveyed by the text – in the form of a numerical rating In contrast to existing approaches to polarity classification, we propose to take the authors of the documents into account Specifically, we present a nearest-neighbour collaborative approach that utilises novel models of user similarity Our evaluation shows that our approach improves on state-of-the-art performance, and yields insights regarding datasets for which such an improvement is achievable.

We explore the problem of how to perform complex strategic planning in an uncertain domain. Although card games such as bridge and poker provide interesting domains for exploring uncertainty, they tend to be strategically simple given perfect information. We present and discuss a strategically complex, easily controlled, imperfect information domain-a variant of chess-which allows us to explore this difficult problem. We also describe a modular hybrid architecture that uses game theory, information theory and utility theory to play this new game. We discuss preliminary results from our architecture and consider possible future directions for the project.

We describe a probabilistic reference disambiguation mechanism developed for a spoken dialogue system mounted on an autonomous robotic agent. Our mechanism processes referring expressions containing intrinsic features of objects (lexical item, colour and size) and locative expressions, which involve more than one concept. The intended objects are identified in the context of the output of a simulated scene analysis system, which returns the colour and size of the seen objects and a distribution for their type. The evaluation of our system shows high resolution performance across a range of spoken referring expressions and simulated vision accuracies.

This paper examines the problem of uncertainty due to instrumentation in user modeling systems within spatial domains. We consider the uncertainty of inferring a user's trajectory within a physical space combined with the uncertainty due to inaccuracies in measuring a user's position. A framework for modeling both types of uncertainties is presented, and applied to a real-world case study from the museum domain. Our results show that this framework may be used to investigate the effects of layout in a gallery, and to explore the degradation in the predictive performance of user models due to measurement error. This information in turn may be used to guide the curation of the space, and the selection of sensing technologies prior to instrumenting the space.

We present a decision-theoretic model which determines whether to pre-send web documents a user is likely to request. Our model balances the cost of transmitting information with the inconvenience or cost to a user caused by having to wait for information. It uses a Markov model built from web-access data to predict which documents are likely to be requested next. A comparative evaluation of our presending scheme identifies circumstances in which our scheme is beneficial for users.

The Lexical Access Problem consists of determining the intended sequence of words corresponding to an input sequence of phonemes (basic speech sounds) that come from a low-level phoneme recognizer. In this paper we present an information-theoretic approach based on the Minimum Message Length Criterion for solving the Lexical Access Problem. We model sentences using phoneme realizations seen in training, and word and part-of-speech information obtained from text corpora. We show results on multiple-speaker, continuous, read speech and discuss a heuristic using equivalence classes of similar sounding words which speeds up the recognition process without significant deterioration in recognition accuracy., Comment: Appears in Proceedings of the Thirteenth Conference on Uncertainty in Artificial Intelligence (UAI1997)

The interpretation of complex discourse, such as arguments, is a difficult task that often requires validation, i.e., a system may need to present its interpretation of a user's discourse for confirmation In this paper, we consider the presentation of discourse interpretations in the context of a probabilistic argumentation system.We first describe our initial approach to the presentation of an interpretation of a user's argument; this interpretation takes the form of a Bayesian subnet.We then report on the results of our preliminary evaluation with users, focusing on their criticisms of our system's output These criticisms motivate a content enhancement procedure that adds information to explain unexpected outcomes and removes superfluous content from an interpretation The discourse generated by this procedure was found to be more acceptable than the discourse generated by our original method.

In this paper we describe Scusi? , the speech interpretation component of a spoken dialogue module designed for an autonomous robotic agent. Scusi? postulates and maintains multiple interpretations of the spoken discourse, and employs a probabilistic formalism to assess and rank hypotheses regarding the meaning of spoken utterances. These constituents in combination enable Scusi? to cope gracefully with ambiguity and speech recognition errors. The results of our evaluation are encouraging, yielding good interpretation performance for utterances of different types and lengths.

The work presented in this paper is the first step in a project which aims to cluster and summarise electronic discussions in the context of help-desk applications. The eventual objective of this project is to use these summaries to assist help-desk users and operators. In this paper, we identify features of electronic discussions that influence the clustering process, and offer a filtering mechanism that removes undesirable influences. We tested the clustering and filtering processes on electronic newsgroup discussions, and evaluated their performance by means of two experiments: coarse-level clustering and simple information retrieval. Our evaluation shows that our filtering mechanism has a significant positive effect on both tasks.

We investigate the effect of paraphrase generation on document retrieval performance. Specifically, we describe experiments where three information sources are used to generate lexical paraphrases of queries posed to the Internet. These information sources are: WordNet, a Webster-based thesaurus, and a combination of Webster and WordNet. Corpus-based information and wordsimilarity information are then used to rank the paraphrases. We evaluated our mechanism using 404 queries whose answers reside in the LA Times subset of the TREC-9 corpus. Our experiments show that query paraphrasing improves retrieval performance, and that performance is influenced both by the number of paraphrases generated for a query and by their quality. Specifically, the best performance was obtained usingWordNet, which improves document recall by 14% and increases the number of questions that can be answered by 8%.

We describe a probabilistic approach for the interpretation of user arguments that integrates three aspects of an interpretation: inferences, suppositions and explanatory extensions. Inferences fill in information that connects the propositions in a user's argument, suppositions postulate new information that is likely believed by the user and is necessary to make sense of his or her argument, and explanatory extensions postulate information the user may have implicitly considered when constructing his or her argument. Our system receives as input an argument entered through a web interface, and produces an interpretation in terms of its underlying knowledge representation--a Bayesian network. Our evaluations show that suppositions and explanatory extensions are necessary components of interpretations, and that users consider appropriate the suppositions and explanatory extensions postulated by our system.

The vast amount of information presented in museums is often overwhelming to a visitor, making it difficult to select personally interesting exhibits. Advances in mobile computing and user modelling have made possible technology that can assist a visitor in this selection process. Such a technology can (1) utilise non-intrusive observations of a visitor's behaviour in the physical space to learn a model of his/her interests, and (2) generate personalised exhibit recommendations based on interest predictions. Due to the physicality of the domain, datasets of visitors' behaviour (i.e. visitor pathways) are difficult to obtain prior to deploying mobile technology in a museum. However, they are necessary to assess different modelling techniques. This paper reports on a methodology that we used to conduct a manual data collection, and describes the dataset we obtained. We also present two collaborative models for predicting a visitor's viewing times of unseen exhibits from his/her viewing times at visited exhibits (viewing time is indicative of interest), and evaluate our models with the dataset we collected. Both models achieve a higher predictive accuracy than a non-personalised baseline.

Collaboration plays a critical role when a group is striving for goals which are difficult or impossible to achieve by an individual. Knowledge about collaborators’ contributions to a task is an important factor when establishing collaboration, in particular when a decision determines the assignment of activities to members of the group. Although there are several systems that implement collaboration, one important problem has not yet received much attention – determining the effect of incomplete and uncertain knowledge of collaborators’ internal resources (i.e. capabilities and knowledge) on the outcomes of the collaboration. We approach this problem by building models of internal resources of individuals and groups of collaborators. These models enable a system to estimate collaborators’ contributions to the task. We then assess the effect of model accuracy on task performance. An empirical evaluation is performed in order to validate this approach.

This work investigates user attitudes towards personalized summaries generated from a coarse-grained user model based on document aspects. We explore user preferences for summaries at differing degrees of fit with their stated interests, the impact of length on user ratings, and the faithfulness of personalized and general summaries.

Hybridising user models can improve predictive accuracy However, research on linearly combining predictive user models (e.g., used in recommender systems) has often made the implicit assumption that the individual models perform uniformly across the user and item space, using static model weights when computing a weighted average of the predictions of the individual models This paper proposes a weighting scheme which combines user- and item-specific weight vectors to compute user- and item-aware model weights The proposed hybridisation approach adaptively estimates online the model parameters that are specific to a target user as information about this user becomes available Hence, it is particularly well-suited for domains where little or no information regarding the target user's preferences or interests is available at the time of offline model training The proposed weighting scheme is evaluated by applying it to a real-world scenario from the museum domain Our results show that in our domain, our hybridisation approach attains a higher predictive accuracy than the individual component models Additionally, our approach outperforms a non-adaptive hybrid model that uses static model weights.