Search

We present a probabilistic approach for the interpretation of arguments that casts the selection of an interpretation as a model selection task. In selecting the best model, our formalism balances conflicting factors: model complexity against data fit, and structure complexity against belief reasonableness. We first describe our basic formalism, which considers interpretations comprising inferential relations, and then show how our formalism is extended to suppositions that account for the beliefs in an argument, and justifications that account for the inferences in an interpretation. Our evaluations with users show that the interpretations produced by our system are acceptable, and that there is strong support for the postulated suppositions and justifications.

The importance of explaining the outcome of a machine learning model, especially a black-box model, is widely acknowledged. Recent approaches explain an outcome by identifying the contributions of input features to this outcome. In environments involving large black-box models or complex inputs, this leads to computationally demanding algorithms. Further, these algorithms often suffer from low stability, with explanations varying significantly across similar examples. In this paper, we propose a Learning to Explain (L2E) approach that learns the behaviour of an underlying explanation algorithm simultaneously from all training examples. Once the explanation algorithm is distilled into an explainer network, it can be used to explain new instances. Our experiments on three classification tasks, which compare our approach to six explanation algorithms, show that L2E is between 5 and 7.5×10ˆ4 times faster than these algorithms, while generating more stable explanations, and having comparable faithfulness to the black-box model.

In many complex, real-world situations, problem solving and decision making require effective reasoning about causation and uncertainty. However, human reasoning in these cases is prone to confusion and error. Bayesian networks (BNs) are an artificial intelligence technology that models uncertain situations, supporting better probabilistic and causal reasoning and decision making. However, to date, BN methodologies and software require (but do not include) substantial upfront training, do not provide much guidance on either the model building process or on using the model for reasoning and reporting, and provide no support for building BNs collaboratively. Here, we contribute a detailed description and motivation for our new methodology and application, Bayesian ARgumentation via Delphi (BARD). BARD utilizes BNs and addresses these shortcomings by integrating (1) short, high-quality e-courses, tips, and help on demand; (2) a stepwise, iterative, and incremental BN construction process; (3) report templates and an automated explanation tool; and (4) a multiuser web-based software platform and Delphi-style social processes. The result is an end-to-end online platform, with associated online training, for groups without prior BN expertise to understand and analyze a problem, build a model of its underlying probabilistic causal structure, validate and reason with the causal model, and (optionally) use it to produce a written analytic report. Initial experiments demonstrate that, for suitable problems, BARD aids in reasoning and reporting. Comparing their effect sizes also suggests BARD's BN-building and collaboration combine beneficially and cumulatively.

Despite recent advances in automatic speech recognition, one of the main stumbling blocks to the widespread adoption of Spoken Dialogue Systems is the lack of reliability of automatic speech recognizers. In this paper, we offer a two-tier error-correction process that harnesses syntactic, semantic and pragmatic information to improve the understanding of spoken referring expressions, specifically descriptions of objects in physical spaces. A syntactic-semantic tier offers generic corrections to perceived ASR errors on the basis of syntactic expectations of a semantic model, and passes the corrected texts to a language understanding system. The output of this system, which consists of pragmatic interpretations, is then refined by a contextual-phonetic tier, which prefers interpretations that are phonetically similar to the mis-heard words. Our results, obtained on a corpus of 341 referring expressions, show that syntactic-semantic error correction significantly improves interpretation performance, and contextual-phonetic refinements yield further improvements.

HighlightsAn interpretation process that considers multiple alternatives.A mechanism for combining uncertainty from a variety of sources.Distance functions with probabilistic semantics that represent similarity measures.Two evaluation experiments to assess our system's performance. In this paper, we present Scusi?, an anytime numerical mechanism for the interpretation of spoken referring expressions. Our contributions are: (1) an anytime interpretation process that considers multiple alternatives at different interpretation stages (speech, syntax, semantics and pragmatics), which enables Scusi? to defer decisions to the end of the interpretation process; (2) a mechanism that combines scores associated with the output of the different interpretation stages, taking into account the uncertainty arising from a variety of sources, such as ambiguity or inaccuracy in a description, speech recognition errors and out-of-vocabulary terms; and (3) distance-based functions with probabilistic semantics that represent lexical similarity between objects' names and similarity between stated requirements and physical properties of objects (viz colour, size and positional relations). We considered two approaches for combining these descriptive attributes, viz multiplicative and additive, and determined whether prioritizing certain interpretation stages and descriptive attributes affects interpretation performance. We conducted two experiments to evaluate different aspects of Scusi?'s performance: Interpretive, where we compared Scusi?'s understanding of descriptions that are mainly ambiguous or inaccurate with people's understanding of these descriptions, and Generative, where we assessed Scusi?'s understanding of naturally occurring spoken descriptions. Our results show that Scusi?'s understanding of the descriptions in the Interpretive trial is comparable to that of people; and that its performance is encouraging when given arbitrary spoken descriptions in diverse scenarios, and excellent for the corresponding written descriptions. In both experiments, Scusi? significantly outperformed a baseline system that maintains only top same-score interpretations.

Contextual sequence mapping is one of the fundamental problems in Natural Language Processing (NLP). Here, instead of relying solely on the information presented in the text, the learning agents have access to a strong external signal given to assist the learning process. In this paper, we propose a novel family of Recurrent Neural Network unit: the Context-dependent Additive Recurrent Neural Network (CARNN) that is designed specifically to address this type of problem. The experimental results on public datasets in the dialog problem (Babi dialog Task 6 and Frame), contextual language model (Switchboard and Penn Tree Bank) and question answering (Trec QA) show that our novel CARNN-based architectures outperform previous methods.

We propose a novel generative neural network architecture for Dialogue Act classification. Building upon the Recurrent Neural Network framework, our model incorporates a novel attentional technique and a label to label connection for sequence learning, akin to Hidden Markov Models. The experiments show that both of these innovations lead our model to outperform strong baselines for dialogue act classification on MapTask and Switchboard corpora. We further empirically analyse the effectiveness of each of the new innovations.

We propose a novel hierarchical Recurrent Neural Network (RNN) for learning sequences of Dialogue Acts (DAs). The input in this task is a sequence of utterances (i.e., conversational contributions) comprising a sequence of tokens, and the output is a sequence of DA labels (one label per utterance). Our model leverages the hierarchical nature of dialogue data by using two nested RNNs that capture long-range dependencies at the dialogue level and the utterance level. This model is combined with an attention mechanism that focuses on salient tokens in utterances. Our experimental results show that our model outperforms strong baselines on two popular datasets, Switchboard and MapTask; and our detailed empirical analysis highlights the impact of each aspect of our model.

Authorship attribution deals with identifying the authors of anonymous texts. Traditionally, research in this field has focused on formal texts, such as essays and novels, but recently more attention has been given to texts generated by on-line users, such as e-mails and blogs. Authorship attribution of such on-line texts is a more challenging task than traditional authorship attribution, because such texts tend to be short, and the number of candidate authors is often larger than in traditional settings. We address this challenge by using topic models to obtain author representations. In addition to exploring novel ways of applying two popular topic models to this task, we test our new model that projects authors and documents to two disjoint topic spaces. Utilizing our model in authorship attribution yields state-of-the-art performance on several data sets, containing either formal texts written by a few authors or informal texts generated by tens to thousands of on-line users. We also present experimental results that demonstrate the applicability of topical author representations to two other problems: inferring the sentiment polarity of texts, and predicting the ratings that users would give to items such as movies.

People are often overwhelmed by the large amount of information provided in museum spaces, which makes it difficult for them to select exhibits of potential interest. As a first step in recommending exhibits where a visitor may wish to spend some time, this article investigates predictive user models for personalised prediction of museum visitors' viewing times at exhibits. We consider two content-based models and a nearest-neighbour collaborative filter, and develop a collaborative model based on the theory of spatial processes which relies on a notion of distance between exhibits. We discuss models of exhibit distance derived from viewing-time similarity, semantic similarity and walking distance. The results from our evaluation with a real-world dataset of visitor pathways collected at Melbourne Museum (Melbourne, Australia) suggest that utilising walking and semantic distances between exhibits enables more accurate predictions of a visitor's viewing times of unseen exhibits than using distances derived from observed exhibit viewing times. Our results also show that all models outperform a non-personalised baseline, that content-based viewing time prediction yields better results than nearest-neighbour collaborative prediction, and that our collaborative model based on spatial processes attains the highest predictive accuracy overall.

While identifying the intention of an utterance has played a major role in natural language understanding, this work is the first to extend intention recognition to the domain of information graphics. A tenet of this work is the belief that information graphics are a form of language. This is supported by the observation that the overwhelming majority of information graphics from popular media sources appear to have some underlying goal or intended message. As Clark noted, language is more than just words. It is any “signal” (or lack of signal when one is expected), where a signal is a deliberate action that is intended to convey a message (Clark, 1996 [15]).As a form of language, information graphics contain communicative signals that can be used in a computational system to identify the message that the graphic conveys. We identify the communicative signals that appear in simple bar charts, and present an implemented Bayesian network methodology for reasoning about these signals and hypothesizing a bar chart's intended message. Once the message conveyed by an information graphic has been inferred, it can then be used to facilitate access to this information resource for a variety of users, including 1) users of digital libraries, 2) visually impaired users, and 3) users of devices where graphics are impractical or inaccessible.

This article presents an investigation of corpus-based methods for the automation of help-desk e-mail responses. Specifically, we investigate this problem along two operational dimensions: (1) information-gathering technique, and (2) granularity of the information. We consider two information-gathering techniques (retrieval and prediction) applied to information represented at two levels of granularity (document-level and sentence-level). Document-level methods correspond to the reuse of an existing response e-mail to address new requests. Sentence-level methods correspond to applying extractive multi-document summarization techniques to collate units of information from more than one e-mail. Evaluation of the performance of the different methods shows that in combination they are able to successfully automate the generation of responses for a substantial portion of e-mail requests in our corpus. We also investigate a meta-selection process that learns to choose one method to address a new inquiry e-mail, thus providing a unified response automation solution.

Museums offer vast amounts of information, but a visitor's receptivity and time are typically limited, providing the visitor with the challenge of selecting the (subjectively) interesting exhibits to view within the available time. Mobile, electronic handheld guides offer the opportunity to improve a visitor's experience by recommending exhibits of interest, and adapting the delivered content. The first step in this personalisation process is the prediction of a visitor's activities and interests. In this paper we study non-intrusive, adaptive user modelling techniques that take into account the physical constraints of the exhibition layout. We present two collaborative models for predicting a visitor's next locations in a museum, and an ensemble model that combines the predictions of these models. The three models were trained and tested on a small dataset of museum visits. Our results are encouraging, with the ensemble model yielding the best performance overall.

Statistical and probabilistic models are concerned with the use of observed sample results to make statements about unknown, dependent parameters. In user modeling, these parameters represent aspects of a user’s behaviour, such as his or her goals, preferences, and forthcoming actions or locations. Recent technological advances, in particular increased computational power, together with anytime, anyplace access to computers, and the information explosion associatedwith the Internet, provide new opportunities for information dissemination and information gathering. On one hand, people have access to large repositories of information in digital form. On the other hand, information providers can find out more about their users’ requirements by logging people’s activities. This mixture of vast electronic content and increased knowledge about people’s actions provides an opportunity to harness statistical and probabilistic models to build user models that support the delivery of personalized content. This usage of statistical and probabilistic models has been manifested in UMUAI for the last ten years. Particularly noteworthy are the articles in the Special Issue on Machine Learning for User Modeling (1998); the survey articles by Zukerman and Albrecht and byWebb et al. in the 10-year anniversary issue, respectively on predictive statistical models for user modeling (Zukerman and Albrecht 2001), and on machine learning for user modeling (Webb et al. 2001); Burke’s survey on recommender systems (Burke 2002); and Pierrakos et al.’s survey on Web usage mining (Pierrakos et al. 2003). These articles identified several challenges that user modeling presents to statistical and probabilistic modeling techniques. We classify these challenges into three categories: (1) limitations of current user modeling approaches, (2) dynamic nature of user modeling data, and (3) efficiency considerations.

We describe a probabilistic approach for the interpretation of user arguments, and investigate the incorporation of different models of a user's beliefs and inferences into this mechanism. Our approach is based on the tenet that the interpretation intended by the user is that with the highest posterior probability. This approach is implemented in a computer-based detective game, where the user explores a virtual scenario, and constructs an argument for a suspect's guilt or innocence. 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. This interpretation may differ from the user's argument in its structure and in its beliefs in the argument propositions. We conducted a synthetic evaluation of the basic interpretation mechanism, and a user-based evaluation which assesses the impact of the different user models. The results of both evaluations were encouraging, with the system generally producing argument interpretations our users found acceptable.

High quality bilingual dictionaries are rarely available for lower-density language pairs, especially for those that are closely related. Using a third language as a pivot to link two other languages is a well-known solution, and usually requires only two input bilingual dictionaries to automatically induce the new one. This approach, however, produces many incorrect translation pairs because the dictionary entries are normally are not transitive due to polysemy and the ambiguous words in the pivot language. Utilizing the complete structures of the input bilingual dictionaries positively influences the result since dropped meanings can be countered. Moreover, an additional input dictionary may provide more complete information for calculating the semantic distance between word senses which is key to suppressing wrong sense matches. This paper proposes an extended constraint optimization model to inducing new dictionaries of closely related languages from multiple input dictionaries, and its formalization based on Integer Linear Programming. Evaluations indicated that the proposal not only outperforms the baseline method, but also shows improvements in performance and scalability as more dictionaries are utilized.

The fields of user modeling and natural language processing have been closely linked since the early days of user modeling. Natural language systems consult user models in order to improve their understanding of users' requirements and to generate appropriate and relevant responses. At the same time, the information natural language systems obtain from their users is expected to increase the accuracy of their user models. In this paper, we review natural language systems for generation, understanding and dialogue, focusing on the requirements and limitations these systems and user models place on each other. We then propose avenues for future research.

The limitations of traditional knowledge representation methods for modeling complex human behaviour led to the investigation of statistical models. Predictive statistical models enable the anticipation of certain aspects of human behaviour, such as goals, actions and preferences. In this paper, we motivate the development of these models in the context of the user modeling enterprise. We then review the two main approaches to predictive statistical modeling, content-based and collaborative, and discuss the main techniques used to develop predictive statistical models. We also consider the evaluation requirements of these models in the user modeling context, and propose topics for future research.

One method for reducing the time required for plan generation is to learn search control rules from experience. The most common approach to learning search control knowledge has been explanationbased learning. An alternative approach is to use inductive learning. An inductive approach does not require a complete and tractable domain theory and has the potential to create more effective rules by learning from more than one example at a time. In this paper we describe Grasshopper, an inductive system that learns search control rules for a classical plan generation system. We also provide an empirical evaluation of Grasshopper by comparing it with an existing explanation-based learning system.

We present an approach to keyhole plan recognition which uses a dynamic belief (Bayesian) network to represent features of the domain that are needed to identify users‘ plans and goals. The application domain is a Multi-User Dungeon adventure game with thousands of possible actions and locations. We propose several network structures which represent the relations in the domain to varying extents, and compare their predictive power for predicting a user‘s current goal, next action and next location. The conditional probability distributions for each network are learned during a training phase, which dynamically builds these probabilities from observations of user behaviour. This approach allows the use of incomplete, sparse and noisy data during both training and testing. We then apply simple abstraction and learning techniques in order to speed up the performance of the most promising dynamic belief networks without a significant change in the accuracy of goal predictions. Our experimental results in the application domain show a high degree of predictive accuracy. This indicates that dynamic belief networks in general show promise for predicting a variety of behaviours in domains which have similar features to those of our domain, while reduced models, obtained by means of learning and abstraction, show promise for efficient goal prediction in such domains.

In recent years, personalised recommendations have gained importance in helping users deal with the abundance of information available online. Personalised recommendations are often based on rating predictions, and thus accurate rating prediction is essential for the generation of useful recommendations. Recently, rating prediction algorithms that are based on matrix factorisation have become increasingly popular, due to their high accuracy and scalability. However, these algorithms still deliver inaccurate rating predictions for new users, who submitted only a few ratings.In this paper, we address the new user problem by introducing several extensions to the basic matrix factorisation algorithm, which take user attributes into account when generating rating predictions. We consider both demographic attributes, explicitly supplied by users, and attributes inferred from user-generated texts. Our results show that employing our text-based user attributes yields personalised rating predictions that are more accurate than our baselines, while not requiring users to explicitly supply any information about themselves and their preferences.

Most Natural Language Generation systems developed to date assume that a user will learn only what is explicitly stated in the discourse. This assumption leads to the generation of discourse that states explicitly all the information to be conveyed, and does not address further inferences from the discourse. In this paper, we describe a student model which provides a qualitative representation of a student's beliefs and inferences, and a content planning mechanism which consults this model in order to address the above problems. Our mechanism performs inferences in backward reasoning mode to generate discourse that conveys the intended information, and in forward reasoning mode to draw conclusions from the presented information. The forward inferences enable our mechanism to address possible incorrect inferences from the discourse, and to omit information that may be easily inferred from the discourse. In addition, our mechanism improves the conciseness of the generated discourse by omitting information known by the student. The domain of our implementation is the explanation of concepts in high school algebra.

Naturally occurring discourse often contains expressions such as “however,”“as I have stated before,” and “next.” These expressions, denoted meta-comments, carry important information which helps a hearer speed up the comprehension process. In this paper, we present a mechanism for the generation of meta-comments and their incorporation into computer-generated discourse. This mechanism is based on a simple model of the anticipated effect of the messages to be conveyed and of the meta-comments on the hearer's comprehension. It was implemented in a system called FIGMENT, which generates commentaries on the solution of algebraic equations. Le discours naturel contient souvent des expressions comme 〈〈 cependant 〉〉 et 〈〈 comme je ľai dit auparavant 〉〉. Ces expressions, appelees meta-commentaires, contiennent des informations importantes qui aident a accelerer le processus de comprehension de ľ auditeur. Dans cet article, ľ auteur presente un mecanisme de generation de metacommentaires et ? integration de ceux-ci au discours cree par ordinateur. Ce mecanisme est base sur un modele simple de ľeffet anticipe des messages devant ětre transmis et de ľ';effet des meta-commentaires sur la comprehension de ľ auditeur. II a ete mis en oeuvre dans un systeme appele FIGMENT, qui genere des commentaires sur la solution ?equations algebriques.

This paper presents a mechanism which infers a user's plans from his/her utterances by directing the inference process towards the more likely interpretations of a speaker's statements among many possible interpretations. Our mechanism uses Bayesian theory of probability to assess the likelihood of an interpretation, and it complements this assessment by taking into consideration two aspects of an interpretation: its coherence and its information content. The coherence of an interpretation is determined by the relationships between the different statements in the discourse. The information content of an interpretation is a measure of how well defined the interpretation is in terms of the actions to be performed on the basis of this interpretation. This measure is used to guide the inference process towards interpretations with higher information content. The information content of an interpretation depends on the specificity and the certainty of the inferences in it, where the certainty of an inference depends on the knowledge on which the inference is based. Our mechanism has been developed for use in task-oriented consultation systems. The particular domain that we have chosen for exploration is that of travel booking.

We describe a probabilistic reference disambiguation mechanism developed for a spoken dialogue system mounted on an autonomous robotic agent. Our mechanism receives as input referring expressions containing intrinsic features of individual concepts (lexical item, size and colour) and features involving more than one concept (ownership and location). It then performs probabilistic comparisons between the given features and features of objects in the domain, yielding a ranked list of candidate referents. Our evaluation shows high reference resolution accuracy across a range of spoken referring expressions.

Human discourse is fraught with rhetorical devices such as contradictions, illustrations, and analogies. These rhetorical devices carry important information which a listener uses to speed up the comprehension process. In this paper, we present a taxonomy of rhetorical devices commonly used in tutoring environment, and model the meaning of a class of rhetorical devices in terms of their anticipated effect on a listener's knowledge. These predictions are then used in planning computer-generated discourse. As a testbed for our ideas, a system called WISHFUL is being implemented to generate commentaries in the domain of high-school algebra within the framework of an intelligent tutoring system.

We are developing a corpus-based approach for the prediction of help-desk responses from features in customers' emails, where responses are represented at two levels of granularity: document and sentence. We present an automatic evaluation of the responses generated by our system, as well as a manual one involving human judges. The automatic evaluation involves textual comparisons between generated responses and responses composed by the help-desk operators. The results show that both levels of granularity produce good responses, addressing inquiries of different kinds. The human-based evaluation measures response in- formativeness, and confirms our conclusion that both levels of granularity produce useful responses.

We describe Scusi?, a multi-stage, spoken language interpretation mechanism designed to be part of a robot- mounted dialogue system. Scusi?'s interpretation process maps spoken utterances to conceptual graphs, and the nodes in these graphs to concepts in the world. Maximum posterior probability is used to rank the (partial) interpretations produced at each stage of this process. We show how the features of our interpretation process yield desirable behaviours that support robust and flexible system performance.

We present a comparative study of corpus-based methods for the automatic synthesis of email responses to help-desk requests. Our methods were developed by considering two operational dimensions: (1) information-gathering technique, and (2) granularity of the information. In particular, we investigate two techniques – retrieval and prediction – applied to information represented at two levels of granularity – sentence level and document level. We also developed a hybrid method that combines prediction with retrieval. Our results show that the different approaches are applicable in different situations, addressing a combined 72% of the requests with either complete or partial responses.

We describe Scusi?, a multi-stage, spoken language interpretation mechanism designed to be part of a robot-mounted dialogue system. Scusi?'s interpretation process maps spoken utterances to conceptual graphs, and the nodes in these graphs to concepts in the world. Maximum posterior probability is used to rank the (partial) interpretations produced at each stage of this process.

In a collaborative environment, knowledge about collaborators' skills is an important factor when determining which team members should perform a task. However, this knowledge may be incomplete or uncertain. In this paper, we extend our ETAPP (Environment-Task-Agents-Policy-Protocol) collaboration framework by modeling team members that exhibit non-deterministic performance, and comparing two alternative ways of using these models to assign agents to tasks. Our simulation-based evaluation shows that performance variability has a large impact on task performance, and that task performance is improved by consulting agent models built from a small number of observations of agents' recent performance.

We present a corpus-based approach for the automatic analysis and synthesis of email responses to help-desk requests. This approach can be used to automatically deal with repetitive requests of low technical content, thus enabling help-desk operators to focus their effort on more difficult requests. We propose a method for extracting high-precision sentences for inclusion in a response, and a measure for predicting the completeness of a planned response. The idea is that complete, high-precision responses may be sent directly to users, while incomplete responses should be passed to operators. Our results show that a small but significant proportion (14%) of our automatically generated responses have a high degree of precision and completeness, and that our measure can reliably predict the completeness of a response.

We investigate two seemingly incompatible approaches for improving document retrieval performance in the context of question answering: query expansion and query reduction. Queries are expanded by generating lexical paraphrases. Syntactic, semantic and corpus-based frequency information is used in this process. Queries are reduced by removing words that may detract from retrieval performance. Features that identify these words were obtained from decision graphs. These approaches were evaluated using a subset of queries from TREC8, 9 and 10. Our evaluation shows that each approach in isolation improves retrieval performance, and both approaches together yield substantial improvements. Specifically, query expansion followed by reduction improved the average number of correct documents retrieved by 21.7% and the average number of queries that can be answered by 15%.

The problem of interpreting Natural Language (NL) discourse is generally of exponential complexity. However, since interactions with users must be conducted in real time, an exhaustive search is not a practical option. In this paper, we present an anytime algorithm that generates "good enough" interpretations of probabilistic NL arguments in the context of a Bayesian network (BN). These interpretations consist of: BN nodes that match the sentences in a given argument, assumptions that justify the beliefs in the argument, and a reasoning structure that adds detail to the argument. We evaluated our algorithm using automatically generated arguments and hand-generated arguments. In both cases, our algorithm generated good interpretations (and often the best interpretation) in real time.

We describe an argument-interpretation mechanism based on the Minimum Message Length Principle [1], and investigate the incorporation of a model of the user's beliefs into this mechanism. 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. This interpretation may differ from the user's argument in its structure and in its beliefs in the argument propositions. The results of our evaluation are encouraging, with the system generally producing plausible interpretations of users' arguments.

We investigate lexical paraphrasing in the context of two distinct applications: document retrieval and node identification. Document retrieval --- the first step in question answering --- retrieves documents that contain answers to user queries. Node identification --- performed in the context of a Bayesian argumentation system --- matches users' Natural Language sentences to nodes in a Bayesian network. Lexical paraphrases are generated using syntactic, semantic and corpus-based information. Our evaluation shows that lexical paraphrasing improves retrieval performance for both applications.

We describe a set of supervised machine learning experiments centering on the construction of statistical models of WH-questions. These models, which are built from shallow linguistic features of questions, are employed to predict target variables which represent a user's informational goals. We report on different aspects of the predictive performance of our models, including the influence of various training and testing factors on predictive performance, and examine the relationships among the target variables.

This paper describes an integrated approach for interpreting a user's responses and generating replies in the framework of a WWW-based Bayesian argumentation system. Our system consults a user model which represents a user's beliefs, inferences and attentional focus, as well as the system's certainty regarding the user's beliefs. The interpretation mechanism takes into account these factors to infer the intended effect of the user's response on the system's argument. The reply-generation mechanism focuses on the identification of discrepancies between the beliefs in the user model and the beliefs held by the system that are relevant to the inferred interpretation.

The automated prediction of a user's interests and requirements is an area of interest to the Artificial Intelligence community. However, current predictive statistical approaches are subject to theoretical and practical limitations which restrict their ability to make useful predictions in domains such as the WWW and computer games that have vast numbers of values for variables of interest. In this paper, we describe an automated abstraction technique which addresses this problem in the context of Dynamic Bayesian Networks. We compare the performance and computational requirements of fine-grained models built with precise variable values with the performance and requirements of a coarse-grained model built with abstracted values. Our results indicate that complex, coarse-grained models offer performance and computational advantages compared to simpler, fine-grained models.

We describe a mechanism which recognizes a user's intentions from short-form rejoinders to arguments generated from Bayesian networks. The mechanism builds candidate reasoning paths that link the user's rejoinder with a previously presented argument, and considers the following factors to select a path: linguistic clues, the impact of the user's rejoinder on the system's argument along the different paths, the user's attentional focus, and the system's confidence in its representation of the user's beliefs. The results of our preliminary evaluation indicate that the interpretations produced by our mechanism are generally appropriate.

We describe several Markov models derived from the behaviour patterns of many users, which predict which documents a user is likely to request next. We then present comparative results of the predictive accuracy of the different models, and, based on these results, build hybrid models which combine the individual models in different ways. These hybrid models generally have a greater predictive accuracy than the individual models. The best models will be incorporated in a system for pre-sending WWW documents.

Our argumentation system NAG uses Bayesian networks in a user model and in a normative model to assemble and assess nice arguments, that is arguments which balance persuasiveness with normative correctness. Attentional focus is simulated in both models to select relevant subnetworks for Bayesian propagation. Bayesian propagation in the user model is modified to represent some human cognitive weaknesses. The subnetworks are expanded in an iterative abductive process until argumentative goals are achieved in both models, when the argument is presented to the user.

We present an approach to goal recognition which uses a Dynamic Belief Network to represent domain features needed to identify users' goals and plans. Different network structures have been developed, and their conditional probability distributions have been automatically acquired from training data. These networks show a high degree of accuracy in predicting users' goals. Our approach allows the use of incomplete, sparse and noisy data during both training and testing. We then apply simple learning techniques to learn significant actions in the domain. This speeds up the performance of the most promising dynamic belief networks without loss in predictive accuracy.

A multimodal presentation planning mechanism must take into consideration the structure of the discourse and the restrictions imposed by partial plans generated in the early stages of the planning process. The latter requirement demands that the planning mechanism be able to transfer plan constraints from one level of planning to the next and to modify partial plans locally at each level. In this paper, we introduce a multi-agent planning mechanism based on the blackboard architecture that satisfies these requirements, and describe the facilities used by agents to communicate with each other in this architecture. In addition, we discuss the planning processes used by the agents in our prototype system.

A method for deriving equivalence classes for lexical access in speech recognition is considered, which automatically derives equivalence classes from training data using unsupervised learning and the Minimum Message Length Criterion. These classes model insertions, deletions and substitutions in an input phoneme string due to mis-recognition and mis-pronunciation, and allow unlikely word candidates to be eliminated quickly. This in turn allows a more detailed examination of the remaining candidates to be carried out efficiently.

During interactions with human consultants, people are used to providing partial and/or inaccurate information, and still be understood and assisted. We attempt to emulate this capability of human consultants; in computer consultation systems. In this paper, we present a mechanism for handling uncertainty in plan recognition during task-oriented consultations. The uncertainty arises while choosing an appropriate interpretation of a user's statements among many possible interpretations. Our mechanism handles this uncertainty by using probability theory to assess the probabilities of the interpretations, and complements this assessment by taking into account the information content of the interpretations. The information content of an interpretation is a measure of how well defined an interpretation is in terms of the actions to be performed on the basis of the interpretation. This measure is used to guide the inference process towards interpretations with a higher information content. The information content of an interpretation depends on the specificity and the strength of the inferences in it, where the strength of an inference depends on the reliability of the information on which the inference is based. Our mechanism has been developed for use in task-oriented consultation systems. The domain that we have chosen for exploration is that of a travel agency.

The computational complexity of planning has motivated significant efforts in machine learning. However, much of this work has concentrated on explanation-based learning techniques. An alternative approach is to use inductive learning. An inductive approach does not require a complete and tractable domain theory to be encoded and has the potential to create more effective rules by learning from more than one example at a time. In this paper, we describe an inductive system for learning search control rules and compare it with an existing explanation-based learning system.

In discourse produced by human speakers, the technical information is often interleaved with rhetorical devices such as contradictions, illustrations and analogies. These rhetorical devices carry important information which assists the listener in assimilating the imparted knowledge. In this paper, we present a taxonomy of rhetorical devices commonly used in a tutoring environment, and model their meaning in terms of their anticipated effect on a listener's knowledge. These predictions are then used in planning computer generated discourse. As a testbed for our ideas, we are implementing a system, called WISHFUL, to generate commentaries in the domain of high-school algebra within the framework of an Intelligent Tutoring System.