We address the problem of dynamic user modelling for referring expression generation in spoken dialogue systems, i.e how a spoken dialogue system should choose referring expressions to refer to domain entities to users with different levels of domain expertise, whose domain knowledge is initially unknown to the system. We approach this problem using a statistical planning framework: Reinforcement Learning techniques in Markov Decision Processes (MDP). We present a new reinforcement learning framework to learn user modelling strategies for adaptive referring expression generation (REG) in resource scarce domains (i.e. where no large corpus exists for learning). As a part of the framework, we present novel user simulation models that are sensitive to the referring expressions used by the system and are able to simulate users with different levels of domain knowledge. Such models are shown to simulate real user behaviour more closely than baseline user simulation models. In contrast to previous approaches to user adaptive systems, we do not assume that the user’s domain knowledge is available to the system before the conversation starts. We show that using a small corpus of non-adaptive dialogues it is possible to learn an adaptive user modelling policy in resource scarce domains using our framework. We also show that the learned user modelling strategies performed better in terms of adaptation than hand-coded baselines policies on both simulated and real users. With real users, the learned policy produced around 20% increase in adaptation in comparison to the best performing hand-coded adaptive baseline. We also show that adaptation to user’s domain knowledge results in improving task success (99.47% for learned policy vs 84.7% for hand-coded baseline) and reducing dialogue time of the conversation (11% relative difference). This is because users found it easier to identify domain objects when the system used adaptive referring expressions during the conversations.

Markov Logic (ML) is a novel approach to Natural Language Processing tasks [Richardson and Domingos, 2006; Riedel, 2008]. It is a Statistical Relational Learning language based on First Order Logic (FOL) and Markov Networks (MN). It allows one to treat a task as structured classification. In this work, we investigate ML for the semantic processing tasks of Spoken Language Understanding (SLU) and Semantic Role Labelling (SRL). Both tasks consist of identifying a semantic representation for the meaning of a given utterance/sentence. However, they differ in nature: SLU is in the field of dialogue systems where the domain is closed and language is spoken [He and Young, 2005], while SRL is for open domains and traditionally for written text [M´arquez et al., 2008]. Robust SLU is a key component of spoken dialogue systems. This component consists of identifying the meaning of the user utterances addressed to the system. Recent statistical approaches to SLU depend on additional resources (e.g., gazetteers, grammars, syntactic treebanks) which are expensive and time-consuming to produce and maintain. On the other hand, simple datasets annotated only with slot-values are commonly used in dialogue system development, and are easy to collect, automatically annotate, and update. However, slot-values leave out some of the fine-grained long distance dependencies present in other semantic representations. In this work we investigate the development of SLU modules with minimum resources with slot-values as their semantic representation. We propose to use the ML to capture long distance dependencies which are not explicitly available in the slot-value semantic representation. We test the adequacy of the ML framework by comparing against a set of baselines using state of the art approaches to semantic processing. The results of this research have been published in Meza-Ruiz et al. [2008a,b]. Furthermore, we address the question of scalability of the ML approach for other NLP tasks involving the identification of semantic representations. In particular, we focus on SRL: the task of identifying predicates and arguments within sentences, together with their semantic roles. The semantic representation built during SRL is more complex than the slot-values used in dialogue systems, in the sense that they include the notion of predicate/argument scope. SRL is defined in the context of open domains under the premises that there are several levels of extra resources (lemmas, POS tags, constituent or dependency parses). In this work, we propose a ML model of SRL and experiment with the different architectures we can describe for the model which gives us an insight into the types of correlations that the ML model can express [Riedel and Meza-Ruiz, 2008; Meza-Ruiz and Riedel, 2009]. Additionally, we tested our minimal resources setup in a state of the art dialogue system: the TownInfo system. In this case, we were given a small dataset of gold standard semantic representations which were system dependent, and we rapidly developed a SLU module used in the functioning dialogue system. No extra resources were necessary in order to reach state of the art results.

A Spoken Dialogue System's (SDS's) dialogue strategy specifies which action it will take depending on its representation of the current dialogue context. Designing it by hand involves anticipating how users will interact with the system, and/or repeated testing and refining, and so can be a difficult, time-consuming task. Since SDSs inevitably make understanding errors, a particularly important issue is how to design ``repair strategies'', the parts of the dialogue strategy which attempt to get the dialogue ``back-on-track'' following these errors. To try to produce better dialogue strategies with less time and effort, previous researchers have modelled a dialogue strategy as a sequential decision problem called a Markov Decision Process (MDP), and then applied Reinforcement Learning (RL) algorithms to example training dialogues to generate dialogue strategies automatically. More recent research has used training dialogues conducted with simulated rather than real users and learned which action to take in all dialogue contexts, (a ``full'' as opposed to a ``partial'' dialogue strategy) - simulated users allow more training dialogues to be generated, and the exploration of new dialogue contexts not present in an original dataset. As yet however, limited insight has been provided as to which dialogue contextual features are important to include in the MDP and why. Indeed, a full dialogue strategy has not been learned from training dialogues with a realistic probabilistic user simulation derived from real user data, and then shown to work well with real users. This thesis investigates the value of adding new linguistically-motivated contextual features to the MDP when using RL to learn full dialogue strategies for SDSs. These new features are recent Dialogue Acts (DAs). DAs indicate the role or intention of an utterance in a dialogue e.g. ``provide-information'', an utterance being a complete unit of a speaker's speech, often bounded by silence. An accurate probabilistic user simulation learned from real user data is used for generating training dialogues, and the recent DAs are shown to improve performance in testing in simulation and with real users. With real users, performance is also better than other competing learned and hand-crafted strategies. Analysis of the strategies, and further simulation experiments show how the DAs improve performance through better repair strategies. The main findings are expected to apply to SDSs in general - indeed our strategies are learned and tested on real users in different domains, (flight-booking versus tourist information). Comparisons are also made to recent research which focuses on handling understanding errors in SDSs, but which does not use RL or user simulations.

<p>The work in this thesis addresses design and development of multimodal dialogue recommender systems for the home context-of-use. In the design part, two investigations on multimodal recommendation dialogue interaction in the home context are reported on. The first study gives implications for the design of dialogue system interaction including personalization and a three-entity multimodal interaction model accommodating dialogue feedback in order to make the interaction more efficient and successful. In the second study a dialogue corpus of movie recommendation dialogues is collected and analyzed, providing a characterization of such dialogues. We identify three initiative types that need to be addressed in a recommender dialogue system implementation: system-driven preference requests, userdriven information requests, and preference volunteering. Through the process of dialogue distilling, a dialogue control strategy covering system-driven preference requests from the corpus is arrived at.</p><p>In the development part, an application-driven development process is adopted where reusable generic components evolve through the iterative and incremental refinement of dialogue systems. The Phase Graph Processor (PGP) design pattern is one such evolved component suggesting a phase-based control of dialogue systems. PGP is a generic and flexible micro architecture accommodating frequent change of requirements inherent of agile, evolutionary system development. As PGP has been used in a series of previous information-providing dialogue system projects, a standard phase graph has been established that covers the second initiative type; user-driven information requests. The phase graph is incrementally refined in order to provide user preference modeling, thus addressing the third initiative type, and multimodality as indicated by the user studies. In the iterative development of the multimodal recommender dialogue system MADFILM the phase graph is coupled with the dialogue control strategy in order to cater for the seamless integration of the three initiative types.</p> / Report code: LiU-TEK-LIC-2004:08.

We study human-chatbot collaborative conversation systems that enable humans to leverage AI chatbot outputs during an online conversation with others. We evaluate response quality in two collaborative systems and compare them with human-only and chatbot-only settings. Both collaborative systems present AI chatbot results as suggestions but encourage the synthesis of human and chatbot responses to different extents. We also examine the influence of chatbot choices, including both retrieval-based and generation-based methods, and the number of suggestions on collaborative systems. Experimental results show that our collaborative systems can significantly improve the efficiency to formulate a response and improve its quality compared with a human-only system while sacrificing the fluency and humanness of the messages. Compared with a chatbot, collaborative systems can provide answers that are more fluent, human-like, and informative. We also found that the retrieval-based chatbots perform better than the generation-based one from all aspects. The optimal number of chatbot suggestions is one, and showing more suggestions has reduced user efficiency. / Master of Science / Artificial Intelligence (AI) systems have become remarkably interactive and accurate with them becoming an integral part of our life. The increasing use of personal assistants like Siri and the application of AI in important real-world tasks such as medical imaging and diagnosis show that AI can perform as good as trained human experts. Organizations today are expanding at a rapid rate and need to service millions of customers concurrently to remain competitive in the market. With the recent success of AI chatbots, the collaboration of Human and AI to augment customer service management is one of the most sought out solutions to this requirement. A service flow where virtual agents and people work together can be a boon to the industry by making the human agents smarter with a bot "whispering" in their ears. We present the design of various collaborative systems we have developed and discuss the improvements in response efficiency and quality due to them in multiple online user experiments. The results of this study can be used to improve conversational chat systems that assist human agents to improve their response time and quality and identify features of the AI agent that are most beneficial for improving the conversation.

Gender is among the most influential of the factors underlying differences in spatial abilities, human communication and interactions with and through computers. Past research has offered important insights into gender differences in navigation and language use. Yet, given the multidimensionality of these domains, many issues remain contentious while others unexplored. Moreover, having been derived from non-interactive, and often artificial, studies, the generalisability of this research to interactive contexts of use, particularly in the practical domain of Human-Computer Interaction (HCI), may be problematic. At the same time, little is known about how gender strategies, behaviours and preferences interact with the features of technology in various domains of HCI, including collaborative systems and systems with natural language interfaces. Targeting these knowledge gaps, the thesis aims to address the central question of how gender differences emerge and operate in spatial navigation dialogues with computer systems. To this end, an empirical study is undertaken, in which, mixed-gender and same-gender pairs communicate to complete an urban navigation task, with one of the participants being under the impression that he/she interacts with a robot. Performance and dialogue data were collected using a custom system that supported synchronous navigation and communication between the user and the robot. Based on this empirical data, the thesis describes the key role of the interaction of gender in navigation performance and communication processes, which outweighed the effect of individual gender, moderating gender differences and reversing predicted patterns of performance and language use. This thesis has produced several contributions; theoretical, methodological and practical. From a theoretical perspective, it offers novel findings in gender differences in navigation and communication. The methodological contribution concerns the successful application of dialogue as a naturalistic, and yet experimentally sound, research paradigm to study gender and spatial language. The practical contributions include concrete design guidelines for natural language systems and implications for the development of gender-neutral interfaces in specific domains of HCI.

In this thesis, the design and use of ontologies as domain knowledge sources in information-providing dialogue systems are investigated. The research is divided into two parts, theoretical investigations that have resulted in a requirements specifications on the design of ontologies to be used in information-providing dialogue systems, and empirical work on the development of a framework for use of ontologies in information-providing dialogue systems. The framework includes three models: A model for ontology-based semantic analysis of questions. A model for ontology-based dialogue management, specifically focus management and clarifications. A model for ontology-based domain knowledge management, specifically transformation of user requests to system oriented concepts used for information retrieval. In this thesis, it is shown that using ontologies to represent and reason on domain knowledge in dialogue systems has several advantages. A deeper semantic analysis is possible in several modules and a more natural and efficient dialogue can be achieved. Another important aspect is that it facilitates portability; to be able to reuse adapt the dialogue system to new tasks and domains, since the domain-specific knowledge is separated form generic features in the dialogue system architecture. Other advantages are that it reduces the complexity of linguistic produced in various domains.

This thesis presents a development of an English spoken dialogue system based on the Alex dialogue system framework. The work describes a component adaptation of the framework for a different domain and language. The system provides public transport information in New York. This work involves creating a statistical model and the deployment of custom Kaldi speech recognizer. Its performance was better in comparison with the Google Speech API. The comparison was based on a subjective user satisfaction acquired by crowdsourcing. Powered by TCPDF (www.tcpdf.org)

Abstract Development of trainable policies for spoken dialogue systems Thanh Le In human­human interaction, speech is the most natural and effective manner of communication. Spoken Dialogue Systems (SDS) have been trying to bring that high level interaction to computer systems, so with SDS, you could talk to machines rather than learn to use mouse and keyboard for performing a task. However, as inaccuracy in speech recognition and inherent ambiguity in spoken language, the dialogue state (user's desire) can never be known with certainty, and therefore, building such a SDS is not trivial. Statistical approaches have been proposed to deal with these uncertainties by maintaining a probability distribution over every possible dialogue state. Based on these distributions, the system learns how to interact with users, somehow to achieve the final goal in the most effective manner. In Reinforcement Learning (RL), the learning process is understood as optimizing a policy of choosing action conditioned on the current belief state. Since the space of dialogue...

The most important component of virtually any dialog system is a dialogue manager. The aim of the dialog manager is to propose an action (a continuation of the dialogue) given the last dialog state. The dialog state summarises all the past user input and the system input and ideally it includes all information necessary for natural progress in the dialog. For the dialog manager to work efficiently, it is important to model the probability distribution over all dialog states as precisely as possible. It is possible that the set of dialog states will be very large, so approximative methods usually must be used. In this thesis we will discuss an implementation of approximate Bayes methods for belief state monitoring. The result is a library for dialog state monitoring in real dialog systems. 1