Dissertations / Theses on the topic 'Task-oriented dialogue'

The Principle of Parsimony states that by and large, agents try to complete tasks using as little effort as possible. This thesis demonstrates that the Principle of Parsimony operates in human task-oriented dialogue by showing the effects of Parsimony in a corpus of human dialogues about a map navigation task and by using the main points of the analysis in order to guide simulated conversations between two computer agents within the JAM system. It makes four major contributions: an analysis of 'communicative posture', or a range of choices in dialogue which can be characterised by decisions about how much effort to spend constructing one's utterances, leading to either careful or risky behaviour about different aspects of communication, an analysis of 'recovery strategies' which allow the participants to recover from failures which have been brought about due to risky postures, a heuristic model of belief which risks failing to capture the full meaning of the dialogue in favour of efficiency in a way which models human belief updating more plausibly than previous models, and a layered agent architecture which allows the simulated agents to make all of their decisions based on the Principle of Parsimony.

This thesis explores the implications of Lindblom's theory of Hyper- and Hypo-articulation (Lindblom, 1983, 1990) for word intelligibility and the likely application of phonological reduction processes in spontaneous discourse, using data from the HCRC Map Task Corpus. Lindblom claims that variability in articulatory clarity is a reflection of speakers' assessments of their listeners' information requirements: speakers hyper-articulate when listeners require maximum acoustic input from with information from other sources. To prevent speakers from over-economising to a point of unintelligibility, hypo-articulation is governed by a constraint of lexical distinctiveness: speakers hypo-articulate only while listeners are able to distinguish the target from competing lexical items. Three main questions are addressed. First, do the informational needs of the listener affect the articulatory clarity of words produced in spontaneous conversation? A series of intelligibility experiments shows that repeated mentions of landmark names are less intelligible than their introductory mentions, independent of which speaker utters either mention, and who can see the landmark on their map. Although the results can be interpreted as supporting Lindblom's view, textual Giveness (Prince, 1981) is shown to depend upon what the speaker knows, rather than what the speaker believes her listener to know. The reduction in clarity associated with an increase in available information is not necessarily listener-oriented as the H & H theory proposes. Secondly, do phonological processes such as word-final /d/-deletion or place assimilation contribute to intelligibility loss? Although reduction processes are found to be more prevalent in tokens from spontaneous discourse than in matched citation forms, they generally fail to account for effects of repetition. An increase in assimilation is found for repeated mentions of nasal-final stimuli in pre-velar position, but no effects is found for assimilation in pre-label position, or for word-final /d/-deletion, nor is an effect found for the duration of schwa in metrically Weak initial syllables of polysyllabic words.

This thesis presents a discussion of proposed structuring principles for dialogue, and tests them empirically using data from the HCRC Map Task Corpus. The concepts of Cooperation (as described by Grice, and as used more generally in Linguistics), Coordination, Collaboration, Parsimony, Risk and Effort are ex¬ amined, and empirically testable hypotheses are developed, with which we are able to evaluate the claims for these principles in the context of task-oriented dialogue. In order to test our hypotheses, we categorise the utterances in our database in terms of Risk and Effort. Unlike Discourse Analysis, Conversation Analysis or Dialogue Games, our approach is evaluative. The intent in our dialogue coding is not only to label what the speakers did, but also to assess it in terms of its appropriateness at that point in the dialogue: the system codes not only what people do, but also what they don't do. Therefore, our system marks both the presence and absence of dialogue attributes. The hypotheses derived from the structuring concepts were statistically tested on the data produced by the coding system. The results produced by the empiri¬ cal tests showed a relationship between dialogue errors and task errors, but not between increased effort and increased task success. The importance of matched effort was also demonstrated, as dialogue pairs who invested similar amounts of effort produced better task results. Dialogue pairs also produced better results over time, which we argue is due to the focusing of effort. Participants work out where their effort should be channelled so that they can increase risk-taking where problems have not occurred, and decrease risk-taking where problems have occurred. These results suggest that interactants' behaviour follows a Principle of Least Individual Effort, which we argue subsumes the Principle of Parsimony and thus the Risk-Effort Trade-Olf. We reject the Principle of Least Collaborative Effort because although the empirical result of high effort not being associated with task success supports this principle in theory, we argue that its motivation is not supported in practice. The empirical work also distinguishes between what we term 'Gricean Cooperation', and folklinguistic notions of Cooperation found in the literature. In general terms, Gricean Cooperation predicted the same type of effort-minimising behaviour as the Principle of Least Individual Effort, and was thus supported by the empirical work. However, the concept of 'helpfulness' suggested by more general uses of Cooperation made predictions which were in conflict with those of the Principle of Least Individual Effort, and were found not to be supported.

Les besoins croissants en formation et en entrainement au travail d’équipe ont motivé l’utilisationd’Environnements de réalité Virtuelle Collaboratifs de Formation (EVCF) qui permettent aux utilisateurs de travailler avec des agents autonomes pour réaliser une activité collective. L’idée directrice est que la coordination efficace entre les membres d’une équipe améliore la productivité et réduit les erreurs individuelles et collectives. Cette thèse traite de la mise en place et du maintien de la coordination au sein d’une équipe de travail composée d’agents et d’humains interagissant dans un EVCF.L’objectif de ces recherches est de doter les agents virtuels de comportements conversationnels permettant la coopération entre agents et avec l’utilisateur dans le but de réaliser un but commun.Nous proposons une architecture d’agents Collaboratifs et Conversationnels, dérivée de l’architecture Belief-Desire-Intention (C2-BDI), qui gère uniformément les comportements délibératifs et conversationnels comme deux comportements dirigés vers les buts de l’activité collective. Nous proposons un modèle intégré de la coordination fondé sur l’approche des modèles mentaux partagés, afin d’établir la coordination au sein de l’équipe de travail composée d’humains et d’agents. Nous soutenons que les interactions en langage naturel entre les membres d’une équipe modifient les modèles mentaux individuels et partagés des participants. Enfin, nous décrivons comment les agents mettent en place et maintiennent la coordination au sein de l’équipe par le biais de conversations en langage naturel. Afin d’établir un couplage fort entre la prise de décision et le comportement conversationnel collaboratif d’un agent, nous proposons tout d’abord une approche fondée sur la modélisation sémantique des activités humaines et de l’environnement virtuel via le modèle mascaret puis, dans un second temps, une modélisation du contexte basée sur l’approche Information State. Ces représentations permettent de traiter de manière unifiée les connaissances sémantiques des agents sur l’activité collective et sur l’environnement virtuel ainsi que des informations qu’ils échangent lors de dialogues.Ces informations sont utilisées par les agents pour la génération et la compréhension du langage naturel multipartite. L’approche Information State nous permet de doter les agents C2BDI de capacités communicatives leur permettant de s’engager pro-activement dans des interactions en langue naturelle en vue de coordonner efficacement leur activité avec les autres membres de l’équipe. De plus, nous définissons les protocoles conversationnels collaboratifs favorisant la coordination entre les membres de l’équipe. Enfin, nous proposons dans cette thèse un mécanisme de prise de décision s’inspirant de l’approche BDI qui lie les comportements de délibération et de conversation des agents. Nous avons mis en oeuvre notre architecture dans trois différents scénarios se déroulant dans des EVCF. Nous montrons que les comportements conversationnels collaboratifs multipartites des agents C2BDI facilitent la coordination effective de l’utilisateur avec les autres membres de l’équipe lors de la réalisation d’une tâche partagée
Growing needs of educational and training requirements motivate the use of collaborative virtual environments for training (CVET) that allows human users to work together with autonomous agents to perform a collective activity. The vision is inspired by the fact that the effective coordination improves productivity, and reduces the individual and team errors. This work addresses the issue of establishing and maintaining the coordination in a mixed human-agent teamwork in the context of CVET. The objective of this research is to provide human-like conversational behavior of the virtual agents in order to cooperate with a user and other agents to achieve shared goals.We propose a belief-desire-intention (BDI) like Collaborative Conversational agent architecture(C2BDI) that treats both deliberative and conversational behaviors uniformly as guided by the goal-directed shared activity. We put forward an integrated model of coordination which is founded on the shared mental model based approaches to establish coordination in a human-agent teamwork. We argue that natural language interaction between team members can affect and modify the individual and shared mental models of the participants. Finally, we describe the cultivation of coordination in a mixed human-agent teamwork through natural language conversation. In order to establish the strong coupling between decision making and the collaborative conversational behavior of the agent, we propose first, the Mascaret based semantic modeling of human activities and the VE, and second, the information state based context model. This representation allows the treatment of semantic knowledge of the collaborative activity and virtual environment, and information exchanged during the dialogue conversation in a unified manner. This knowledge can be used by the agent for multiparty natural language processing (understanding and generation) in the context of the CEVT. To endow the communicative capabilities to C2BDI agent, we put forward the information state based approach for the natural language processing of the utterances. We define collaborative conversation protocols that ensure the coordination between team members. Finally, in this thesis, we propose a decision making mechanism, which is inspired by the BDI based approach and provides the interleaving between deliberation and conversational behavior of the agent. We have applied the proposed architecture to three different scenarios in the CVET. We found that the multiparty collaborative conversational behavior of C2BDI agent is more constructive and facilitates the user to effectively coordinate with other team members to perform a shared task

The thesis seeks to develop an account of collaborative activities within the framework of ecological realism—an approach to psychology developed by James J. Gibson in the course of work on visual perception. Two main questions are addressed; one ontological, and one methodological. The ontological question is: given that collaborative activities take place within an environment, what kinds of structure must this environment contain? The response emphasizes the importance of relations which exist between entities, and which connect a given perceiver-actor with the other objects and individuals in its surroundings, and with the relations between those entities. It is held that activities take place within a field of relations. This description draws on the radical empiricist doctrine that relations are real, are external, and are directly perceivable. The present proposal insists that, in addition to being directly perceivable, relations can also be directly acted upon: throwing a ball for a dog is acting on a relation between dog and ball in space. The relational field account of collaboration naturally extends to an account of speaking: people, through their history of acting in an environment populated by other speakers, come to stand in a set of relations with objects and events around them, and these relations can be directly acted upon by others through the use of verbal actions. Verbal actions serve to direct the attention of others to relevant aspects of the environment, and this allows us as speakers to coordinate and manage one another’s activity. The methodological question is this: granting that the environment may be structured as a field of relations, how are we to conduct our empirical investigations, such that we can ask precise questions which lead to useful insights about how a given collaborative activity is carried out in practice? The central issue here concerns the concept of the task. Psychologists are in the habit of using this term quite loosely, to denote the actions of an individual or a group, in a laboratory or outside. This creates confusion in discussions of collaborative phenomena: who is the agent of a ‘collaborative task’? The definition offered here states that a task is a researcher-defined unit of study that corresponds to a change in the structure of the environment that has a characteristic pattern and that is meaningful from the first-person perspective of a particular actor. On this definition, the task is a tool that allows ecological psychologists to carve up the problem space into specific, tractable questions; the task is the equivalent of the cognitivist’s mental module. Task-oriented psychology encourages us to ask the question: which specific resources is the individual making use of in controlling this particular activity? The methodology is developed through an examination of the alarm calling behaviour of vervet monkeys, which is explained in terms of actions on the relational field, and through an analysis of corpus data from a laboratory-based collaborative assembly game. The relational field model promises to provide a way of studying social and collaborative activities on ecological realist principles. The concluding chapter identifies two particular areas in which the model might fruitfully be developed: in the study of learning, and in the theory of designing objects and spaces for interaction.

Task-oriented dialogue systems are a popular way for organisations to generate extra value both internally and for customers. Modern approaches for these dialogue systems that use neural networks to enable training directly on written dialogues are very data hungry, which complicates their implementation. Crowdsourcing is an attractive solution for generating this type of training data, but the method also comes with several difficulties. We introduce a new method for generating training data based on parallel crowdsourcing of dialogues, as well as crowdsourced quality review. We use this method to collect a small dataset that takes place within the domain bus driver-traveler. We believe that this method offers an efficient way to collect new, high-quality datasets. Hybrid Code Networks is a model for dialogue systems that combines a neural network with domain-specific knowledge, and thus requires a significantly smaller amount of training data than other similar dialogue systems to achieve comparable performance. By combining Hybrid Code Networks with our new method for generating training data, we believe that the threshold for implementing task-oriented dialogue systems on domains with insufficient training data can be lowered. We implement Hybrid Code Networks and train the implementation on the collected dataset and achieve good results.
Uppgiftsorienterade dialogsystem är ett populärt sätt för företag att generera extra värde både internt och för kunder. Moderna modeller för dessa dialogsystem som använder neurala nätverk för att möjliggöra träning direkt på skriftliga dialoger är väldigt datahungriga, vilket försvårar implementationen av dessa. Crowdsourcing är en attraktiv lösning för att generera denna typ av träningsdata, men metoden kommer även med flera svårigheter. Vi introducerar en ny metod för generering av träningsdata som bygger på parallell crowdsourcing av dialoger, samt crowdsourcad kvalitetsgranskning. Vi använder denna metod för att samla in ett litet dataset som utspelar sig inom domänen busschaufför-resenär. Vi menar att denna metod erbjuder ett effektivt sätt att samla in nya, högkvalitativa dataset. Hybrid Code Networks är en modell för dialogsystem som kombinerar ett neuralt nätverk med domänspecifik kunskap, och som på så sätt kräver en betydligt mindre mängd träningsdata än andra liknande dialogsystem för att uppnå jämförbar prestanda. Genom att kombinera Hybrid Code Networks med vår nya metod för generering av träningsdata menar vi att man kan sänka tröskeln för att implementera uppgiftsorienterade dialogsystem på domäner med otillräcklig träningsdata. Vi implementerar Hybrid Code Networks och tränar implementationen på det insamlade datasetet, och uppnår goda resultat.

Ellipsis is a phenomenon whereby constituents which are normally obligatory in the grammar are omitted in actual discourse. It is found in all types of discourse, from everyday conversation to poetry. The omitted constituents can range from one word to an entire clause, and recovery of the ellipted item depends sometimes on the linguistic and sometimes on the non-linguistic context. From a practical point of view, the contribution of ellipsis in the context is twofold. First, it is one of several important means of achieving cohesion in a text. Secondly, ellipsis contributes to communicative appropriateness determined by the type of linguistic activity (e.g., narrative, casual conversation), the mode of communication (e.g., written / spoken) and the relationship between participants. The aim of this research is to provide a description of the functions of elliptical utterances – textual and interpersonal – in English and Japanese, based on a cross-linguistic analysis of dialogues in the English and Japanese map task corpora. In order to analyse ellipsis in relation to its two key functions, elliptical clauses in the map task dialogues were examined. I discuss how ellipsis is used to realise cohesion in the map task dialogues. The findings challenge the well-known claim that topics are established by full noun phrases, which are subsequently realised by pronouns (English) and null pronouns (Japanese). Rather, the results suggest that full noun phrases are used for topic continuity in both languages. Constituents which are ellipted in an utterance are identified and related to the moves types which the utterance realises within the exchange structure. The ellipted elements will be categorised according to the constituent types (Subject, Finite, Predicator, Complement and Adjunct), using the systemic functional approach. This analysis reveals that whereas in the English dialogues the most common types of ellipsis are that of Subject and Finite elements, in the Japanese dialogues the most common type is that of Subject. Types of ellipsis are also correlated with speech acts in the dialogues. The relation between types of ellipsis and particular speech acts associated with them is strikingly similar in the English and Japanese dialogues, despite the notable difference in grammar and pragmatics between the two languages. This analysis also shows how these types of ellipsis are associated with interpersonal effects in particular speech acts: ellipsis of Subject and Finite can contribute to a sharp contrast in the question and answer sequence, while Subject ellipsis in Japanese can contribute to modifying the command-like force in giving instructions. These effects can be summed up as epistemic and deontic modality respectively. Ultimately, it is argued that some types of ellipsis can serve as modality expressions. Additionally, in comparison to the way of realising the speech act of giving instructions in the English dialogues, it emerges that the Japanese speakers exploit ellipsis, which seems to be associated with lowering the degree of the speaker’s commitment to the proposition. As implications for pedagogical settings, I present pedagogical descriptions of ellipsis for Japanese learners of English and English learners of Japanese. Since the description is for specific learners, the approach which takes the difference in grammar and pragmatics between the two languages is made possible. Although descriptions state some detailed facts of ellipsis in English and Japanese, primarily highlighted is the importance of raising awareness of elliptical forms for particular functions in particular contexts. As ellipsis is a product of forms, functions and contexts, it is a most remarkable feature of spoken language. Spoken language is claimed by some researchers to show similar linguistic features among languages because of the restrictions inherent in the medium on communication. In the form of pedagogical description, I show the similarities and differences in ellipsis which derive from the grammar and pragmatics of each language, which are observed in the preceding linguistic research. Through the presentation of the findings which are modified for learners, learners will know how languages show convergence and divergence cross-linguistically.

A longstanding objective of human-computer interaction research is to develop better dialog systems for end users. The subset of user modelling research specifically, aims to provide dialog researchers with models of user behaviour to aid with the design and improvement of dialog systems. Where dialog systems are commercially deployed, they are often to be used by a vast number of users, where sub-optimal performance could lead to an immediate financial loss for the service provider, and even user alienation. Thus, there is a strong incentive to make dialog systems as functional as possible immediately, and crucially prior to their release to the public. Models of user behaviour fill this gap, by simulating the role of human users in the lab, without the losses associated with sub-optimal system performance. User models can also tremendously aid design decisions, by serving as tools for exploratory analysis of real user behaviour, prior to designing dialog software. User modelling is the central problem of this thesis. We focus on a particular kind of dialogs termed task-oriented dialogs (those centred around solving an explicit task) because they represent the frontier of current dialog research and commercial deployment. Users taking part in these dialogs behave according to a set of user goals, which specify what they wish to accomplish from the interaction, and tend to exhibit variability of behaviour given the same set of goals. Our objective is to capture and reproduce (at the semantic utterance level) the range of behaviour that users exhibit while being consistent with their goals. We approach the problem as an instance of generative probabilistic modelling, with explicit user goals, and induced entirely from data. We argue that doing so has numerous practical and theoretical benefits over previous approaches to user modelling which have either lacked a model of user goals, or have been not been driven by real dialog data. A principal problem with user modelling development thus far has been the difficulty in evaluation. We demonstrate how treating user models as probabilistic models alleviates some of these problems through the ability to leverage a whole raft of techniques and insights from machine learning for evaluation. We demonstrate the efficacy of our approach by applying it to two different kinds of task-oriented dialog domains, which exhibit two different sub-problems encountered in real dialog corpora. The first are informational (or slot-filling) domains, specifically those concerning flight and bus route information. In slot-filling domains, user goals take categorical values which allow multiple surface realisations, and are corrupted by speech recognition errors. We address this issue by adopting a topic model representation of user goals which allows us capture both synonymy and phonetic confusability in a unified model. We first evaluate our model intrinsically using held-out probability and perplexity, and demonstrate substantial gains over an alternative string-goal representations, and over a non-goal-directed model. We then show in an extrinsic evaluation that features derived from our model lead to substantial improvements over strong baseline in the task of discriminating between real dialogs (consistent dialogs) and dialogs comprised of real turns sampled from different dialogs (inconsistent dialogs). We then move on to a spatial navigational domain in which user goals are spatial trajectories across a landscape. The disparity between the representation of spatial routes as raw pixel coordinates and their grounding as semantic utterances creates an interesting challenge compared to conventional slot-filling domains. We derive a feature-based representation of spatial goals which facilitates reasoning and admits generalisation to new routes not encountered at training time. The probabilistic formulation of our model allows us to capture variability of behaviour given the same underlying goal, a property frequently exhibited by human users in the domain. We first evaluate intrinsically using held-out probability and perplexity, and find a substantial reduction in uncertainty brought by our spatial representation. We further evaluate extrinsically in a human judgement task and find that our model’s behaviour does not differ significantly from the behaviour of real users. We conclude by sketching two novel ideas for future work: the first is to deploy the user models as transition functions for MDP-based dialog managers; the second is to use the models as a means of restricting the search space for optimal policies, by treating optimal behaviour as a subset of the (distributions over) plausible behaviour which we have induced.

Philosophiae Doctor - PhD
Between 1974 and 1982 the MOE introduced in two phases the diversification programme [SSCDP] which sought to establish practical subjects in the secondary school curriculum. This study examines the sustainability of implementation efforts beyond project expiry. It was hypothesised that SSCDP is not working as originally intended. The broad research problem was framed thus: What implementation response arises from an open-ended innovation policy? Subsidiary questions are: 1. How far have the policy-makers communicated the meaning of SSCDP and what factors account for mismatches between policy intentions and innovation practice? 2. What is the response of Project schools and what factors explain variation in response? 3. What is their significance for the sustainability of SSCDP? The analysis draws key concepts from the innovation literature on models and strategies of planned change; relationships in the implementation hierarchy; determinants of and orientations to the implementation process. Centred around qualitative research methods, the investigation utilises data from project documents, semi-structured interviews and from observations during school visits. Findings show an overall low level of implementation that varies among project schools. This is attributed to: Poor interpretation of SSCDP goals; Deficiencies in the implementation management; Idiosyncratic school behaviours. The study concludes that the 'practitioner-policy-maker' discrepancy is significant, hence the gap between policy intents and innovation practice. The gap is not regarded so much as an ultimate failure of the programme but as a necessary condition that allows for mutual adaptation between the innovation and its setting. This is reflected in the varied patterns of implementation response, classified as the: faithful; negotiators; selective adaptors; expansionists; and reductionist. As a policy-oriented study aiming at providing an 'improvement value', the findings lead to a proposal of improvements in the strategies of managing change in three areas: shifting focus from an adoption to an implementation perspective. Recognising implementation as a process dependent on a mutual linkage relationship among participants. Recognising schools as important bearers of change. These three are crucial factors in the implementation-sustainability relationship.

Traditional development of goal-oriented conversational agents typically require a lot of domain-specific handcrafting, which precludes scaling up to different domains; end-to-end systems would escape this limitation because they can be trained directly from dialogues. The very promising success recently obtained in end-to-end chatbots development could carry over to goal-oriented settings: applying deep learning models for building robust and scalable goal-oriented dialog systems directly from corpora of conversations is a challenging task and an open research area. For this reason, I decided that it would have been more relevant in the context of a master's thesis to experiment and get acquainted with these new promising methodologies - although not yet ready for production - rather than investing time in hand-crafting dialogue rules for a domain-specific solution. My thesis work had the following macro objectives: (i) investigate the latest research works concerning goal-oriented conversational agents development; (ii) choose a reference study, understand it and implement it with an appropriate technology; (iii) apply what learnt to a particular domain of interest. As a reference framework I chose the end-to-end memory networks (MemN2N) (Sukhbaatar et al., 2015) because it has proven to be particularly promising and has been used as a baseline for many recent works. Not having real dialogues available for training though, I took care of synthetically generating a corpora of conversations, taking a cue from the Dialog bAbI dataset for restaurant reservations (Bordes et al., 2016) and adapting it to the new domain of interest of risk awareness. Finally, I built a simple prototype which exploited the pre-trained dialog model in order to advise users about risk through an anthropomorphic talking avatar interface.

碩士
國立交通大學
土木工程系所
106
In this study, an attempt was made to develop a conversational bot for assisting people evacuate in case of fire. Whether civilians could adequately evacuate in the event of a fire depends on the dynamics between their behavior and the characteristics of buildings and burning fires. When a fire occurs, people have limited knowledge about fire, which makes it difficult to assess the scenario in time while the fire is spreading. However, it is of paramount importance whether people can effectively escape within a limited time and space. In this study, an evacuation decision-making model is proposed, based on literature review on evacuation simulations based on pedestrian flow correlation theory the experience of fire survival experts. Moreover, the interaction between the evacuation decision-making model, dialogue-based system, decision tree recommendation method, and the applicability and limitation of the human-computer interaction mode are discussed. Under the limitation mode of human-computer interaction, this study used the time of evacuation as the main limitation of successful escape. Using Natural Language Processing and Conversational User Interfaces, Messenger was used as a platform to explore the impact of dialogue-type robots on fire evacuee patterns in the event of fire. It is hoped that the results of this study will be useful for establishing or reviewing guidelines or standard procedures for fire evacuation and fire drills.

Cette thèse par article se compose de quatre articles qui contribuent au domaine de l’apprentissage profond, en particulier dans la compréhension et l’apprentissage des ap- proches neuronales des systèmes de dialogue. Le premier article fait un pas vers la compréhension si les architectures de dialogue neuronal couramment utilisées capturent efficacement les informations présentes dans l’historique des conversations. Grâce à une série d’expériences de perturbation sur des ensembles de données de dialogue populaires, nous constatons que les architectures de dialogue neuronal couramment utilisées comme les modèles seq2seq récurrents et basés sur des transformateurs sont rarement sensibles à la plupart des perturbations du contexte d’entrée telles que les énoncés manquants ou réorganisés, les mots mélangés, etc. Le deuxième article propose d’améliorer la qualité de génération de réponse dans les systèmes de dialogue de domaine ouvert en modélisant conjointement les énoncés avec les attributs de dialogue de chaque énoncé. Les attributs de dialogue d’un énoncé se réfèrent à des caractéristiques ou des aspects discrets associés à un énoncé comme les actes de dialogue, le sentiment, l’émotion, l’identité du locuteur, la personnalité du locuteur, etc. Le troisième article présente un moyen simple et économique de collecter des ensembles de données à grande échelle pour modéliser des systèmes de dialogue orientés tâche. Cette approche évite l’exigence d’un schéma d’annotation d’arguments complexes. La version initiale de l’ensemble de données comprend 13 215 dialogues basés sur des tâches comprenant six domaines et environ 8 000 entités nommées uniques, presque 8 fois plus que l’ensemble de données MultiWOZ populaire.
This thesis by article consists of four articles which contribute to the field of deep learning, specifically in understanding and learning neural approaches to dialog systems. The first article takes a step towards understanding if commonly used neural dialog architectures effectively capture the information present in the conversation history. Through a series of perturbation experiments on popular dialog datasets, wefindthatcommonly used neural dialog architectures like recurrent and transformer-based seq2seq models are rarely sensitive to most input context perturbations such as missing or reordering utterances, shuffling words, etc. The second article introduces a simple and cost-effective way to collect large scale datasets for modeling task-oriented dialog systems. This approach avoids the requirement of a com-plex argument annotation schema. The initial release of the dataset includes 13,215 task-based dialogs comprising six domains and around 8k unique named entities, almost 8 times more than the popular MultiWOZ dataset. The third article proposes to improve response generation quality in open domain dialog systems by jointly modeling the utterances with the dialog attributes of each utterance. Dialog attributes of an utterance refer to discrete features or aspects associated with an utterance like dialog-acts, sentiment, emotion, speaker identity, speaker personality, etc. The final article introduces an embedding-free method to compute word representations on-the-fly. This approach significantly reduces the memory footprint which facilitates de-ployment in on-device (memory constraints) devices. Apart from being independent of the vocabulary size, we find this approach to be inherently resilient to common misspellings.