Modelling dialogue management as a reinforcement learning task enables a system to learn to act optimally by maximising a reward function. This reward function is designed to induce the system behaviour required for goal-oriented applications, which usually means fulfilling the user’s goal as efficiently as possible. However, in real-world spoken dialogue systems, the reward is hard to measure, because the goal of the conversation is often known only to the user. Certainly, the system can ask the user if the goal has been satisfied, but this can be intrusive. Furthermore, in practice, the reliability of the user’s response has been found to be highly variable. In addition, due to the sparsity of the reward signal and the large search space, reinforcement learning-based dialogue policy optimisation is often slow. This thesis presents several approaches to address these problems. To better evaluate a dialogue for policy optimisation, two methods are proposed. First, a recurrent neural network-based predictor pre-trained from off-line data is proposed to estimate task success during subsequent on-line dialogue policy learning to avoid noisy user ratings and problems related to not knowing the user’s goal. Second, an on-line learning framework is described where a dialogue policy is jointly trained alongside a reward function modelled as a Gaussian process with active learning. This mitigates the noisiness of user ratings and minimises user intrusion. It is shown that both off-line and on-line methods achieve practical policy learning in real-world applications, while the latter provides a more general joint learning system directly from users. To enhance the policy learning speed, the use of reward shaping is explored and shown to be effective and complementary to the core policy learning algorithm. Furthermore, as deep reinforcement learning methods have the potential to scale to very large tasks, this thesis also investigates the application to dialogue systems. Two sample-efficient algorithms, trust region actor-critic with experience replay (TRACER) and episodic natural actor-critic with experience replay (eNACER), are introduced. In addition, a corpus of demonstration data is utilised to pre-train the models prior to on-line reinforcement learning to handle the cold start problem. Combining these two methods, a practical approach is demonstrated to effectively learn deep reinforcement learning-based dialogue policies in a task-oriented information seeking domain. Overall, this thesis provides solutions which allow truly on-line and continuous policy learning in spoken dialogue systems.

Spoken dialogue systems provide a natural conversational interface to computer applications. In recent years, the substantial improvements in the performance of speech recognition engines have helped shift the research focus to the next component of the dialogue system pipeline: the one in charge of language understanding. The role of this module is to translate user inputs into accurate representations of the user goal in the form that can be used by the system to interact with the underlying application. The challenges include the modelling of linguistic variation, speech recognition errors and the effects of dialogue context. Recently, the focus of language understanding research has moved to making use of word embeddings induced from large textual corpora using unsupervised methods. The work presented in this thesis demonstrates how these methods can be adapted to overcome the limitations of language understanding pipelines currently used in spoken dialogue systems. The thesis starts with a discussion of the pros and cons of language understanding models used in modern dialogue systems. Most models in use today are based on the delexicalisation paradigm, where exact string matching supplemented by a list of domain-specific rephrasings is used to recognise users' intents and update the system's internal belief state. This is followed by an attempt to use pretrained word vector collections to automatically induce domain-specific semantic lexicons, which are typically hand-crafted to handle lexical variation and account for a plethora of system failure modes. The results highlight the deficiencies of distributional word vectors which must be overcome to make them useful for downstream language understanding models. The thesis next shifts focus to overcoming the language understanding models' dependency on semantic lexicons. To achieve that, the proposed Neural Belief Tracking (NBT) model forsakes the use of standard one-hot n-gram representations used in Natural Language Processing in favour of distributed representations of user utterances, dialogue context and domain ontologies. The NBT model makes use of external lexical knowledge embedded in semantically specialised word vectors, obviating the need for domain-specific semantic lexicons. Subsequent work focuses on semantic specialisation, presenting an efficient method for injecting external lexical knowledge into word vector spaces. The proposed Attract-Repel algorithm boosts the semantic content of existing word vectors while simultaneously inducing high-quality cross-lingual word vector spaces. Finally, NBT models powered by specialised cross-lingual word vectors are used to train multilingual belief tracking models. These models operate across many languages at once, providing an efficient method for bootstrapping language understanding models for lower-resource languages with limited training data.

Made available in DSpace on 2016-08-17T14:53:29Z (GMT). No. of bitstreams: 1 DISSERTACAO Arikleyton de Oliveira Ferreira.pdf: 1952997 bytes, checksum: 4c3733cd1aefc31e6f18a8068828d271 (MD5) Previous issue date: 2014-08-04 / Conselho Nacional de Desenvolvimento Científico e Tecnológico / Applications for mobile computing environments usually have several accessibility limitations due to the dependency on the interaction with the user through the device display, which hinders its use to people who have limitations to read, write (type) and/or have little fluency in the use of technology. In this master thesis we propose a middleware that provides support for developing mobile applications with accessibility features through spoken dialogue systems. These systems are able to hold a conversation with the user, providing a natural interaction interface that does not require prior learning. Thus, mobile applications can use the middleware to provide accessibility to the user that overcomes the physical or visual contact needs. The proposed middleware was developed in the context of the MobileHealthNet project, where it will help mobile applications with focus in the health domain to reach users with different profiles, with particular attention to underserved and remote communities. To perform the middleware evaluation, we used a case study based on a mobile application for evaluating the health condition of patients with atrial fibrillation. The evaluation involved 10 individuals, and the results obtained were very positive. / Aplicações para ambientes computacionais móveis usualmente apresentam diversas limitações de acessibilidade por dependerem da interação com o usuário através da tela dos dispositivos móveis, o que dificulta seu uso às pessoas que possuem limitações para ler, escrever (digitar) e que tenham pouca fluência no uso de tecnologias. Neste trabalho de mestrado propomos um middleware que fornece suporte ao desenvolvimento de aplicações móveis com recurso de acessibilidade através do diálogo falado. Essa modalidade de acesso é capaz de manter uma conversa com o usuário, proporcionando uma interface de interação natural que não requer prévio aprendizado. Assim, aplicações móveis podem utilizar o middleware para proporcionar acessibilidade ao usuário que supera a necessidade do contato físico ou visual, pois eles podem apenas dialogar entre si. O middleware proposto está inserido no contexto do projeto MobileHealthNet, onde auxiliará aplicações móveis focadas ao domínio da saúde a atingir usuários com diferentes perfis, com especial atenção a moradores de comunidades carentes e distantes. No processo de avalidação do middleware proposto foi utilizado um estudo de caso de uma aplicação dedicada a acompanhar o estado de saúde de pacientes portadores de fibrilação atrial, realizando-se uma avaliação com 10 sujeitos na qual obteve-se resultados bastante positivos.

The quality of automatic speech recognition has increased dramatically the last few years, but the performance for low and middle resource languages such as Swedish is still far from optimal. In this project a language model trained on large written corpora called KB-BERT is utilized to improve the quality of transcriptions for Swedish. The large language model is inserted as a repairing module after the automatic speech recognition, aiming to repair the original output into a transcription more closely resembling the ground truth by using a sequence to sequence translating approach. Two automatic speech recognition models are used to transcribe the speech, one of the models are developed in this project using the Kaldi framework, the other model is Microsoft’s Azure Speech to text platform. The performance of the translator is evaluated with four different datasets, three consisting of read speech and one of spontaneous speech. The spontaneous speech and one of the read datasets include both native and non-native speakers. The performance is measured by three different metrics, word error rate, a weighted word error rate and a semantic similarity. The repairs improve the transcriptions of two of the read speech datasets significantly, decreasing the word error rate from 13.69% to 3.05% and from 36.23% to 21.17%. The repairs improve the word error rate from 44.38% to 44.06% on the data with spontaneous speech, and fail on the last read dataset, instead increasing the word error rate. The lower performance on the latter is likely due to lack of data. / Automatisk taligenkänning har förbättrats de senaste åren, men för små språk såsom svenska är prestandan fortfarande långt ifrån optimal. Det här projektet använder KB-BERT, en neural språkmodell tränad på stora mängder skriven text, för att förbättra kvalitén på transkriptioner av svenskt tal. Transkriptionerna kommer från två olika taligenkänningsmodeller, dels en utvecklad i det här projektet med hjälp av mjukvarubiblioteket Kaldi, dels Microsoft Azures plattform för tal till text. Transkriptionerna repareras med hjälp av en sequence-to-sequence översättningsmodell, och KB-BERT används för att initiera modellen. Översättningen sker från den urpsrungliga transkriptionen från en av tal-till-text-modellerna till en transkription som är mer lik den korrekta, faktiska transkriptionen. Kvalitéen på reparationerna evalueras med tre olika metriker, på fyra olika dataset. Tre av dataseten är läst tal och det fjärde spontant, och det spontana talet samt ett av de lästa dataseten kommer både från talare som har svenska som modersmål, och talare som har det som andraspråk. De tre metrikerna är word error rate, en viktad word error rate, samt ett mått på semantisk likhet. Reparationerna förbättrar transkriptionerna från två av de lästa dataseten markant, och sänker word error rate från 13.69% till 3.05% och från 36.23% till 21.17%. På det spontana talet sänks word error rate från 44.38% till 44.06%. Reparationerna misslyckas på det fjärde datasetet, troligen på grund av dess lilla storlek.

No description available.

Maintaining cognitive health is often a pressing concern for aging adults, and given the world's shifting age demographics, it is impractical to assume that older adults will be able to rely on individualized human support for doing so. Recently, interest has turned toward technology as an alternative. Companion robots offer an attractive vehicle for facilitating cognitive exercise, but the language technologies guiding their interactions are still nascent; in elder-focused human-robot systems proposed to date, interactions have been limited to motion or buttons and canned speech. The incapacity of these systems to autonomously participate in conversational discourse limits their ability to engage users at a cognitively meaningful level. I addressed this limitation by developing a platform for human-robot book discussions, designed to promote cognitive exercise by encouraging users to consider the authors' underlying intentions in employing creative metaphors. The choice of book discussions as the backdrop for these conversations has an empirical basis in neuro- and social science research that has found that reading often, even in late adulthood, has been correlated with a decreased likelihood to exhibit symptoms of cognitive decline. The more targeted focus on novel metaphors within those conversations stems from prior work showing that processing novel metaphors is a cognitively challenging task, for young adults and even more so in older adults with and without dementia. A central contribution arising from the work was the creation of the first computational method for modelling metaphor novelty in word pairs. I show that the method outperforms baseline strategies as well as a standard metaphor detection approach, and additionally discover that incorporating a sentence-based classifier as a preliminary filtering step when applying the model to new books results in a better final set of scored word pairs. I trained and evaluated my methods using new, large corpora from two sources, and release those corpora to the research community. In developing the corpora, an additional contribution was the discovery that training a supervised regression model to automatically aggregate the crowdsourced annotations outperformed existing label aggregation strategies. Finally, I show that automatically-generated questions adhering to the Questioning the Author strategy are comparable to human-generated questions in terms of naturalness, sensibility, and question depth; the automatically-generated questions score slightly higher than human-generated questions in terms of clarity. I close by presenting findings from a usability evaluation in which users engaged in thirty-minute book discussions with a robot using the platform, showing that users find the platform to be likeable and engaging.

[pt] Spoken Dialogue Systems (SDS s) são sistemas baseados em computadores desenvolvidos para fornecerem informações e realizar tarefas utilizando o diálogo como forma de interação. Eles são capazes de reconhecimento de voz, interpretação, gerenciamento de diálogo e são capazes de ter uma voz como saída de dados, tentando reproduzir uma interação natural falada entre um usuário humano e um sistema. SDS s provém diferentes serviços, todos através de linguagem falada com um sistema. Mesmo com todo o desenvolvimento nesta área, há escassez de informações sobre como avaliar a qualidade de tais sistemas com o propósito de otimização do mesmo. Com dois destes sistemas, BoRIS e INSPIRE, usados para reservas de restaurantes e gerenciamento de casas inteligentes, diversos experimentos foram conduzidos no passado, onde tais sistemas foram utilizados para resolver tarefas específicas. Os participantes avaliaram a qualidade do sistema em uma série de questões. Além disso, todas as interações foram gravadas e anotadas por um especialista.O desenvolvimento de métodos para avaliação de performance é um tópico aberto de pesquisa na área de SDS s. Seguindo a idéia do modelo PARADISE (PARAdigm for DIalogue System Evaluation – desenvolvido pro Walker e colaboradores na AT&T em 1998), diversos experimentos foram conduzidos para desenvolver modelos de previsão de performance de sistemas de reconhecimento de voz e linguagem falada. O objetivo desta dissertação de mestrado é desenvolver modelos que permitam a previsão de dimensões de qualidade percebidas por um usuário humano, baseado em parâmetros instrumentalmente mensuráveis utilizando dados coletados nos experimentos realizados com os sistemas BoRIS e INSPIRE , dois sistemas de reconhecimento de voz (o primeiro para busca de restaurantes e o segundo para Smart Homes). Diferentes algoritmos serão utilizados para análise (Regressão linear, Árvores de Regressão, Árvores de Classificação e Redes Neurais) e para cada um dos algoritmos, uma ferramenta diferente será programada em MATLAB, para poder servir de base para análise de experimentos futuros, sendo facilmente modificado para sistemas e parâmetros novos em estudos subsequentes.A idéia principal é desenvolver ferramentas que possam ajudar na otimização de um SDS sem o envolvimento direto de um usuário humano ou servir de ferramenta para estudos futuros na área. / [en] Spoken Dialogue Systems (SDS s) are computer-based systems developed to provide information and carry out tasks using speech as the interaction mode. They are capable of speech recognition, interpretation, management of dialogue and have speech output capabilities, trying to reproduce a more or less natural spoken interaction between a human user and the system. SDS s provide several different services, all through spoken language. Even with all this development, there is scarcity of information on ways to assess and evaluate the quality of such systems with the purpose of optimization. With two of these SDS s ,BoRIS and INSPIRE, (used for Restaurant Booking Services and Smart Home Systems), extensive experiments were conducted in the past, where the systems were used to resolve specific tasks. The evaluators rated the quality of the system on a multitude of scales. In addition to that, the interactions were recorded and annotated by an expert. The development of methods for performance evaluation is an open research issue in this area of SDS s. Following the idea of the PARADISE model (PARAdigm for DIalogue System Evaluation model, the most well-known model for this purpose (developed by Walker and co-workers at AT&T in 1998), several experiments were conducted to develop predictive models of spoken dialogue performance. The objective of this dissertation is to develop and assess models which allow the prediction of quality dimensions as perceived by the human user, based on instrumentally measurable variables using all the collected data from the BoRIS and INSPIRE systems. Different types of algorithms will be compared to their prediction performance and to how generic they are. Four different approaches will be used for these analyses: Linear regression, Regression Trees, Classification Trees and Neural Networks. For each of these methods, a different tool will be programmed using MATLAB, that can carry out all experiments from this work and be easily modified for new experiments with data from new systems or new variables on future studies. All the used MATLAB programs will be made available on the attached CD with an operation manual for future users as well as a guide to modify the existing programs to work on new data. The main idea is to develop tools that would help on the optimization of a spoken dialogue system without a direct involvement of the human user or serve as tools for future studies in this area.

Cette thèse présente une série de mesures prises pour étudier l’apprentissage de représentations (par exemple, l’apprentissage profond) afin de mettre en place des systèmes de dialogue et des agents de conversation virtuels. La thèse est divisée en deux parties générales. La première partie de la thèse examine l’apprentissage des représentations pour les modèles de dialogue génératifs. Conditionnés sur une séquence de tours à partir d’un dialogue textuel, ces modèles ont la tâche de générer la prochaine réponse appropriée dans le dialogue. Cette partie de la thèse porte sur les modèles séquence-à-séquence, qui est une classe de réseaux de neurones profonds génératifs. Premièrement, nous proposons un modèle d’encodeur-décodeur récurrent hiérarchique ("Hierarchical Recurrent Encoder-Decoder"), qui est une extension du modèle séquence-à-séquence traditionnel incorporant la structure des tours de dialogue. Deuxièmement, nous proposons un modèle de réseau de neurones récurrents multi-résolution ("Multiresolution Recurrent Neural Network"), qui est un modèle empilé séquence-à-séquence avec une représentation stochastique intermédiaire (une "représentation grossière") capturant le contenu sémantique abstrait communiqué entre les locuteurs. Troisièmement, nous proposons le modèle d’encodeur-décodeur récurrent avec variables latentes ("Latent Variable Recurrent Encoder-Decoder"), qui suivent une distribution normale. Les variables latentes sont destinées à la modélisation de l’ambiguïté et l’incertitude qui apparaissent naturellement dans la communication humaine. Les trois modèles sont évalués et comparés sur deux tâches de génération de réponse de dialogue: une tâche de génération de réponses sur la plateforme Twitter et une tâche de génération de réponses de l’assistance technique ("Ubuntu technical response generation task"). La deuxième partie de la thèse étudie l’apprentissage de représentations pour un système de dialogue utilisant l’apprentissage par renforcement dans un contexte réel. Cette partie porte plus particulièrement sur le système "Milabot" construit par l’Institut québécois d’intelligence artificielle (Mila) pour le concours "Amazon Alexa Prize 2017". Le Milabot est un système capable de bavarder avec des humains sur des sujets populaires à la fois par la parole et par le texte. Le système consiste d’un ensemble de modèles de récupération et de génération en langage naturel, comprenant des modèles basés sur des références, des modèles de sac de mots et des variantes des modèles décrits ci-dessus. Cette partie de la thèse se concentre sur la tâche de sélection de réponse. À partir d’une séquence de tours de dialogues et d’un ensemble des réponses possibles, le système doit sélectionner une réponse appropriée à fournir à l’utilisateur. Une approche d’apprentissage par renforcement basée sur un modèle appelée "Bottleneck Simulator" est proposée pour sélectionner le candidat approprié pour la réponse. Le "Bottleneck Simulator" apprend un modèle approximatif de l’environnement en se basant sur les trajectoires de dialogue observées et le "crowdsourcing", tout en utilisant un état abstrait représentant la sémantique du discours. Le modèle d’environnement est ensuite utilisé pour apprendre une stratégie d’apprentissage du renforcement par le biais de simulations. La stratégie apprise a été évaluée et comparée à des approches concurrentes via des tests A / B avec des utilisateurs réel, où elle démontre d’excellente performance. / This thesis presents a series of steps taken towards investigating representation learning (e.g. deep learning) for building dialogue systems and conversational agents. The thesis is split into two general parts. The first part of the thesis investigates representation learning for generative dialogue models. Conditioned on a sequence of turns from a text-based dialogue, these models are tasked with generating the next, appropriate response in the dialogue. This part of the thesis focuses on sequence-to-sequence models, a class of generative deep neural networks. First, we propose the Hierarchical Recurrent Encoder-Decoder model, which is an extension of the vanilla sequence-to sequence model incorporating the turn-taking structure of dialogues. Second, we propose the Multiresolution Recurrent Neural Network model, which is a stacked sequence-to-sequence model with an intermediate, stochastic representation (a "coarse representation") capturing the abstract semantic content communicated between the dialogue speakers. Third, we propose the Latent Variable Recurrent Encoder-Decoder model, which is a variant of the Hierarchical Recurrent Encoder-Decoder model with latent, stochastic normally-distributed variables. The latent, stochastic variables are intended for modelling the ambiguity and uncertainty occurring naturally in human language communication. The three models are evaluated and compared on two dialogue response generation tasks: a Twitter response generation task and the Ubuntu technical response generation task. The second part of the thesis investigates representation learning for a real-world reinforcement learning dialogue system. Specifically, this part focuses on the Milabot system built by the Quebec Artificial Intelligence Institute (Mila) for the Amazon Alexa Prize 2017 competition. Milabot is a system capable of conversing with humans on popular small talk topics through both speech and text. The system consists of an ensemble of natural language retrieval and generation models, including template-based models, bag-of-words models, and variants of the models discussed in the first part of the thesis. This part of the thesis focuses on the response selection task. Given a sequence of turns from a dialogue and a set of candidate responses, the system must select an appropriate response to give the user. A model-based reinforcement learning approach, called the Bottleneck Simulator, is proposed for selecting the appropriate candidate response. The Bottleneck Simulator learns an approximate model of the environment based on observed dialogue trajectories and human crowdsourcing, while utilizing an abstract (bottleneck) state representing high-level discourse semantics. The learned environment model is then employed to learn a reinforcement learning policy through rollout simulations. The learned policy has been evaluated and compared to competing approaches through A/B testing with real-world users, where it was found to yield excellent performance.

Chat-bots are computer programs designed to conduct textual or audible conversations with a single user. The job of a chat-bot is to be able to find the best response for any request the user issues. The best response is considered to answer the question and contain relevant information while following grammatical and lexical rules. Modern chat-bots often have trouble accomplishing all these tasks. State-of-the-art approaches, such as deep learning, and large datasets help chat-bots tackle this problem better. While there is a number of different approaches that can be applied for different kind of bots, datasets of suitable size are not always available. In this work, we introduce and evaluate a method of expanding the size of datasets. This will allow chat-bots, in combination with a good learning algorithm, to achieve higher precision while handling their tasks. The expansion method uses the continual learning approach that allows the bot to expand its own dataset while holding conversations with its users. In this work we test continual learning with IBM Watson Assistant chat-bot as well as a custom case study chat-bot implementation. We conduct the testing using a smaller and a larger datasets to find out if continual learning stays effective as the dataset size increases. The results show that the more conversations the chat-bot holds, the better it gets at guessing the intent of the user. They also show that continual learning works well for larger and smaller datasets, but the effect depends on the specifics of the chat-bot implementation. While continual learning makes good results better, it also turns bad results into worse ones, thus the chat-bot should be manually calibrated should the precision of the original results, measured before the expansion, decrease.