Following the success of Natural Language Processing (NLP) transformers pretrained via self-supervised learning, similar models have been proposed recently for speech processing such as Wav2Vec2, HuBERT and UniSpeech-SAT. An interesting yet unexplored area of application of these models is Spoken Dialogue Systems, where the users' audio signals are typically just mapped to word-level features derived from an Automatic Speech Recogniser (ASR), and then processed using NLP techniques to generate system responses. This paper reports a comprehensive comparison of dialogue policies trained using ASR-based transcriptions and extended with the aforementioned audio processing transformers in the DSTC2 task. Whilst our dialogue policies are trained with supervised and policy-based deep reinforcement learning, they are assessed using both automatic task completion metrics and a human evaluation. Our results reveal that using audio embeddings is more beneficial than detrimental in most of our trained dialogue policies, and that the benefits are stronger for supervised learning than reinforcement learning.

Learning robotic tasks in the real world is still highly challenging and effective practical solutions remain to be found. Traditional methods used in this area are imitation learning and reinforcement learning, but they both have limitations when applied to real robots. Combining reinforcement learning with pre-collected demonstrations is a promising approach that can help in learning control policies to solve robotic tasks. In this paper, we propose an algorithm that uses novel techniques to leverage offline expert data using offline and online training to obtain faster convergence and improved performance. The proposed algorithm (AWET) weights the critic losses with a novel agent advantage weight to improve over the expert data. In addition, AWET makes use of an automatic early termination technique to stop and discard policy rollouts that are not similar to expert trajectories-to prevent drifting far from the expert data. In an ablation study, AWET showed improved and promising performance when compared to state-of-the-art baselines on four standard robotic tasks.

Neural transformer architectures have gained a lot of interest for text-based dialogue management in the last few years. They have shown high learning capabilities for open domain dialogue with huge amounts of data and also for domain adaptation in task-oriented setups. But the potential benefits of exploiting the users' audio signal have rarely been explored in such frameworks. In this work, we combine text dialogue history representations generated by a GPT-2 model with audio embeddings obtained by the recently released Wav2Vec2 transformer model. We jointly fine-tune these models to learn dialogue policies via supervised learning and two policy gradient-based reinforcement learning algorithms. Our experimental results, using the DSTC2 dataset and a simulated user model capable of sampling audio turns, reveal that audio embeddings lead to overall higher task success (than without using audio embeddings) with statistically significant results across evaluation metrics and training algorithms.

Robots learning a new manipulation task from a small amount of demonstrations are increasingly demanded in different workspaces. A classifier model assessing the quality of actions can predict the successful completion of a task, which can be used by intelligent agents for action-selection. This paper presents a novel classifier that learns to classify task completion only from a few demonstrations. We carry out a comprehensive comparison of different neural classifiers, e.g. fully connected-based, fully convolutional-based, sequence2sequence-based, and domain adaptation-based classification. We also present a new dataset including five robot manipulation tasks, which is publicly available. We compared the performances of our novel classifier and the existing models using our dataset and the MIME dataset. The results suggest domain adaptation and timing-based features improve success prediction. Our novel model, i.e. fully convolutional neural network with domain adaptation and timing features, achieves an average classification accuracy of 97.3% and 95.5% across tasks in both datasets whereas state-of-the-art classifiers without domain adaptation and timing-features only achieve 82.4% and 90.3%, respectively.

Training chatbots using the reinforcement learning paradigm is challenging due to high-dimensional states, infinite action spaces and the difficulty in specifying the reward function. We address such problems using clustered actions instead of infinite actions, and a simple but promising reward function based on human-likeness scores derived from human-human dialogue data. We train Deep Reinforcement Learning (DRL) agents using chitchat data in raw text—without any manual annotations. Experimental results using different splits of training data report the following. First, that our agents learn reasonable policies in the environments they get familiarised with, but their performance drops substantially when they are exposed to a test set of unseen dialogues. Second, that the choice of sentence embedding size between 100 and 300 dimensions is not significantly different on test data. Third, that our proposed human-likeness rewards are reasonable for training chatbots as long as they use lengthy dialogue histories of ≥10 sentences.

In recent years many deep neural networks have been proposed to solve Reading Comprehension (RC) tasks.Most of these models suffer from reasoning over long documents and do not trivially generalize to cases where the answer is not present as a span in a given document.We present a novel neural-based architecture that is capable of extracting relevant regions based on a given question-document pair and generating a well-formed answer.To show the effectiveness of our architecture, we conducted several experiments on the recently proposed and challenging RC dataset 'Nar-rativeQA'.The proposed architecture outperforms state-of-the-art results (Tay et al., 2018) by 12.62% (ROUGE-L) relative improvement.

Standard deep reinforcement learning methods such as Deep Q-Networks (DQN) for multiple tasks (domains) face scalability problems due to large search spaces. This paper proposes a three-stage method for multi-domain dialogue policy learning-termed NDQN, and applies it to an information-seeking spoken dialogue system in the domains of restaurants and hotels. In this method, the first stage does multi-policy learning via a network of DQN agents; the second makes use of compact state representations by compressing raw inputs; and the third stage applies a pre-training phase for bootstraping the behaviour of agents in the network. Experimental results comparing DQN (baseline) versus NDQN (proposed) using simulations report that the proposed method exhibits better scalability and is promising for optimising the behaviour of multi-domain dialogue systems. An additional evaluation reports that the NDQN agents outperformed a K-Nearest Neighbour baseline in task success and dialogue length, yielding more efficient and successful dialogues.

Deep reinforcement learning dialogue systems are attractive because they can jointly learn their feature representations and policies without manual feature engineering. But its application is challenging due to slow learning. We propose a two-stage method for accelerating the induction of single or multi-domain dialogue policies. While the first stage reduces the amount of weight updates over time, the second stage uses very limited minibatches (of as much as two learning experiences) sampled from experience replay memories. The former frequently updates the weights of the neural nets at early stages of training, and decreases the amount of updates as training progresses by performing updates during exploration and by skipping updates during exploitation. The learning process is thus acceleratedthrough less weight updates in both stages. An empirical evaluation in three domains (restaurants, hotels and tv guide) confirms that the proposed method trains policies 5 times faster than a baseline without the proposed method. Our findings are useful for training larger-scale neural-based spoken dialogue systems.

Strategic conversational agents often need to trade resources with their opponent conversants -- and trading strategically can lead to better results. While rule-based or supervised agents can be used for such a purpose, here we explore a learning approach based on automatically labelled examples from human players for automatic trading in the game of Settlers of Catan. Our experiments are based on data collected from human players trading in text-based natural language. We compare the performance of Bayes Nets, Conditional Random Fields, and Random Forests on the task of ranking trading offers, trained from both manually labelled and automatically labelled data. Our experimental results show that our best agent trained on automatic labels outperformed its counterpart trained on manual labels (with moderate annotator agreement) in terms of (a) predicting human trading negotiations better, and (b) winning more games.

Abstract Natural Language Generation systems in interactive settings often face a multitude of choices, given that the communicative effect of each utterance they generate depends crucially on the interplay between its physical circumstances, addressee and interaction history. This is particularly true in interactive and situated settings. In this paper we present a novel approach forsituated Natural Language Generationin dialogue that is based onhierarchical reinforcement learningand learns the best utterance for a context by optimisation through trial and error. The model is trained from human–human corpus data and learns particularly to balance the trade-off betweenefficiencyanddetailin giving instructions: the user needs to be given sufficient information to execute their task, but without exceeding their cognitive load. We present results from simulation and a task-based human evaluation study comparing two different versions of hierarchical reinforcement learning: One operates using a hierarchy of policies with a large state space and local knowledge, and the other additionally shares knowledge across generation subtasks to enhance performance. Results show that sharing knowledge across subtasks achieves better performance than learning in isolation, leading to smoother and more successful interactions that are better perceived by human users.

Surface realisations typically depend on their target style and audience.A challenge in estimating a stylistic realiser from data is that humans vary significantly in their subjective perceptions of linguistic forms and styles, leading to almost no correlation between ratings of the same utterance.We address this problem in two steps.First, we estimate a mapping function between the linguistic features of a corpus of utterances and their human style ratings.Users are partitioned into clusters based on the similarity of their ratings, so that ratings for new utterances can be estimated, even for new, unknown users.In a second step, the estimated model is used to re-rank the outputs of a number of surface realisers to produce stylistically adaptive output.Results confirm that the generated styles are recognisable to human judges and that predictive models based on clusters of users lead to better rating predictions than models based on an average population of users.

Work on training semantic slot labellers for use in Natural Language Processing applications has typically either relied on large amounts of labelled input data, or has assumed entirely unlabelled inputs. The former technique tends to be costly to apply, while the latter is often not as accurate as its supervised counterpart. Here, we present a semi-supervised learning approach that automatically labels the semantic slots in a set of training data and aims to strike a balance between the dependence on labelled data and prediction accuracy. The essence of our algorithm is to cluster clauses based on a similarity function that combines lexical and semantic information. We present experiments that compare different similarity functions for both our semi-supervised setting and a fully unsupervised baseline. While semi-supervised learning expectedly outperforms unsupervised learning, our results show that (1) this effect can be observed based on very few training data instances and that increasing the size of the training data does not lead to better performance, and (2) that lexical and semantic information contribute differently in different domains so that clustering based on both types of information offers the best generalisation.

Helen Hastie, Marie-Aude Aufaure, Panos Alexopoulos, Hugues Bouchard, Catherine Breslin, Heriberto Cuayáhuitl, Nina Dethlefs, Milica Gašić, James Henderson, Oliver Lemon, Xingkun Liu, Peter Mika, Nesrine Ben Mustapha, Tim Potter, Verena Rieser, Blaise Thomson, Pirros Tsiakoulis, Yves Vanrompay, Boris Villazon-Terrazas, Majid Yazdani, Steve Young, Yanchao Yu. Proceedings of the 15th Annual Meeting of the Special Interest Group on Discourse and Dialogue (SIGDIAL). 2014.

Training a statistical surface realiser typically relies on labelled training data or parallel data sets, such as corpora of paraphrases. The procedure for obtaining such data for new domains is not only time-consuming, but it also restricts the incorporation of new semantic slots during an interaction, i.e. using an online learning scenario for automatically extended domains. Here, we present an alternative approach to statistical surface realisation from unlabelled data through automatic semantic slot labelling. The essence of our algorithm is to cluster clauses based on a similarity function that combines lexical and semantic information. Annotations need to be reliable enough to be utilised within a spoken dialogue system. We compare different similarity functions and evaluate our surface realiser—trained from unlabelled data—in a human rating study. Results confirm that a surface realiser trained from automatic slot labels can lead to outputs of comparable quality to outputs trained from human-labelled inputs.

This special issue highlights research articles that apply machine learning to robots and other systems that interact with users through more than one modality, such as speech, gestures, and vision. For example, a robot may coordinate its speech with its actions, taking into account (audio-)visual feedback during their execution. Machine learning provides interactive systems with opportunities to improve performance not only of individual components but also of the system as a whole. However, machine learning methods that encompass multiple modalities of an interactive system are still relatively hard to find. The articles in this special issue represent examples that contribute to filling this gap.

Conversational systems and robots that use reinforcement learning for policy optimization in large domains often face the problem of limited scalability. This problem has been addressed either by using function approximation techniques that estimate the approximate true value function of a policy or by using a hierarchical decomposition of a learning task into subtasks. We present a novel approach for dialogue policy optimization that combines the benefits of both hierarchical control and function approximation and that allows flexible transitions between dialogue subtasks to give human users more control over the dialogue. To this end, each reinforcement learning agent in the hierarchy is extended with a subtask transition function and a dynamic state space to allow flexible switching between subdialogues. In addition, the subtask policies are represented with linear function approximation in order to generalize the decision making to situations unseen in training. Our proposed approach is evaluated in an interactive conversational robot that learns to play quiz games. Experimental results, using simulation and real users, provide evidence that our proposed approach can lead to more flexible (natural) interactions than strict hierarchical control and that it is preferred by human users.

Dialogue act recognition and simulation are traditionally considered separate processes. Here, we argue that both can be fruitfully treated as interleaved processes within the same probabilistic model, leading to a synchronous improvement of performance in both. To demonstrate this, we train multiple Bayes Nets that predict the timing and content of the next user utterance. A specific focus is on providing support for barge-ins. We describe experiments using the Let's Go data that show an improvement in classification accuracy (+5%) in Bayesian dialogue act recognition involving barge-ins using partial context compared to using full context. Our results also indicate that simulated dialogues with user barge-in are more realistic than simulations without barge-in events.

This work presents preliminary observations from a study of children (N=19, age 5-12) interacting in multiple sessions with a humanoid robot in a scenario involving game activities. The main purpose of the study was to see how their perception of the robot, their engagement, and their enjoyment of the robot as a companion evolve across multiple interactions, separated by one-two weeks. However, an interesting phenomenon was observed during the experiment: most of the children soon adapted to the behaviors of the robot, in terms of speech timing, speed and tone, verbal input formulation, nodding, gestures, etc. We describe the experimental setup and the system, and our observations and preliminary analysis results, which open interesting questions for further research.

We present a conversational learning agent that helps users navigate through complex and challenging spatial environments. The agent exhibits adaptive behaviour by learning spatiallyaware dialogue actions while the user carries out the navigation task. To this end, we use Hierarchical Reinforcement Learning with relational representations to efficiently optimize dialogue actions tightly-coupled with spatial ones, and Bayesian networks to model the user’s beliefs of the navigation environment. Since these beliefs are continuously changing, we induce the agent’s behaviour in real time. Experimental results, using simulation, are encouraging by showing efficient adaptation to the user’s navigation knowledge, specifically to the generated route and the intermediate locations to negotiate with the user. Index Terms: spoken dialogue systems, situated interaction, reinforcement learning, hierarchical control, Bayesian networks

The reinforcement learning paradigm has been adopted for inferring optimized and adaptive spoken dialogue agents. Such agents are typically learnt and tested without combining competing agents that may yield better performance at some points in the conversation. This paper presents an approach that learns dialogue behaviour from competing agents—switching from one policy to another competing one—on a previously proposed hierarchical learning framework. This policy-switching approach was investigated using a simulated flight booking dialogue system based on different types of information request. Experimental results reported that the induced agent using the proposed policy-switching approach yielded 8.2% fewer system actions than three baselines with a fixed type of information request. This result suggests that the proposed approach is useful for learning adaptive and scalable spoken dialogue agents. Index Terms: spoken dialogue systems, adaptive behaviours, hierarchical reinforcement learning, policy-switching.

This paper addresses the problem of dialogue optimization on large search spaces. For such a purpose, in this paper we propose to learn dialogue strategies using multiple Semi-Markov Decision Processes and hierarchical reinforcement learning. This approach factorizes state variables and actions in order to learn a hierarchy of policies. Our experiments are based on a simulated flight booking dialogue system and compare flat versus hierarchical reinforcement learning. Experimental results show that the proposed approach produced a dramatic search space reduction (99.36 than flat reinforcement learning with a very small loss in optimality (on average 0.3 system turns). Results also report that the learnt policies outperformed a hand-crafted one under three different conditions of ASR confidence levels. This approach is appealing to dialogue optimization due to faster learning, reusable subsolutions, and scalability to larger problems.

Learning dialogue strategies using the reinforcement learning framework is problematic due to its expensive computational cost. In this paper we propose an algorithm that reduces a state-action space to one which includes only valid state-actions. We performed experiments on full and reduced spaces using three systems (with 5, 9 and 20 slots) in the travel domain using a simulated environment. The task was to learn multi-goal dialogue strategies optimizing single and multiple confirmations. Average results using strategies learnt on reduced spaces reveal the following benefits against full spaces: 1) less computer memory (94 % reduction), 2) faster learning (93 % faster convergence) and better performance (8.4 % less time steps and 7.7 % higher reward). Index Terms: reinforcement learning, spoken dialogue systems. 1.

In this paper we propose partially specified dialogue strategies for dialogue strategy optimization, where part of the strategy is specified deterministically and the rest optimized with reinforcement learning (RL). To do this we apply RL with hierarchical abstract machines (HAMs). We also propose to build simulated users using HAMs, incorporating a combination of hierarchical deterministic and probabilistic behaviour. We performed experiments using a single-goal flight booking dialogue system, and compare two dialogue strategies (deterministic and optimized) using three types of simulated user (novice, experienced and expert). Our results show that HAMs are promising for both dialogue optimization and simulation, and provide evidence that indeed partially specified dialogue strategies can outperform deterministic ones (on average 4.7 fewer system turns) with faster learning than the traditional RL framework.

In this paper we present VoiceBuilder, a framework for automating the process of developing speech applications. Our framework allows speech User Interface (UI) specialists to introduce UI’s in two ways: a stand-alone GUI application, and a web-based interface; in which speech UI’s are stored in a markup language previously proposed called SUIML [1], supporting either system initiative or mixed initiative dialogue strategies. For automatic coding, we propose an algorithm based on a macro-processor that generates VoiceXML code by parsing SUIML documents. This algorithm was designed to generate various kinds of code with a minimal initial effort. We performed experiments considering both system initiative and mixed initiative dialogue strategies with three different speech applications: auto-attendant, e-mail reader, and flight reservations. VoiceBuilder is very useful for building speech applications in new domains, requires no programming effort and could be incorporated into several voice toolkits.