Lin, Chien-Wei, Auvray, Vincent, Elkind, Daniel, Biswas, Arijit, Fazel-Zarandi, Maryam, Belgamwar, Nehal, Chandra, Shubhra, Zhao, Matt, Metallinou, Angeliki, Chung, Tagyoung, Zhu, Charlie Shucheng, Adhikari, Suranjit, Hakkani-Tur, Dilek
Goal-oriented dialog systems enable users to complete specific goals like requesting information about a movie or booking a ticket. Typically the dialog system pipeline contains multiple ML models, including natural language understanding, state tracking and action prediction (policy learning). These models are trained through a combination of supervised or reinforcement learning methods and therefore require collection of labeled domain specific datasets. However, collecting annotated datasets with language and dialog-flow variations is expensive, time-consuming and scales poorly due to human involvement. In this paper, we propose an approach for automatically creating a large corpus of annotated dialogs from a few thoroughly annotated sample dialogs and the dialog schema. Our approach includes a novel goal-sampling technique for sampling plausible user goals and a dialog simulation technique that uses heuristic interplay between the user and the system (Alexa), where the user tries to achieve the sampled goal. We validate our approach by generating data and training three different downstream conversational ML models. We achieve 18 ? 50% relative accuracy improvements on a held-out test set compared to a baseline dialog generation approach that only samples natural language and entity value variations from existing catalogs but does not generate any novel dialog flow variations. We also qualitatively establish that the proposed approach is better than the baseline. Moreover, several different conversational experiences have been built using this method, which enables customers to have a wide variety of conversations with Alexa.
Designing task-oriented dialogue systems is a challenging research topic, since it needs not only to generate utterances fulfilling user requests but also to guarantee the comprehensibility. Many previous works trained end-to-end (E2E) models with supervised learning (SL), however, the bias in annotated system utterances remains as a bottleneck. Reinforcement learning (RL) deals with the problem through using non-differentiable evaluation metrics (e.g., the success rate) as rewards. Nonetheless, existing works with RL showed that the comprehensibility of generated system utterances could be corrupted when improving the performance on fulfilling user requests. In our work, we (1) propose modelling the hierarchical structure between dialogue policy and natural language generator (NLG) with the option framework, called HDNO; (2) train HDNO with hierarchical reinforcement learning (HRL), as well as suggest alternating updates between dialogue policy and NLG during HRL inspired by fictitious play, to preserve the comprehensibility of generated system utterances while improving fulfilling user requests; and (3) propose using a discriminator modelled with language models as an additional reward to further improve the comprehensibility. We test HDNO on MultiWoz 2.0 and MultiWoz 2.1, the datasets on multi-domain dialogues, in comparison with word-level E2E model trained with RL, LaRL and HDSA, showing a significant improvement on the total performance evaluated with automatic metrics.
Spoken dialog systems have seen applications in many domains, including medical for automatic conversational diagnosis. State-of-the-art dialog managers are usually driven by deep reinforcement learning models, such as deep Q networks (DQNs), which learn by interacting with a simulator to explore the entire action space since real conversations are limited. However, the DQN-based automatic diagnosis models do not achieve satisfying performances when adapted to new, unseen diseases with only a few training samples. In this work, we propose the Prototypical Q Networks (ProtoQN) as the dialog manager for the automatic diagnosis systems. The model calculates prototype embeddings with real conversations between doctors and patients, learning from them and simulator-augmented dialogs more efficiently. We create both supervised and few-shot learning tasks with the Muzhi corpus. Experiments showed that the ProtoQN significantly outperformed the baseline DQN model in both supervised and few-shot learning scenarios, and achieves state-of-the-art few-shot learning performances.
Reinforcement-based training methods have emerged as the most popular choice to train an efficient and effective dialog policy. However, these methods are suffering from sparse and unstable reward signals usually returned from the user simulator at the end of the dialog. Besides, the reward signal is manually designed by human experts which requires domain knowledge. A number of adversarial learning methods have been proposed to learn the reward function together with the dialog policy. However, to alternatively update the dialog policy and the reward model on the fly, the algorithms to update the dialog policy are limited to policy gradient-based algorithms, such as REINFORCE and PPO. Besides, the alternative training of the dialog agent and the reward model can easily get stuck in local optimum or result in mode collapse. In this work, we propose to decompose the previous adversarial training into two different steps. We first train the discriminator with an auxiliary dialog generator and then incorporate this trained reward model to a common reinforcement learning method to train a high-quality dialog agent. This approach is applicable to both on-policy and off-policy reinforcement learning methods. By conducting several experiments, we show the proposed methods can achieve remarkable task success and its potential to transfer knowledge from existing domains to a new domain.
Language systems have been of great interest to the research community and have recently reached the mass market through various assistant platforms on the web. Reinforcement Learning methods that optimize dialogue policies have seen successes in past years and have recently been extended into methods that personalize the dialogue, e.g. take the personal context of users into account. These works, however, are limited to personalization to a single user with whom they require multiple interactions and do not generalize the usage of context across users. This work introduces a problem where a generalized usage of context is relevant and proposes two Reinforcement Learning (RL)-based approaches to this problem. The first approach uses a single learner and extends the traditional POMDP formulation of dialogue state with features that describe the user context. The second approach segments users by context and then employs a learner per context. We compare these approaches in a benchmark of existing non-RL and RL-based methods in three established and one novel application domain of financial product recommendation. We compare the influence of context and training experiences on performance and find that learning approaches generally outperform a handcrafted gold standard.
In this paper we survey the methods and concepts developed for the evaluation of dialogue systems. Evaluation is a crucial part during the development process. Often, dialogue systems are evaluated by means of human evaluations and questionnaires. However, this tends to be very cost and time intensive. Thus, much work has been put into finding methods, which allow to reduce the involvement of human labour. In this survey, we present the main concepts and methods. For this, we differentiate between the various classes of dialogue systems (task-oriented dialogue systems, conversational dialogue systems, and question-answering dialogue systems). We cover each class by introducing the main technologies developed for the dialogue systems and then by presenting the evaluation methods regarding this class.
Can we learn a control policy able to adapt its behaviour in real time so as to take any desired amount of risk? The general Reinforcement Learning framework solely aims at optimising a total reward in expectation, which may not be desirable in critical applications. In stark contrast, the Budgeted Markov Decision Process (BMDP) framework is a formalism in which the notion of risk is implemented as a hard constraint on a failure signal. Existing algorithms solving BMDPs rely on strong assumptions and have so far only been applied to toy-examples. In this work, we relax some of these assumptions and demonstrate the scalability of our approach on two practical problems: a spoken dialogue system and an autonomous driving task. On both examples, we reach similar performances as Lagrangian Relaxation methods with a significant improvement in sample and memory efficiency.
One of the core challenges in Visual Dialogue problems is asking the question that will provide the most useful information towards achieving the required objective. Encouraging an agent to ask the right questions is difficult because we don't know a-priori what information the agent will need to achieve its task, and we don't have an explicit model of what it knows already. We propose a solution to this problem based on a Bayesian model of the uncertainty in the implicit model maintained by the visual dialogue agent, and in the function used to select an appropriate output. By selecting the question that minimises the predicted regret with respect to this implicit model the agent actively reduces ambiguity. The Bayesian model of uncertainty also enables a principled method for identifying when enough information has been acquired, and an action should be selected. We evaluate our approach on two goal-oriented dialogue datasets, one for visual-based collaboration task and the other for a negotiation-based task. Our uncertainty-aware information-seeking model outperforms its counterparts in these two challenging problems.
Reinforcement learning methods have been used for learning dialogue policies from the experience of conversations. However, learning an effective dialogue policy frequently requires prohibitively many conversations. This is partly because of the sparse rewards in dialogues, and the relatively small number of successful dialogues in early learning phase. Hindsight experience replay (HER) enables an agent to learn from failure, but the vanilla HER is inapplicable to dialogue domains due to dialogue goals being implicit (c.f., explicit goals in manipulation tasks). In this work, we develop two complex HER methods providing different trade-offs between complexity and performance. Experiments were conducted using a realistic user simulator. Results suggest that our HER methods perform better than standard and prioritized experience replay methods (as applied to deep Q-networks) in learning rate, and that our two complex HER methods can be combined to produce the best performance.
In spoken dialogue systems, we aim to deploy artificial intelligence to build automated dialogue agents that can converse with humans. A part of this effort is the policy optimisation task, which attempts to find a policy describing how to respond to humans, in the form of a function taking the current state of the dialogue and returning the response of the system. In this paper, we investigate deep reinforcement learning approaches to solve this problem. Particular attention is given to actor-critic methods, off-policy reinforcement learning with experience replay, and various methods aimed at reducing the bias and variance of estimators. When combined, these methods result in the previously proposed ACER algorithm that gave competitive results in gaming environments. These environments however are fully observable and have a relatively small action set so in this paper we examine the application of ACER to dialogue policy optimisation. We show that this method beats the current state-of-the-art in deep learning approaches for spoken dialogue systems. This not only leads to a more sample efficient algorithm that can train faster, but also allows us to apply the algorithm in more difficult environments than before. We thus experiment with learning in a very large action space, which has two orders of magnitude more actions than previously considered. We find that ACER trains significantly faster than the current state-of-the-art.