Task-oriented dialogue systems (TODS) have become crucial for users to interact with machines and computers using natural language. One of its key components is the dialogue manager, which guides the conversation towards a good goal for the user by providing the best possible response. Previous works have proposed rule-based systems (RBS), reinforcement learning (RL), and supervised learning (SL) as solutions for the correct dialogue management; in other words, select the best response given input by the user. However, this work argues that the leading cause of DMs not achieving maximum performance resides in the quality of the datasets rather than the models employed thus far; this means that dataset errors, like mislabeling, originate a large percentage of failures in dialogue management. We studied the main errors in the most widely used datasets, Multiwoz 2.1 and SGD, to demonstrate this hypothesis. To do this, we have designed a synthetic dialogue generator to fully control the amount and type of errors introduced in the dataset. Using this generator, we demonstrated that errors in the datasets contribute proportionally to the performance of the models
Popular dialog datasets such as MultiWOZ are created by providing crowd workers an instruction, expressed in natural language, that describes the task to be accomplished. Crowd workers play the role of a user and an agent to generate dialogs to accomplish tasks involving booking restaurant tables, calling a taxi etc. In this paper, we present a data creation strategy that uses the pre-trained language model, GPT2, to simulate the interaction between crowd workers by creating a user bot and an agent bot. We train the simulators using a smaller percentage of actual crowd-generated conversations and their corresponding instructions. We demonstrate that by using the simulated data, we achieve significant improvements in low-resource settings on two publicly available datasets - the MultiWOZ dataset and the Persona chat dataset.
We propose a novel methodology to address dialog learning in the context of goal-oriented conversational systems. The key idea is to quantize the dialog space into clusters and create a language model across the clusters, thus allowing for an accurate choice of the next utterance in the conversation. The language model relies on n-grams associated with clusters of utterances. This quantized-dialog language model methodology has been applied to the end-to-end goal-oriented track of the latest Dialog System Technology Challenges (DSTC6). The objective is to find the correct system utterance from a pool of candidates in order to complete a dialog between a user and an automated restaurant-reservation system. Our results show that the technique proposed in this paper achieves high accuracy regarding selection of the correct candidate utterance, and outperforms other state-of-the-art approaches based on neural networks.
Modern statistical machine learning (SML) methods share a major limitation with the early approaches to AI: there is no scalable way to adapt them to new domains. Human learning solves this in part by leveraging a rich, shared, updateable world model. Such scalability requires modularity: updating part of the world model should not impact unrelated parts. We have argued that such modularity will require both "correctability" (so that errors can be corrected without introducing new errors) and "interpretability" (so that we can understand what components need correcting). To achieve this, one could attempt to adapt state of the art SML systems to be interpretable and correctable; or one could see how far the simplest possible interpretable, correctable learning methods can take us, and try to control the limitations of SML methods by applying them only where needed. Here we focus on the latter approach and we investigate two main ideas: "Teacher Assisted Learning", which leverages crowd sourcing to learn language; and "Factored Dialog Learning", which factors the process of application development into roles where the language competencies needed are isolated, enabling non-experts to quickly create new applications. We test these ideas in an "Automated Personal Assistant" (APA) setting, with two scenarios: that of detecting user intent from a user-APA dialog; and that of creating a class of event reminder applications, where a non-expert "teacher" can then create specific apps. For the intent detection task, we use a dataset of a thousand labeled utterances from user dialogs with Cortana, and we show that our approach matches state of the art SML methods, but in addition provides full transparency: the whole (editable) model can be summarized on one human-readable page. For the reminder app task, we ran small user studies to verify the efficacy of the approach.
In this work, we propose an adversarial learning method for reward estimation in reinforcement learning (RL) based task-oriented dialog models. Most of the current RL based task-oriented dialog systems require the access to a reward signal from either user feedback or user ratings. Such user ratings, however, may not always be consistent or available in practice. Furthermore, online dialog policy learning with RL typically requires a large number of queries to users, suffering from sample efficiency problem. To address these challenges, we propose an adversarial learning method to learn dialog rewards directly from dialog samples. Such rewards are further used to optimize the dialog policy with policy gradient based RL. In the evaluation in a restaurant search domain, we show that the proposed adversarial dialog learning method achieves advanced dialog success rate comparing to strong baseline methods. We further discuss the covariate shift problem in online adversarial dialog learning and show how we can address that with partial access to user feedback.