Dialogue Turn Research Articles

An emotion-sensitive dialogue policy for task-oriented dialogue system

Reinforcement learning (RL) is an effective method in training dialogue policies to steer the conversation towards successful task completion. However, most RL-based methods only rely on semantic inputs that lack empathy as they ignore the user emotional information. Moreover, these methods suffer from delayed rewards caused by the user simulator returning valuable results only at dialogue end. Recently, some methods have been proposed to learn the reward function together with user emotions, but they omit considering user emotion in each dialogue turn. In this paper, we proposed an emotion-sensitive dialogue policy model (ESDP), it incorporates user emotions information into dialogue policy and selects the optimal action by the combination of top-k actions with the user emotions. The user emotion information in each turn is used as an immediate reward for the current dialogue state to solve sparse rewards and the dependency on termination. Extensive experiments validate that our method outperforms the baseline approaches when combined with different Q-Learning algorithms, and also surpasses other popular existing dialog policies’ performance.

RSP-DST: Revisable State Prediction for Dialogue State Tracking

Task-oriented dialogue systems depend on dialogue state tracking to keep track of the intentions of users in the course of conversations. Although recent models in dialogue state tracking exhibit good performance, the errors in predicting the value of each slot at the current dialogue turn of these models are easily carried over to the next turn, and unlikely to be revised in the next turn, resulting in error propagation. In this paper, we propose a revisable state prediction for dialogue state tracking, which constructs a two-stage slot value prediction process composed of an original prediction and a revising prediction. The original prediction process jointly models the previous dialogue state and dialogue context to predict the original dialogue state of the current dialogue turn. Then, in order to avoid the errors existing in the original dialogue state continuing to the next dialogue turn, a revising prediction process utilizes the dialogue context to revise errors, alleviating the error propagation. Experiments are conducted on MultiWOZ 2.0, MultiWOZ 2.1, and MultiWOZ 2.4 and results indicate that our model outperforms previous state-of-the-art works, achieving new state-of-the-art performances with 56.35, 58.09, and 75.65% joint goal accuracy, respectively, which has a significant improvement (2.15, 1.73, and 2.03%) over the previous best results.

A randomized prospective study of a hybrid rule- and data-driven virtual patient

AbstractRandomized prospective studies represent the gold standard for experimental design. In this paper, we present a randomized prospective study to validate the benefits of combining rule-based and data-driven natural language understanding methods in a virtual patient dialogue system. The system uses a rule-based pattern matching approach together with a machine learning (ML) approach in the form of a text-based convolutional neural network, combining the two methods with a simple logistic regression model to choose between their predictions for each dialogue turn. In an earlier, retrospective study, the hybrid system yielded a nearly 50% error reduction on our initial data, in part due to the differential performance between the two methods as a function of label frequency. Given these gains, and considering that our hybrid approach is unique among virtual patient systems, we compare the hybrid system to the rule-based system by itself in a randomized prospective study. We evaluate 110 unique medical student subjects interacting with the system over 5,296 conversation turns, to verify whether similar gains are observed in a deployed system. This prospective study broadly confirms the findings from the earlier one but also highlights important deficits in our training data. The hybrid approach still improves over either rule-based or ML approaches individually, even handling unseen classes with some success. However, we observe that live subjects ask more out-of-scope questions than expected. To better handle such questions, we investigate several modifications to the system combination component. These show significant overall accuracy improvements and modest F1 improvements on out-of-scope queries in an offline evaluation. We provide further analysis to characterize the difficulty of the out-of-scope problem that we have identified, as well as to suggest future improvements over the baseline we establish here.

Hard-style Selective Context Utilization for dialogue generation based on what user just said

Dialogue is a process of information exchanging, where global background is stable while local focuses are transiting. Thus, at the ongoing dialogue turn, there are both relevant and irrelevant semantics existing in dialogue contexts. How to filter out noises and selectively utilize context can pave the way to successful dialogue generation. Current work on dialogue context utilization either processes contexts as vanilla monologue text ignoring dynamic conversation flows, or depends on weighted strategies to fuse all contexts where irrelevant utterances cannot be filter out even may overwhelm relevant ones. To deal with this, this paper proposes a Hard-style Selective Context Utilization method (HardSCU). We first define and measure the information density of the last utterance (query) of a dialogue, marking it as “strong” or “weak”. For a dialogue with strong query, HardSCU directly inputs the query into a RNN-based or T5-based encoder–decoder framework to generate a response; for a dialogue with weak query, HardSCU conducts a selective context utilization for dialogue generation, where a semantic interaction module introduces relevant semantics of context to enrich the query and the co-reference relations existing in dialogue are extracted to promote the learning process of response decoder. Extensive experiments on two benchmark conversation corpora verify that our HardSCU method can outperform competitive baselines on generating appropriate responses for chit-chat-bots with yielding strong robustness to the variations of dialogue lengths.

UBAR: Towards Fully End-to-End Task-Oriented Dialog System with GPT-2

This paper presents our task-oriented dialog system UBAR which models task-oriented dialogs on a dialog session level. Specifically, UBAR is acquired by fine-tuning the large pre-trained unidirectional language model GPT-2 on the sequence of the entire dialog session which is composed of user utterance, belief state, database result, system act, and system response of every dialog turn. Additionally, UBAR is evaluated in a more realistic setting, where its dialog context has access to user utterances and all content it generated such as belief states, system acts, and system responses. Experimental results on the MultiWOZ datasets show that UBAR achieves state-of-the-art performances in multiple settings, improving the combined score of response generation, policy optimization, and end-to-end modeling by 4.7, 3.5, and 9.4 points respectively. Thorough analyses demonstrate that the session-level training sequence formulation and the generated dialog context are essential for UBAR to operate as a fully end-to-end task-oriented dialog system in real life. We also examine the transfer ability of UBAR to new domains with limited data and provide visualization and a case study to illustrate the advantages of UBAR in modeling on a dialog session level.

A novel focus encoding scheme for addressee detection in multiparty interaction using machine learning algorithms

Addressee detection is a fundamental task for seamless dialogue management and turn taking in human-agent interaction. Though addressee detection is implicit in dyadic interaction, it becomes a challenging task when more than two participants are involved. This article proposes multiple addressee detection models based on smart feature selection and focus encoding schemes. The models are trained using different machine learning and deep learning algorithms. This research work improves existing baseline accuracies for addressee prediction on two datasets. In addition, the article explores the impact of different focus encoding schemes in several addressee detection cases. Finally, an implementation strategy for addressee detection model in real-time is discussed.

Speech Emotion Recognition Considering Nonverbal Vocalization in Affective Conversations

In real-life communication, nonverbal vocalization such as laughter, cries or other emotion interjections, within an utterance play an important role for emotion expression. In previous studies, only few emotion recognition systems consider nonverbal vocalization, which naturally exists in our daily conversation. In this work, both verbal and nonverbal sounds within an utterance are considered for emotion recognition of real-life affective conversations. Firstly, a support vector machine (SVM)-based verbal and nonverbal sound detector is developed. A prosodic phrase auto-tagger is further employed to extract the verbal/nonverbal sound segments. For each segment, the emotion and sound feature embeddings are respectively extracted using the deep residual networks (ResNets). Finally, a sequence of the extracted feature embeddings for the entire dialog turn are fed to an attentive long short-term memory (LSTM)-based sequence-to-sequence model to output an emotional sequence as recognition result. The NNIME corpus (The NTHU-NTUA Chinese interactive multimodal emotion corpus), which consists of verbal and nonverbal sounds, was adopted for system training and testing. 4766 single speaker dialogue turns in the audio data of the NNIME corpus were selected for evaluation. The experimental results showed that nonverbal vocalization was helpful for speech emotion recognition. For comparison, the proposed method based on decision-level fusion achieved an accuracy of 61.92% for speech emotion recognition outperforming the traditional methods as well as the feature-level and model-level fusion approaches.

End-to-End latent-variable task-oriented dialogue system with exact log-likelihood optimization

We propose an end-to-end dialogue model based on a hierarchical encoder-decoder, which employed a discrete latent variable to learn underlying dialogue intentions. The system is able to model the structure of utterances dominated by statistics of the language and the dependencies among utterances in dialogues without manual dialogue state design. We argue that the latent discrete variable interprets the intentions that guide machine responses generation. We also propose a model which can be refined autonomously with reinforcement learning, due to that intention selection at each dialogue turn can be formulated as a sequential decision-making process. Our experiments show that exact MLE optimized model is much more robust than neural variational inference on dialogue success rate with limited BLEU sacrifice.

Cascaded deep neural network models for dialog state tracking

Dialog state tracking (DST) maintains and updates dialog states at each time step as the dialog progresses. It is necessary to include dialog historical information in DST. Previous word-based DST models took historical utterances as a word sequence and used n-grams in the sequence as inputs of models. It suffered from the problem of data sparseness. This paper proposes a cascaded deep neural network framework for DST. It alleviates the problem of data sparseness by making use of the hierarchical structure in dialog. The bottom layer of the cascaded framework, implemented by an Long Short Term Memory (LSTM) or a Convolutional Neural Network (CNN), encodes the word sequence into a sentence embedding in each dialog turn, and the upper layer integrates the representation of each turn gradually to get the dialog state using an LSTM. The cascaded models integrate natural language understanding into DST, and the entire network is trained as a whole. The experimental results on the DSTC2 dataset indicate that the proposed models, LSTM+LSTM and CNN + LSTM, can achieve better performance than existing models.

Constrained Markov Bayesian Polynomial for Efficient Dialogue State Tracking

Dialogue state tracking (DST) is a process to estimate the distribution of the dialogue states at each dialogue turn given the interaction history. Although data-driven statistical approaches are of most interest, there have been attempts of using rule-based methods for DST, due to their simplicity, efficiency and portability. However, the performance of these methods are usually not competitive to data-driven tracking approaches and it is not possible to improve the DST performance when training data are available. In this paper, a novel hybrid framework, constrained Markov Bayesian polynomial (CMBP), is proposed to formulate rule-based DST in a general way and allow data-driven rule generation. Here, a DST rule is defined as a polynomial function of a set of probabilities satisfying certain linear constraints. Prior knowledge is encoded in these constraints. Under reasonable assumptions, CMBP optimization can be converted to a constrained integer linear programming problem. The integer coefficient CMBP model is further extended to CMBP with real coefficients by applying grid search. CMBP was evaluated on the data corpora of the first, the second, and the third Dialog State Tracking Challenge (DSTC-1/2/3). Experiments showed that CMBP has good generalization ability and can significantly outperform both traditional rule-based approaches and data-driven statistical approaches with similar feature set. Compared with the state-of-the-art statistical DST approaches with much richer features, CMBP is also competitive.

Evolvable dialogue state tracking for statistical dialogue management

Statistical dialogue management is the core of cognitive spoken dialogue systems (SDS) and has attracted great research interest. In recent years, SDS with the ability of evolution is of particular interest and becomes the cuttingedge of SDS research. Dialogue state tracking (DST) is a process to estimate the distribution of the dialogue states at each dialogue turn, given the previous interaction history. It plays an important role in statistical dialogue management. To provide a common testbed for advancing the research of DST, international DST challenges (DSTC) have been organised and well-attended by major SDS groups in the world. This paper reviews recent progresses on rule-based and statistical approaches during the challenges. In particular, this paper is focused on evolvable DST approaches for dialogue domain extension. The two primary aspects for evolution, semantic parsing and tracker, are discussed. Semantic enhancement and a DST framework which bridges rule-based and statistical models are introduced in detail. By effectively incorporating prior knowledge of dialogue state transition and the ability of being data-driven, the new framework supports reliable domain extension with little data and can continuously improve with more data available. Thismakes it excellent candidate for DST evolution. Experiments show that the evolvable DST approaches can achieve the state-of-the-art performance and outperform all previously submitted trackers in the third DSTC.

A Recursive Dialogue Game for Personalized Computer-Aided Pronunciation Training

Learning languages in addition to the native language is very important for all people in the globalized world today, and computer-aided pronunciation training (CAPT) is attractive since the software can be used anywhere at any time, and repeated as many times as desired. In this paper, we introduce the immersive interaction scenario offered by spoken dialogues to CAPT by proposing a recursive dialogue game to make CAPT personalized. A number of tree-structured sub-dialogues are linked sequentially and recursively as the script for the game. The system policy at each dialogue turn is to select in real-time along the dialogue the best training sentence for each specific individual learner within the dialogue script, considering the learner's learning status and the future possible dialogue paths in the script, such that the learner can have the scores for all pronunciation units considered reaching a predefined standard in a minimum number of turns. The purpose here is that those pronunciation units poorly produced by the specific learner can be offered with more practice opportunities in the future sentences along the dialogue, which enables the learner to improve the pronunciation without having to repeat the same training sentences many times. This makes the learning process for each learner completely personalized. The dialogue policy is modeled by Markov decision process (MDP) with high-dimensional continuous state space, and trained with fitted value iteration using a huge number of simulated learners. These simulated leaners have the behavior similar to real learners, and were generated from a corpus of real learner data. The experiments demonstrated very promising results and a real cloud-based system is also successfully implemented.

Unsegmented Dialogue Act Annotation and Decoding with N-Gram Transducers

Most studies on dialogue corpora, as well as most dialogue systems, employ dialogue acts as the basic units for interpreting discourse structure, user input and system actions. The definition of the discourse structure and the dialogue strategy consequently require the tagging of dialogue corpora in terms of dialogue acts. The tagging problem presents two basic variants: a batch variant (annotation of whole dialogues, in order to define dialogue strategy or study discourse structure) and an online variant (decoding of the dialogue act sequence of a given turn, in order to interpret user intentions). In the two variants is unusual having the segmentation of each turn into the dialogue meaningful units (segments) to which a dialogue act is assigned. In this paper we present the use of the N-Gram Transducer technique for tagging dialogues, without needing to provide a prior segmentation, in these two different variants (dialogue annotation and turn decoding). Experiments were performed in two corpora of different nature and results show that N-Gram Transducer models are suitable for these tasks and provide good performance.

MULTIMODAL FUSION AS COMMUNICATIVE ACTS DURING HUMAN–ROBOT INTERACTION

Research on dialog systems is a very active area in social robotics. During the last two decades, these systems have evolved from those based only on speech recognition and synthesis to the current and modern systems, which include new components and multimodality. By multimodal dialogue we mean the interchange of information among several interlocutors, not just using their voice as the mean of transmission but also all the available channels such as gestures, facial expressions, touch, sounds, etc. These channels add information to the message to be transmitted in every dialogue turn. The dialogue manager (IDiM) is one of the components of the robotic dialog system (RDS) and is in charge of managing the dialogue flow during the conversational turns. In order to do that, it is necessary to coherently treat the inputs and outputs of information that flow by different communication channels: audio, vision, radio frequency, touch, etc. In our approach, this multichannel input of information is temporarily fused into communicative acts (CAs). Each CA groups the information that flows through the different input channels into the same pack, transmitting a unique message or global idea. Therefore, this temporary fusion of information allows the IDiM to abstract from the channels used during the interaction, focusing only on the message, not on the way it is transmitted. This article presents the whole RDS and the description of how the multimodal fusion of information is made as CAs. Finally, several scenarios where the multimodal dialogue is used are presented.

Does History Help? An Experiment on How Context Affects Crowdsourcing Dialogue Annotation

Crowds of people can potentially solve some problems faster than individuals. Crowd sourced data can be leveraged to benefit the crowd by providing information or solutions faster than traditional means. Many tasks needed for developing dialogue systems such as annotation can benefit from crowdsourcing as well. We investigate how to outsource dialogue data annotation through Amazon Mechanical Turk. We are in particular interested in empirically analyzing how much context from previous parts of the dialogue (e.g. previous dialogue turns) is needed to be provided before the target part (dialogue turn) is presented to the annotator. The answer to this question is essentially important for leveraging crowd sourced data for appropriate and efficient response and coordination. We study the effect of presenting different numbers of previous data (turns) to the Turkers in annotating sentiments of dyadic negotiation dialogs on the inter annotator reliability and comparison to the gold standard.

On the dynamic adaptation of language models based on dialogue information

We present an approach to adapt dynamically the language models (LMs) used by a speech recognizer that is part of a spoken dialogue system. We have developed a grammar generation strategy that automatically adapts the LMs using the semantic information that the user provides (represented as dialogue concepts), together with the information regarding the intentions of the speaker (inferred by the dialogue manager, and represented as dialogue goals). We carry out the adaptation as a linear interpolation between a background LM, and one or more of the LMs associated to the dialogue elements (concepts or goals) addressed by the user. The interpolation weights between those models are automatically estimated on each dialogue turn, using measures such as the posterior probabilities of concepts and goals, estimated as part of the inference procedure to determine the actions to be carried out. We propose two approaches to handle the LMs related to concepts and goals. Whereas in the first one we estimate a LM for each one of them, in the second one we apply several clustering strategies to group together those elements that share some common properties, and estimate a LM for each cluster. Our evaluation shows how the system can estimate a dynamic model adapted to each dialogue turn, which helps to significantly improve the performance of the speech recognition, which leads to an improvement in both the language understanding and the dialogue management tasks.

Effective handling of dialogue state in the hidden information state POMDP-based dialogue manager

Effective dialogue management is critically dependent on the information that is encoded in the dialogue state. In order to deploy reinforcement learning for policy optimization, dialogue must be modeled as a Markov Decision Process. This requires that the dialogue state must encode all relevent information obtained during the dialogue prior to that state. This can be achieved by combining the user goal, the dialogue history, and the last user action to form the dialogue state. In addition, to gain robustness to input errors, dialogue must be modeled as a Partially Observable Markov Decision Process (POMDP) and hence, a distribution over all possible states must be maintained at every dialogue turn. This poses a potential computational limitation since there can be a very large number of dialogue states. The Hidden Information State model provides a principled way of ensuring tractability in a POMDP-based dialogue model. The key feature of this model is the grouping of user goals into partitions that are dynamically built during the dialogue. In this article, we extend this model further to incorporate the notion of complements. This allows for a more complex user goal to be represented, and it enables an effective pruning technique to be implemented that preserves the overall system performance within a limited computational resource more effectively than existing approaches.

Estimating the number of segments for improving dialogue act labelling

AbstractIn dialogue systems it is important to label the dialogue turns with dialogue-related meaning. Each turn is usually divided into segments and these segments are labelled with dialogue acts (DAs). A DA is a representation of the functional role of the segment. Each segment is labelled with one DA, representing its role in the ongoing discourse. The sequence of DAs given a dialogue turn is used by the dialogue manager to understand the turn. Probabilistic models that perform DA labelling can be used on segmented or unsegmented turns. The last option is more likely for a practical dialogue system, but it provides poorer results. In that case, a hypothesis for the number of segments can be provided to improve the results. We propose some methods to estimate the probability of the number of segments based on the transcription of the turn. The new labelling model includes the estimation of the probability of the number of segments in the turn. We tested this new approach with two different dialogue corpora:SwitchBoardandDihana. The results show that this inclusion significantly improves the labelling accuracy.

Generating with Discourse Combinatory Categorial Grammar

This article introduces Discourse Combinatory Categorial Grammar (DCCG) and shows how it can be used to generate multisentence paraphrases, flexibly incorporating both intra- and intersentential discourse connectives. DCCG employs a simple, practical approach to extending Combinatory Categorial Grammar (CCG) to encompass coverage of discourse-level phenomena, which furthermore makes it possible to generate clauses with multiple connectives and — in contrast to approaches based on Rhetorical Structure Theory — with rhetorical dependencies that do not form a tree. To do so, it borrows from Discourse Lexicalized Tree Adjoining Grammar (D-LTAG) the distinction between structural connectives and anaphoric discourse adverbials. Unlike D-LTAG, however, DCCG treats both sentential and discourse phenomena in the same grammar, rather than employing a separate discourse grammar. A key ingredient of this single-grammar approach is cue threading, a tightly constrained technique for extending the semantic scope of a discourse connective beyond the sentence. As DCCG requires no additions to the CCG formalism, it can be used to generate paraphrases of an entire dialogue turn using the OpenCCG realizer as-is, without the need to revise its architecture. In addition, from an interpretation perspective, a single grammar enables easier management of ambiguity across discourse and sentential levels using standard dynamic programming techniques, whereas D-LTAG has required a potentially complex interaction of sentential and discourse grammars to manage the same ambiguity. As a proof-of-concept, the article demonstrates how OpenCCG can be used with a DCCG to generate multi-sentence paraphrases that reproduce and extend those in the SPaRKy Restaurant Corpus.

Low-level grounding in a multimodal mobile service robot conversational system using graphical models

The main task of a service robot with a voice-enabled communication interface is to engage a user in dialogue providing an access to the services it is designed for. In managing such interaction, inferring the user goal (intention) from the request for a service at each dialogue turn is the key issue. In service robot deployment conditions speech recognition limitations with noisy speech input and inexperienced users may jeopardize user goal identification. In this paper, we introduce a grounding state-based model motivated by reducing the risk of communication failure due to incorrect user goal identification. The model exploits the multiple modalities available in the service robot system to provide evidence for reaching grounding states. In order to handle the speech input as sufficiently grounded (correctly understood) by the robot, four proposed states have to be reached. Bayesian networks combining speech and non-speech modalities during user goal identification are used to estimate probability that each grounding state has been reached. These probabilities serve as a base for detecting whether the user is attending to the conversation, as well as for deciding on an alternative input modality (e.g., buttons) when the speech modality is unreliable. The Bayesian networks used in the grounding model are specially designed for modularity and computationally efficient inference. The potential of the proposed model is demonstrated comparing a conversational system for the mobile service robot RoboX employing only speech recognition for user goal identification, and a system equipped with multimodal grounding. The evaluation experiments use component and system level metrics for technical (objective) and user-based (subjective) evaluation with multimodal data collected during the conversations of the robot RoboX with users.