Task-oriented Dialogue Systems Research Articles

Combining replay and LoRA for continual learning in natural language understanding

Large language models have significantly improved dialogue systems through enhanced capabilities in understanding queries and generating responses. Despite these enhancements, task-oriented dialogue systems—which power many intelligent assistants—face challenges when adapting to new domains and applications. This challenge arises from a phenomenon known as catastrophic forgetting, where models forget previously acquired knowledge when learning new tasks. This paper addresses this issue through continual learning techniques to preserve previously learned knowledge while seamlessly integrating new tasks and domains. We propose Experience Replay Informative-Low Rank Adaptation or ERI-LoRA, a hybrid continual learning method for natural language understanding in dialogue systems that effectively combines the replay-based methods with parameter-efficient techniques. Our experiments on intent detection and slot-filling tasks demonstrate that ERI-LoRA significantly outperforms competitive baselines in continual learning. The results of our catastrophic forgetting experiments demonstrate that ERI-LoRA maintains robust memory stability in the model, demonstrating its effectiveness in mitigating these effects.

Optimizing pipeline task-oriented dialogue systems using post-processing networks

Many studies have proposed methods for optimizing the dialogue performance of an entire pipeline task-oriented dialogue system by jointly training modules in the system using reinforcement learning. However, these methods are limited in that they can only be applied to modules implemented using trainable neural-based methods. To solve this problem, we propose a method for optimizing the dialogue performance of a pipeline system that consists of modules implemented with arbitrary methods for dialogue. With our method, neural-based components called post-processing networks (PPNs) are installed inside such a system to post-process the output of each module. All PPNs are updated to improve the overall dialogue performance of the system by using reinforcement learning, not necessitating that each module be differentiable. Through dialogue simulations and human evaluations on two well-studied task-oriented dialogue datasets, CamRest676 and MultiWOZ, we show that our method can improve the dialogue performance of pipeline systems consisting of various modules. In addition, a comprehensive analysis of the results of the MultiWOZ experiments reveals the patterns of post-processing by PPNs that contribute to the overall dialogue performance of the system.

A multi-agent collaborative algorithm for task-oriented dialogue systems

A multi-agent collaborative algorithm for task-oriented dialogue systems

Enhancing Task-Oriented Dialogue Systems through Synchronous Multi-Party Interaction and Multi-Group Virtual Simulation

This paper presents two innovative approaches: a synchronous multi-party dialogue system that engages in simultaneous interactions with multiple users, and multi-group simulations involving virtual user groups to evaluate the resilience of this system. Unlike most other chatbots that communicate with each user independently, our system facilitates information gathering from multiple users and executes 17 administrative tasks for group requests adeptly by leveraging a state machine-based framework for complete control over dialogue flow and a large language model (LLM) for robust context understanding. Assessing such a unique dialogue system poses challenges, as it requires many groups of users to interact with the system concurrently for an extended duration. To address this, we simulate various virtual groups using an LLM, each comprising 10–30 users who may belong to multiple groups, in order to evaluate the efficacy of our system; each user is assigned a persona and allowed to interact freely without scripts. As a result, our system shows average success rates of 87% for task completion and 89% for natural language understanding. Comparatively, our virtual simulation, which has an average success rate of 80%, is juxtaposed with a group of 15 human users, depicting similar task diversity and error trends. To our knowledge, it is the first work to show the LLM’s potential in both task execution and the simulation of a synchronous dialogue system to fully automate administrative tasks.

A novel model based on a transformer for intent detection and slot filling

Building task-oriented dialogue systems has become a topic of interest in the research community and industry. The task-oriented dialogue system is a closed-domain dialogue system that can perform specific tasks for users. The natural language understanding module of a task-oriented dialogue system is crucial because it is related to a task-oriented dialogue system that provides correctional services for users. The natural language understanding module of a task-oriented dialogue system performs two tasks: intent detection and slot filling. The intent detection task can be regarded as a text classification task; a classification model is trained to predict the intention of the user from the user’s input information. The slot filling task can be regarded as a sequence analysis task; a sequence analysis model is trained to predict the details of the user’s intention. In this paper, we proposed a novel model based on a transformer encoder for intent detection and slot filling. It follows the encoder-decoder structure, including a vanilla Transformer encoder, a bidirectional LSTM encoder, a linear classification decoder for intent detection, and a conditional random field decoder for slot filling. The experimental results on two public datasets show that our proposed model outperforms the existing methods based on the Transformer and can be combined with BERT to achieve better intent detection and slot filling results.

A multi-agent curiosity reward model for task-oriented dialogue systems

In practical decision-making dialogues, reinforcement learning methods face hurdles due to delays and sparse reward feedback for agents, and in some cases, lack of rewards altogether. These issues can impede efficient learning of dialogue strategies and compromise the performance of the model. To address this challenge, this paper introduces the Multi-Agent Curiosity Reward Model (MACRM) for task-oriented dialog systems. Firstly, in terms of dialog reward mechanisms, a forward dynamics model generates curiosity rewards, which are integrated with extrinsic rewards from the dialog environment feedback to mitigate the problem of sparse rewards resulting from inadequate agent exploration. Secondly, regarding the dialogue strategy training mechanism, an exploration-exploitation approach inspired by organismic exploration is adopted. This approach involves fully exploring the dialogue environment in the early stages and optimally exploiting learned knowledge later, thereby balancing exploration and exploitation and enhancing dialogue strategy learning efficiency. To assess the proposed model's effectiveness, experiments are conducted using the MultiWOZ corpus across three reward environments: (1) extrinsic rewards only, (2) curiosity rewards only, and (3) a combination of both. The experimental results demonstrate that agents employing MACRM exhibit faster learning of dialogue strategies compared to those relying on a single exploratory reward method, effectively addressing reward sparsity and delay issues in practical decision-making scenarios.

Knowledge-aware audio-grounded generative slot filling for limited annotated data

Manually annotating fine-grained slot-value labels for task-oriented dialogue (ToD) systems is an expensive and time-consuming endeavour. This motivates research into slot-filling methods that operate with limited amounts of labelled data. Moreover, the majority of current work on ToD is based solely on text as the input modality, neglecting the additional challenges of imperfect automatic speech recognition (ASR) when working with spoken language. In this work, we propose a Knowledge-Aware Audio-Grounded generative slot filling framework, termed KA2G, that focuses on few-shot and zero-shot slot filling for ToD with speech input. KA2G achieves robust and data-efficient slot filling for speech-based ToD by (1) framing it as a text generation task, (2) grounding text generation additionally in the audio modality, and (3) conditioning on available external knowledge (e.g. a predefined list of possible slot values). We show that combining both modalities within the KA2G framework improves the robustness against ASR errors. Further, the knowledge-aware slot-value generator in KA2G, implemented via a pointer generator mechanism, particularly benefits few-shot and zero-shot learning. Experiments, conducted on the standard speech-based single-turn SLURP dataset and a multi-turn dataset extracted from a commercial ToD system, display strong and consistent gains over prior work, especially in few-shot and zero-shot setups.

Mixture-of-languages Routing for Multilingual Dialogues

We consider multilingual dialogue systems and ask how the performance of a dialogue system can be improved by using information that is available in other languages than the language in which a conversation is being conducted. We adopt a collaborative chair-experts framework, where each expert agent can be either monolingual or cross-lingual, and a chair agent follows a mixture-of-experts procedure for globally optimizing multilingual task-oriented dialogue systems. We propose a mixture-of-languages routing framework that includes four functional components, i.e., input embeddings of multilingual dialogues, language model, pairwise alignment between the representation of every two languages, and mixture-of-languages. We quantify language characteristics of unity and diversity using a number of similarity metrics, i.e., genetic similarity, and word and sentence similarity based on embeddings. Our main finding is that the performance of multilingual task-oriented dialogue systems can be greatly impacted by three key aspects, i.e., data sufficiency, language characteristics, and model design in a mixture-of-languages routing framework.

TCLNet: Turn-level contrastive learning network with reranking for dialogue state tracking

Dialogue state tracking (DST) plays a crucial role in task-oriented dialogue systems, as it interprets and tracks user intentions throughout the dialogue. The accuracy of DST directly impacts system efficiency and user experience. While existing generation-based DST models have shown promising results, effectively modelling long dialogue sequences remains a challenge. Furthermore, the greedy search decoding strategy employed by most previous works for speed is suboptimal because the probability of ground truth tokens is not always the highest. In this study, we propose a novel learning framework that addresses these issues by incorporating turn-level contrastive learning and a reranking module. Specifically, we adopt turn-level contrastive learning to progressively rectify the intermediate states of lengthy dialogues by contrasting data points at a finer granularity. Additionally, the integration of the reranking module empowers the model to make more reliable decisions at each time step by further considering alternative tokens with high probabilities. Experimental results demonstrate that our approach consistently enhances the model’s capability to handle long dialogue sequences and achieves superior performance over strong baselines on both the MultiWOZ 2.1 and MultiWOZ 2.2 datasets.

Prompting large language models for user simulation in task-oriented dialogue systems

Large Language Models (LLMs) have gained widespread popularity due to their instruction-following abilities. In this study, we evaluate their ability in simulating user interactions for task-oriented dialogue (TOD) systems. Our findings demonstrate that prompting LLMs reveals their promising capabilities for training and testing dialogue policies, eliminating the need for domain expertise in crafting complex rules or relying on large annotated data, as required by traditional simulators. The results show that the dialogue system trained with the ChatGPT simulator achieves a success rate of 59%, comparable to a 62% success rate of the dialogue system trained with the manual-rules, agenda-based user simulator (ABUS). Furthermore, the dialogue system trained with the ChatGPT simulator demonstrates better generalization ability compared to the dialogue system trained with the ABUS. Its success rate outperforms that of the dialogue system trained with the ABUS by 4% on GenTUS, 5% on the ChatGPT Simulator, and 3% on the Llama simulator. Nevertheless, LLM-based user simulators provide challenging environment, lexically rich, diverse, and random responses. Llama simulator outperforms the human reference in all lexical diversity metrics with a margin of 0.66 in SE, 0.39 in CE, 0.01 in MSTTR, 0.04 in HDD, and 0.55 in MTLD, while the ChatGPT simulator achieves comparable results. This ultimately contributes to enhancing the system’s ability to generalize more effectively.

Position-aware Interactive Attention Network for multi-intent spoken language understanding

Spoken Language Understanding (SLU) is a crucial component of task-oriented dialog systems. In recent years, there has been increasing attention on multi-intent SLU due to its relevance to complex real-world applications. Most existing joint models only utilize multi-intent information to guide the slot filling, with only a small number of models achieving bidirectional interaction. Additionally, unlike traditional single-intent SLU, multi-intent SLU is scope-dependent, where each intent in a sentence has its specific dependency range. However, current bidirectional joint models often employ a pipeline approach to implement interaction between the two sub-tasks, which fails to fully leverage the semantic information between them and can lead to error propagation. Moreover, attention-based multi-intent joint models do not adequately model the positional relationships between words, resulting in suboptimal overall performance. In this paper, we propose a novel multi-intent SLU model called Position-aware Interactive Attention Network (PIAN), which consists of interactive attention and rotary position embedding. The aim of PIAN is to fully exploit the correlation between the two sub-tasks and capturing the positional dependencies between words. This facilitates mutual guidance and enhancement between the two sub-tasks. Additionally, to address the issue of uncoordinated slot problem generated by traditional slot filling decoders, we employ an MCRF slot filling decoder to constrain the slot labels. We evaluate our model on two public multi-intent SLU datasets, and the experimental results demonstrate that our model achieves state-of-the-art performance on key metrics. Code for this paper is publicly available at https://github.com/puhahahahahaha/SLU_with_Co_PRoE.

STOD: Towards Scalable Task-Oriented Dialogue System on MultiWOZ-API

Task-oriented dialogue systems (TODs) enable users to complete specific goals and are widely used in practice. Although existing models have achieved delightful performance for single-domain dialogues, scalability to new domains is far from well explored. Traditional dialogue systems rely on domain-specific information like dialogue state and database (DB), which limits the scalability of such systems. In this paper, we propose a Scalable Task-Oriented Dialogue modeling framework (STOD). Instead of labeling multiple dialogue components, which have been adopted by previous work, we only predict structured API queries to interact with DB and generate responses based on the complete DB results. Further, we construct a new API-schema-based TOD dataset MultiWOZ-API with API query and DB result annotation based on MultiWOZ 2.1. We then propose MSTOD and CSTOD for multi-domain and cross-domain TOD systems, respectively. We perform extensive qualitative experiments to verify the effectiveness of our proposed framework. We find the following. (1) Scalability across multiple domains: MSTOD achieves 2% improvements than the previous state-of-the-art in the multi-domain TOD. (2) Scalability to new domains: our framework enables satisfying generalization capability to new domains, a significant margin of 10% to existing baselines.

Modeling the impact of out-of-schema questions in task-oriented dialog systems

Existing work on task-oriented dialog systems generally assumes that the interaction of users with the system is restricted to the information stored in a closed data schema. However, in practice users may ask ‘out-of-schema’ questions, that is, questions that the system cannot answer, because the information does not exist in the schema. Failure to answer these questions may lead the users to drop out of the chat before reaching the success state (e.g. reserving a restaurant). A key challenge is that the number of these questions may be too high for a domain expert to answer them all. We formulate the problem of out-of-schema question detection and selection that identifies the most critical out-of-schema questions to answer, in order to maximize the expected success rate of the system. We propose a two-stage pipeline to solve the problem. In the first stage, we propose a novel in-context learning (ICL) approach to detect out-of-schema questions. In the second stage, we propose two algorithms for out-of-schema question selection (OQS): a naive approach that chooses a question based on its frequency in the dropped-out conversations, and a probabilistic approach that represents each conversation as a Markov chain and a question is picked based on its overall benefit. We propose and publish two new datasets for the problem, as existing datasets do not contain out-of-schema questions or user drop-outs. Our quantitative and simulation-based experimental analyses on these datasets measure how our methods can effectively identify out-of-schema questions and positively impact the success rate of the system.

Leveraging intent–entity relationships to enhance semantic accuracy in NLU models

Natural Language Understanding (NLU) components are used in Dialog Systems (DS) to perform intent detection and entity extraction. In this work, we introduce a technique that exploits the inherent relationships between intents and entities to enhance the performance of NLU systems. The proposed method involves the utilization of a carefully crafted set of rules that formally express these relationships. By utilizing these rules, we effectively address inconsistencies within the NLU output, leading to improved accuracy and reliability. We implemented the proposed method using the Rasa framework as an NLU component and used our own conversational dataset AWPS to evaluate the improvement. Then, we validated the results in other three commonly used datasets: ATIS, SNIPS, and NLU-Benchmark. The experimental results show that the proposed method has a positive impact on the semantic accuracy metric, reaching an improvement of 12.6% in AWPS when training with a small amount of data. Furthermore, the practical application of the proposed method can easily be extended to other Task-Oriented Dialog Systems (T-ODS) to boost their performance and enhance user satisfaction.

From easy to hard: Improving personalized response generation of task-oriented dialogue systems by leveraging capacity in open-domain dialogues

A task-oriented dialogue system (TOD) is an important application of artificial intelligence. In the past few years, works on personalized TODs have attracted increased research attention and have seen much progress. The main challenge of such dialogue systems is finding ways to exploit user profiles under conditions of fixed and monolithic dialogue style. However, most of the existing works overlook the observation that the personalization capability of the dialogue system is fundamental and generic, and they treat all attributes of the user profile equally throughout the dialogue flows, which makes them inadequate for developing a well-performing personalized TOD. In this paper, we propose a two-stage learning framework equipped with GPT2 as a backbone to alleviate the above two problems. In the first learning phase, we finetune the GPT2 model on the personalized open-domain dialogues to preliminarily acquire personalization power. Then, we transfer this power to personalized task-oriented dialogues for the second stage of learning. After that, we enable the proposed model to be capable of the desired personalization capacity. Moreover, we present a dynamic profile fusion mechanism and an auxiliary task that detects which attributes contribute to the current utterance to facilitate the model’s performance. Eventually, we rewrite the attribute descriptions of user profiles in sentences to mitigate the consistency gap between the open-domain and task-oriented dialogues. The experimental results show that the proposed model achieves superior results compared to the state-of-the-art models on two versions of the Personalized bAbI dataset.

Winnie: Task-Oriented Dialog System with Structure-Aware Contrastive Learning and Enhanced Policy Planning

Pre-trained encoder-decoder models are widely applied in Task-Oriented Dialog (TOD) systems on the session level, mainly focusing on modeling the dialog semantic information. Dialogs imply structural information indicating the interaction among user utterances, belief states, database search results, system acts and responses, which is also crucial for TOD systems. In addition, for the system acts, additional pre-training and datasets are considered to improve their accuracies, undoubtedly introducing a burden. Therefore, a novel end-to-end TOD system named Winnie is proposed in this paper to improve the TOD performance. First, to make full use of the intrinsic structural information, supervised contrastive learning is adopted to narrow the gap in the representation space between text representations of the same category and enlarge the overall continuous representation margin between text representations of different categories in dialog context. Then, a system act classification task is introduced for policy optimization during fine-tuning. Empirical results show that Winnie substantially improves the performance of the TOD system. By introducing the supervised contrastive and system act classification losses, Winnie achieves state-of-the-art results on benchmark datasets, including MultiWOZ2.2, In-Car, and Camrest676. Their end-to-end combined scores are improved by 3.2, 1.9, and 1.1 points, respectively.

From Retrieval to Generation: A Simple and Unified Generative Model for End-to-End Task-Oriented Dialogue

Retrieving appropriate records from the external knowledge base to generate informative responses is the core capability of end-to-end task-oriented dialogue systems (EToDs). Most of the existing methods additionally train the retrieval model or use the memory network to retrieve the knowledge base, which decouples the knowledge retrieval task from the response generation task, making it difficult to jointly optimize and failing to capture the internal relationship between the two tasks. In this paper, we propose a simple and unified generative model for task-oriented dialogue systems, which recasts the EToDs task as a single sequence generation task and uses maximum likelihood training to train the two tasks in a unified manner. To prevent the generation of non-existent records, we design the prefix trie to constrain the model generation, which ensures consistency between the generated records and the existing records in the knowledge base. Experimental results on three public benchmark datasets demonstrate that our method achieves robust performance on generating system responses and outperforms the baseline systems. To facilitate future research in this area, the code is available at https://github.com/dzy1011/Uni-ToD.

TA&AT: Enhancing Task-Oriented Dialog with Turn-Level Auxiliary Tasks and Action-Tree Based Scheduled Sampling

Task-oriented dialog systems have witnessed substantial progress due to conversational pre-training techniques. Yet, two significant challenges persist. First, most systems primarily utilize the latest turn's state label for the generator. This practice overlooks the comprehensive value of state labels in boosting the model's understanding for future generations. Second, an overreliance on generated policy often leads to error accumulation, resulting in suboptimal responses when adhering to incorrect actions. To combat these challenges, we propose turn-level multi-task objectives for the encoder. With the guidance of essential information from labeled intermediate states, we establish a more robust representation for both understanding and generation. For the decoder, we introduce an action tree-based scheduled sampling technique. Specifically, we model the hierarchical policy as trees and utilize the similarity between trees to sample negative policy based on scheduled sampling, hoping the model to generate invariant responses under perturbations. This method simulates potential pitfalls by sampling similar negative policy, bridging the gap between task-oriented dialog training and inference. Among methods without continual pre-training, our approach achieved state-of-the-art (SOTA) performance on the MultiWOZ dataset series and was also competitive with pre-trained SOTA methods.

OSTOD: One-Step Task-Oriented Dialogue with activated state and retelling response

As of present, the progress of conversational AI research has been greatly propelled by large-scale pre-trained language models. In particular, task-oriented dialogue systems have gained widespread attention owing to their immense potential in helping individuals accomplish diverse objectives, such as booking hotels, making restaurant reservations, and purchasing train tickets. In the past, task-oriented dialogue systems were typically viewed as a multi-step process that included spoken language understanding, dialogue state tracking, dialogue policy learning, and natural language generation. More recently, large-scale pre-trained language models enables the development of end-to-end neural pipeline task-oriented dialogue systems, which combine multiple steps into a single model, allowing for joint optimization and preventing error propagation. However, in order to explicitly retrieve information from databases to ensure the interpretability of the system, almost all end-to-end neural pipeline methods inevitably require predicting dialogue state as an intermediate result specialized for the domain or task, which results in significant challenges for generalization. To solve the problem above, we propose One-Step Task-Oriented Dialogue (OSTOD) in this paper, which models task-oriented dialogue by synchronously generating activated states and retelling responses, where activated states refer to slot values that contribute to database access, and retelling responses are system responses that contain activated state information. Specifically, first, automatic methods are designed to build data containing activated states and retelling responses. Then, a joint generation model that synchronously predicts activated states and retelling responses in a single step is proposed for task-oriented dialogue modelling. Based on empirical results obtained from the MultiWOZ 2.0 and MultiWOZ 2.1 datasets, our OSTOD model demonstrates comparable performance to state-of-the-art baselines. Moreover, our model exhibits exceptional generalization capabilities in few-shot learning and domain transfer scenarios.

An open intent detection model optimized for datasets based on the Bert large model

Within current task-oriented dialogue systems, the focus of intent detection predominantly centers on closed domains. Nevertheless, in real-world usage scenarios, a substantial proportion of interactions fall into the open-domain category. User intentions frequently transcend predefined boundaries, giving rise to a multitude of out-of-domain intents, which pose a formidable challenge to existing models, ultimately leading to diminished recognition rates and accuracy. The demand for open intent detection models is increasing in today's society to address this issue effectively. This paper proposes a method to optimize datasets, thereby enhancing the training accuracy of open intent detection models. Specifically, this paper employs the Adaptive Decision Boundary Learning algorithm, which is currently popular in open intent detection. Leveraging this algorithm, this paper suggests using the K-means clustering algorithm to refine the intent labels within the dataset. This process helps identify and remove outliers in the dataset, making the distinction between known domain and open-domain intent labels more precise. Experimental results on two datasets, banking77 and stackoverflow, demonstrate the effectiveness of our approach in significantly improving model accuracy.