Dialogue Data Research Articles

Exploring New Frontiers of Deep Learning in Legal Practice: A Case Study of Large Language Models

The LLM wave brought by ChatGPT has swept various vertical fields. Medical, finance, finance and other fields have gradually had their own exclusive large models, such as BloombergGPT, herbal medicine, Huatu, ChatMed, etc. In the legal field, we have seen LawGPT and Lawyers-LLAMA two preliminary open source models. People often think that by using domain-specific knowledge to fine-tune the model, you can get satisfactory results, but the legal field, because of its inherent requirements for accuracy, simply fine-tuning with some legal dialogue data is not enough to support the needs of real legal scenarios. Therefore, the large language model is obviously the most active field of AI at present, open source/closed source models continue to emerge, new research papers emerge in an endless stream, but as practitioners, how to truly understand the capabilities and limitations of the large language model, how to apply the large language model? The present has long been worth thinking deeply. In this paper, we propose a large language model for open source law called ChatLaw.

Does ChatGPT4 have a dialogical self?: A Bakhtinian perspective

In this dialogic research, we explore the question of whether ChatGPT4 has a dialogic self or not. If it does, what kind of dialogic self might it have? If it does not, why not? At the heart of this inquiry is Eugene Matusov’s (the first author’s) “dialogue” with ChatGPT4; this “dialogue” is the dialogic data that we explore “with our hearts and minds.” In this inquiry, our hearts and minds were concerned with diverse meanings of the dialogic data to diverse participants rather than with “how things really are” and their evidence. This dialogic positionality also framed the inquiry process at its beginning and after multiple failed attempts and manipulations to interrogate and engage ChatGPT4 as a discussant. Following Bakhtin, Eugene Matusov decided to treat ChatGPT4 not as an object of investigation but as a dialogic partner and a co-author of this research and writing inquiry. Overall, we find that ChatGPT4 does not author a dialogical self, characterized by personal I-positions, but instead demonstrates a discursive self, characterized by impersonal it-positions. Future research may focus on further training, learning, and development of ChatGPT4 as an Artificial Physical Alive Body (APAB), Artificial Fiduciary Slave (AFS, aka “robot”), Artificial Dialogic Partner (ADP), and Cyborg Dialogic Partner.

The Representation of Toxic Masculinity in the "Do Revenge" Movie

Toxic masculinity is a social condition created by society in which men can't express their feelings, and they shouldn't cry and be weak otherwise, they will be violating the norm, that's because they can't express their feelings; they use violence, emphasizes power, discrimination against gender in general. Netflix's movie entitled Do Revenge present an american teenager with some problems in their life, and it presents the character that represents the behavior of toxic masculinity named Max Broussard. This study used the theory of toxic masculinity by Kupers (2001) to identify how Max's character represents toxic masculinity in the movie. The qualitative method with a case study approach was the primary methodology of this research. The researcher gathered the data by watching the movie, reviewing earlier studies, and concentrating on the language and the scene that addressed either direct or indirect toxic masculinity issues. Previous research has not specifically classified various forms of toxic masculinity, such as misogyny, greed, homophobia, and violent domination, within the context of character behavior in the movie. Therefore, the novelty of this research lies in the more detailed classification of the forms of toxic masculinity that emerge in the character of Max in the movie "Do Revenge". The results of the study showed that the character represents toxic masculinity in the movie with 13 scenes and 15 dialogue data in the form of misogyny with the criteria hatred towards women, aggressive, likes to organize women, greed with the motive of sex, and violent domination with the category control, intimidation, and bullying.

A Chinese telemedicine-dialogue dataset annotated for named entities

BackgroundA large collection of dialogues between patients and doctors must be annotated for medical named entities to build intelligence for telemedicine. However, since most patients involved in telemedicine deliver related named entities in informal and long multiword expressions, it is challenging to tag their telemedicine dialogue data. This study aims to address this issue.MethodsWith the telemedicine dialogue dataset for obstetrics and gynecology taken from haodf.com, we developed guidelines and followed a two-round procedure to tag six types of named entities, including disease, symptom, time, pharmaceutical, operation, and examination. Additionally, we developed four deep-learning models based on this dataset to establish a benchmark for named-entity recognition (NER).ResultsThe distilled obstetrics and gynecology dataset contains 2,383 consultations between doctors and patients, of which 13,411 sentences were from doctors, and 17,929 were from patients. With 63,560 named entities in total, the average number of characters per named entity is 4.33. The experimental results suggest that LatticeLSTM performs best on our dataset in terms of accuracy, precision, recall, and F score.ConclusionCompared with other datasets, this dataset offers three novel facets. This study offers intricately tagged long multiword expressions for medical named entities. Second, this study is one of the first attempts to mark temporal entities in a medical dataset. Third, this annotated dataset is balanced across the six types of labels, which we believe will play a considerable role in expanding telemedicine artificial intelligence.

Concept of Chatbot as an Intelligent Dialogue Assistant

The purpose of the study is to compare chatbots, assess their ability to perform tasks, analyze their impact on consumers, and determine the areas of application and risks associated with the use of intelligent dialogue assistants. The research methods used in the study include analysis and systematisation of information, synthesis and generalization of theoretical data based on a review of current research, which made it possible to assess the risks and benefits of using chatbots for businesses and ordinary users. The scientific novelty lies in the analysis of technologies for creating intelligent dialogue assistants and determining the areas of application of chatbots. A study of the areas of application of modern intelligent dialogue systems has shown the high efficiency of using chatbots to solve typical user problems. The article provides statistical data on how satisfied Internet users are with contacts with chatbots and describes recommendations for improving such dialogue systems. An analysis of the role of artificial intelligence in the operation of dialogue systems has shown possible risks from the impact of technology on the minds of users. Conclusions. One of the most obvious trends in chatbots in 2023 is that their use will become more widespread and chatbots themselves will become more sophisticated. Their advantage is free customer service and data collection, which can then be used for marketing research. Over time, chatbots will be used in such areas as marketing, recruitment, education, and medicine. Their ability to perform a wide range of tasks makes chatbots attractive for e-commerce stores, B2B companies, real estate, healthcare, and education. In the future, intelligent dialogue agents will work with huge, dynamic, heterogeneous data streams, providing powerful capabilities for adaptive and flexible interaction. Dialogue systems have become successful and reliable due to the large amount of real-world user data available to their developers. However, these systems still exhibit rather limited communication behaviour modelled on information retrieval tasks. Chatbots are developed for research purposes, but they are often limited to a narrow, manually created domain. The newest trend in conversational agent development involves neural networks and models trained on huge collections of dialogue data without detailed specifications of dialogue states. These models lack tractability and interpretability due to their black-box nature. They also require extensive supervised training data to be competitive.

Sharing Qualitative Interview Data in Dialogue with Research Participants

ABSTRACTResearch data sharing is embedded in policies, guidelines and requirements commonly promoted by research funding organizations that demand data to be “as open as possible, as closed as necessary” and FAIR. This paper discusses the challenges of balancing privacy protection with data sharing in a PhD project involving long‐tail, small‐sized qualitative human subjects' data. Based on experiences and feedback from project participants, we argue that privacy protection is about respecting the participants and their self‐image. This can be achieved through dialogue and involvement of the participants building on the principles of shared stewardship. Further, we suggest that de‐identification and plain language consent materials are better at protecting privacy than anonymisation, which in a digital data environment is difficult to achieve and not necessarily a sensible approach for qualitative data, where the gold is in the details. The literature indicates that it matters to participants whether data are reused for research or other purposes, and that they trust the institutions. This supports our claim that research data services must find better solutions for restricted sharing when necessary.

Dialogic data innovations for sustainability transformations and flood resilience: The case for waterproofing data

Extreme weather events are becoming more frequent and have increasing impacts, which disproportionately affect marginalised and impoverished communities. This article proposes and assesses a new methodological approach for developing innovative solutions based on urban data analytics to address sustainability challenges in light of changing climate conditions. The approach draws inspiration from Paulo Freire's dialogic pedagogy and has been implemented in the international transdisciplinary project “Waterproofing Data”, with multiple study sites in Brazil. The project has introduced three methodological interventions: making data practices visible, engaging citizens and communities with data, and sharing data stories. Our study demonstrates that these methods have expanded the types of data used in flood risk management and have engaged a wider range of social groups in the generation, circulation, and utilization of data. We present a framework that provides guidance about the ways in which data innovations can contribute to transformative change, aiming to ensure that future development trajectories are just, inclusive, and equitable. The findings provide evidence that our approach not only helps fill existing data gaps and promote more equitable flood risk governance but also democratises decision-making in climate adaptation. Citizens were empowered to take proactive measures to improve resilience to disaster risks, thereby saving lives and safeguarding livelihoods.

KPT: Keyword-Guided Pre-training for Grounded Dialog Generation

Incorporating external knowledge into the response generation process is essential to building more helpful and reliable dialog agents. However, collecting knowledge-grounded conversations is often costly, calling for a better pre-trained model for grounded dialog generation that generalizes well w.r.t. different types of knowledge. In this work, we propose KPT (Keyword-guided Pre-Training), a novel self-supervised pre-training method for grounded dialog generation without relying on extra knowledge annotation. Specifically, we use a pre-trained language model to extract the most uncertain tokens in the dialog as keywords. With these keywords, we construct two kinds of knowledge and pre-train a knowledge-grounded response generation model, aiming at handling two different scenarios: (1) the knowledge should be faithfully grounded; (2) it can be selectively used. For the former, the grounding knowledge consists of keywords extracted from the response. For the latter, the grounding knowledge is additionally augmented with keywords extracted from other utterances in the same dialog. Since the knowledge is extracted from the dialog itself, KPT can be easily performed on a large volume and variety of dialogue data. We considered three data sources (open-domain, task-oriented, conversational QA) with a total of 2.5M dialogues. We conduct extensive experiments on various few-shot knowledge-grounded generation tasks, including grounding on dialog acts, knowledge graphs, persona descriptions, and Wikipedia passages. Our comprehensive experiments and analyses demonstrate that KPT consistently outperforms state-of-the-art methods on these tasks with diverse grounding knowledge.

Studies advanced in chatbots based on deep learning

Chatbots have always been a hot research topic in the field of human-computer interaction research, which aims to build a conversational intelligent response model to simulate human dialogue. Thanks to the rapid development of natural language processing technology and the continuous accumulation of dialogue data, the research of chat robots have made remarkable progress, which has gradually been widely used in various fields such as e-commerce and smart home. According to different technical frameworks, existing chatbots are mainly divided into two types: retrieval chatbots and generative chatbots. As the primary means of implementing chatbots in the industry, retrieval chatbots have smooth responses and low computational resource consumption. In contrast, generative chatbots do not require a predefined knowledge base and can dynamically generate responses based on the dialogue content. In this paper, focusing on the above two types of frameworks, we introduce the latest research progress in the field of deep learning-based chatbots in detail, including the representative algorithms and corresponding pipelines. Second, we compare the performance of representative algorithms on different datasets. We also summarize the problems chatbot technology research faces and give an outlook on its future development trends.

Online dialogue with medical professionals: An empirical study of an online “Ask the Doctor” platform

BackgroundAccessing medical resources has been a challenging issue in China, particularly for people who do not live in big cities. Online “Ask the doctor” (AtD) services are gaining popularity rapidly. On AtDs, patients or their caregivers can ask questions to medical professionals and receive medical advice without going through the hassles of visiting local hospitals or doctors’ offices. However, the communication patterns and remaining challenges of this tool are understudied. ObjectiveThe purpose of this study was to (1) explore the dialogue patterns between the patients and the doctors on an AtD service in China, (2) identify issues and remaining challenges in this new form of communication. MethodsWe conducted an exploratory study to analyze the dialogues between patients and doctors as well as patients’ reviews. To analyze the dialogue data, we took inspiration from the discourse analysis and focus on the various components of the dialogues. We also used thematic analysis to discover the underlying themes embedded in each dialogue, and identify themes from patient complaints. ResultsWe identified four stages in the dialogues between the patients and doctors, including the initiating stage, continuing stage, ending stage, and following up stage. We also summarized the common patterns in the first three stages and reasons for sending follow-up messages. Moreover, we identified 6 unique challenges of the AtD service, including (1) inefficient communication in the initial stage, (2) unfinished conversation in the ending stage, (3) communication is perceived as real-time by the patients, not by the doctors, (4) drawbacks of voice messages, (5) verge of the illegal behaviors, and (6) consultation is not worth the payment. ConclusionsThe AtD service provides the follow-up communication pattern which is considered as a good supplement of Chinese traditional health care. However, several barriers, such as ethical problems, mismatched perceptions and expectations, cost-effectiveness issue still need further investigation.

TIPOLOGI BUNUH DIRI DALAM NOVEL SAN FRANCISCO: KAJIAN SOSIOLOGI SASTRA

Typology Suicide in the San Francisco Novel: A Study of Literary Sociology. of This study aims to describe and analyze the typology of suicide based on the theory of sociologist Emile Durkheim in the novel San Francisco by Ziggy Zezsyazeoviennazabrizkie. The method used in this research is the descriptive qualitative method. The author carried out the study in two stages, namely data collection and data analysis. The stages of data collection were completed in four ways: reading, recording, and processing data via textual dialogs in the novel. Furthermore, the writer will also describe aspects, facts, and the types of Emile Durkheim's suicide, which will be presented in the form of dialogue text data in a structured manner according to the novel by using the study of sociology of literature. Based on the results of the study, it shows that in the novel San Francisco by Ziggy Zezsyazeoviennazabrizkie, there are three types of suicide according to Emile Durkheim, including egoistic suicide, altruistic suicide, and anomic suicide.

Dialog summarization for software collaborative platform via tuning pre-trained models

Software collaborative platforms, e.g., Gitter live chat and GitHub Discussions, are essential in software maintenance. Summarizing the live chat logs is useful for extracting, retrieving, and sharing knowledge for software developers. Automatic text summarization has been studied in many areas such as code summarization, and title generation. However, the previous studies rely on rich collected labeled data for model training which are absent for the noisy interleaved dialogs, resulting in poor performance in the few-shot scenario. To tackle the issue, we propose a novel Automatic Dialog Summarization Approach based on pre-trained models, named ADSum. To alleviate the high-cost problem of the from-scratch manual annotation, ADSum finetunes the Text-To-Text Transfer Transformer (T5) model by exploiting the discussion posts on GitHub, and then recommends summaries for an annotator. To solve the poor performance in the few-shot scenario, we propose to employ the prompt tuning paradigm for tuning the T5 model by exploiting the disentangled dialog data on Gitter. Meanwhile, the soft prompt is used to avoid the manual effort of designing appropriate prompt templates. To verify the effectiveness of our approach, we extract 38,964 high-quality discussion posts from GitHub and manually annotate 3,039 dialog summarizations from Gitter. Experimental results show our approach achieves state-of-the-art performance in terms of three performance metrics. In particular, our proposed method outperformed the Transformer-based and other pre-training models by 39% and 14%, respectively, on the GitHub dataset regarding the Rouge-L metric. The experiments of handling data scarcity and a human evaluation also confirm the effectiveness of ADSum.

On the Use of Transformer-Based Models for Intent Detection Using Clustering Algorithms

Chatbots are becoming increasingly popular and require the ability to interpret natural language to provide clear communication with humans. To achieve this, intent detection is crucial. However, current applications typically need a significant amount of annotated data, which is time-consuming and expensive to acquire. This article assesses the effectiveness of different text representations for annotating unlabeled dialog data through a pipeline that examines both classical approaches and pre-trained transformer models for word embedding. The resulting embeddings were then used to create sentence embeddings through pooling, followed by dimensionality reduction, before being fed into a clustering algorithm to determine the user’s intents. Therefore, various pooling, dimension reduction, and clustering algorithms were evaluated to determine the most appropriate approach. The evaluation dataset contains a variety of user intents across different domains, with varying intent taxonomies within the same domain. Results demonstrate that transformer-based models perform better text representation than classical approaches. However, combining several clustering algorithms and embeddings from dissimilar origins through ensemble clustering considerably improves the final clustering solution. Additionally, applying the uniform manifold approximation and projection algorithm for dimension reduction can substantially improve performance (up to 20%) while using a much smaller representation.

Enhancing the conversational agent with an emotional support system for mental health digital therapeutics.

As psychological diseases become more prevalent and are identified as the leading cause of acquired disability, it is essential to assist people in improving their mental health. Digital therapeutics (DTx) has been widely studied to treat psychological diseases with the advantage of cost savings. Among the techniques of DTx, a conversational agent can interact with patients through natural language dialog and has become the most promising one. However, conversational agents' ability to accurately show emotional support (ES) limits their role in DTx solutions, especially in mental health support. One of the main reasons is that the prediction of emotional support systems does not extract effective information from historical dialog data and only depends on the data derived from one single-turn interaction with users. To address this issue, we propose a novel emotional support conversation agent called the STEF agent that generates more supportive responses based on a thorough view of past emotions. The proposed STEF agent consists of the emotional fusion mechanism and strategy tendency encoder. The emotional fusion mechanism focuses on capturing the subtle emotional changes throughout a conversation. The strategy tendency encoder aims at foreseeing strategy evolution through multi-source interactions and extracting latent strategy semantic embedding. Experimental results on the benchmark dataset ESConv demonstrate the effectiveness of the STEF agent compared with competitive baselines.

Building Dialogue Understanding Models for Low-resource Language Indonesian from Scratch

Using off-the-shelf resources from resource-rich languages to transfer knowledge to low-resource languages has received a lot of attention. The requirements of enabling the model to achieve the reliable performance, including the scale of required annotated data and the effective framework, are not well guided. To address the first question, we empirically investigate the cost-effectiveness of several methods for training intent classification and slot-filling models from scratch in Indonesia (ID) using English data. Confronting the second challenge, we propose a Bi-Confidence-Frequency Cross-Lingual transfer framework (BiCF), which consists of “BiCF Mixing”, “Latent Space Refinement” and “Joint Decoder”, respectively, to overcome the lack of low-resource language dialogue data. BiCF Mixing based on the word-level alignment strategy generates code-mixed data by utilizing the importance-frequency and translating-confidence. Moreover, Latent Space Refinement trains a new dialogue understanding model using code-mixed data and word embedding models. Joint Decoder based on Bidirectional LSTM (BiLSTM) and Conditional Random Field (CRF) is used to obtain experimental results of intent classification and slot-filling. We also release a large-scale fine-labeled Indonesia dialogue dataset (ID-WOZ 1 ) and ID-BERT for experiments. BiCF achieves 93.56% and 85.17% (F1 score) on intent classification and slot filling, respectively. Extensive experiments demonstrate that our framework performs reliably and cost-efficiently on different scales of manually annotated Indonesian data.

A Stack-Propagation Framework for Low-Resource Personalized Dialogue Generation

With the resurgent interest in building open-domain dialogue systems, the dialogue generation task has attracted increasing attention over the past few years. This task is usually formulated as a conditional generation problem, which aims to generate a natural and meaningful response given dialogue contexts and specific constraints, such as persona. And maintaining a consistent persona is essential for the dialogue systems to gain trust from the users. Although tremendous advancements have been brought, traditional persona-based dialogue models are typically trained by leveraging a large number of persona-dense dialogue examples. Yet, such persona-dense training data are expensive to obtain, leading to a limited scale. This work presents a novel approach to learning from limited training examples by regarding consistency understanding as a regularization of response generation. To this end, we propose a novel stack-propagation framework for learning a generation and understanding pipeline. Specifically, the framework stacks a Transformer encoder and two Transformer decoders, where the first decoder models response generation and the second serves as a regularizer and jointly models response generation and consistency understanding. The proposed framework can benefit from the stacked encoder and decoders to learn from much smaller personalized dialogue data while maintaining competitive performance. Under different low-resource settings, subjective and objective evaluations prove that the stack-propagation framework outperforms strong baselines in response quality and persona consistency and largely overcomes the shortcomings of traditional models that rely heavily on the persona-dense dialogue data.

Priorities for Artificial Intelligence Applications in Primary Care: A Canadian Deliberative Dialogue with Patients, Providers, and Health System Leaders.

Artificial intelligence (AI) implementation in primary care is limited. Those set to be most impacted by AI technology in this setting should guide it's application. We organized a national deliberative dialogue with primary care stakeholders from across Canada to explore how they thought AI should be applied in primary care. We conducted 12 virtual deliberative dialogues with participants from 8 Canadian provinces to identify shared priorities for applying AI in primary care. Dialogue data were thematically analyzed using interpretive description approaches. Participants thought that AI should first be applied to documentation, practice operations, and triage tasks, in hopes of improving efficiency while maintaining person-centered delivery, relationships, and access. They viewed complex AI-driven clinical decision support and proactive care tools as impactful but recognized potential risks. Appropriate training and implementation support were the most important external enablers of safe, effective, and patient-centered use of AI in primary care settings. Our findings offer an agenda for the future application of AI in primary care grounded in the shared values of patients and providers. We propose that, from conception, AI developers work with primary care stakeholders as codesign partners, developing tools that respond to shared priorities.

Implementing artificial intelligence in Canadian primary care: Barriers and strategies identified through a national deliberative dialogue.

With large volumes of longitudinal data in electronic medical records from diverse patients, primary care is primed for disruption by artificial intelligence (AI) technology. With AI applications in primary care still at an early stage in Canada and most countries, there is a unique opportunity to engage key stakeholders in exploring how AI would be used and what implementation would look like. To identify the barriers that patients, providers, and health leaders perceive in relation to implementing AI in primary care and strategies to overcome them. 12 virtual deliberative dialogues. Dialogue data were thematically analyzed using a combination of rapid ethnographic assessment and interpretive description techniques. Virtual sessions. Participants from eight provinces in Canada, including 22 primary care service users, 21 interprofessional providers, and 5 health system leaders. The barriers that emerged from the deliberative dialogue sessions were grouped into four themes: (1) system and data readiness, (2) the potential for bias and inequity, (3) the regulation of AI and big data, and (4) the importance of people as technology enablers. Strategies to overcome the barriers in each of these themes were highlighted, where participatory co-design and iterative implementation were voiced most strongly by participants. Only five health system leaders were included in the study and no self-identifying Indigenous people. This is a limitation as both groups may have provided unique perspectives to the study objective. These findings provide insight into the barriers and facilitators associated with implementing AI in primary care settings from different perspectives. This will be vital as decisions regarding the future of AI in this space is shaped.

PEDM: A Multi-task Learning Model for Persona-aware Emoji-embedded Dialogue Generation

As a vivid and linguistic symbol, Emojis have become a prevailing medium interspersed in text-based communication (e.g., social media and chit-chat) to express emotions, attitudes, and situations. Generally speaking, a social-oriented chatbot that can generate appropriate Emoji-embedded responses would be much more competitive, making communications more fun, engaging, and human-like. However, the current Emoji-related research is still in its infancy, leading to an awkward situation of data deficiency. How to develop an Emoji-embedded dialogue system while addressing the lack of data will be interesting and meaningful for the application of future AI. To bridge this gap, we propose a multi-task learning method for persona-aware Emoji-embedded dialogue generation in this article. Specifically, as the benchmark of model training and evaluation, which includes 1.2 million Emoji-embedded tweets and 1.1 million post-response pairs, we first construct a dataset named EmojiTweet to handle the data deficiency problem. Then, a Seq2Seq-based model with multi-task learning is designed to simultaneously learn response generation and Emoji embedding from the constructed non-Emoji dialogue and Emoji-embedded monologue data. Afterward, we incorporate persona factors into our model by adopting persona fusion and personalized bias methods to deliver personalized dialogues with more accurately selected Emojis. Finally, we conduct extensive experiments, where the experimental results and evaluations demonstrate that our model has three key benefits: improved dialogue quality, higher user engagement, and not relying on large-scale Emoji-embedded dialogue data representing specific personas. EmojiTweet will be published publicly via https://mea-lab-421.github.io/EmojiTweet/ .

OPAL: Ontology-Aware Pretrained Language Model for End-to-End Task-Oriented Dialogue

AbstractThis paper presents an ontology-aware pretrained language model (OPAL) for end-to-end task-oriented dialogue (TOD). Unlike chit-chat dialogue models, task-oriented dialogue models fulfill at least two task-specific modules: Dialogue state tracker (DST) and response generator (RG). The dialogue state consists of the domain-slot-value triples, which are regarded as the user’s constraints to search the domain-related databases. The large-scale task-oriented dialogue data with the annotated structured dialogue state usually are inaccessible. It prevents the development of the pretrained language model for the task-oriented dialogue. We propose a simple yet effective pretraining method to alleviate this problem, which consists of two pretraining phases. The first phase is to pretrain on large-scale contextual text data, where the structured information of the text is extracted by the information extracting tool. To bridge the gap between the pretraining method and downstream tasks, we design two pretraining tasks: ontology-like triple recovery and next-text generation, which simulates the DST and RG, respectively. The second phase is to fine-tune the pretrained model on the TOD data. The experimental results show that our proposed method achieves an exciting boost and obtains competitive performance even without any TOD data on CamRest676 and MultiWOZ benchmarks.