Existing works on dialogue discourse parsing mostly utilize encoder-only models and sophisticated decoding strategies to extract structures. Despite recent advances in Large Language Models (LLMs), there has been little work applying directly these models on discourse parsing. To fully utilize the rich semantic and discourse knowledge in LLMs, we explore the feasibility of transforming discourse parsing into a generation task using a text-to-text paradigm. Our approach is intuitive and requires no modification of the LLM architecture. Experimental results on STAC and Molweni datasets show that a sequence-to-sequence model such as T0 can perform reasonably well. Notably, our improved transition-based sequence-to-sequence system achieves new state-of-the-art performance on Molweni, demonstrating the effectiveness of the proposed method. Furthermore, our systems can generate richer discourse structures such as directed acyclic graphs, whereas previous methods are limited to trees.

More and more corpora are being annotated with Rhetorical Structure Theory (RST) trees, often in a multi-layer scenario, as analyzing RST annotations in combination with other layers can lead to a deeper understanding of texts. To date, prior work on RST for the analysis of diplomatic language however, is scarce. We are interested in political speeches and investigate what rhetorical strategies diplomats use to communicate critique or deal with disputes. To this end, we present a new dataset with RST annotations of 82 diplomatic speeches aligned to existing Conflict annotations (UNSC-RST). We explore ways of using rhetorical trees to analyze an annotated multi-layer corpus, looking at both the relation distribution and the tree structure of speeches. In preliminary analyses we already see patterns that are characteristic for particular topics or countries.

There are many strategies used to simplify texts. In this paper, we focus specifically on the act of inserting information or elaborative simplification. Adding information is done for various reasons, such as providing definitions for concepts, making relations between concepts more explicit, and providing background information that is a prerequisite for the main content. As all of these reasons have the main goal of ensuring coherence, we first conduct a corpus analysis of simplified German-language texts that have been annotated with Rhetorical Structure Theory (RST). We focus specifically on how additional information is incorporated into the RST annotation for a text. We then transfer these insights to automatic simplification using Large Language Models (LLMs), as elaborative simplification is a nuanced task which LLMs still seem to struggle with.

We propose a framework for analyzing discourse by combining two interdependent concepts from sociolinguistic theory: face acts and politeness. While politeness has robust existing tools and data, face acts are less resourced. We introduce a new corpus created by annotating Wikipedia talk pages with face acts and we use this to train a face act tagger. We then employ our framework to study how face and politeness interact with gender and power in discussions between Wikipedia editors. Among other findings, we observe that female Wikipedians are not only more polite, which is consistent with prior studies, but that this difference corresponds with significantly more language directed at humbling aspects of their own face. Interestingly, the distinction nearly vanishes once limiting to editors with administrative power.

Reference resolution is an important problem, one that is essential to understand and successfully handle contexts of different kinds. This context includes both previous turns and context that pertains to non-conversational entities, such as entities on the user’s screen or those running in the background. While LLMs have been shown to be extremely powerful for a variety of tasks, their use in reference resolution, particularly for non-conversational entities, remains underutilized. This paper demonstrates how LLMs can be used to create an effective system to resolve references of various types, by showing how reference resolution can be converted into a language modeling problem, despite involving forms of entities like those on screen that are not traditionally conducive to being reduced to a text-only modality. We demonstrate large improvements over an existing system with similar functionality across different types of references, with our smallest model obtaining absolute gains of over 5% for on-screen references. We also benchmark against GPT-3.5 and GPT-4, with our smallest model achieving performance comparable to that of GPT-4, and our larger models substantially outperforming it.

LLM-driven dialog systems are used in a diverse set of applications, ranging from healthcare to customer service. However, given their generalization capability, it is difficult to ensure that these chatbots stay within the boundaries of the specialized domains, potentially resulting in inaccurate information and irrelevant responses. This paper introduces an unsupervised approach for automatically inducing domain-specific dialog flows that can be used to constrain LLM-based chatbots. We introduce two variants of dialog flow based on the availability of in-domain conversation instances. Through human and automatic evaluation over 24 dialog domains, we demonstrate that our high-quality data-guided dialog flows achieve better domain coverage, thereby overcoming the need for extensive manual crafting of such flows.

Open domain spoken dialogue systems need to controllably generate many different dialogue acts (DAs) to allow Natural Language Generation (NLG) to create interesting and engaging conversational interactions with users. We aim to create an NLG engine that can produce a variety of DAs that make substantive knowledge-grounded contributions to a conversation. Training such an NLG typically requires dialogue corpora that are labelled for DAs, which are expensive to produce and vulnerable to quality issues. Here, we present a prompt-based learning approach to transfer DAs from one domain, video games, to 7 new domains. For each novel domain, we first crawl WikiData to create Meaning Representations that systematically vary both the number of attributes and hops on the WikiData Knowledge Graph. The proposed method involves a self-training step to create prompt examples for each domain followed by an overgeneration and ranking step. The result is a novel, high-quality dataset, Wiki-Dialogue, of 71K knowledge-grounded utterances, covering 9 DAs and the Art, Movies, Music, Sports, TV, Animal, and Boardgames domains, whose combined DA and semantic accuracy is 89%. We assess the corpus quality using both automatic and human evaluations and find it high. The corpus is found to be safe, lexically rich, and large in vocabulary, when compared to similar datasets.

In today’s industrial landscape, seamless collaboration between humans and machines is essential and requires a shared knowledge of the operational domain. In this framework, the technical knowledge for operator assistance has traditionally been derived from static sources such as technical documents. However, experienced operators hold invaluable know-how that can significantly contribute to support other operators. This work focuses on enhancing the operator assistance tasks in the manufacturing industry by leveraging spoken natural language interaction. More specifically, a Human-in-the-Loop (HIL) incremental learning approach is proposed to integrate this expertise into a domain knowledge graph (KG) dynamically, along with the use of in-context learning for Large Language Models (LLMs) to benefit other capabilities of the system. Preliminary results of the experimentation carried out in an industrial scenario, where the graph size was increased in a 25%, demonstrate that the incremental enhancing of the KG benefits the dialogue system’s performance.

The paper describes methods for anticipating follow-up questions in exploratory information search. There are two main cases: information stored in knowledge graphs, and information in unstructured texts such as Wikipedia. In the first case, follow-up questions are anticipated by extracting subgraphs relevant to user queries, passing the subgraphs to an LLM to generate responses. In the second case, entities and their relationships are extracted from the texts and added to short-term knowledge graphs relevant to initial queries. Follow-up questions are then anticipated by extracting subgraphs relevant to subsequent queries and passing the subgraphs to the LLM, as in the first case. The short-term graphs in dialogue memory are often sufficient to answer follow-up questions. If they are not, the described steps are repeated as required.

Knowledge models are fundamental to dialogue systems for enabling conversational interactions, which require handling domain-specific knowledge. Ensuring effective communication in information-providing conversations entails aligning user understanding with the knowledge available to the system. However, dialogue systems often face challenges arising from semantic inconsistencies in how information is expressed in natural language compared to how it is represented within the system’s internal knowledge. To address this problem, we study the potential of large language models for conversational grounding, a mechanism to bridge information gaps by establishing shared knowledge between dialogue participants. Our approach involves annotating human conversations across five knowledge domains to create a new dialogue corpus called BridgeKG. Through a series of experiments on this dataset, we empirically evaluate the capabilities of large language models in classifying grounding acts and identifying grounded information items within a knowledge graph structure. Our findings offer insights into how these models use in-context learning for conversational grounding tasks and common prediction errors, which we illustrate with examples from challenging dialogues. We discuss how the models handle knowledge graphs as a semantic layer between unstructured dialogue utterances and structured information items.

Educational dialogue systems have been used to support students and teachers for decades. Such systems rely on explicit pedagogically motivated dialogue rules. With the ease of integrating large language models (LLMs) into dialogue systems, applications have been arising that directly use model responses without the use of human-written rules, raising concerns about their use in classroom settings. Here, we explore how to constrain LLM outputs to generate appropriate and supportive teacher-like responses. We present results comparing the effectiveness of different constraint variations in a zero-shot prompting setting on a large mathematics classroom corpus. Generated outputs are evaluated with human annotation for Fluency, Relevance, Helpfulness, and Adherence to the provided constraints. Including all constraints in the prompt led to the highest values for Fluency and Helpfulness, and the second highest value for Relevance. The annotation results also demonstrate that the prompts that result in the highest adherence to constraints do not necessarily indicate higher perceived scores for Fluency, Relevance, or Helpfulness. In a direct comparison, all of the non-baseline LLM responses were ranked higher than the actual teacher responses in the corpus over 50% of the time.

A Time-Offset Interaction Application (TOIA) is a software system that allows people to engage in face-to-face dialogue with previously recorded videos of other people. There are two TOIA usage modes: (a) creation mode, where users pre-record video snippets of themselves representing their answers to possible questions someone may ask them, and (b) interaction mode, where other users of the system can choose to interact with created avatars. This paper presents the HelloThere corpus that has been collected from two user studies involving several people who recorded avatars and many more who engaged in dialogues with them. The interactions with avatars are annotated by people asking them questions through three modes (card selection, text search, and voice input) and rating the appropriateness of their answers on a 1 to 5 scale. The corpus, made available to the research community, comprises 26 avatars’ knowledge bases and 317 dialogues between 64 interrogators and the avatars in text format.

Active participation in a conversation is key to building common ground, since understanding is jointly tailored by producers and recipients. Overhearers are deprived of the privilege of performing grounding acts and can only conjecture about intended meanings. Still, data generation and annotation, modelling, training and evaluation of NLP dialogue models place reliance on the overhearing paradigm. How much of the underlying grounding processes are thereby forfeited? As we show, there is evidence pointing to the impossibility of properly modelling human meta-communicative acts with data-driven learning models. In this paper, we discuss this issue and provide a preliminary analysis on the variability of human decisions for requesting clarification. Most importantly, we wish to bring this topic back to the community’s table, encouraging discussion on the consequences of having models designed to only “listen in’”.

This study addresses the interaction challenges encountered by spoken dialogue systems (SDSs) when engaging with users who exhibit distinct conversational behaviors, particularly minors, in scenarios where data are scarce. We propose a novel data augmentation framework to enhance SDS performance for user groups with limited resources. Our approach leverages a large language model (LLM) to extract speaker styles and a pre-trained language model (PLM) to simulate dialogue act history. This method generates enriched and personalized dialogue data, facilitating improved interactions with unique user demographics. Extensive experiments validate the efficacy of our methodology, highlighting its potential to foster the development of more adaptive and inclusive dialogue systems.

Recent approaches for empathetic response generation mainly focus on emotional resonance and user understanding, without considering the system’s personality. Consistent personality is evident in real human expression and is important for creating trustworthy systems. To address this problem, we propose StyEmp, which aims to stylize the empathetic response generation with a consistent personality. Specifically, it incorporates a multi-grained prefix mechanism designed to capture the intricate relationship between a system’s personality and its empathetic expressions. Furthermore, we introduce a personality reinforcement module that leverages contrastive learning to calibrate the generation model, ensuring that responses are both empathetic and reflective of a distinct personality. Automatic and human evaluations on the EMPATHETICDIALOGUES benchmark show that StyEmp outperforms competitive baselines in terms of both empathy and personality expressions. Our code is available at https://github.com/fuyahuii/StyEmp.

Live streaming, a dynamic medium that merges real-time audiovisual content with interactive text-based chat, presents unique challenges for maintaining viewer engagement and ensuring streamers’ well-being. This study introduces a multi-criteria evaluation framework designed to identify response-worthy chats during live streaming. We proposed a system that evaluates chats based on sentiment polarity and intensity, contextual relevance, and topic uniqueness. We also constructed a dataset annotated by human reviewers who validates the framework, demonstrating a closer alignment with human preferences compared to single-criterion baselines. This framework not only supports the development of more responsive and engaging live streaming environments but also contributes to the broader field of dialog systems by highlighting the distinct needs of real-time, large-scale conversational contexts.

The demand for mental health services has risen substantially in recent years, leading to challenges in meeting patient needs promptly. Virtual agents capable of emulating motivational interviews (MI) have emerged as a potential solution to address this issue, offering immediate support that is especially beneficial for therapy modalities requiring multiple sessions. However, developing effective patient simulation methods for training MI dialog systems poses challenges, particularly in generating syntactically and contextually correct, and diversified dialog acts while respecting existing patterns and trends in therapy data. This paper investigates data-driven approaches to simulate patients for training MI dialog systems. We propose a novel method that leverages time series models to generate diverse and contextually appropriate patient dialog acts, which are then transformed into utterances by a conditioned large language model. Additionally, we introduce evaluation measures tailored to assess the quality and coherence of simulated patient dialog. Our findings highlight the effectiveness of dialog act-conditioned approaches in improving patient simulation for MI, offering insights for developing virtual agents to support mental health therapy.

Hedges allow speakers to mark utterances as provisional, whether to signal non-prototypicality or “fuzziness”, to indicate a lack of commitment to an utterance, to attribute responsibility for a statement to someone else, to invite input from a partner, or to soften critical feedback in the service of face management needs. Here we focus on hedges in an experimentally parameterized corpus of 63 Roadrunner cartoon narratives spontaneously produced from memory by 21 speakers for co-present addressees, transcribed to text (Galati and Brennan, 2010). We created a gold standard of hedges annotated by human coders (the Roadrunner-Hedge corpus) and compared three LLM-based approaches for hedge detection: fine-tuning BERT, and zero and few-shot prompting with GPT-4o and LLaMA-3. The best-performing approach was a fine-tuned BERT model, followed by few-shot GPT-4o. After an error analysis on the top performing approaches, we used an LLM-in-the-Loop approach to improve the gold standard coding, as well as to highlight cases in which hedges are ambiguous in linguistically interesting ways that will guide future research. This is the first step in our research program to train LLMs to interpret and generate collateral signals appropriately and meaningfully in conversation.

This paper investigates the enhancement of Dialogue Systems by integrating the creative capabilities of Large Language Models. While traditional Dialogue Systems focus on understanding user input and selecting appropriate system actions, Language Models excel at generating natural language text based on prompts. Therefore, we propose to improve controllability and coherence of interactions by guiding a Language Model with control signals that enable explicit control over the system behaviour. To address this, we tested and evaluated our concept in 815 conversations with over 3600 dialogue exchanges on a dataset. Our experiment examined the quality of generated system responses using two strategies: An unguided strategy where task data was provided to the models, and a controlled strategy in which a simulated Dialogue Controller provided appropriate system actions. The results show that the average BLEU score and the classification of dialogue acts improved in the controlled Natural Language Generation.

Ambiguous Candidate Identification(ACI) in multimodal dialogue is the task of identifying all potential objects that a user’s utterance could be referring to in a visual scene, in cases where the reference cannot be uniquely determined. End-to-end models are the dominant approach for this task, but have limited real-world applicability due to unrealistic inference-time assumptions such as requiring predefined catalogues of items. Focusing on a more generalized and realistic ACI setup, we demonstrate that a modular approach, which first emphasizes language-only reasoning over dialogue context before performing vision-language fusion, significantly outperforms end-to-end trained baselines. To mitigate the lack of annotations for training the language-only module (student), we propose a pseudo-labelling strategy with a prompted Large Language Model (LLM) as the teacher.

When engaging in conversations, dialogue agents in a virtual simulation environment may exhibit their own emotional states that are unrelated to the immediate conversational context, a phenomenon known as self-emotion. This study explores how such self-emotion affects the agents’ behaviors in dialogue strategies and decision-making within a large language model (LLM)-driven simulation framework. In a dialogue strategy prediction experiment, we analyze the dialogue strategy choices employed by agents both with and without self-emotion, comparing them to those of humans. The results show that incorporating self-emotion helps agents exhibit more human-like dialogue strategies. In an independent experiment comparing the performance of models fine-tuned on GPT-4 generated dialogue datasets, we demonstrate that self-emotion can lead to better overall naturalness and humanness. Finally, in a virtual simulation environment where agents have free discussions, we show that self-emotion of agents can significantly influence the decision-making process of the agents, leading to approximately a 50% change in decisions.

Explainable artificial intelligence (XAI) is a rapidly evolving field that seeks to create AI systems that can provide human-understandable explanations for their decision-making processes. However, these explanations rely on model and data-specific information only. To support better human decision-making, integrating domain knowledge into AI systems is expected to enhance understanding and transparency. In this paper, we present an approach for combining XAI explanations with domain knowledge within a dialogue system. We concentrate on techniques derived from the field of computational argumentation to incorporate domain knowledge and corresponding explanations into human-machine dialogue. We implement the approach in a prototype system for an initial user evaluation, where users interacted with the dialogue system to receive predictions from an underlying AI model. The participants were able to explore different types of explanations and domain knowledge. Our results indicate that users tend to more effectively evaluate model performance when domain knowledge is integrated. On the other hand, we found that domain knowledge was not frequently requested by the user during dialogue interactions.

Affect recognition, encompassing emotions, moods, and feelings, plays a pivotal role in human communication. In the realm of conversational artificial intelligence, the ability to discern and respond to human affective cues is a critical factor for creating engaging and empathetic interactions. This study investigates the capacity of large language models (LLMs) to recognise human affect in conversations, with a focus on both open-domain chit-chat dialogues and task-oriented dialogues. Leveraging three diverse datasets, namely IEMOCAP (Busso et al., 2008), EmoWOZ (Feng et al., 2022), and DAIC-WOZ (Gratch et al., 2014), covering a spectrum of dialogues from casual conversations to clinical interviews, we evaluate and compare LLMs’ performance in affect recognition. Our investigation explores the zero-shot and few-shot capabilities of LLMs through in-context learning as well as their model capacities through task-specific fine-tuning. Additionally, this study takes into account the potential impact of automatic speech recognition errors on LLM predictions. With this work, we aim to shed light on the extent to which LLMs can replicate human-like affect recognition capabilities in conversations.

Customer-support services increasingly rely on automation, whether fully or with human intervention. Despite optimising resources, this may result in mechanical protocols and lack of human interaction, thus reducing customer loyalty. Our goal is to enhance interpretability and provide guidance in communication through novel tools for easier analysis of message trends and sentiment variations. Monitoring these contributes to more informed decision-making, enabling proactive mitigation of potential issues, such as protocol deviations or customer dissatisfaction. We propose a generic approach for dialogue flow discovery that leverages clustering techniques to identify dialogue states, represented by related utterances. State transitions are further analyzed to detect prevailing sentiments. Hence, we discover sentiment-aware dialogue flows that offer an interpretability layer to artificial agents, even those based on black-boxes, ultimately increasing trustworthiness. Experimental results demonstrate the effectiveness of our approach across different dialogue datasets, covering both human-human and human-machine exchanges, applicable in task-oriented contexts but also to social media, highlighting its potential impact across various customer-support settings.

When customers present ambiguous references, service staff typically need to clarify the customers’ specific intentions. To advance research in this area, we collected 1,000 real-world consumer dialogues with ambiguous references. This dataset will be used for subsequent studies to identify ambiguous references and generate responses. Our analysis of the dataset revealed common strategies employed by service staff, including directly asking clarification questions (CQ) and listing possible options before asking a clarification question (LCQ). However, we found that merely using CQ often fails to fully satisfy customers. In contrast, using LCQ, as well as recommending specific products after listing possible options, proved more effective in resolving ambiguous references and enhancing customer satisfaction.

Despite the challenges posed by data sparsity in discourse parsing for dialogues, unsupervised methods have been underexplored. Leveraging recent advances in Large Language Models (LLMs), in this paper we investigate an unsupervised coherence-based method to build discourse structures for multi-party dialogues using open-source LLMs fine-tuned on conversational data. Specifically, we propose two algorithms that extract dialogue structures by identifying their most coherent sub-dialogues: DS-DP employs a dynamic programming strategy, while DS-FLOW applies a greedy approach. Evaluation on the STAC corpus demonstrates a micro-F1 score of 58.1%, surpassing prior unsupervised methods. Furthermore, on a cleaned subset of the Molweni corpus, the proposed method achieves a micro-F1 score of 74.7%, highlighting its effectiveness across different corpora.

The challenge of defining a slot schema to represent the state of a task-oriented dialogue system is addressed by Slot Schema Induction (SSI), which aims to automatically induce slots from unlabeled dialogue data. Whereas previous approaches induce slots by clustering value spans extracted directly from the dialogue text, we demonstrate the power of discovering slots using a generative approach. By training a model to generate slot names and values that summarize key dialogue information with no prior task knowledge, our SSI method discovers high-quality candidate information for representing dialogue state. These discovered slot-value candidates can be easily clustered into unified slot schemas that align well with human-authored schemas. Experimental comparisons on the MultiWOZ and SGD datasets demonstrate that Generative Dialogue State Inference (GenDSI) outperforms the previous state-of-the-art on multiple aspects of the SSI task.

This paper presents the development and capabilities of a spoken dialogue robot that uses respiration to enhance human-robot dialogue. By employing a respiratory estimation technique that uses video input, the dialogue robot captures user respiratory information during dialogue. This information is then used to prevent speech collisions between the user and the robot and to present synchronized pseudo-respiration with the user, thereby enhancing the smoothness and engagement of human-robot dialogue.

We developed an interface to explain news articles through dialogue by considering the user’s comprehension level. The interface generates several pertinent questions based on the ongoing dialogue and news article, and users advance the conversation by selecting a question. Based on the user’s selected questions, the interface estimates their comprehension level of the news article and adjusts the difficulty of the generated questions accordingly. This enables a personalized dialogue tailored to each user’s comprehension needs. The results of the baseline comparison experiments confirmed the usefulness of the interface.

Recent dialogue systems typically operate through turn-based spoken interactions between users and agents. These systems heavily depend on accurate Automatic Speech Recognition (ASR), as transcription errors can significantly degrade performance in downstream dialogue tasks. To alleviate this challenge, robust ASR is required, and one effective method is to utilize the dialogue context from user and agent interactions for transcribing the subsequent user utterance. This method incorporates the transcription of the user’s speech and the agent’s response as model input, using the accumulated context generated by each turn. However, this context is susceptible to ASR errors because the ASR model generates it auto-regressively. Such noisy context can further degrade the benefits of context input, resulting in suboptimal ASR performance. In this paper, we introduce context noise representation learning to enhance robustness against noisy context, ultimately improving dialogue speech recognition accuracy. To maximize the advantage of context awareness, our approach involves decoder pre-training with text-based dialogue data and noise representation learning for a context encoder. Evaluated on DSTC11 (MultiWoZ 2.1 audio dialogues), it achieves a 24% relative reduction in Word Error Rate (WER) compared to wav2vec2.0 baselines and a 13% reduction compared to Whisper-large-v2. Notably, in noisy environments where user speech is barely audible, our method proves its effectiveness by utilizing contextual information for accurate transcription. Tested on audio data with strong noise level (Signal Noise Ratio of 0dB), our approach shows up to a 31% relative WER reduction compared to the wav2vec2.0 baseline, providing a reassuring solution for real-world noisy scenarios.

A common approach for sequence tagging tasks based on contextual word representations is to train a machine learning classifier directly on these embedding vectors. This approach has two shortcomings. First, such methods consider single input sequences in isolation and are unable to put an individual embedding vector in relation to vectors outside the current local context of use. Second, the high performance of these models relies on fine-tuning the embedding model in conjunction with the classifier, which may not always be feasible due to the size or inaccessibility of the underlying feature-generation model. It is thus desirable, given a collection of embedding vectors of a corpus, i.e. a datastore, to find features of each vector that describe its relation to other, similar vectors in the datastore. With this in mind, we introduce complexity measures of the local topology of the latent space of a contextual language model with respect to a given datastore. The effectiveness of our features is demonstrated through their application to dialogue term extraction. Our work continues a line of research that explores the manifold hypothesis for word embeddings, demonstrating that local structure in the space carved out by word embeddings can be exploited to infer semantic properties.

The advancements in time-efficient data collection techniques such as active learning (AL) has become salient for user intent classification performance in task-oriented dialog systems (TODS). In realistic settings, however, traditional AL techniques often fail to efficiently select targeted in-distribution (IND) data when encountering newly acquired out-of-distribution (OOD) user intents in the unlabeled pool. In this paper, we introduce a novel AL framework viz., AOSAL for TODS that combines a distance-based OOD detector using adaptive false positive rate threshold with an informativeness measure (e.g., entropy) to strategically select informative IND data points in the unlabeled pool. Specifically, we utilize the adaptive OOD detector to classify and filter out OOD samples from the unlabeled pool, then prioritize the acquisition of classified IND instances based on their informativeness scores. To validate our approach, we conduct experiments that display our framework’s flexibility and performance over multiple distance-based approaches and informativeness measures against deep AL baselines on benchmark text datasets. The results suggest that our AOSAL approach consistently outperforms the baselines on IND classification and OOD detection, advancing knowledge on improving robustness of task-oriented dialog systems.

State-of-the-art task-oriented dialogue systems typically rely on task-specific ontologies for fulfilling user queries. The majority of task-oriented dialogue data, such as customer service recordings, comes without ontology and annotation. Such ontologies are normally built manually, limiting the application of specialised systems. Dialogue ontology construction is an approach for automating that process and typically consists of two steps: term extraction and relation extraction. In this work, we focus on relation extraction in a transfer learning set-up. To improve the generalisation, we propose an extension to the decoding mechanism of large language models. We adapt Chain-of-Thought (CoT) decoding, recently developed for reasoning problems, to generative relation extraction. Here, we generate multiple branches in the decoding space and select the relations based on a confidence threshold. By constraining the decoding to ontology terms and relations, we aim to decrease the risk of hallucination. We conduct extensive experimentation on two widely used datasets and find improvements in performance on target ontology for source fine-tuned and one-shot prompted large language models.

In this paper, we propose a multimodal dialogue system designed to elicit spontaneous speech samples from second language learners for reliable oral proficiency assessment. The primary challenge in utilizing dialogue systems for language testing lies in obtaining ratable speech samples that demonstrates the user’s full capabilities of interactional skill. To address this, we developed a virtual agent capable of conducting extended interactions, consisting of a 15-minute interview and 10-minute roleplay. The interview component is a system-led dialogue featuring questions that aim to elicit specific language functions from the user. The system dynamically adjusts the topic difficulty based on real-time assessments to provoke linguistic breakdowns as evidence of their upper limit of proficiency. The roleplay component is a mixed-initiative, collaborative conversation aimed at evaluating the user’s interactional competence. Two experiments were conducted to evaluate our system’s reliability in assessing oral proficiency. In experiment 1, we collected a total of 340 interview sessions, 45-72% of which successfully elicited upper linguistic limit by adjusting the topic difficulty levels. In experiment 2, based on the ropleplay dataset of 75 speakers, the interactional speech elicited by our system was found to be as ratable as those by human examiners, indicated by the reliability index of interactional ratings. These results demonstrates that our system can elicit ratable interactional performances comparable to those elicited by human interviewers. Finally, we report on the deployment of our system with over 10,000 university students in a real-world testing scenario.

Chatbots have become popular in educational settings, revolutionizing how students interact with material and how teachers teach. We present Curriculum-Driven EduBot, a framework for developing a chatbot that combines the interactive features of chatbots with the systematic material of English textbooks to assist students in enhancing their conversational skills. We begin by extracting pertinent topics from textbooks and using large language models to generate dialogues related to these topics. We then fine-tune an open-source LLM using our generated conversational data to create our curriculum-driven chatbot. User studies demonstrate that EduBot outperforms ChatGPT in leading curriculum-based dialogues and adapting its dialogue to match the user’s English proficiency level. By combining traditional textbook methodologies with conversational AI, our approach offers learners an interactive tool that aligns with their curriculum and provides user-tailored conversation practice. This facilitates meaningful student-bot dialogues and enriches the overall learning experience within the curriculum’s pedagogical framework.

Building a dialogue agent that can seamlessly interact with humans in multi-modal regimes, requires two fundamental abilities: (1) understanding emotion and dialogue acts within situated user scenarios, and (2) grounding perceived visual cues to dialogue contexts. However, recent works have uncovered shortcomings of existing dialogue agents in understanding emotions and dialogue acts, and in ground- ing visual cues effectively. In this work, we investigate whether additional dialogue data with only visual descriptions can help dialogue agents effectively align visual and textual features, and enhance the ability of dialogue agents to ground perceived visual cues to dialogue contexts. To this end, in the absence of a suitable dataset, we propose a synthetic visual description generation pipeline, and con- tribute a large-scale synthetic visual description dataset. In addition, we propose a general training procedure for effectively leveraging these synthetic data. We conduct comprehensive analyses to evaluate the impact of synthetic data on two benchmarks: MELD and IEMOCAP. Our findings suggest that synthetic visual descriptions can serve as an effective way to enhance a dialogue agents’ grounding ability, and that the training scheme affects the extent to which these descriptions improve the agent’s performance.

User reviews on e-commerce and review sites are crucial for making purchase decisions, although creating detailed reviews is time-consuming and labor-intensive. In this study, we propose a novel use of dialogue systems to facilitate user review creation by generating reviews from information gathered during interview dialogues with users. To validate our approach, we implemented our system using GPT-4 and conducted comparative experiments from the perspectives of system users and review readers. The results indicate that participants who used our system rated their interactions positively. Additionally, reviews generated by our system required less editing to achieve user satisfaction compared to those by the baseline. We also evaluated the reviews from the readers’ perspective and found that our system-generated reviews are more helpful than those written by humans. Despite challenges with the fluency of the generated reviews, our method offers a promising new approach to review writing.

Scripted dialogues such as movie and TV subtitles constitute a widespread source of training data for conversational NLP models. However, there are notable linguistic differences between these dialogues and spontaneous interactions, especially regarding the occurrence of communicative feedback such as backchannels, acknowledgments, or clarification requests. This paper presents a quantitative analysis of such feedback phenomena in both subtitles and spontaneous conversations. Based on conversational data spanning eight languages and multiple genres, we extract lexical statistics, classifications from a dialogue act tagger, expert annotations and labels derived from a fine-tuned Large Language Model (LLM). Our main empirical findings are that (1) communicative feedback is markedly less frequent in subtitles than in spontaneous dialogues and (2) subtitles contain a higher proportion of negative feedback. We also show that dialogues generated by standard LLMs lie much closer to scripted dialogues than spontaneous interactions in terms of communicative feedback.

In task-oriented dialogue systems, intent classification is crucial for accurately understanding user queries and providing appropriate services. This study explores the use of intent descriptions with large language models for unseen domain intent classification. By examining the effects of description quality, quantity, and input length management, we identify practical guidelines for optimizing performance. Our experiments using FLAN-T5 3B demonstrate that 1) high-quality descriptions for both training and testing significantly improve accuracy, 2) diversity in training descriptions doesn’t greatly affect performance, and 3) off-the-shelf rankers selecting around ten intent options reduce input length without compromising performance. We emphasize that high-quality testing descriptions have a greater impact on accuracy than training descriptions. These findings provide practical guidelines for using intent descriptions with large language models to achieve effective and efficient intent classification in low-resource settings.

As conversational Text-to-Speech (TTS) technologies become increasingly realistic and expressive, understanding the impact of prosodic variation on speech perception and social dynamics is crucial for enhancing conversational systems. This study explores the influence of prosodic features on listener responses to indirect requests using a specifically designed conversational TTS engine capable of controlling prosody, and generating speech across three different speaker profiles: female, male, and gender-ambiguous. We conducted two experiments to analyse how naturalistic variations in speech rate and vocal energy (projection) impact the likelihood of request compliance and perceived politeness. In the first experiment, we examined how prosodic modifications affect the perception of politeness in permission- and service requests. In the second experiment participants compared pairs of spoken requests, each rendered with different prosodic features, and chose which they were more likely to grant. Results indicate that both faster speech rates and higher projection increased the willingness to comply, though the extent of this influence varied by speaker gender. Higher projection in service request increases the chance of being granted more than in permission requests. Politeness has a demonstrated positive impact on the likelihood of requests being granted, this effect is stronger for the male voice compared to female and gender-ambiguous voices.

Collaboration is an integral part of human dialogue. Typical task-oriented dialogue games assign asymmetric roles to the participants, which limits their ability to elicit naturalistic role-taking in collaboration and its negotiation. We present a novel and simple online setup that favors balanced collaboration: a two-player 2D object placement game in which the players must negotiate the goal state themselves. We show empirically that human players exhibit a variety of role distributions, and that balanced collaboration improves task performance. We also present an LLM-based baseline agent which demonstrates that automatic playing of our game is an interesting challenge for artificial systems.

This paper introduces a new method that improves the performance of Automatic speech recognition (ASR) engines, e.g., Whisper in practical cases. Different from prior methods that usually require both speech data and its transcription for decoding, our method only uses jargon as the context for decoding. To do that, the method first represents the jargon in a trie tree structure for efficient storing and traversing. The method next forces the decoding of Whisper to more focus on the jargon by adjusting the probability of generated tokens with the use of the trie tree. To further improve the performance, the method utilizes the prompting method that uses the jargon as the context. Final tokens are generated based on the combination of prompting and decoding. Experimental results on Japanese and English datasets show that the proposed method helps to improve the performance of Whisper, specially for domain-specific data. The method is simple but effective and can be deployed to any encoder-decoder ASR engines in actual cases. The code and data are also accessible (https://shorturl.at/nMsaY).

Large language models demonstrate remarkable proficiency in various tasks across multiple languages. However, their potential in code-switching remains underexplored, particularly in cultural and educational contexts. Code-switching or translanguaging plays a crucial role in bilingual education, facilitating comprehension and engagement among students with varied linguistic proficiencies. In this work, we present a pedagogy-inspired framework that introduces traditional classroom practices of code-switching to intelligent tutoring systems. Specifically, we develop fine-grained instructional strategies tailored to multilingual and educational needs. We conduct experiments involving both LLM-based evaluation and expert analysis to assess the effectiveness of translanguaging in tutoring dialogues. Our experimental results indicate that strategic code-switching can significantly enhance the learning experience. This work not only advances dialogic tutors in language learning, but also extends LLMs to better accommodate multilingual interaction.

Despite being heralded as the new standard for dialogue evaluation, the closed-source nature of GPT-4 poses challenges for the community. Motivated by the need for lightweight, open source, and multilingual dialogue evaluators, this paper introduces GenResCoh (Generated Responses targeting Coherence). GenResCoh is a novel LLM generated dataset comprising over 130k negative and positive responses and accompanying explanations seeded from XDailyDialog and XPersona covering English, French, German, Italian, and Chinese. Leveraging GenResCoh, we propose ECoh (Evaluation of Coherence), a family of evaluators trained to assess response coherence across multiple languages. Experimental results demonstrate that ECoh achieves multilingual detection capabilities superior to the teacher model (GPT-3.5-Turbo) on GenResCoh, despite being based on a much smaller architecture. Furthermore, the explanations provided by ECoh closely align in terms of quality with those generated by the teacher model.

Research on hate speech has predominantly revolved around the detection and interpretation from textual inputs, leaving verbal content largely unexplored. Moreover, while there has been some limited exploration into hate speech detection within verbal acoustic speech inputs, the aspect of interpretability has been overlooked. As such, we introduce a new task within the audio hate speech detection task domain - we specifically aim to identify specific time frames of hate speech within audio utterances. Towards this, we propose two different approaches, cascading and End-to-End (E2E). The first cascading approach initially converts audio to transcripts, identifies hate speech within these transcripts, and subsequently locates the corresponding audio time frames. Conversely, the second E2E approach processes audio utterances directly, which allows it to pinpoint hate speech within specific time frames. Moreover, due to the lack of explainable audio hate speech datasets that include frame-level rationales, we curated a synthetic audio dataset to train our models. We further validate these models on actual human speech utterances and we find that the E2E approach outperforms the cascading method in terms of audio frame Intersection over Union (IoU) metric. Furthermore, we observe that the inclusion of frame-level rationales significantly enhances hate speech detection accuracy for both E2E and cascading approaches.

To create conversational systems with human-like listener behavior, generating short feedback responses (e.g., “mhm”, “ah”, “wow”) appropriate for their context is crucial. These responses convey their communicative function through their lexical form and their prosodic realization. In this paper, we transplant the prosody of feedback responses from human-human U.S. English telephone conversations to a target speaker using two synthesis techniques (TTS and signal processing). Our evaluation focuses on perceived naturalness, contextual appropriateness and preservation of communicative function. Results indicate TTS-generated feedback were perceived as more natural than signal-processing-based feedback, with no significant difference in appropriateness. However, the TTS did not consistently convey the communicative function of the original feedback.

The availability of corpora annotated for discourse relations is limited and discourse relation classification performance varies greatly depending on both language and domain. This is a problem for downstream applications that are intended for a language (i.e., not English) or a domain (i.e., not financial news) with comparatively low coverage for discourse annotations. In this paper, we experiment with a state-of-the-art model for discourse relation classification, originally developed for English, extend it to a multi-lingual setting (testing on Italian, Portuguese and Turkish), and employ a simple, yet effective method to mark out-of-domain training instances. By doing so, we aim to contribute to better generalization and more robust discourse relation classification performance across both language and domain.

The standard language modeling (LM) loss by itself has been shown to be inadequate for effective dialogue modeling. As a result, various training approaches, such as auxiliary loss functions and leveraging human feedback, are being adopted to enrich open-domain dialogue systems. One such auxiliary loss function is Bag-of-Words (BoW) loss, defined as the cross-entropy loss for predicting all the words/tokens of the next utterance. In this work, we propose a novel auxiliary loss named Bag-of-Keywords (BoK) loss to capture the central thought of the response through keyword prediction and leverage it to enhance the generation of meaningful and interpretable responses in open-domain dialogue systems. BoK loss upgrades the BoW loss by predicting only the keywords or critical words/tokens of the next utterance, intending to estimate the core idea rather than the entire response. We incorporate BoK loss in both encoder-decoder (T5) and decoder-only (DialoGPT) architecture and train the models to minimize the weighted sum of BoK and LM (BoK-LM) loss. We perform our experiments on two popular open-domain dialogue datasets, DailyDialog and Persona-Chat. We show that the inclusion of BoK loss improves the dialogue generation of backbone models while also enabling post-hoc interpretability. We also study the effectiveness of BoK-LM loss as a reference-free metric and observe comparable performance to the state-of-the-art metrics on various dialogue evaluation datasets.

Most harmful dialogue detection models are developed for high-resourced languages. Consequently, users who speak under-resourced languages cannot fully benefit from these models in terms of usage, development, detection and mitigation of harmful dialogue utterances. Our work aims at detecting harmful utterances in under-resourced African languages. We leverage transfer learning using pretrained models trained with multilingual embeddings to develop a cross-lingual model capable of detecting harmful content across various African languages. We first fine-tune a harmful dialogue detection model on a selected African dialogue dataset. Additionally, we fine-tune a model on a combined dataset in some African languages to develop a multilingual harmful dialogue detection model. We then evaluate the cross-lingual model’s ability to generalise to an unseen African language by performing harmful dialogue detection in an under-resourced language not present during pretraining or fine-tuning. We evaluate our models on the test datasets. We show that our best performing models achieve impressive results in terms of F1 score. Finally, we discuss the results and limitations of our work.

Large language models (LLMs) tuned for chat have recently been adopted for few-shot end-to-end task-oriented dialogue (TOD), with some success. To further assess this method, we conduct experiments on two, more complex, task-oriented benchmarks that integrate elements of chitchat into the conversation. We enhance a few-shot baseline by adding zero-shot chitchat detection and implementing function calling for dialogue state tracking (DST). We focus on this step in the task-oriented pipeline as it comes first, and errors due to added chitchat at this stage have the most impact on end-to-end performance. We find that this prompting method shows increased resilience to mixed-mode inputs and our enhanced pipeline allows for natural inter-mode conversations, as assessed through human evaluation. Our findings also suggest that the performance gap between few-shot prompting for TOD and supervised task-specific models is narrowing.

In daily conversations, people often encounter problems prompting conversational repair to enhance mutual understanding. By employing an automatic coreference solver, alongside examining repetition, we identify various linguistic features that distinguish turns when the addressee initiates repair from those when they do not. Our findings reveal distinct patterns that characterize the repair sequence and each type of repair initiation.

We use Gaussian Process Regression to predict different types of ratings provided by users after interacting with various task-oriented dialogue systems. We compare the performance of domain-independent dialogue features (e.g., duration, number of filled slots, number of confirmed slots, word error rate) with pre-trained dialogue embeddings. These pre-trained dialogue embeddings are computed by averaging over sentence embeddings in a dialogue. Sentence embeddings are created using various models based on sentence transformers (appearing on the Hugging Face Massive Text Embedding Benchmark leaderboard) or by averaging over BERT word embeddings (varying the BERT layers used). We also compare pre-trained embeddings extracted from human transcriptions with pre-trained embeddings extracted from speech recognition outputs, to determine the robustness of these models to errors. Our results show that overall, for most types of user satisfaction ratings and advanced/recent (or sometimes less advanced/recent) pre-trained embedding models, using only pre-trained embeddings outperforms using only domain-independent features. However, this pattern varies depending on the type of rating and the embedding model used. Also, pre-trained embeddings are found to be robust to speech recognition errors, more advanced/recent embedding models do not always perform better than less advanced/recent ones, and larger models do not necessarily outperform smaller ones. The best prediction performance is achieved by combining pre-trained embeddings with domain-independent features.

In spontaneous natural debate, questions play a variety of crucial roles: they allow speakers to introduce new topics, seek other speakers’ opinions or indeed confront them. A three-class question typology has previously been demonstrated to effectively capture details pertaining to the nature of questions and the different functions associated with them in a debate setting. We adopt this classification and investigate the performance of several machine learning approaches on this task by incorporating various sets of lexical, dialogical and argumentative features. We find that BERT demonstrates the best performance on the task, followed by a Random Forest model enriched with pragmatic features.

While recent years have seen a surge of interest in the automatic processing of memes, much of the work in this area has focused on determining whether a meme contains malicious content. This paper proposes the new task of intent description generation: generating a description of the author’s intentions when creating the meme. To stimulate future work on this task, we (1) annotated a corpus of memes with the intents being perceived by the reader as well as the background knowledge needed to infer the intents and (2) established baseline performance on the intent description generation task using state-of-the-art large language models. Our results suggest the importance of background knowledge retrieval in intent description generation for memes.

The shortage of clinical workforce presents significant challenges in mental healthcare, limiting access to formal diagnostics and services. We aim to tackle this shortage by integrating a customized large language model (LLM) into the workflow, thus promoting equity in mental healthcare for the general population. Although LLMs have showcased their capability in clinical decision-making, their adaptation to severe conditions like Post-traumatic Stress Disorder (PTSD) remains largely unexplored. Therefore, we collect 411 clinician-administered diagnostic interviews and devise a novel approach to obtain high-quality data. Moreover, we build a comprehensive framework to automate PTSD diagnostic assessments based on interview contents by leveraging two state-of-the-art LLMs, GPT-4 and Llama-2, with potential for broader clinical diagnoses. Our results illustrate strong promise for LLMs, tested on our dataset, to aid clinicians in diagnostic validation. To the best of our knowledge, this is the first AI system that fully automates assessments for mental illness based on clinician-administered interviews.

We demonstrate DialBB, a dialogue system development framework, which we have been building as an educational material for dialogue system technology. Building a dialogue system requires the adoption of an appropriate architecture depending on the application and the integration of various technologies. However, this is not easy for those who have just started learning dialogue system technology. Therefore, there is a demand for educational materials that integrate various technologies to build dialogue systems, because traditional dialogue system development frameworks were not designed for educational purposes. DialBB enables the development of dialogue systems by combining modules called building blocks. After understanding sample applications, learners can easily build simple systems using built-in blocks and can build advanced systems using their own developed blocks.

The evolution of large language models has enabled fluent dialogue, increasing interest in the coexistence of humans and avatars. An essential aspect of achieving this coexistence involves developing sophisticated dialogue systems that can influence user behavior. In this background, we propose an effective multimodal dialogue system designed to promote consensus building with humans. Our system employs a slot-filling strategy to guide discussions and attempts to influence users with suggestions through emotional expression and intent conveyance via its avatar. These innovations have resulted in our system achieving the highest performance in a competition evaluating consensus building between humans and dialogue systems. We hope that our research will promote further discussion on the development of dialogue systems that enhance consensus building in human collaboration.

Persona-aware dialogue systems can improve the consistency of the system’s responses, users’ trust and user enjoyment. Filtering nonpersona-like utterances is important for constructing persona-aware dialogue systems. This paper presents the PersonaCLR model for capturing a given utterance’s intensity of persona characteristics. We trained the model with contrastive learning based on the sameness of the utterances’ speaker. Contrastive learning enables PersonaCLR to evaluate the persona characteristics of a given utterance, even if the target persona is not included in training data. For training and evaluating our model, we also constructed a new dataset of 2,155 character utterances from 100 Japanese online novels. Experimental results indicated that our model outperforms existing methods and a strong baseline using a large language model. Our source code, pre-trained model, and dataset are available at https://github.com/1never/PersonaCLR.

Dialogue systems for mental health care aim to provide appropriate support to individuals experiencing mental distress. While extensive research has been conducted to deliver adequate emotional support, existing studies cannot identify individuals who require professional medical intervention and cannot offer suitable guidance. We introduce the Diagnostic Emotional Support Conversation task for an advanced mental health management system. We develop the DESC dataset to assess depression symptoms while maintaining user experience by utilizing task-specific utterance generation prompts and a strict filtering algorithm. Evaluations by professional psychological counselors indicate that DESC has a superior ability to diagnose depression than existing data. Additionally, conversational quality evaluation reveals that DESC maintains fluent, consistent, and coherent dialogues.

Emotions are indispensable in human communication, but are often overlooked in task-oriented dialogue (ToD) modelling, where the task success is the primary focus. While existing works have explored user emotions or similar concepts in some ToD tasks, none has so far included emotion modelling into a fully-fledged ToD system nor conducted interaction with human or simulated users. In this work, we incorporate emotion into the complete ToD processing loop, involving understanding, management, and generation. To this end, we extend the EmoWOZ dataset (Feng et al., 2022) with system affective behaviour labels. Through interactive experimentation involving both simulated and human users, we demonstrate that our proposed framework significantly enhances the user’s emotional experience as well as the task success.

Dialogue systems need to accurately understand the user’s mental state to generate appropriate responses, but accurately discerning such states solely from text or speech can be challenging. To determine which information is necessary, we first collected human-human multimodal dialogues using heterogeneous sensors, resulting in a dataset containing various types of information including speech, video, physiological signals, gaze, and body movement. Additionally, for each time step of the data, users provided subjective evaluations of their emotional valence while reviewing the dialogue videos. Using this dataset and focusing on physiological signals, we analyzed the relationship between the signals and the subjective evaluations through Granger causality analysis. We also investigated how sensor signals differ depending on the polarity of the valence. Our findings revealed several physiological signals related to the user’s emotional valence.

Goal-oriented dialogue systems interact with humans to accomplish specific tasks. However, sometimes these systems fail to establish a common ground with users, leading to task failures. In such cases, it is crucial not to just end with failure but to correct and recover the dialogue to turn it into a success for building a robust goal-oriented dialogue system. Effective recovery from task failures in a goal-oriented dialogue involves not only successful recovery but also accurately understanding the situation of the failed task to minimize unnecessary interactions and avoid frustrating the user. In this study, we analyze the capabilities of GPT-4V in recovering failure tasks by comparing its performance with that of humans using Guess What?! Game. The results show that GPT-4V employs less efficient recovery strategies, such as asking additional unnecessary questions, than humans. We also found that while humans can occasionally ask questions that doubt the accuracy of the interlocutor’s answer during task recovery, GPT-4V lacks this capability.

Humans talk in daily conversations while aligning and negotiating the expressed meanings or common ground. Despite the impressive conversational abilities of the large generative language models, they do not consider the individual differences in contextual understanding in a shared situated environment. In this work, we propose MindDial, a novel conversational framework that can generate situated free-form responses to align and negotiate common ground. We design an explicit mind module that can track three-level beliefs – the speaker’s belief, the speaker’s prediction of the listener’s belief, and the belief gap between the first two. Then the next response is generated to resolve the belief difference and take task-related action. Our framework is applied to both prompting and fine-tuning-based models, and is evaluated across scenarios involving both common ground alignment and negotiation. Experiments show that models with mind modeling can generate more human-like responses when aligning and negotiating common ground. The ablation study further validates the three-level belief design can aggregate information and improve task outcomes in both cooperative and negotiating settings.

In the development of dialog systems the discovery of the set of target intents to identify is a crucial first step that is often overlooked. Most intent detection works assume that a labelled dataset already exists, however creating these datasets is no trivial task and usually requires humans to manually analyse, decide on intent labels and tag accordingly. The field of Open Intent Discovery addresses this problem by automating the process of grouping utterances and providing the user with the discovered intents. Our Open Intent Discovery framework allows for the user to choose from a range of different techniques for each step in the discovery process, including the ability to extend previous works with a human-readable label generation stage. We also provide an analysis of the relationship between dataset features and optimal combination of techniques for each step to help others choose without having to explore every possible combination for their unlabelled data.

The proliferation of social media has increased the visibility and effects of hate speech. To address this, NLP solutions have been developed to identify both explicit and implicit forms of hate speech. Typically, these approaches evaluate the toxicity of utterances in isolation, ignoring the context. Drawing on pragmatics, our study examines how contextual factors can influence the perceived toxicity of utterances, thereby anchoring assessments in a more nuanced semantic framework. We present Toximatics, a dataset that includes context-dependent utterances and it’s toxicity score. We also introduce a novel synthetic data generation pipeline designed to create context-utterance pairs at scale with controlled polarity. This pipeline can enhance existing hate speech datasets by adding contextual information to utterances, either preserving or altering their polarity, and also generate completely new pairs from seed statements. We utilised both features to create Toximatics. To address biases in state-of-the-art hate datasets, which often skew towards specific sensitive topics such as politics, race, and gender, we propose a method to generate neutral utterances typical of various social settings. These are then contextualized to show how neutrality can shift to toxicity or benignity depending on the surrounding context. The evaluation results clearly indicate that the current models are underperforming on this dataset.