Zhen-Hua Ling

Yufei Shi, Qian Chen, Wen Wang et al.

We propose UniVocal, a unified framework that implicitly infers vocal modes from text context to pioneer Speech-Singing Code-Switching (SCS) Synthesis - a task where transitions are autonomously driven by textual semantics, akin to seamless human language blending. Unlike single-mode generation or systems relying on switching-control tags, our proposed UniVocal implicitly infers vocal modes solely from text context. To achieve this, we employ a data-efficient two-stage curriculum learning strategy that progressively trains a competitive TTS system to acquire the desired SCS capability. Addressing data scarcity, we introduce a scalable pipeline to synthesize diverse code-switching data that is both semantically and acoustically natural, alongside a new multi-scenario benchmark, SCSBench. To address limitations of semantic tokenizers in capturing acoustic details, we also introduce refined cent token and Chain-of-Thought (CoT) generation for planning prosody before content generation, effectively enhancing empathetic speech generation and singing melody. Experimental results demonstrate that UniVocal achieves state-of-the-art performance on SCSBench while maintaining competitive performance on regular speech and singing tasks. Audio samples are available at https://project-univocal-demo.github.io/demo/. The code and dataset are released at https://github.com/FunAudioLLM/FunResearch/tree/main/UniVocal.

Chao-Hong Tan, Jia-Chen Gu, Chongyang Tao et al. · microsoft-research

Generating natural and informative texts has been a long-standing problem in NLP. Much effort has been dedicated into incorporating pre-trained language models (PLMs) with various open-world knowledge, such as knowledge graphs or wiki pages. However, their ability to access and manipulate the task-specific knowledge is still limited on downstream tasks, as this type of knowledge is usually not well covered in PLMs and is hard to acquire. To address the problem, we propose augmenting TExt Generation via Task-specific and Open-world Knowledge (TegTok) in a unified framework. Our model selects knowledge entries from two types of knowledge sources through dense retrieval and then injects them into the input encoding and output decoding stages respectively on the basis of PLMs. With the help of these two types of knowledge, our model can learn what and how to generate. Experiments on two text generation tasks of dialogue generation and question generation, and on two datasets show that our method achieves better performance than various baseline models.

Xiao-Hang Jiang, Han-Jie Guo, Ying-Si Liang et al.

Ambient clinical scribes increasingly combine Automatic Speech Recognition with Large Language Models to automate documentation. However, traditional metrics like Word Error Rate mask systemic safety degradation. We present a paired acoustic stress test to isolate the causal impact of noise on clinical reasoning. For the same dialogues, we inject diverse noise types while keeping the downstream model configuration frozen. Crucially, we uncover a dangerous disconnect between signal fidelity and clinical safety. Stationary ambient noise increased the Word Error Rate by a negligible 0.71 percentage points yet nearly doubled the rate of unsafe outputs. Our analysis reveals that minor acoustic perturbations can invert clinical meaning without substantially inflating error rates. Furthermore, we demonstrate a lightweight mitigation strategy that mitigates safety degradation under noisy conditions without requiring model fine tuning.

Jun-Yu Ma, Beiduo Chen, Jia-Chen Gu et al.

Zero-shot cross-lingual named entity recognition (NER) aims at transferring knowledge from annotated and rich-resource data in source languages to unlabeled and lean-resource data in target languages. Existing mainstream methods based on the teacher-student distillation framework ignore the rich and complementary information lying in the intermediate layers of pre-trained language models, and domain-invariant information is easily lost during transfer. In this study, a mixture of short-channel distillers (MSD) method is proposed to fully interact the rich hierarchical information in the teacher model and to transfer knowledge to the student model sufficiently and efficiently. Concretely, a multi-channel distillation framework is designed for sufficient information transfer by aggregating multiple distillers as a mixture. Besides, an unsupervised method adopting parallel domain adaptation is proposed to shorten the channels between the teacher and student models to preserve domain-invariant features. Experiments on four datasets across nine languages demonstrate that the proposed method achieves new state-of-the-art performance on zero-shot cross-lingual NER and shows great generalization and compatibility across languages and fields.

Jia-Chen Gu, Chao-Hong Tan, Chongyang Tao et al.

Recently, various response generation models for two-party conversations have achieved impressive improvements, but less effort has been paid to multi-party conversations (MPCs) which are more practical and complicated. Compared with a two-party conversation where a dialogue context is a sequence of utterances, building a response generation model for MPCs is more challenging, since there exist complicated context structures and the generated responses heavily rely on both interlocutors (i.e., speaker and addressee) and history utterances. To address these challenges, we present HeterMPC, a heterogeneous graph-based neural network for response generation in MPCs which models the semantics of utterances and interlocutors simultaneously with two types of nodes in a graph. Besides, we also design six types of meta relations with node-edge-type-dependent parameters to characterize the heterogeneous interactions within the graph. Through multi-hop updating, HeterMPC can adequately utilize the structural knowledge of conversations for response generation. Experimental results on the Ubuntu Internet Relay Chat (IRC) channel benchmark show that HeterMPC outperforms various baseline models for response generation in MPCs.

Jun-Yu Ma, Jia-Chen Gu, Zhen-Hua Ling et al.

Recent studies have demonstrated that large language models (LLMs) store massive factual knowledge within their parameters. But existing LLMs are prone to hallucinate unintended text due to false or outdated knowledge. Since retraining LLMs is resource intensive, there has been a growing interest in the concept of model editing. Despite the emergence of benchmarks and approaches, these unidirectional editing and evaluation have failed to explore the reversal curse. Intuitively, if "The capital of France is" is edited to be a counterfact "London" within a model, then it should be able to naturally reason and recall the reverse fact, i.e., "London is the capital of" followed by "France" instead of "England". In this paper, we study bidirectional language model editing, aiming to provide rigorous model editing evaluation to assess if edited LLMs can recall the editing knowledge bidirectionally. A new evaluation metric of reversibility is introduced, and a benchmark dubbed as Bidirectional Assessment for Knowledge Editing (BAKE) is constructed to evaluate the reversibility of edited models in recalling knowledge in the reverse direction of editing. We surprisingly observe that while current editing methods and LLMs can effectively recall editing facts in the direction of editing, they suffer serious deficiencies when evaluated in the reverse direction. To mitigate the reversal curse, a method named Bidirectionally Inversible Relationship moDeling (BIRD) is proposed. A set of editing objectives that incorporate bidirectional relationships between subject and object into the updated model weights are designed. Experiments show that BIRD improves the performance of four representative LLMs of different sizes via question answering and judgement.

Chao-Hong Tan, Jia-Chen Gu, Zhen-Hua Ling

Large Language Models (LLMs) have emerged as influential instruments within the realm of natural language processing; nevertheless, their capacity to handle multi-party conversations (MPCs) -- a scenario marked by the presence of multiple interlocutors involved in intricate information exchanges -- remains uncharted. In this paper, we delve into the potential of generative LLMs such as ChatGPT and GPT-4 within the context of MPCs. An empirical analysis is conducted to assess the zero-shot learning capabilities of ChatGPT and GPT-4 by subjecting them to evaluation across three MPC datasets that encompass five representative tasks. The findings reveal that ChatGPT's performance on a number of evaluated MPC tasks leaves much to be desired, whilst GPT-4's results portend a promising future. Additionally, we endeavor to bolster performance through the incorporation of MPC structures, encompassing both speaker and addressee architecture. This study provides an exhaustive evaluation and analysis of applying generative LLMs to MPCs, casting a light upon the conception and creation of increasingly effective and robust MPC agents. Concurrently, this work underscores the challenges implicit in the utilization of LLMs for MPCs, such as deciphering graphical information flows and generating stylistically consistent responses.

Beiduo Chen, Jun-Yu Ma, Jiajun Qi et al.

This paper describes the system developed by the USTC-NELSLIP team for SemEval-2022 Task 11 Multilingual Complex Named Entity Recognition (MultiCoNER). We propose a gazetteer-adapted integration network (GAIN) to improve the performance of language models for recognizing complex named entities. The method first adapts the representations of gazetteer networks to those of language models by minimizing the KL divergence between them. After adaptation, these two networks are then integrated for backend supervised named entity recognition (NER) training. The proposed method is applied to several state-of-the-art Transformer-based NER models with a gazetteer built from Wikidata, and shows great generalization ability across them. The final predictions are derived from an ensemble of these trained models. Experimental results and detailed analysis verify the effectiveness of the proposed method. The official results show that our system ranked 1st on three tracks (Chinese, Code-mixed and Bangla) and 2nd on the other ten tracks in this task.

Zheng-Yan Sheng, Yang Ai, Yan-Nian Chen et al.

This paper presents a novel task, zero-shot voice conversion based on face images (zero-shot FaceVC), which aims at converting the voice characteristics of an utterance from any source speaker to a newly coming target speaker, solely relying on a single face image of the target speaker. To address this task, we propose a face-voice memory-based zero-shot FaceVC method. This method leverages a memory-based face-voice alignment module, in which slots act as the bridge to align these two modalities, allowing for the capture of voice characteristics from face images. A mixed supervision strategy is also introduced to mitigate the long-standing issue of the inconsistency between training and inference phases for voice conversion tasks. To obtain speaker-independent content-related representations, we transfer the knowledge from a pretrained zero-shot voice conversion model to our zero-shot FaceVC model. Considering the differences between FaceVC and traditional voice conversion tasks, systematic subjective and objective metrics are designed to thoroughly evaluate the homogeneity, diversity and consistency of voice characteristics controlled by face images. Through extensive experiments, we demonstrate the superiority of our proposed method on the zero-shot FaceVC task. Samples are presented on our demo website.

Fei Liu, Yang Ai, Hui-Peng Du et al.

Audio super-resolution aims to recover missing high-frequency details from bandwidth-limited low-resolution audio, thereby improving the naturalness and perceptual quality of the reconstructed signal. However, most existing methods directly operate in the waveform or time-frequency domain, which not only involves high-dimensional generation spaces but is also largely limited to speech tasks, leaving substantial room for improvement on more complex audio types such as sound effects and music. To mitigate these limitations, we introduce LatentFlowSR, a new audio super-resolution approach that leverages conditional flow matching (CFM) within a latent representation space. Specifically, we first train a noise-robust autoencoder, which encodes low-resolution audio into a continuous latent space. Conditioned on the low-resolution latent representation, a CFM mechanism progressively generates the corresponding high-resolution latent representation from a Gaussian prior with a one-step ordinary differential equation (ODE) solver. The resulting high-resolution latent representation is then decoded by the pretrained autoencoder to reconstruct the high-resolution audio. Experimental results demonstrate that LatentFlowSR consistently outperforms baseline methods across various audio types and super-resolution settings. These results indicate that the proposed method possesses strong high-frequency reconstruction capability and robust generalization performance, providing compelling evidence for the effectiveness of latent-space modeling in audio super-resolution. All relevant code will be made publicly available upon completion of the paper review process.

Hao-Xiang Xu, Chong Deng, Jiaqing Liu et al.

Large Language Models (LLMs) extend their capabilities through function-calling (FC), which relies on training data with high quality, diversity, and broad coverage of scenario. However, obtaining and annotating real function-calling data is challenging, while synthetic data from existing pipelines often suffers from unreliable APIs, limited tool scalability, insufficient diversity, and weak quality control. To address these, we present GenesisFunc, an automated pipeline for generating FC training data. Starting from reliable tools in widely used public benchmarks, our GenesisFunc employs a multi-agent framework to support a dialogue generation system that produces conversations spanning diverse scenarios, while maintaining both diversity and quality throughout the process. The accuracy of the data is further reinforced through a multi-stage evaluation system. We fine-tune an 8B LLM on the synthetic dataset and show through extensive experiments that it outperforms similarly sized open-source models in in-domain FC performance and out-of-domain generalization, while reaching FC capabilities comparable to some of the latest API-based models. In addition, our method demonstrates strong potential to scale effectively across downstream tools, underscoring its real-world applicability.

Zhengyan Sheng, Yang Ai, Li-Juan Liu et al.

Despite recent advancements in speech generation with text prompt providing control over speech style, voice attributes in synthesized speech remain elusive and challenging to control. This paper introduces a novel task: voice attribute editing with text prompt, with the goal of making relative modifications to voice attributes according to the actions described in the text prompt. To solve this task, VoxEditor, an end-to-end generative model, is proposed. In VoxEditor, addressing the insufficiency of text prompt, a Residual Memory (ResMem) block is designed, that efficiently maps voice attributes and these descriptors into the shared feature space. Additionally, the ResMem block is enhanced with a voice attribute degree prediction (VADP) block to align voice attributes with corresponding descriptors, addressing the imprecision of text prompt caused by non-quantitative descriptions of voice attributes. We also establish the open-source VCTK-RVA dataset, which leads the way in manual annotations detailing voice characteristic differences among different speakers. Extensive experiments demonstrate the effectiveness and generalizability of our proposed method in terms of both objective and subjective metrics. The dataset and audio samples are available on the website.

Jun-Yu Ma, Zhen-Hua Ling, Ningyu Zhang et al.

Despite their exceptional capabilities, large language models (LLMs) are prone to generating unintended text due to false or outdated knowledge. Given the resource-intensive nature of retraining LLMs, there has been a notable increase in the development of knowledge editing. However, current approaches and evaluations rarely explore the perturbation of editing on neighboring knowledge. This paper studies whether updating new knowledge to LLMs perturbs the neighboring knowledge encapsulated within them. Specifically, we seek to figure out whether appending a new answer into an answer list to a factual question leads to catastrophic forgetting of original correct answers in this list, as well as unintentional inclusion of incorrect answers. A metric of additivity is introduced and a benchmark dubbed as Perturbation Evaluation of Appending Knowledge (PEAK) is constructed to evaluate the degree of perturbation to neighboring knowledge when appending new knowledge. Besides, a plug-and-play framework termed Appending via Preservation and Prevention (APP) is proposed to mitigate the neighboring perturbation by maintaining the integrity of the answer list. Experiments demonstrate the effectiveness of APP coupling with four editing methods on four LLMs. The code and data are available at https://github.com/mjy1111/PEAK.

Jinghao He, Zhengyan Sheng, Liping Chen et al.

This paper focuses on explaining the timbre conveyed by speech signals and introduces a task termed voice timbre attribute detection (vTAD). In this task, voice timbre is explained with a set of sensory attributes describing its human perception. A pair of speech utterances is processed, and their intensity is compared in a designated timbre descriptor. Moreover, a framework is proposed, which is built upon the speaker embeddings extracted from the speech utterances. The investigation is conducted on the VCTK-RVA dataset. Experimental examinations on the ECAPA-TDNN and FACodec speaker encoders demonstrated that: 1) the ECAPA-TDNN speaker encoder was more capable in the seen scenario, where the testing speakers were included in the training set; 2) the FACodec speaker encoder was superior in the unseen scenario, where the testing speakers were not part of the training, indicating enhanced generalization capability. The VCTK-RVA dataset and open-source code are available on the website https://github.com/vTAD2025-Challenge/vTAD.

Yin-Long Liu, Rui Feng, Jia-Hong Yuan et al.

Compared to other clinical screening techniques, speech-and-language-based automated Alzheimer's disease (AD) detection methods are characterized by their non-invasiveness, cost-effectiveness, and convenience. Previous studies have demonstrated the efficacy of fine-tuning pre-trained language models (PLMs) for AD detection. However, the objective of this traditional fine-tuning method, which involves inputting only transcripts, is inconsistent with the masked language modeling (MLM) task used during the pre-training phase of PLMs. In this paper, we investigate prompt-based fine-tuning of PLMs, converting the classification task into a MLM task by inserting prompt templates into the transcript inputs. We also explore the impact of incorporating pause information from forced alignment into manual transcripts. Additionally, we compare the performance of various automatic speech recognition (ASR) models and select the Whisper model to generate ASR-based transcripts for comparison with manual transcripts. Furthermore, majority voting and ensemble techniques are applied across different PLMs (BERT and RoBERTa) using different random seeds. Ultimately, we obtain maximum detection accuracy of 95.8% (with mean 87.9%, std 3.3%) using manual transcripts, achieving state-of-the-art performance for AD detection using only transcripts on the ADReSS test set.

Shi-Qi Yan, Jia-Chen Gu, Yun Zhu et al.

Large language models (LLMs) inevitably exhibit hallucinations since the accuracy of generated texts cannot be secured solely by the parametric knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a practicable complement to LLMs, it relies heavily on the relevance of retrieved documents, raising concerns about how the model behaves if retrieval goes wrong. To this end, we propose the Corrective Retrieval Augmented Generation (CRAG) to improve the robustness of generation. Specifically, a lightweight retrieval evaluator is designed to assess the overall quality of retrieved documents for a query, returning a confidence degree based on which different knowledge retrieval actions can be triggered. Since retrieval from static and limited corpora can only return sub-optimal documents, large-scale web searches are utilized as an extension for augmenting the retrieval results. Besides, a decompose-then-recompose algorithm is designed for retrieved documents to selectively focus on key information and filter out irrelevant information in them. CRAG is plug-and-play and can be seamlessly coupled with various RAG-based approaches. Experiments on four datasets covering short- and long-form generation tasks show that CRAG can significantly improve the performance of RAG-based approaches.

Lu Dong, Haiyu Zhang, Han Lin et al.

Video temporal grounding (VTG) aims to locate precise segments in videos based on language queries, which is a fundamental challenge in video understanding. While recent Multimodal Large Language Models (MLLMs) have shown promise in tackling VTG through reinforcement learning (RL), they overlook the challenges arising from both the quality and difficulty of training samples. (1) Partially annotated samples. Many samples contain relevant segments beyond the annotated interval, introducing ambiguous supervision. (2) Hard-to-ground samples. Samples with poor zero-shot performance produce consistently low and indistinguishable rewards during RL training, exhibiting no clear preference among multiple outputs and thus hindering learning efficiency. To address these challenges, we propose VideoTG-R1, a novel curriculum RL framework with reflected boundary annotations, enabling data-efficient training. Specifically, we propose a Boundary Reflection Agent that utilizes MLLMs to predict query-relevant timestamps outside the annotated intervals, allowing us to identify and filter out partially annotated samples, thereby reducing ambiguity. Furthermore, we introduce a Difficulty Estimation Agent to assess the training difficulty of each sample and design a curriculum RL strategy that dynamically masks the videos of hard-to-ground samples according to the training steps, easing the training difficulty and providing clearer preference. Experiments on the VTG and grounded VideoQA tasks demonstrate the effectiveness of our method. Remarkably, with only 10% of the training samples and 21% of the computational budget, VideoTG-R1 outperforms full-data counterparts under both group relative policy optimization (GRPO) and supervised fine-tuning (SFT). The code is available at https://github.com/ldong1111/VideoTG-R1.

Jie Zhang, Qing-Tian Xu, Qiu-Shi Zhu et al.

Time-domain single-channel speech enhancement (SE) still remains challenging to extract the target speaker without any prior information on multi-talker conditions. It has been shown via auditory attention decoding that the brain activity of the listener contains the auditory information of the attended speaker. In this paper, we thus propose a novel time-domain brain-assisted SE network (BASEN) incorporating electroencephalography (EEG) signals recorded from the listener for extracting the target speaker from monaural speech mixtures. The proposed BASEN is based on the fully-convolutional time-domain audio separation network. In order to fully leverage the complementary information contained in the EEG signals, we further propose a convolutional multi-layer cross attention module to fuse the dual-branch features. Experimental results on a public dataset show that the proposed model outperforms the state-of-the-art method in several evaluation metrics. The reproducible code is available at https://github.com/jzhangU/Basen.git.

Yang Ai, Zhen-Hua Ling

This paper presents a neural vocoder named HiNet which reconstructs speech waveforms from acoustic features by predicting amplitude and phase spectra hierarchically. Different from existing neural vocoders such as WaveNet, SampleRNN and WaveRNN which directly generate waveform samples using single neural networks, the HiNet vocoder is composed of an amplitude spectrum predictor (ASP) and a phase spectrum predictor (PSP). The ASP is a simple DNN model which predicts log amplitude spectra (LAS) from acoustic features. The predicted LAS are sent into the PSP for phase recovery. Considering the issue of phase warping and the difficulty of phase modeling, the PSP is constructed by concatenating a neural source-filter (NSF) waveform generator with a phase extractor. We also introduce generative adversarial networks (GANs) into both ASP and PSP. Finally, the outputs of ASP and PSP are combined to reconstruct speech waveforms by short-time Fourier synthesis. Since there are no autoregressive structures in both predictors, the HiNet vocoder can generate speech waveforms with high efficiency. Objective and subjective experimental results show that our proposed HiNet vocoder achieves better naturalness of reconstructed speech than the conventional STRAIGHT vocoder, a 16-bit WaveNet vocoder using open source implementation and an NSF vocoder with similar complexity to the PSP and obtains similar performance with a 16-bit WaveRNN vocoder. We also find that the performance of HiNet is insensitive to the complexity of the neural waveform generator in PSP to some extend. After simplifying its model structure, the time consumed for generating 1s waveforms of 16kHz speech using a GPU can be further reduced from 0.34s to 0.19s without significant quality degradation.

Jia-Chen Gu, Hao-Xiang Xu, Jun-Yu Ma et al. · stanford

Model editing is a technique that edits the large language models (LLMs) with updated knowledge to alleviate hallucinations without resource-intensive retraining. While current model editing methods can effectively modify a model's behavior within a specific area of interest, they often overlook the potential unintended side effects on the general abilities of LLMs such as reasoning, natural language inference, and question answering. In this paper, we raise concerns that model editing's improvements on factuality may come at the cost of a significant degradation of the model's general abilities. We systematically analyze the side effects by evaluating four popular editing methods on three LLMs across eight representative tasks. Our extensive empirical experiments show that it is challenging for current editing methods to simultaneously improve factuality of LLMs and maintain their general abilities. Our analysis reveals that the side effects are caused by model editing altering the original model weights excessively, leading to overfitting to the edited facts. To mitigate this, a method named RECT is proposed to regularize the edit update weights by imposing constraints on their complexity based on the RElative Change in weighT. Evaluation results show that RECT can significantly mitigate the side effects of editing while still maintaining over 94% editing performance.

Zhengyan Sheng, Jinghao He, Liping Chen et al.

Voice timbre refers to the unique quality or character of a person's voice that distinguishes it from others as perceived by human hearing. The Voice Timbre Attribute Detection (VtaD) 2025 challenge focuses on explaining the voice timbre attribute in a comparative manner. In this challenge, the human impression of voice timbre is verbalized with a set of sensory descriptors, including bright, coarse, soft, magnetic, and so on. The timbre is explained from the comparison between two voices in their intensity within a specific descriptor dimension. The VtaD 2025 challenge starts in May and culminates in a special proposal at the NCMMSC2025 conference in October 2025 in Zhenjiang, China.

Zhengyan Sheng, Zhihao Du, Heng Lu et al.

Recent advancements in personalized speech generation have brought synthetic speech increasingly close to the realism of target speakers' recordings, yet multimodal speaker generation remains on the rise. This paper introduces UniSpeaker, a unified approach for multimodality-driven speaker generation. Specifically, we propose a unified voice aggregator based on KV-Former, applying soft contrastive loss to map diverse voice description modalities into a shared voice space, ensuring that the generated voice aligns more closely with the input descriptions. To evaluate multimodality-driven voice control, we build the first multimodality-based voice control (MVC) benchmark, focusing on voice suitability, voice diversity, and speech quality. UniSpeaker is evaluated across five tasks using the MVC benchmark, and the experimental results demonstrate that UniSpeaker outperforms previous modality-specific models. Speech samples are available at \url{https://UniSpeaker.github.io}.

Lu Dong, Haiyu Zhang, Hongjie Zhang et al.

The task of weakly supervised temporal sentence grounding (WSTSG) aims to detect temporal intervals corresponding to a language description from untrimmed videos with only video-level video-language correspondence. For an anchor sample, most existing approaches generate negative samples either from other videos or within the same video for contrastive learning. However, some training samples are highly similar to the anchor sample, directly regarding them as negative samples leads to difficulties for optimization and ignores the correlations between these similar samples and the anchor sample. To address this, we propose Positive Sample Mining (PSM), a novel framework that mines positive samples from the training set to provide more discriminative supervision. Specifically, for a given anchor sample, we partition the remaining training set into semantically similar and dissimilar subsets based on the similarity of their text queries. To effectively leverage these correlations, we introduce a PSM-guided contrastive loss to ensure that the anchor proposal is closer to similar samples and further from dissimilar ones. Additionally, we design a PSM-guided rank loss to ensure that similar samples are closer to the anchor proposal than to the negative intra-video proposal, aiming to distinguish the anchor proposal and the negative intra-video proposal. Experiments on the WSTSG and grounded VideoQA tasks demonstrate the effectiveness and superiority of our method.

Hao-Xiang Xu, Jun-Yu Ma, Zhen-Hua Ling et al.

Large language models (LLMs) struggle with hallucinations due to false or outdated knowledge. Given the high resource demands of retraining these models, there is an increasing focus on developing model editing. However, the general abilities of LLMs across downstream tasks are prone to significant degradation during sequential editing. This paper statistically observes that the parameter matrix after editing exhibits a significant deviation compared to its previous state as the number of edits increases. This serious deviation affects the original knowledge associations within LLMs and leads to the degradation of their general abilities. To this end, a framework termed Editing Anchor Compression (EAC) is proposed to constrain the deviation of the parameter matrix during sequential editing. It compresses the editing information by selecting editing anchors that are important in encoding new relations without deviating too much from the original matrix, thereby preserving the general abilities. Experiments of applying EAC to two popular editing methods on three LLMs across four tasks are conducted. Evaluation results show that EAC effectively minimizes unreasonable deviations caused by model editing, preserving over 70% of the general abilities while better retaining the editing knowledge compared to the original counterpart methods.

Shi-Qi Yan, Quan Liu, Zhen-Hua Ling

While Retrieval-Augmented Generation (RAG) has exhibited promise in utilizing external knowledge, its generation process heavily depends on the quality and accuracy of the retrieved context. Large language models (LLMs) struggle to evaluate the correctness of non-parametric knowledge retrieved externally when it differs from internal memorization, leading to knowledge conflicts during response generation. To this end, we introduce the Retrieval Preference Optimization (RPO), a lightweight and effective alignment method to adaptively leverage multi-source knowledge based on retrieval relevance. An implicit representation of retrieval relevance is derived and incorporated into the reward model to integrate retrieval evaluation and response generation into a single model, solving the problem that previous methods necessitate the additional procedure to assess the retrieval quality. Notably, RPO is the only RAG-dedicated alignment approach that quantifies the awareness of retrieval relevance in training, overcoming mathematical obstacles. Experiments on four datasets demonstrate that RPO outperforms RAG by 4-10% in accuracy without any extra component, exhibiting its robust generalization.

Chenyang Guo, Liping Chen, Zhuhai Li et al.

Neural networks are commonly known to be vulnerable to adversarial attacks mounted through subtle perturbation on the input data. Recent development in voice-privacy protection has shown the positive use cases of the same technique to conceal speaker's voice attribute with additive perturbation signal generated by an adversarial network. This paper examines the reversibility property where an entity generating the adversarial perturbations is authorized to remove them and restore original speech (e.g., the speaker him/herself). A similar technique could also be used by an investigator to deanonymize a voice-protected speech to restore criminals' identities in security and forensic analysis. In this setting, the perturbation generative module is assumed to be known in the removal process. To this end, a joint training of perturbation generation and removal modules is proposed. Experimental results on the LibriSpeech dataset demonstrated that the subtle perturbations added to the original speech can be predicted from the anonymized speech while achieving the goal of privacy protection. By removing these perturbations from the anonymized sample, the original speech can be restored. Audio samples can be found in \url{https://voiceprivacy.github.io/Perturbation-Generation-Removal/}.

Chao Wang, Rui-Chen Zheng, Yang Ai et al.

The integration of speech into Large Language Models (LLMs) has substantially expanded their capabilities, but often at the cost of weakening their core textual competence. This degradation limits the ability of speech-enabled LLMs to fully exploit their pre-trained text-based knowledge. In this work, we analyze the underlying mechanisms of this issue through a focused study of the widely used encoder-adaptor paradigm. We propose an analytical framework based on parameter importance estimation, which reveals that fine-tuning for speech introduces a textual importance distribution shift: the layer-wise allocation of parameters critical to textual reasoning is disrupted. Building on this insight, we investigate two mitigation strategies: layer-wise learning rate scheduling and Low-Rank Adaptation (LoRA), both aim to preserve the original parameter distribution. Experimental results show that both approaches better maintain textual competence than full fine-tuning, while also improving downstream spoken question answering performance. Furthermore, our analysis offers a principled explanation for the effectiveness of the proposed mitigation strategies, linking their benefits to the structural properties of textual knowledge in LLMs.

Liping Chen, Jinghao He, Zhengyan Sheng et al.

The first voice timbre attribute detection challenge is featured in a special session at NCMMSC 2025. It focuses on the explainability of voice timbre and compares the intensity of two speech utterances in a specified timbre descriptor dimension. The evaluation was conducted on the VCTK-RVA dataset. Participants developed their systems and submitted their outputs to the organizer, who evaluated the performance and sent feedback to them. Six teams submitted their outputs, with five providing descriptions of their methodologies.

Rui Wang, Liping Chen, Kong AiK Lee et al.

Voice anonymization has been developed as a technique for preserving privacy by replacing the speaker's voice in a speech signal with that of a pseudo-speaker, thereby obscuring the original voice attributes from machine recognition and human perception. In this paper, we focus on altering the voice attributes against machine recognition while retaining human perception. We referred to this as the asynchronous voice anonymization. To this end, a speech generation framework incorporating a speaker disentanglement mechanism is employed to generate the anonymized speech. The speaker attributes are altered through adversarial perturbation applied on the speaker embedding, while human perception is preserved by controlling the intensity of perturbation. Experiments conducted on the LibriSpeech dataset showed that the speaker attributes were obscured with their human perception preserved for 60.71% of the processed utterances.

Jia-Chen Gu, Chao-Hong Tan, Caiyuan Chu et al.

Modeling multi-party conversations (MPCs) with graph neural networks has been proven effective at capturing complicated and graphical information flows. However, existing methods rely heavily on the necessary addressee labels and can only be applied to an ideal setting where each utterance must be tagged with an addressee label. To study the scarcity of addressee labels which is a common issue in MPCs, we propose MADNet that maximizes addressee deduction expectation in heterogeneous graph neural networks for MPC generation. Given an MPC with a few addressee labels missing, existing methods fail to build a consecutively connected conversation graph, but only a few separate conversation fragments instead. To ensure message passing between these conversation fragments, four additional types of latent edges are designed to complete a fully-connected graph. Besides, to optimize the edge-type-dependent message passing for those utterances without addressee labels, an Expectation-Maximization-based method that iteratively generates silver addressee labels (E step), and optimizes the quality of generated responses (M step), is designed. Experimental results on two Ubuntu IRC channel benchmarks show that MADNet outperforms various baseline models on the task of MPC generation, especially under the more common and challenging setting where part of addressee labels are missing.

Jun-Yu Ma, Jia-Chen Gu, Zhen-Hua Ling et al.

Zero-shot cross-lingual information extraction(IE) aims at constructing an IE model for some low-resource target languages, given annotations exclusively in some rich-resource languages. Recent studies based on language-universal features have shown their effectiveness and are attracting increasing attention. However, prior work has neither explored the potential of establishing interactions between language-universal features and contextual representations nor incorporated features that can effectively model constituent span attributes and relationships between multiple spans. In this study, a syntax-augmented hierarchical interactive encoder (SHINE) is proposed to transfer cross-lingual IE knowledge. The proposed encoder is capable of interactively capturing complementary information between features and contextual information, to derive language-agnostic representations for various IE tasks. Concretely, a multi-level interaction network is designed to hierarchically interact the complementary information to strengthen domain adaptability. Besides, in addition to the well-studied syntax features of part-of-speech and dependency relation, a new syntax feature of constituency structure is introduced to model the constituent span information which is crucial for IE. Experiments across seven languages on three IE tasks and four benchmarks verify the effectiveness and generalization ability of the proposed method.

Chao-Hong Tan, Jia-Chen Gu, Zhen-Hua Ling

Diffusion models have emerged as the new state-of-the-art family of deep generative models, and their promising potentials for text generation have recently attracted increasing attention. Existing studies mostly adopt a single encoder architecture with partially noising processes for conditional text generation, but its degree of flexibility for conditional modeling is limited. In fact, the encoder-decoder architecture is naturally more flexible for its detachable encoder and decoder modules, which is extensible to multilingual and multimodal generation tasks for conditions and target texts. However, the encoding process of conditional texts lacks the understanding of target texts. To this end, a spiral interaction architecture for encoder-decoder text diffusion (DiffuSIA) is proposed. Concretely, the conditional information from encoder is designed to be captured by the diffusion decoder, while the target information from decoder is designed to be captured by the conditional encoder. These two types of information flow run through multilayer interaction spirally for deep fusion and understanding. DiffuSIA is evaluated on four text generation tasks, including paraphrase, text simplification, question generation, and open-domain dialogue generation. Experimental results show that DiffuSIA achieves competitive performance among previous methods on all four tasks, demonstrating the effectiveness and generalization ability of the proposed method.

Jia-Chen Gu, Zhen-Hua Ling, Quan Liu et al.

Addressing the issues of who saying what to whom in multi-party conversations (MPCs) has recently attracted a lot of research attention. However, existing methods on MPC understanding typically embed interlocutors and utterances into sequential information flows, or utilize only the superficial of inherent graph structures in MPCs. To this end, we present a plug-and-play and lightweight method named graph-induced fine-tuning (GIFT) which can adapt various Transformer-based pre-trained language models (PLMs) for universal MPC understanding. In detail, the full and equivalent connections among utterances in regular Transformer ignore the sparse but distinctive dependency of an utterance on another in MPCs. To distinguish different relationships between utterances, four types of edges are designed to integrate graph-induced signals into attention mechanisms to refine PLMs originally designed for processing sequential texts. We evaluate GIFT by implementing it into three PLMs, and test the performance on three downstream tasks including addressee recognition, speaker identification and response selection. Experimental results show that GIFT can significantly improve the performance of three PLMs on three downstream tasks and two benchmarks with only 4 additional parameters per encoding layer, achieving new state-of-the-art performance on MPC understanding.

Zheng-Yan Sheng, Yang Ai, Zhen-Hua Ling

Lip-to-Speech (Lip2Speech) synthesis, which predicts corresponding speech from talking face images, has witnessed significant progress with various models and training strategies in a series of independent studies. However, existing studies can not achieve voice control under zero-shot condition, because extra speaker embeddings need to be extracted from natural reference speech and are unavailable when only the silent video of an unseen speaker is given. In this paper, we propose a zero-shot personalized Lip2Speech synthesis method, in which face images control speaker identities. A variational autoencoder is adopted to disentangle the speaker identity and linguistic content representations, which enables speaker embeddings to control the voice characteristics of synthetic speech for unseen speakers. Furthermore, we propose associated cross-modal representation learning to promote the ability of face-based speaker embeddings (FSE) on voice control. Extensive experiments verify the effectiveness of the proposed method whose synthetic utterances are more natural and matching with the personality of input video than the compared methods. To our best knowledge, this paper makes the first attempt on zero-shot personalized Lip2Speech synthesis with a face image rather than reference audio to control voice characteristics.

Jun-Yu Ma, Jia-Chen Gu, Jiajun Qi et al.

This paper describes the system developed by the USTC-NELSLIP team for SemEval-2023 Task 2 Multilingual Complex Named Entity Recognition (MultiCoNER II). A method named Statistical Construction and Dual Adaptation of Gazetteer (SCDAG) is proposed for Multilingual Complex NER. The method first utilizes a statistics-based approach to construct a gazetteer. Secondly, the representations of gazetteer networks and language models are adapted by minimizing the KL divergence between them at both the sentence-level and entity-level. Finally, these two networks are then integrated for supervised named entity recognition (NER) training. The proposed method is applied to XLM-R with a gazetteer built from Wikidata, and shows great generalization ability across different tracks. Experimental results and detailed analysis verify the effectiveness of the proposed method. The official results show that our system ranked 1st on one track (Hindi) in this task.

Lu Dong, Zhi-Qiang Guo, Chao-Hong Tan et al.

Neural network models have achieved state-of-the-art performance on grapheme-to-phoneme (G2P) conversion. However, their performance relies on large-scale pronunciation dictionaries, which may not be available for a lot of languages. Inspired by the success of the pre-trained language model BERT, this paper proposes a pre-trained grapheme model called grapheme BERT (GBERT), which is built by self-supervised training on a large, language-specific word list with only grapheme information. Furthermore, two approaches are developed to incorporate GBERT into the state-of-the-art Transformer-based G2P model, i.e., fine-tuning GBERT or fusing GBERT into the Transformer model by attention. Experimental results on the Dutch, Serbo-Croatian, Bulgarian and Korean datasets of the SIGMORPHON 2021 G2P task confirm the effectiveness of our GBERT-based G2P models under both medium-resource and low-resource data conditions.

Chao-Hong Tan, Qian Chen, Wen Wang et al.

Transformer-based models have achieved great success in various NLP, vision, and speech tasks. However, the core of Transformer, the self-attention mechanism, has a quadratic time and memory complexity with respect to the sequence length, which hinders applications of Transformer-based models to long sequences. Many approaches have been proposed to mitigate this problem, such as sparse attention mechanisms, low-rank matrix approximations and scalable kernels, and token mixing alternatives to self-attention. We propose a novel Pooling Network (PoNet) for token mixing in long sequences with linear complexity. We design multi-granularity pooling and pooling fusion to capture different levels of contextual information and combine their interactions with tokens. On the Long Range Arena benchmark, PoNet significantly outperforms Transformer and achieves competitive accuracy, while being only slightly slower than the fastest model, FNet, across all sequence lengths measured on GPUs. We also conduct systematic studies on the transfer learning capability of PoNet and observe that PoNet achieves 95.7% of the accuracy of BERT on the GLUE benchmark, outperforming FNet by 4.5% relative. Comprehensive ablation analysis demonstrates effectiveness of the designed multi-granularity pooling and pooling fusion for token mixing in long sequences and efficacy of the designed pre-training tasks for PoNet to learn transferable contextualized language representations.

Jia-Chen Gu, Zhen-Hua Ling, Yu Wu et al.

Personas are useful for dialogue response prediction. However, the personas used in current studies are pre-defined and hard to obtain before a conversation. To tackle this issue, we study a new task, named Speaker Persona Detection (SPD), which aims to detect speaker personas based on the plain conversational text. In this task, a best-matched persona is searched out from candidates given the conversational text. This is a many-to-many semantic matching task because both contexts and personas in SPD are composed of multiple sentences. The long-term dependency and the dynamic redundancy among these sentences increase the difficulty of this task. We build a dataset for SPD, dubbed as Persona Match on Persona-Chat (PMPC). Furthermore, we evaluate several baseline models and propose utterance-to-profile (U2P) matching networks for this task. The U2P models operate at a fine granularity which treat both contexts and personas as sets of multiple sequences. Then, each sequence pair is scored and an interpretable overall score is obtained for a context-persona pair through aggregation. Evaluation results show that the U2P models outperform their baseline counterparts significantly.

Jia-Chen Gu, Chongyang Tao, Zhen-Hua Ling et al.

Recently, various neural models for multi-party conversation (MPC) have achieved impressive improvements on a variety of tasks such as addressee recognition, speaker identification and response prediction. However, these existing methods on MPC usually represent interlocutors and utterances individually and ignore the inherent complicated structure in MPC which may provide crucial interlocutor and utterance semantics and would enhance the conversation understanding process. To this end, we present MPC-BERT, a pre-trained model for MPC understanding that considers learning who says what to whom in a unified model with several elaborated self-supervised tasks. Particularly, these tasks can be generally categorized into (1) interlocutor structure modeling including reply-to utterance recognition, identical speaker searching and pointer consistency distinction, and (2) utterance semantics modeling including masked shared utterance restoration and shared node detection. We evaluate MPC-BERT on three downstream tasks including addressee recognition, speaker identification and response selection. Experimental results show that MPC-BERT outperforms previous methods by large margins and achieves new state-of-the-art performance on all three downstream tasks at two benchmarks.

Jia-Chen Gu, Hui Liu, Zhen-Hua Ling et al.

Persona can function as the prior knowledge for maintaining the consistency of dialogue systems. Most of previous studies adopted the self persona in dialogue whose response was about to be selected from a set of candidates or directly generated, but few have noticed the role of partner in dialogue. This paper makes an attempt to thoroughly explore the impact of utilizing personas that describe either self or partner speakers on the task of response selection in retrieval-based chatbots. Four persona fusion strategies are designed, which assume personas interact with contexts or responses in different ways. These strategies are implemented into three representative models for response selection, which are based on the Hierarchical Recurrent Encoder (HRE), Interactive Matching Network (IMN) and Bidirectional Encoder Representations from Transformers (BERT) respectively. Empirical studies on the Persona-Chat dataset show that the partner personas neglected in previous studies can improve the accuracy of response selection in the IMN- and BERT-based models. Besides, our BERT-based model implemented with the context-response-aware persona fusion strategy outperforms previous methods by margins larger than 2.7% on original personas and 4.6% on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset.

Yu-Ping Ruan, Zhen-Hua Ling

This paper presents an emotion-regularized conditional variational autoencoder (Emo-CVAE) model for generating emotional conversation responses. In conventional CVAE-based emotional response generation, emotion labels are simply used as additional conditions in prior, posterior and decoder networks. Considering that emotion styles are naturally entangled with semantic contents in the language space, the Emo-CVAE model utilizes emotion labels to regularize the CVAE latent space by introducing an extra emotion prediction network. In the training stage, the estimated latent variables are required to predict the emotion labels and token sequences of the input responses simultaneously. Experimental results show that our Emo-CVAE model can learn a more informative and structured latent space than a conventional CVAE model and output responses with better content and emotion performance than baseline CVAE and sequence-to-sequence (Seq2Seq) models.

Chao-Hong Tan, Xiaoyu Yang, Zi'ou Zheng et al.

Task-oriented conversational modeling with unstructured knowledge access, as track 1 of the 9th Dialogue System Technology Challenges (DSTC 9), requests to build a system to generate response given dialogue history and knowledge access. This challenge can be separated into three subtasks, (1) knowledge-seeking turn detection, (2) knowledge selection, and (3) knowledge-grounded response generation. We use pre-trained language models, ELECTRA and RoBERTa, as our base encoder for different subtasks. For subtask 1 and 2, the coarse-grained information like domain and entity are used to enhance knowledge usage. For subtask 3, we use a latent variable to encode dialog history and selected knowledge better and generate responses combined with copy mechanism. Meanwhile, some useful post-processing strategies are performed on the model's final output to make further knowledge usage in the generation task. As shown in released evaluation results, our proposed system ranks second under objective metrics and ranks fourth under human metrics.

Run-Ze Wang, Zhen-Hua Ling, Jing-Bo Zhou et al.

The task of multi-turn text-to-SQL semantic parsing aims to translate natural language utterances in an interaction into SQL queries in order to answer them using a database which normally contains multiple table schemas. Previous studies on this task usually utilized contextual information to enrich utterance representations and to further influence the decoding process. While they ignored to describe and track the interaction states which are determined by history SQL queries and are related with the intent of current utterance. In this paper, two kinds of interaction states are defined based on schema items and SQL keywords separately. A relational graph neural network and a non-linear layer are designed to update the representations of these two states respectively. The dynamic schema-state and SQL-state representations are then utilized to decode the SQL query corresponding to current utterance. Experimental results on the challenging CoSQL dataset demonstrate the effectiveness of our proposed method, which achieves better performance than other published methods on the task leaderboard.

Jing-Xuan Zhang, Li-Juan Liu, Yan-Nian Chen et al.

With the development of automatic speech recognition (ASR) and text-to-speech synthesis (TTS) technique, it's intuitive to construct a voice conversion system by cascading an ASR and TTS system. In this paper, we present a ASR-TTS method for voice conversion, which used iFLYTEK ASR engine to transcribe the source speech into text and a Transformer TTS model with WaveNet vocoder to synthesize the converted speech from the decoded text. For the TTS model, we proposed to use a prosody code to describe the prosody information other than text and speaker information contained in speech. A prosody encoder is used to extract the prosody code. During conversion, the source prosody is transferred to converted speech by conditioning the Transformer TTS model with its code. Experiments were conducted to demonstrate the effectiveness of our proposed method. Our system also obtained the best naturalness and similarity in the mono-lingual task of Voice Conversion Challenge 2020.

Jing-Xuan Zhang, Zhen-Hua Ling, Li-Rong Dai

This paper presents an adversarial learning method for recognition-synthesis based non-parallel voice conversion. A recognizer is used to transform acoustic features into linguistic representations while a synthesizer recovers output features from the recognizer outputs together with the speaker identity. By separating the speaker characteristics from the linguistic representations, voice conversion can be achieved by replacing the speaker identity with the target one. In our proposed method, a speaker adversarial loss is adopted in order to obtain speaker-independent linguistic representations using the recognizer. Furthermore, discriminators are introduced and a generative adversarial network (GAN) loss is used to prevent the predicted features from being over-smoothed. For training model parameters, a strategy of pre-training on a multi-speaker dataset and then fine-tuning on the source-target speaker pair is designed. Our method achieved higher similarity than the baseline model that obtained the best performance in Voice Conversion Challenge 2018.

Yang Ai, Xin Wang, Junichi Yamagishi et al.

This paper presents a reverberation module for source-filter-based neural vocoders that improves the performance of reverberant effect modeling. This module uses the output waveform of neural vocoders as an input and produces a reverberant waveform by convolving the input with a room impulse response (RIR). We propose two approaches to parameterizing and estimating the RIR. The first approach assumes a global time-invariant (GTI) RIR and directly learns the values of the RIR on a training dataset. The second approach assumes an utterance-level time-variant (UTV) RIR, which is invariant within one utterance but varies across utterances, and uses another neural network to predict the RIR values. We add the proposed reverberation module to the phase spectrum predictor (PSP) of a HiNet vocoder and jointly train the model. Experimental results demonstrate that the proposed module was helpful for modeling the reverberation effect and improving the perceived quality of generated reverberant speech. The UTV-RIR was shown to be more robust than the GTI-RIR to unknown reverberation conditions and achieved a perceptually better reverberation effect.

Jia-Chen Gu, Zhen-Hua Ling, Quan Liu et al.

The challenges of building knowledge-grounded retrieval-based chatbots lie in how to ground a conversation on its background knowledge and how to match response candidates with both context and knowledge simultaneously. This paper proposes a method named Filtering before Iteratively REferring (FIRE) for this task. In this method, a context filter and a knowledge filter are first built, which derive knowledge-aware context representations and context-aware knowledge representations respectively by global and bidirectional attention. Besides, the entries irrelevant to the conversation are discarded by the knowledge filter. After that, iteratively referring is performed between context and response representations as well as between knowledge and response representations, in order to collect deep matching features for scoring response candidates. Experimental results show that FIRE outperforms previous methods by margins larger than 2.8% and 4.1% on the PERSONA-CHAT dataset with original and revised personas respectively, and margins larger than 3.1% on the CMU_DoG dataset in terms of top-1 accuracy. We also show that FIRE is more interpretable by visualizing the knowledge grounding process.

Yang Ai, Zhen-Hua Ling

In our previous work, we have proposed a neural vocoder called HiNet which recovers speech waveforms by predicting amplitude and phase spectra hierarchically from input acoustic features. In HiNet, the amplitude spectrum predictor (ASP) predicts log amplitude spectra (LAS) from input acoustic features. This paper proposes a novel knowledge-and-data-driven ASP (KDD-ASP) to improve the conventional one. First, acoustic features (i.e., F0 and mel-cepstra) pass through a knowledge-driven LAS recovery module to obtain approximate LAS (ALAS). This module is designed based on the combination of STFT and source-filter theory, in which the source part and the filter part are designed based on input F0 and mel-cepstra, respectively. Then, the recovered ALAS are processed by a data-driven LAS refinement module which consists of multiple trainable convolutional layers to get the final LAS. Experimental results show that the HiNet vocoder using KDD-ASP can achieve higher quality of synthetic speech than that using conventional ASP and the WaveRNN vocoder on a text-to-speech (TTS) task.

Tianda Li, Jia-Chen Gu, Xiaodan Zhu et al.

Disentanglement is a problem in which multiple conversations occur in the same channel simultaneously, and the listener should decide which utterance is part of the conversation he will respond to. We propose a new model, named Dialogue BERT (DialBERT), which integrates local and global semantics in a single stream of messages to disentangle the conversations that mixed together. We employ BERT to capture the matching information in each utterance pair at the utterance-level, and use a BiLSTM to aggregate and incorporate the context-level information. With only a 3% increase in parameters, a 12% improvement has been attained in comparison to BERT, based on the F1-Score. The model achieves a state-of-the-art result on the a new dataset proposed by IBM and surpasses previous work by a substantial margin.

Jia-Chen Gu, Tianda Li, Quan Liu et al.

In this paper, we study the problem of employing pre-trained language models for multi-turn response selection in retrieval-based chatbots. A new model, named Speaker-Aware BERT (SA-BERT), is proposed in order to make the model aware of the speaker change information, which is an important and intrinsic property of multi-turn dialogues. Furthermore, a speaker-aware disentanglement strategy is proposed to tackle the entangled dialogues. This strategy selects a small number of most important utterances as the filtered context according to the speakers' information in them. Finally, domain adaptation is performed to incorporate the in-domain knowledge into pre-trained language models. Experiments on five public datasets show that our proposed model outperforms the present models on all metrics by large margins and achieves new state-of-the-art performances for multi-turn response selection.