Fine-grained image analysis is widely recognized as highly challenging, since distinguishing individual differences within a certain category, species, or type often depends on tiny, subtle patterns. However, learning fine-grained semantic categories from these subtle part patterns is inherently fragile, as they can easily be overwhelmed by the dominant patterns resting in the coarse-category information. Therefore, how to enhance the relation between the fine-grained semantics and these subtle patterns is the key. To push this frontier, a novel semantic-part alignment (SPA) learning scheme is proposed in this paper. Its general idea is to firstly measure the relevance of each part to the fine-grained semantics, and then regularize the fine-grained visual representation learning. Specifically, it consists of three key components, namely, joint semantic-part modeling, semantic-part set modeling, and optimal semantic-part transport. The joint semantic-part modeling associates each part in an image with the fine-grained semantics in a latent space. Then, the optimal semantic-part transport component is devised to enhance the relation between fine-grained semantic embeddings and the discriminative part embeddings. Notably, the proposed SPA is plug-in-and-play, easy-to-implement, and insensitive to the latent embedding dimension and loss weight. Experiments show the proposed method can substantially boost performance on multiple fine-grained image analysis tasks across various baselines.

Domain generalization aims to learn a representation from the source domain, which can be generalized to arbitrary unseen target domains. A fundamental challenge for visual domain generalization is the domain gap caused by the dramatic style variation whereas the image content is stable. The realm of selective state space, exemplified by VMamba, demonstrates its global receptive field in representing the content. However, the way exploiting the domain-invariant property for selective state space is rarely explored. In this paper, we propose a novel Flow Factorized State Space model, dubbed as DGFamba, for visual domain generalization. To maintain domain consistency, we innovatively map the style-augmented and the original state embeddings by flow factorization. In this latent flow space, each state embedding from a certain style is specified by a latent probability path. By aligning these probability paths in the latent space, the state embeddings are able to represent the same content distribution regardless of the style differences. Extensive experiments conducted on various visual domain generalization settings show its state-of-the-art performance.

Fine-grained domain generalization (FGDG) aims to learn a fine-grained representation that can be well generalized to unseen target domains when only trained on the source domain data. Compared with generic domain generalization, FGDG is particularly challenging in that the fine-grained category can be only discerned by some subtle and tiny patterns. Such patterns are particularly fragile under the cross-domain style shifts caused by illumination, color and etc. To push this frontier, this paper presents a novel Hyperbolic State Space Hallucination (HSSH) method. It consists of two key components, namely, state space hallucination (SSH) and hyperbolic manifold consistency (HMC). SSH enriches the style diversity for the state embeddings by firstly extrapolating and then hallucinating the source images. Then, the pre- and post- style hallucinate state embeddings are projected into the hyperbolic manifold. The hyperbolic state space models the high-order statistics, and allows a better discernment of the fine-grained patterns. Finally, the hyperbolic distance is minimized, so that the impact of style variation on fine-grained patterns can be eliminated. Experiments on three FGDG benchmarks demonstrate its state-of-the-art performance.

Zhuang Li, Yuncheng Hua, Thuy-Trang Vu, Haolan Zhan, Lizhen Qu, Gholamreza Haffari. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Night-time scene segmentation is a critical yet challenging task in the real-world applications, primarily due to the complicated lighting conditions. However, existing methods lack sufficient generalization ability to unseen nighttime scenes with varying illumination. In light of this issue, we focus on investigating generalizable paradigms for night-time scene segmentation and propose an efficient fine-tuning scheme, dubbed NightAdapter, alleviating the domain gap across various scenes. Interestingly, different properties embedded in the day-time and night-time features can be characterized by the bands after discrete sine transform, which can be categorized into illumination-sensitive/-insensitive bands. Hence, our NightAdapter is powered by two appealing designs: (1) Illumination-Insensitive Band Adaptation that provides a foundation for understanding the prior, enhancing the robustness to illumination shifts; (2) Illumination-Sensitive Band Adaptation that fine-tunes the randomized frequency bands, mitigating the domain gap between the day-time and various night-time scenes. As a consequence, illumination-insensitive enhancement improves the domain invariance, while illumination-sensitive diminution strengthens the domain shift between different scenes. NightAdapter yields significant improvements over the state-of-the-art methods under various day-to-night, night-to-night, and in-domain night segmentation experiments. Source code is available at https://github.com/BiQiWHU/NightAdapter.

Task automation has been greatly empowered by the recent advances in Large Language Models (LLMs) via Python code, where the tasks range from software engineering development to general-purpose reasoning. While current benchmarks have shown that LLMs can solve tasks using programs like human developers, the majority of their evaluations are limited to short and self-contained algorithmic tasks or standalone function calls. Solving challenging and practical tasks requires the capability of utilizing **diverse function calls as tools** to efficiently implement functionalities like data analysis and web development. In addition, using multiple tools to solve a task needs compositional reasoning by accurately understanding **complex instructions**. Fulfilling both of these characteristics can pose a great challenge for LLMs. To assess how well LLMs can solve challenging and practical tasks via programs, we introduce BigCodeBench, a benchmark that challenges LLMs to invoke multiple function calls as tools from 139 libraries and 7 domains for 1,140 fine-grained tasks. To evaluate LLMs rigorously, each task encompasses 5.6 test cases with an average branch coverage of 99%. In addition, we propose a natural-language-oriented variant of BigCodeBench, BigCodeBench-Instruct, that automatically transforms the original docstrings into short instructions containing only essential information. Our extensive evaluation of 60 LLMs shows that **LLMs are not yet capable of following complex instructions to use function calls precisely, with scores up to 60%, significantly lower than the human performance of 97%**. The results underscore the need for further advancements in this area.

Image restoration is a fundamental task in computer vision and machine learning, which learns a mapping between the clear images and the degraded images under various conditions (e.g., blur, low-light, haze). Yet, most existing image restoration methods are highly restricted by the requirement of degraded and clear image pairs, which limits the generalization and feasibility to enormous real-world scenarios without paired images. To address this bottleneck, we propose a Degradation-aware Dynamic Schr\"{o}dinger Bridge (DDSB) for unpaired image restoration. Its general idea is to learn a Schr\"{o}dinger Bridge between clear and degraded image distribution, while at the same time emphasizing the physical degradation priors to reduce the accumulation of errors during the restoration process. A Degradation-aware Optimal Transport (DOT) learning scheme is accordingly devised. Training a degradation model to learn the inverse restoration process is particularly challenging, as it must be applicable across different stages of the iterative restoration process. A Dynamic Transport with Consistency (DTC) learning objective is further proposed to reduce the loss of image details in the early iterations and therefore refine the degradation model. Extensive experiments on multiple image degradation tasks show its state-of-the-art performance over the prior arts.

Domain generalization aims to train models that perform robustly on unseen target domains without access to target data. The realm of vision-language foundation model has opened a new venue owing to its inherent out-of-distribution generalization capability. However, the static alignment to class-level textual anchors remains insufficient to handle the dramatic distribution discrepancy from diverse domain-specific visual features. In this work, we propose a novel cross-domain Schrödinger Bridge (SB) method, namely SBGen, to handle this challenge, which explicitly formulates the stochastic semantic evolution, to gain better generalization to unseen domains. Technically, the proposed \texttt{SBGen} consists of three key components: (1) \emph{text-guided domain-aware feature selection} to isolate semantically aligned image tokens; (2) \emph{stochastic cross-domain evolution} to simulate the SB dynamics via a learnable time-conditioned drift; and (3) \emph{stochastic domain-agnostic interpolation} to construct semantically grounded feature trajectories. Empirically, \texttt{SBGen} achieves state-of-the-art performance on domain generalization in both classification and segmentation. This work highlights the importance of modeling domain shifts as structured stochastic processes grounded in semantic alignment.

High-quality annotation of fine-grained visual categories demands great expert knowledge, which is taxing and time consuming. Alternatively, learning fine-grained visual representation from enormous unlabeled images (e.g., species, brands) by self-supervised learning becomes a feasible solution. However, recent investigations find that existing self-supervised learning methods are less qualified to represent fine-grained categories. The bottleneck lies in that the pre-trained class-agnostic representation is built from every patch-wise embedding, while fine-grained categories are only determined by several key patches of an image. In this paper, we propose a Cross-level Multi-instance Distillation (CMD) framework to tackle this challenge. Our key idea is to consider the importance of each image patch in determining the fine-grained representation by multiple instance learning. To comprehensively learn the relation between informative patches and fine-grained semantics, the multi-instance knowledge distillation is implemented on both the region/image crop pairs from the teacher and student net, and the region-image crops inside the teacher / student net, which we term as intra-level multi-instance distillation and inter-level multi-instance distillation. Extensive experiments on several commonly used datasets, including CUB-200-2011, Stanford Cars and FGVC Aircraft, demonstrate that the proposed method outperforms the contemporary methods by up to 10.14% and existing state-of-the-art self-supervised learning approaches by up to 19.78% on both top-1 accuracy and Rank-1 retrieval metric. Source code is available at https://github.com/BiQiWHU/CMD.

Sociocultural norms serve as guiding principles for personal conduct in social interactions, emphasizing respect, cooperation, and appropriate behavior, which is able to benefit tasks including conversational information retrieval, contextual information retrieval, and retrieval-enhanced machine learning. We propose a scalable approach for constructing a Sociocultural Norm ( Scn ) Base using large language models (LLMs) for socially aware dialogues. We construct a comprehensive and publicly accessible Chinese Sociocultural NormBase ( ChineseNormBase ). Our approach utilizes socially aware dialogues, enriched with contextual frames, as the primary data source to constrain the generating process and reduce the hallucinations. This enables extracting of high-quality and nuanced natural-language norm statements, leveraging the pragmatic implications of utterances with respect to the situation. As real dialogue annotated with gold frames are not readily available, we propose using synthetic data. Our empirical results show (i) the quality of the Scn s derived from synthetic data is comparable to that from real dialogues annotated with gold frames, and (ii) the quality of the Scn s extracted from real data, annotated with either silver (predicted) or gold frames, surpasses that without the frame annotations. We further show the effectiveness of the extracted Scn s in a Retrieval-Augmented Generation (RAG)-based model to reason about multiple downstream dialogue tasks.

Machine learning models have made incredible progress, but they still struggle when applied to examples from unseen domains.This study focuses on a specific problem of domain generalization, where a model is trained on one source domain and tested on multiple target domains that are unseen during training.We propose IMO: Invariant features Masks for Out-of-Distribution text classification, to achieve OOD generalization by learning domain-invariant features.During training, IMO employs a greedy algorithm to learn sparse representations for each layer in a top-down manner.It performs better than the opposite direction and learning of sparse representations for all layers simultaneously.Our comprehensive experiments show that IMO substantially outperforms strong baselines such as prompt-based methods and large language models, in terms of various evaluation metrics and settings.1

The rapid development of Vision Foundation Model (VFM) brings inherent out-domain generalization for a variety of down-stream tasks. Among them, domain generalized semantic segmentation (DGSS) holds unique challenges as the cross-domain images share common pixel-wise content information but vary greatly in terms of the style. In this paper, we present a novel Spectral-dEcomposed Token (SET) learning framework to advance the frontier. Delving into further than existing fine-tuning token & frozen backbone paradigm, the proposed SET especially focuses on the way learning style-invariant features from these learnable tokens. Particularly, the frozen VFM features are first decomposed into the phase and amplitude components in the frequency space, which mainly contain the information of content and style, respectively, and then separately processed by learnable tokens for task-specific information extraction. Particularly, the frozen VFM features are first decomposed into the phase and amplitude components in the frequency space, which mainly contain the information of content and style, respectively, and then separately processed by learnable tokens for task-specific information extraction.After the decomposition, style variation primarily impacts the token-based feature enhancement within the amplitude branch. To address this issue, we further develop an attention optimization method to bridge the gap between style-affected representation and static tokens during inference. Extensive cross-domain experiments show its state-of-the-art performance.

Disease grading is a crucial task in medical image analysis. Due to the continuous progression of diseases, i.e., the variability within the same level and the similarity between adjacent stages, accurate grading is highly challenging. Furthermore, in real-world scenarios, models trained on limited source domain datasets should also be capable of handling data from unseen target domains. Due to the cross-domain variants, the feature distribution between source and unseen target domains can be dramatically different, leading to a substantial decrease in model performance. To address these challenges in cross-domain disease grading, we propose a Severity-aware Recurrent Modeling (Samba) method in this paper. As the core objective of most staging tasks is to identify the most severe lesions, which may only occupy a small portion of the image, we propose to encode image patches in a sequential and recurrent manner. Specifically, a state space model is tailored to store and transport the severity information by hidden states. Moreover, to mitigate the impact of cross-domain variants, an Expectation-Maximization (EM) based state recalibration mechanism is designed to map the patch embeddings into a more compact space. We model the feature distributions of different lesions through the Gaussian Mixture Model (GMM) and reconstruct the intermediate features based on learnable severity bases. Extensive experiments show the proposed Samba outperforms the VMamba baseline by an average accuracy of 23.5\%, 5.6\% and 4.1\% on the cross-domain grading of fatigue fracture, breast cancer and diabetic retinopathy, respectively. Source code is available at \url{https://github.com/BiQiWHU/Samba}.

The emerging vision foundation model (VFM) has inherited the ability to generalize to unseen images. Nevertheless, the key challenge of domain-generalized semantic segmentation (DGSS) lies in the domain gap attributed to the cross-domain styles, i.e., the variance of urban landscape and environment dependencies. Hence, maintaining the style-invariant property with varying domain styles becomes the key bottleneck in harnessing VFM for DGSS. The frequency space after Haar wavelet transformation provides a feasible way to decouple the style information from the domain-invariant content, since the content and style information are retained in the low- and high- frequency components of the space, respectively. To this end, we propose a novel Frequency-Adapted (FADA) learning scheme to advance the frontier. Its overall idea is to separately tackle the content and style information by frequency tokens throughout the learning process. Particularly, the proposed FADA consists of two branches, i.e., low- and high- frequency branches. The former one is able to stabilize the scene content, while the latter one learns the scene styles and eliminates its impact to DGSS. Experiments conducted on various DGSS settings show the state-of-the-art performance of our FADA and its versatility to a variety of VFMs. Source code is available at \url{https://github.com/BiQiWHU/FADA}.

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 grounding 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 contribute 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 IEMO-CAP.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.

Existing neural methods have shown great potentials towards generating informative text from structured tabular data as well as maintaining high content fidelity. However, few of them shed light on generating personalized expressions, which often requires well-aligned persona-table-text datasets that are difficult to obtain. To overcome these obstacles, we explore personalized table-to-text generation under a zero-shot setting, by assuming no well-aligned persona-table-text triples are required during training. To this end, we firstly collect a set of unpaired persona information and then propose a semi-supervised approach with contrastive persona distillation (S2P-CPD) to generate personalized context. Specifically, tabular data and persona information are firstly represented as latent variables separately. Then, we devise a latent space fusion technique to distill persona information into the table representation. Besides, a contrastive-based discriminator is employed to guarantee the style consistency between the generated context and its corresponding persona. Experimental results on two benchmarks demonstrate S2P-CPD’s ability on keeping both content fidelity and personalized expressions.

The principal task in neural dialogue generation is to learn to generate coherent and diverse responses conditioned on dialogue history, and thus imitate the intrinsic conversational skills of real humans. Conventional training approaches for dialogue generation (e.g., Maximum Likely Estimation (MLE)) focus on capturing coherent syntactic patterns, but fall short of showcasing diversity in expression. In this paper, we utilize contrastive learning for dialogue, which extends the conventional MLE-based dialogue training with a contrastive counterpart, where dialogue generation models can be explicitly trained on multiple positive and negative pairs. However, naive contrastive learning with discrete manipulations is sub-optimal, since they tend to sample examples that can be easily distinguished. To migrate this problem, we propose a novel contrastive learning paradigm with continuous semantic perturbations (CL-CSP), which constructs positive examples within an adjacency semantic region and pushes those negative examples far away from the semantic region. Such “soft” positive and negative pairs constructed by CL-CSP guides the model to distinguish positive instances from negative ones better. We conduct extensive experiments to demonstrate that our proposed method improves the diversity of response generation on two benchmarks, showing that our proposed techniques are crucial to achieving superior performance.

In the current landscape of pervasive smartphones and tablets, apps frequently exist across both platforms.Although apps share most graphic user interfaces (GUIs) and functionalities across phones and tablets, developers often rebuild from scratch for tablet versions, escalating costs and squandering existing design resources.Researchers are attempting to collect data and employ deep learning in automated GUIs development to enhance developers' productivity.There are currently several publicly accessible GUI page datasets for phones, but none for pairwise GUIs between phones and tablets.This poses a significant barrier to the employment of deep learning in automated GUI development.In this paper, we introduce the Papt dataset, a pioneering pairwise GUI dataset tailored for Android phones and tablets, encompassing 10,035 phone-tablet GUI page pairs sourced from 5,593 unique app pairs.We propose novel pairwise GUI collection approaches for constructing this dataset and delineate its advantages over currently prevailing datasets in the field.Through preliminary experiments on this dataset, we analyze the present challenges of utilizing deep learning in automated GUI development.

Content Warning: this paper may contain content that is offensive or upsetting.

In product description generation (PDG), the user-cared aspect is critical for the recommendation system, which can not only improve user's experiences but also obtain more clicks. High-quality customer reviews can be considered as an ideal source to mine user-cared aspects. However, in reality, a large number of new products (known as long-tailed commodities) cannot gather sufficient amount of customer reviews, which brings a big challenge in the product description generation task. Existing works tend to generate the product description solely based on item information, i.e., product attributes or title words, which leads to tedious contents and cannot attract customers effectively. To tackle this problem, we propose an adaptive posterior network based on Transformer architecture that can utilize user-cared information from customer reviews. Specifically, we first extend the self-attentive Transformer encoder to encode product titles and attributes. Then, we apply an adaptive posterior distillation module to utilize useful review information, which integrates user-cared aspects to the generation process. Finally, we apply a Transformer-based decoding phase with copy mechanism to automatically generate the product description. Besides, we also collect a large-scare Chinese product description dataset to support our work and further research in this field. Experimental results show that our model is superior to traditional generative models in both automatic indicators and human evaluation.

Being able to reply with a related, fluent, and informative response is an indispensable requirement for building high-quality conversational agents. In order to generate better responses, some approaches have been proposed, such as feeding extra information by collecting large-scale datasets with human annotations, designing neural conversational models (NCMs) with complex architecture and loss functions, or filtering out untrustworthy samples based on a dialogue attribute, e.g., Relatedness or Genericness. In this paper, we follow the third research branch and present a data filtering method for open-domain dialogues, which identifies untrustworthy samples from training data with a quality measure that linearly combines seven dialogue attributes. The attribute weights are obtained via Bayesian Optimization (BayesOpt) that aims to optimize an objective function for dialogue generation iteratively on the validation set. Then we score training samples with the quality measure, sort them in descending order, and filter out those at the bottom. Furthermore, to accelerate the "filter-train-evaluate'' iterations involved in BayesOpt on large-scale datasets, we propose a training framework that integrates maximum likelihood estimation (MLE) and negative training method (NEG). The training method updates parameters of a trained NCMs on two small sets with newly maintained and removed samples, respectively. Specifically, MLE is applied to maximize the log-likelihood of newly maintained samples, while NEG is used to minimize the log-likelihood of newly removed ones. Experimental results on two datasets show that our method can effectively identify untrustworthy samples, and NCMs trained on the filtered datasets achieve better performance.

Knowledge-grounded dialogue generation has achieved promising performance with the engagement of external knowledge sources.Typical approaches towards this task usually perform relatively independent two sub-tasks, i.e., knowledge selection and knowledge-aware response generation.In this paper, in order to improve the diversity of both knowledge selection and knowledge-aware response generation, we propose a collaborative latent variable (CoLV) model to integrate these two aspects simultaneously in separate yet collaborative latent spaces, so as to capture the inherent correlation between knowledge selection and response generation.During generation, our proposed model firstly draws knowledge candidate from the latent space conditioned on the dialogue context, and then samples a response from another collaborative latent space conditioned on both the context and the selected knowledge.Experimental results on two widely-used knowledge-grounded dialogue datasets show that our model outperforms previous methods on both knowledge selection and response generation.

Open-domain multi-turn conversations mainly have three features, which are hierarchical semantic structure, redundant information, and long-term dependency. Grounded on these, selecting relevant context becomes a challenge step for multiturn dialogue generation. However, existing methods cannot differentiate both useful words and utterances in long distances from a response. Besides, previous work just performs context selection based on a state in the decoder, which lacks a global guidance and could lead some focuses on irrelevant or unnecessary information. In this paper, we propose a novel model with hierarchical self-attention mechanism and distant supervision to not only detect relevant words and utterances in short and long distances, but also discern related information globally when decoding. Experimental results on two public datasets of both automatic and human evaluations show that our model significantly outperforms other baselines in terms of fluency, coherence, and informativeness.

Open-domain dialogue generation in natural language processing (NLP) is by default a pure-language task, which aims to satisfy human need for daily communication on open-ended topics by producing related and informative responses. In this paper, we point out that hidden images, named as visual impressions (VIs), can be explored from the text-only data to enhance dialogue understanding and help generate better responses. Besides, the semantic dependency between an dialogue post and its response is complicated, e.g., few word alignments and some topic transitions. Therefore, the visual impressions of them are not shared, and it is more reasonable to integrate the response visual impressions (RVIs) into the decoder, rather than the post visual impressions (PVIs). However, both the response and its RVIs are not given directly in the test process. To handle the above issues, we propose a framework to explicitly construct VIs based on pure-language dialogue datasets and utilize them for better dialogue understanding and generation. Specifically, we obtain a group of images (PVIs) for each post based on a pre-trained word-image mapping model. These PVIs are used in a co-attention encoder to get a post representation with both visual and textual information. Since the RVIs are not provided during testing, we design a cascade decoder that consists of two sub-decoders. The first sub-decoder predicts the content words in response, and applies the word-image mapping model to get corresponding RVIs. Then, the second sub-decoder generates the response based on the post and RVIs. Experimental results on two open-domain dialogue datasets show that our proposed approach achieves superior performance over competitive baselines in terms of fluency, relatedness, and diversity.

Haolan Zhan, Hainan Zhang, Hongshen Chen, Zhuoye Ding, Yongjun Bao, Yanyan Lan. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

Recommendation reason generation, aiming at showing the selling points of products for customers, plays a vital role in attracting customers' attention as well as improving user experience. A simple and effective way is to extract keywords directly from the knowledge-base of products, i.e., attributes or title, as the recommendation reason. However, generating recommendation reason from product knowledge doesn't naturally respond to users' interests. Fortunately, on some E-commerce websites, there exists more and more user-generated content (user-content for short), i.e., product question-answering (QA) discussions, which reflect user-cared aspects. Therefore, in this paper, we consider generating the recommendation reason by taking into account not only the product attributes but also the customer-generated product QA discussions. In reality, adequate user-content is only possible for the most popular commodities, whereas large sums of long-tail products or new products cannot gather a sufficient number of user-content. To tackle this problem, we propose a user-inspired multi-source posterior transformer (MSPT), which induces the model reflecting the users' interests with a posterior multiple QA discussions module, and generating recommendation reasons containing the product attributes as well as the user-cared aspects. Experimental results show that our model is superior to traditional generative models. Additionally, the analysis also shows that our model can focus more on the user-cared aspects than baselines.

Multi-turn conversations consist of complex semantic structures, and it is still a challenge to generate coherent and diverse responses given previous utterances.It's practical that a conversation takes place under a background, meanwhile, the query and response are usually most related and they are consistent in topic but also different in content.However, little work focuses on such hierarchical relationship among utterances.To address this problem, we propose a Conversational Semantic Relationship RNN (CSRR) model to construct the dependency explicitly.The model contains latent variables in three hierarchies.The discourse-level one captures the global background, the pair-level one stands for the common topic information between query and response, and the utterance-level ones try to represent differences in content.Experimental results show that our model significantly improves the quality of responses in terms of fluency, coherence and diversity compared to baseline methods.