Fangda Ye, Kuicai Dong, Xie Zhifei, Yuxin Hu, Yihang Yin, Shurui Huang, Shikai Dong, Chen Zhang, Jianzhu Bao, Shuicheng Yan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Deep Research systems have revolutionized how LLMs solve complex questions through iterative reasoning and evidence gathering. However, current systems remain fundamentally constrained to textual web data, overlooking the vast knowledge embedded in multimodal documents: scientific papers, technical reports, and financial documents where critical information exists in figures, tables, charts, and equations. Processing such documents demands sophisticated parsing to preserve visual semantics, intelligent chunking to maintain structural coherence, and adaptive retrieval across modalities, which are capabilities absent in existing systems. In response, we present Doc-Researcher, a unified system that bridges this gap through three integrated components: (i) deep multimodal parsing that preserves layout structure and visual semantics while creating multi-granular representations from chunk to document level, (ii) systematic retrieval architecture supporting text-only, vision-only, and hybrid paradigms with dynamic granularity selection, and (iii) iterative multi-agent workflows that decompose complex queries, progressively accumulate evidence, and synthesize comprehensive answers across documents and modalities. To enable rigorous evaluation, we introduce M4DocBench, the first benchmark for Multi-modal, Multi-hop, Multi-document, and Multi-turn deep research. Featuring 158 expert-annotated questions with complete evidence chains across 304 documents, M4DocBench tests capabilities that existing benchmarks cannot assess. Experiments demonstrate that Doc-Researcher achieves 50.6% accuracy, 3.4× better than state-of-the-art baselines, validating that effective document research requires not just better retrieval, but fundamentally deep parsing that preserve multimodal integrity and support iterative research. Our work establishes a new paradigm for conducting deep research on multimodal document collections.

With the diversification of 5G networks, the accurate classification of network traffic is of great significance to network management and optimization. Based on the study of classical volumes and neural networks, this paper proposes a lightweight, low convolutional neural network traffic classification algorithm-LWLC-CNN, to complete network traffic classification. Experimental results show that the proposed algorithm significantly reduces the computer's training capacity and model size, and has a good effect on network traffic classification.

Rationality and emotion are two fundamental elements of humans. Endowing agents with rationality and emotion has been one of the major milestones in AI. However, in the field of conversational AI, most existing models only specialize in one aspect and neglect the other, which often leads to dull or unrelated responses. In this paper, we hypothesize that combining rationality and emotion into conversational agents can improve response quality. To test the hypothesis, we focus on one fundamental aspect of rationality, i.e., commonsense, and propose CARE, a novel model for commonsense-aware emotional response generation. Specifically, we first propose a framework to learn and construct commonsense-aware emotional latent concepts of the response given an input message and a desired emotion. We then propose three methods to collaboratively incorporate the latent concepts into response generation. Experimental results on two large-scale datasets support our hypothesis and show that our model can produce more accurate and commonsense-aware emotional responses and achieve better human ratings than state-of-the-art models that only specialize in one aspect.

Virtual assistants aim to build a human-like conversational agent. However, current human-machine conversations still cannot make users feel intelligent enough to build a continued dialog over time. Some responses from agents are usually inconsistent, uninformative, less-engaging and even memoryless. In recent years, most researchers have tried to employ conversation context and external knowledge, e.g. wiki pages and knowledge graphs, into the model which only focuses on solving some special conversation problems in local perspectives. Few researchers are dedicated to the whole capability of the conversational agent which is endowed with abilities of not only passively reacting the conversation but also proactively leading the conversation.

Empathetic conversational models have been shown to improve user satisfaction and task outcomes in numerous domains.In Psychology, persona has been shown to be highly correlated to personality, which in turn influences empathy.In addition, our empirical analysis also suggests that persona plays an important role in empathetic conversations.To this end, we propose a new task towards persona-based empathetic conversations and present the first empirical study on the impact of persona on empathetic responding.Specifically, we first present a novel large-scale multi-domain dataset for persona-based empathetic conversations.We then propose CoBERT, an efficient BERT-based response selection model that obtains the state-of-the-art performance on our dataset.Finally, we conduct extensive experiments to investigate the impact of persona on empathetic responding.Notably, our results show that persona improves empathetic responding more when CoBERT is trained on empathetic conversations than non-empathetic ones, establishing an empirical link between persona and empathy in human conversations.

Massive public resume data emerging on the internet indicates individual-related characteristics in terms of profile and career experiences. Resume Analysis (RA) provides opportunities for many applications, such as recruitment trend predict, talent seeking and evaluation. Existing RA studies either largely rely on the knowledge of domain experts, or leverage classic statistical or data mining models to identify and filter explicit attributes based on pre-defined rules. However, they fail to discover the latent semantic information from semi-structured resume text, i.e., individual career progress trajectory and social-relations, which are otherwise vital to comprehensive understanding of people’s career evolving patterns. Besides, when dealing with large numbers of resumes, how to properly visualize such semantic information to reduce the information load and to support better human cognition is also challenging. To tackle these issues, we propose a visual analytics system called ResumeVis to mine and visualize resume data. First, a text mining-based approach is presented to extract semantic information. Then, a set of visualizations are devised to represent the semantic information in multiple perspectives. Through interactive exploration on ResumeVis performed by domain experts, the following tasks can be accomplished: to trace individual career evolving trajectory; to mine latent social-relations among individuals; and to hold the full picture of massive resumes’ collective mobility. Case studies with over 2,500 government officer resumes demonstrate the effectiveness of our system.

Many important data mining problems can be modeled as learning a (bidirectional) multidimensional mapping between two data domains. Based on the generative adversarial networks (GANs), particularly conditional ones, cross-domain joint distribution matching is an increasingly popular kind of methods addressing such problems. Though significant advances have been achieved, there are still two main disadvantages of existing models, i.e., the requirement of large amount of paired training samples and the notorious instability of training. In this paper, we propose a multi-view adversarially learned inference (ALI) model, termed as MALI, to address these issues. Unlike the common practice of learning direct domain mappings, our model relies on shared latent representations of both domains and can generate arbitrary number of paired faking samples, benefiting from which usually very few paired samples (together with sufficient unpaired ones) is enough for learning good mappings. Extending the vanilla ALI model, we design novel discriminators to judge the quality of generated samples (both paired and unpaired), and provide theoretical analysis of our new formulation. Experiments on image-to-image translation, image-to-attribute generation (multi-label classification), attribute-to-image generation tasks demonstrate that our semi-supervised learning framework yields significant performance improvements over existing ones. Results on cross-modality retrieval show that our latent space based method can achieve competitive similarity search performance in relative fast speed, compared to those methods that compute similarities in the high-dimensional data space.

Stock comments from analysts contain important consulting information for investors to foresee stock volatility and market trends. Existing studies on stock comments usually focused on capturing coarse-grained opinion polarities or understanding market fundamentals. However, investors are often overwhelmed and confused by massive comments with huge noises and ambiguous opinions. Therefore, it is an emerging need to have a fine-grained stock comment analysis tool to identify more reliable stock comments. To this end, this paper provides a solution called StockAssIstant for modeling the reliability of stock comments by considering multiple factors, such as stock price trends, comment content, and the performances of analysts, in a holistic manner. Specifically, we first analyze the pattern of analysts' opinion dynamics from historical comments. Then, we extract key features from the time-series constructed by using the semantic information in comment text, stock prices and the historical behaviors of analysts. Based on these features, we propose an ensemble learning based approach for measuring the reliability of comments. Finally, we conduct extensive experiments and provide a trading simulation on real-world stock data. The experimental results and the profit achieved by the simulated trading in 12-month period clearly validate the effectiveness of our approach for modeling the reliability of stock comments.

Learning an appropriate distance metric plays a substantial role in the success of many learning machines. Conventional metric learning algorithms have limited utility when the training and test samples are drawn from related but different domains (i.e., source domain and target domain). In this letter, we propose two novel metric learning algorithms for domain adaptation in an information-theoretic setting, allowing for discriminating power transfer and standard learning machine propagation across two domains. In the first one, a cross-domain Mahalanobis distance is learned by combining three goals: reducing the distribution difference between different domains, preserving the geometry of target domain data, and aligning the geometry of source domain data with label information. Furthermore, we devote our efforts to solving complex domain adaptation problems and go beyond linear cross-domain metric learning by extending the first method to a multiple kernel learning framework. A convex combination of multiple kernels and a linear transformation are adaptively learned in a single optimization, which greatly benefits the exploration of prior knowledge and the description of data characteristics. Comprehensive experiments in three real-world applications (face recognition, text classification, and object categorization) verify that the proposed methods outperform state-of-the-art metric learning and domain adaptation methods.

As a fundamental constituent of machine learning, domain adaptation generalizes a learning model from a source domain to a different (but related) target domain. In this paper, we focus on semi-supervised domain adaptation and explicitly extend the applied range of unlabeled target samples into the combination of distribution alignment and adaptive classifier learning. Specifically, our extension formulates the following aspects in a single optimization: 1) learning a cross-domain predictive model by developing the Fredholm integral based kernel prediction framework; 2) reducing the distribution difference between two domains; 3) exploring multiple kernels to induce an optimal learning space. Correspondingly, such an extension is distinguished with allowing for noise resiliency, facilitating knowledge transfer and analyzing diverse data characteristics. It is emphasized that we prove the differentiability of our formulation and present an effective optimization procedure based on the reduced gradient, guaranteeing rapid convergence. Comprehensive empirical studies verify the effectiveness of the proposed method.

Transit advertising provides frequent exposure to a large number of residents in different urban regions. However, traditional methods are generally manual and qualitatively based on a rough estimation, such as the number of passengers taken by a bus or functional regions covered. How to accurately put an advertisement on appropriate transportations becomes an important task for potential business value. In this paper, our goal is to recommend top-k bus routes for an ad, which can maximize advertising effectiveness, by mining multiple data sources, such as Smart Card Transaction (SCT) data, geographic data and point of interests (POIs) data. We propose a framework that captures the mobility patterns of passengers and the characteristics of bus stations so as to evaluate the influence of an ad for the passengers quantitatively. Specifically, we first extract the passenger trajectories to model passenger mobility from SCT data and then model each bus station over three characteristics (i.e., hub degree, topic distribution, district region) by geographic data and nearby POIs. According to three characteristics of bus stations, we formulate three application scenarios and design the corresponding methods to make top-k bus routes recommendation. Extensive experiments on real-world public transportation data validate the effectiveness of our proposed approach.

Sequence pattern mining is an important mining task in data mining. However, most researches focus on improving the efficiency of algorithms, more and more attentions are paid to independently analyzing each frequent sequential pattern, which could bring biases to the final decisions. To understand the whole situation with sequence patterns, this paper proposes a systematic approach called Timing-IdeaGraph to build a directed cognition graph. Firstly, with consideration of big data on event sequences, an efficient algorithm is applied to capture frequent sequential patterns. Next, duplicate patterns are removed. After that, we merge relevant patterns and visualize them into a directed cognition graph. In order to make the approach human-centric, we propose an algorithm to identify bridge events and patterns which would proactively trigger human's deep cognition, e.g., creative design, for better decision making. Two real case studies are introduced to show how to use Timing-IdeaGraph in a computer supported cooperative environment.

The real-time information on the Web changes dynamically and surge quickly, which cause considerable difficulty in access to interested information. How to mine hot events, how to analyze the correlation of events and how to organize information structurally are challenging tasks. In this paper, to address these problems, we propose STeller, an approach to mine context-aware story — a series of correlated events. Firstly, we cluster similar pieces of information text into a meme—a piece of information and all its variants. This is also the process of information flow tracking. We view a meme as a fine-grained event. Then we use three novel efficient similarity metrics to measure content similarity and correlation of events. The social stream can be transformed into co-occurrence graph and we define the context-aware story as a novel dense subgraph type called (λ,d)-Clique. Lastly, two corresponding dense subgraph mining algorithms are developed to extract (λ,d)-Clique structure. We also perform detailed experiments on real news data and the results demonstrate the value of our work.

Learning an appropriate feature representation across source and target domains is one of the most effective solutions to domain adaptation problems. Conventional cross-domain feature learning methods rely on the Reproducing Kernel Hilbert Space (RKHS) induced by a single kernel. Recently, Multiple Kernel Learning (MKL), which bases classifiers on combinations of kernels, has shown improved performance in the tasks without distribution difference between domains. In this paper, we generalize the framework of MKL for cross-domain feature learning and propose a novel Transfer Feature Representation (TFR) algorithm. TFR learns a convex combination of multiple kernels and a linear transformation in a single optimization which integrates the minimization of distribution difference with the preservation of discriminating power across domains. As a result, standard machine learning models trained in the source domain can be reused for the target domain data. After rewritten into a differentiable formulation, TFR can be optimized by a reduced gradient method and reaches the convergence. Experiments in two real-world applications verify the effectiveness of our proposed method.

Idea Graph is a core component of Idea Discovery, which discovers idea by converting unstructured data into a scenario graph. Since the scenario graph is complex and heterogeneous, the layouts generated by general graph layout algorithms can't well support human cognition. To tackle this issue, a novel graph layout algorithm named CiFDAL is proposed, which is a hybrid of circular layout algorithm and Force-Directed Algorithm (FDA). Overall, it places clusters and key nodes on two nested circles, and different hierarchies distinguish their different importance. Specifically, it uses FDA to ensure the related nodes being closer in distance. Additionally, it improves the classical random initialized FDA by adopting a better initial position set and iterative node-position exchange technique. Experimental results demonstrate the superiority of our algorithm by comparing with several benchmarks implemented in a famous visualization tool - GraphViz.

Recently, Retrospective Event Detection attracts much attention. Most researches focus on detecting coarse-grained events or frequent events. However, they neglect to discover fine grained events or important rare events which are significant for human decision making. The important rare events are different from the frequent events revealing common patterns, their features are unnoticed and cannot be effectively extracted from the historical data. To tackle this issue, a systematic approach called RCFGED is proposed to simultaneously detect coarse and fine-grained events by incorporating Chance Discovery theory with topic modeling. The approach employs a Chance Discovery algorithm called Idea Graph, which mines the latent term relations for converting the corpus into a term graph. Then, a semantic-relation extraction approach is proposed based on topic modeling to enrich the graph. Lastly, a graph analytical method is employed to detect the events from the graph. An experiment demonstrates the superiority of RCFGED by comparing with benchmarks.

Supervised metric learning plays a substantial role in statistical classification. Conventional metric learning algorithms have limited utility when the training data and testing data are drawn from related but different domains (i.e., source domain and target domain). Although this issue has got some progress in feature-based transfer learning, most of the work in this area suffers from non-trivial optimization and pays little attention to preserving the discriminating information. In this paper, we propose a novel metric learning algorithm to transfer knowledge from the source domain to the target domain in an information-theoretic setting, where a shared Mahalanobis distance across two domains is learnt by combining three goals together: 1) reducing the distribution difference between different domains; 2) preserving the geometry of target domain data; 3) aligning the geometry of source domain data with its label information. Based on this combination, the learnt Mahalanobis distance effectively transfers the discriminating power and propagates standard classifiers across these two domains. More importantly, our proposed method has closed-form solution and can be efficiently optimized. Experiments in two real-world applications demonstrate the effectiveness of our proposed method.

This paper proposes a novel bootstrapping based framework jointed with automatic refinement to extract opinion words and targets. We employ a reasonable set of opinion seed words and pre-defined rules to start bootstrapping. We leverage statistical word co-occurrence and dependency patterns for propagation between opinion words and targets. A Sentiment Graph Model (SGM) is constructed to evaluate these opinion relations. Furthermore, we employ Automatic Rule Refinement (ARR) to refine the rules to extract false results. By using false results pruning and ARR process, we can efficiently alleviate the error propagation problem in traditional bootstrapping-based methods. Preliminary evaluation shows the effectiveness of our method.

A complex or ill-structured problem cannot be solved only replying on computer since they require human's nonlinear thinking and even innovation. In this paper, taking respective advantages of computer and human, we propose a novel computational knowledge environment (CKE) where human is able to cognize tacit and new knowledge for complex or ill-structured problem solving, through fusing potential knowledge discovered from the collected information data supported by computer. Following the CKE framework, we develop a creativity technique called InnoSpider which greatly support human's perception, intuition, reasoning and cognition for quick decision making. A case study have verified the effectiveness of CKE, and have indicated InnoSpider can drive human to make high-quality decisions in a shorter time.

In recent years, Chance Discovery as an extension of data mining has been presented to discover rare but significant events, i.e., chances, for human decision making from large amounts of data. KeyGraph or IdeaGraph as a chance mining algorithm can capture these chances by converting the unstructured data into a scenario graph. However, they both fail to eliminate the interference of frequent events when uncovering rare events, causing a bottleneck of capturing important rare events. In this paper, we propose an improved algorithm of IdeaGraph to address this issue. It takes rare events as the essential components to preserve them from being filtered when forming a cluster. On base of that, it conducts cluster refining such as pruning and ranking to optimize the construction of a scenario graph. Additionally, it provides an enhanced method to evaluate important rare events by measuring the significance of an event on the perspectives of the co-occurring frequency and the probability distribution. An experiment demonstrates the superiority of our algorithm on capturing important rare events by comparing with benchmark algorithms.