Jin Ma
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Jin Ma has a total of 39 co-authors. Here are the top co-authors by number of shared publications:
- Dezhi Ye
Tencent, Beijing, China
6 shared publications - Jie Liu
Tencent, Beijing, China
6 shared publications - Bowen Tian
Tencent, Beijing, China
5 shared publications - Haijin Liang
Tencent, Beijing, China
4 shared publications - Jiabin Fan
Tencent, Beijing, China
4 shared publications - Junwei Hu
Tencent, Beijing, China
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 8410-8418
UniRank: A Unified Framework for Efficient Multi-Objective LLM Ranking in Industrial Search
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 8410-8418
Open AccessUniRank: A Unified Framework for Efficient Multi-Objective LLM Ranking in Industrial Search
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 14548-14558
mR$^{2}$AG: Multimodal Retrieval-Reflection-Augmented Generation for Knowledge-Based VQA
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Multimedia
- Vol. 28, 14548-14558
mR$^{2}$AG: Multimodal Retrieval-Reflection-Augmented Generation for Knowledge-Based VQA
- Open Access
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5433-5438
Can LLMs Really Help Query Understanding In Web Search? A Practical Perspective
- 10 November 2025
- Published by Association for Computing Machinery (ACM)
- p. 5433-5438
Open AccessCan LLMs Really Help Query Understanding In Web Search? A Practical Perspective
- Open Access
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 5171-5181
Applying Large Language Model For Relevance Search In Tencent
- 3 August 2025
- Published by Association for Computing Machinery (ACM)
- p. 5171-5181
Open AccessApplying Large Language Model For Relevance Search In Tencent
- Open Access
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4208-4212
Span Confusion is All You Need for Chinese Spelling Correction
Chinese Spelling Correction is to detect errors in a statement and correct them. Researchers have recently begun to investigate the use of pre-train language models to perform this task. They replace tokens with similar characters based on a confusion set rather than mask tokens when pre-training. A significant limitation of this method is that pre-training with single tokens cannot effectively simulate natural errors, such as full spelling. Furthermore, the confusion set consists of single characters, which lacks the contextual relationship between characters. This leads to a lower quality of generated pseudo-labeled samples. To this end, we contribute a novel approach of constructing confusion corpus, which can automatically generate high-quality spelling errors. The construction of our confusion set is based on user behavior patterns, aiming to identify characters that are easily confused in similar contexts. Then we introduce a Span Confusion Pre-Train(SCPre) strategy for Chinese Spelling Correction. The span confusion strategy replaces characters, words, and phrases in the text according to a large confusion set. Upon the constructed corpus, different models are trained and evaluated for CSC with respect to real-world datasets. Quantitative experiments on datasets show that our approach empirically achieves unprecedented performance.
- 21 October 2024
- Published by Association for Computing Machinery (ACM)
- p. 4208-4212
Open AccessSpan Confusion is All You Need for Chinese Spelling Correction
Chinese Spelling Correction is to detect errors in a statement and correct them. Researchers have recently begun to investigate the use of pre-train language models to perform this task. They replace tokens with similar characters based on a confusion set rather than mask tokens when pre-training. A significant limitation of this method is that pre-training with single tokens cannot effectively simulate natural errors, such as full spelling. Furthermore, the confusion set consists of single characters, which lacks the contextual relationship between characters. This leads to a lower quality of generated pseudo-labeled samples. To this end, we contribute a novel approach of constructing confusion corpus, which can automatically generate high-quality spelling errors. The construction of our confusion set is based on user behavior patterns, aiming to identify characters that are easily confused in similar contexts. Then we introduce a Span Confusion Pre-Train(SCPre) strategy for Chinese Spelling Correction. The span confusion strategy replaces characters, words, and phrases in the text according to a large confusion set. Upon the constructed corpus, different models are trained and evaluated for CSC with respect to real-world datasets. Quantitative experiments on datasets show that our approach empirically achieves unprecedented performance.
- Open Access
- 24 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 6127-6136
Enhancing Asymmetric Web Search through Question-Answer Generation and Ranking
This paper addresses the challenge of the semantic gap between user queries and web content, commonly referred to as asymmetric text matching, within the domain of web search. By leveraging BERT for reading comprehension, current algorithms enable significant advancements in query understanding, but still encounter limitations in effectively resolving the asymmetrical ranking problem due to model comprehension and summarization constraints. To tackle this issue, we propose the QAGR (Question-Answer Generation and Ranking) method, comprising an offline module called QAGeneration and an online module called QARanking. The QAGeneration module utilizes large language models (LLMs) to generate high-quality question-answering pairs for each web page. This process involves two steps: generating question-answer pairs and performing verification to eliminate irrelevant questions, resulting in high-quality questions associated with their respective documents. The QARanking module combines and ranks the generated questions and web page content. To ensure efficient online inference, we design the QARanking model as a homogeneous dual-tower model, incorporating query intent to drive score fusion while balancing keyword matching and asymmetric matching. Additionally, we conduct a preliminary screening of questions for each document, selecting only the top-N relevant questions for further relevance calculation. Empirical results demonstrate the substantial performance improvement of our proposed method in web search. We achieve over 8.7% relative offline relevance improvement and over 8.5% online engagement gain compared to the state-of-the-art web search system. Furthermore, we deploy QAGR to online web search engines and share our deployment experience, including production considerations and ablation experiments. This research contributes to advancing the field of asymmetric web search and provides valuable insights for enhancing search engine performance.
- 24 August 2024
- Published by Association for Computing Machinery (ACM)
- p. 6127-6136
Open AccessEnhancing Asymmetric Web Search through Question-Answer Generation and Ranking
This paper addresses the challenge of the semantic gap between user queries and web content, commonly referred to as asymmetric text matching, within the domain of web search. By leveraging BERT for reading comprehension, current algorithms enable significant advancements in query understanding, but still encounter limitations in effectively resolving the asymmetrical ranking problem due to model comprehension and summarization constraints. To tackle this issue, we propose the QAGR (Question-Answer Generation and Ranking) method, comprising an offline module called QAGeneration and an online module called QARanking. The QAGeneration module utilizes large language models (LLMs) to generate high-quality question-answering pairs for each web page. This process involves two steps: generating question-answer pairs and performing verification to eliminate irrelevant questions, resulting in high-quality questions associated with their respective documents. The QARanking module combines and ranks the generated questions and web page content. To ensure efficient online inference, we design the QARanking model as a homogeneous dual-tower model, incorporating query intent to drive score fusion while balancing keyword matching and asymmetric matching. Additionally, we conduct a preliminary screening of questions for each document, selecting only the top-N relevant questions for further relevance calculation. Empirical results demonstrate the substantial performance improvement of our proposed method in web search. We achieve over 8.7% relative offline relevance improvement and over 8.5% online engagement gain compared to the state-of-the-art web search system. Furthermore, we deploy QAGR to online web search engines and share our deployment experience, including production considerations and ablation experiments. This research contributes to advancing the field of asymmetric web search and provides valuable insights for enhancing search engine performance.
- Open Access
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2775-2784
VTLayout: A Multi-Modal Approach for Video Text Layout
The rapid explosion of video distribution is accompanied by a massive amount of video text, which encompasses rich information about the video content. While previous research has primarily focused on text extraction from videos like text detection, tracking, recognition and end to end spotting, the layout of video text has received limited attention. As different text categories convey distinct meanings, video text layout is critical for video understanding tasks such as video summarization and shooting environment comprehension. To bridge the gap between video OCR and understanding, we explore the study of video text layout in this work. We first optimize the layout annotation of the BOVText, a bilingual, open-world video text dataset, by expanding text categories and defining five clear categories: scene, subtitle, title, logo, and other. Additionally, we rectify the original unreasonable layout annotation based on these definitions. We also propose a Video-level Text Layout model (VTLayout) to address the layout problem, which fuses textual, visual, and spatial-temporal embedding of video text trajectories. To the best of our knowledge, this is the first method to tackle text layout on video level. Our method outperforms image-level layout methods across all text categories and exhibits faster inference speed. This study underscores the significance of video text layout in video understanding and offers an effective solution to this challenge. Our annotation is available at https://github.com/TencentARC/VTLayout.
- 26 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2775-2784
Open AccessVTLayout: A Multi-Modal Approach for Video Text Layout
The rapid explosion of video distribution is accompanied by a massive amount of video text, which encompasses rich information about the video content. While previous research has primarily focused on text extraction from videos like text detection, tracking, recognition and end to end spotting, the layout of video text has received limited attention. As different text categories convey distinct meanings, video text layout is critical for video understanding tasks such as video summarization and shooting environment comprehension. To bridge the gap between video OCR and understanding, we explore the study of video text layout in this work. We first optimize the layout annotation of the BOVText, a bilingual, open-world video text dataset, by expanding text categories and defining five clear categories: scene, subtitle, title, logo, and other. Additionally, we rectify the original unreasonable layout annotation based on these definitions. We also propose a Video-level Text Layout model (VTLayout) to address the layout problem, which fuses textual, visual, and spatial-temporal embedding of video text trajectories. To the best of our knowledge, this is the first method to tackle text layout on video level. Our method outperforms image-level layout methods across all text categories and exhibits faster inference speed. This study underscores the significance of video text layout in video understanding and offers an effective solution to this challenge. Our annotation is available at https://github.com/TencentARC/VTLayout.
- Open Access
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2999-3008
Improving Query Correction Using Pre-train Language Model In Search Engines
Query correction is a task that automatically detects and corrects errors in what users type into a search engine. Misspelled queries can lead to user dissatisfaction and churn. However, correcting a user query accurately is not an easy task. One major challenge is that a correction model must be capable of high-level language comprehension. Recently, pre-trained language models (PLMs) have been successfully applied to text correction tasks, but few works have been done on query correction. However, it is nontrivial to directly apply these PLMs to query correction in large-scale search systems due to the following challenging issues: 1) Expensive deployment. Deploying such a model requires expensive computations. 2) Lacking domain knowledge. A neural correction model needs massive training data to activate its power. To this end, we introduce KSTEM, a Knowledge-based Sequence To Edit Model for Chinese query correction. KSTEM transforms the sequence generation task into sequence tagging by mapping errors into five categories: KEEP, REPLACE, SWAP, DELETE, and INSERT, reducing computational complexity. Additionally, KSTEM adopts 2D position encoding, which is composed of the internal and external order of the words. Meanwhile, to compensate for the lack of domain knowledge, we propose a task-specific training paradigm for query correction, including edit strategy-based pre-training, user click-based post pre-train, and human label-based fine-tuning. Finally, we apply KSTEM to the industrial search system. Extensive offline and online experiments show that KSTEM significantly improves query correction performance. We hope that our experience will benefit frontier researchers.
- 21 October 2023
- Published by Association for Computing Machinery (ACM)
- p. 2999-3008
Open AccessImproving Query Correction Using Pre-train Language Model In Search Engines
Query correction is a task that automatically detects and corrects errors in what users type into a search engine. Misspelled queries can lead to user dissatisfaction and churn. However, correcting a user query accurately is not an easy task. One major challenge is that a correction model must be capable of high-level language comprehension. Recently, pre-trained language models (PLMs) have been successfully applied to text correction tasks, but few works have been done on query correction. However, it is nontrivial to directly apply these PLMs to query correction in large-scale search systems due to the following challenging issues: 1) Expensive deployment. Deploying such a model requires expensive computations. 2) Lacking domain knowledge. A neural correction model needs massive training data to activate its power. To this end, we introduce KSTEM, a Knowledge-based Sequence To Edit Model for Chinese query correction. KSTEM transforms the sequence generation task into sequence tagging by mapping errors into five categories: KEEP, REPLACE, SWAP, DELETE, and INSERT, reducing computational complexity. Additionally, KSTEM adopts 2D position encoding, which is composed of the internal and external order of the words. Meanwhile, to compensate for the lack of domain knowledge, we propose a task-specific training paradigm for query correction, including edit strategy-based pre-training, user click-based post pre-train, and human label-based fine-tuning. Finally, we apply KSTEM to the industrial search system. Extensive offline and online experiments show that KSTEM significantly improves query correction performance. We hope that our experience will benefit frontier researchers.
- Open Access
- 18 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 2681-2690
T2Ranking: A Large-scale Chinese Benchmark for Passage Ranking
Passage ranking involves two stages: passage retrieval and passage re-ranking, which are important and challenging topics for both academics and industries in the area of Information Retrieval (IR). However, the commonly-used datasets for passage ranking usually focus on the English language. For non-English scenarios, such as Chinese, the existing datasets are limited in terms of data scale, fine-grained relevance annotation and false negative issues. To address this problem, we introduce T2Ranking, a large-scale Chinese benchmark for passage ranking. T2Ranking comprises more than 300K queries and over 2M unique passages from real-world search engines. Expert annotators are recruited to provide 4-level graded relevance scores (fine-grained) for query-passage pairs instead of binary relevance judgments (coarse-grained). To ease the false negative issues, more passages with higher diversities are considered when performing relevance annotations, especially in the test set, to ensure a more accurate evaluation. Apart from the textual query and passage data, other auxiliary resources are also provided, such as query types and XML files of documents which passages are generated from, to facilitate further studies. To evaluate the dataset, commonly used ranking models are implemented and tested on T2Ranking as baselines. The experimental results show that T2Ranking is challenging and there is still scope for improvement. The full data and all codes are available at https://github.com/THUIR/T2Ranking/.
- 18 July 2023
- Published by Association for Computing Machinery (ACM)
- p. 2681-2690
Open AccessT2Ranking: A Large-scale Chinese Benchmark for Passage Ranking
Passage ranking involves two stages: passage retrieval and passage re-ranking, which are important and challenging topics for both academics and industries in the area of Information Retrieval (IR). However, the commonly-used datasets for passage ranking usually focus on the English language. For non-English scenarios, such as Chinese, the existing datasets are limited in terms of data scale, fine-grained relevance annotation and false negative issues. To address this problem, we introduce T2Ranking, a large-scale Chinese benchmark for passage ranking. T2Ranking comprises more than 300K queries and over 2M unique passages from real-world search engines. Expert annotators are recruited to provide 4-level graded relevance scores (fine-grained) for query-passage pairs instead of binary relevance judgments (coarse-grained). To ease the false negative issues, more passages with higher diversities are considered when performing relevance annotations, especially in the test set, to ensure a more accurate evaluation. Apart from the textual query and passage data, other auxiliary resources are also provided, such as query types and XML files of documents which passages are generated from, to facilitate further studies. To evaluate the dataset, commonly used ranking models are implemented and tested on T2Ranking as baselines. The experimental results show that T2Ranking is challenging and there is still scope for improvement. The full data and all codes are available at https://github.com/THUIR/T2Ranking/.