|
Wang Minghui exemplifies the modern AI researcher who successfully transitions from academic excellence to industry leadership. Currently serving as an Advertising Software Engineer at Beijing ByteDance Technology Co., Ltd., Wang has established himself as a key contributor to both cutting-edge NLP research and practical AI applications. With a research portfolio that has garnered 84 citations and achieved an h-index of 3, his work in dialogue generation and knowledge-enhanced AI systems demonstrates remarkable ability to advance theoretical understanding while delivering measurable commercial impact—including a 10% cumulative increase in TikTok's advertiser value.
Interviewer: Mr. Wang, your career journey from Peking University's NLP research to ByteDance's advertising technology represents a fascinating transition. How has your academic background informed your current industry work?
Wang Minghui: My journey from academic research at Peking University to commercial AI development at ByteDance has been incredibly enriching. The theoretical foundations I developed during my Master's in Software Engineering, particularly in dialogue generation and knowledge-enhanced systems, directly translate to advertising technology challenges.
At ByteDance, I optimize advertising algorithms to estimate CTR and CVR more accurately by introducing new signals and algorithm improvements. The same principles I applied in academic research—understanding user intent, processing multi-source information, and generating contextually relevant responses—are fundamental to creating effective advertising systems.
My research experience with multi-source heterogeneous knowledge integration has been particularly valuable. In advertising, we process vast amounts of user behavior data, content signals, and contextual information simultaneously, mirroring the knowledge fusion challenges I tackled during my academic research.
Interviewer: Your collaborative research on dialogue generation has received significant attention, particularly your work on multi-source heterogeneous knowledge systems. Could you elaborate on your contributions to this breakthrough research?
Wang Minghui: My involvement in multi-source heterogeneous knowledge research represents some of the most impactful work in modern dialogue generation. As a key contributor, I helped develop frameworks that revolutionize how AI systems integrate diverse knowledge sources—knowledge graphs, knowledge text, and knowledge tables—to improve dialogue quality.
Our work "More is Better: Enhancing Open-Domain Dialogue Generation via Multi-Source Heterogeneous Knowledge," which received 28 citations at EMNLP 2021, addressed a fundamental limitation in existing dialogue systems: their reliance on single knowledge sources. I contributed to developing the MSKE-Dialog model, which uses innovative Relevance Gates and Selection Gates to resolve topic conflicts between different knowledge sources.
The follow-up research, "Improving the Applicability of Knowledge-Enhanced Dialogue Generation Systems," published at WSDM 2022 with 26 citations, built upon these foundations. My contributions focused on developing robust encoder-decoder architectures that handle the complexity of integrating heterogeneous knowledge sources while maintaining computational efficiency.
These research experiences directly inform my current work at ByteDance, where I'm applying large language models to help advertising systems understand advertisements more accurately and predict user engagement patterns.
Interviewer: Your work on hierarchical Infobox accessing for dialogue generation, published at IJCAI 2021, has received considerable academic recognition. How does this research advance the field?
Wang Minghui: The hierarchical Infobox research, with 21 citations, represents a significant advancement in how dialogue systems leverage structured knowledge. As a contributing researcher, I helped develop novel approaches to encoding Infobox information at multiple granularities—Intra-attribute, Inter-attribute, and Dialogue context levels.
The key innovation lies in our hierarchical encoding strategy and the Infobox-Dialogue interaction graph neural network. This enables dialogue systems to access and utilize structured information more effectively than traditional approaches. We also developed a hierarchical attention mechanism that allows the system to focus on relevant Infobox attributes at different granularities depending on dialogue context.
This research has practical applications beyond academic dialogue systems. In my current role at ByteDance, similar hierarchical information processing techniques are crucial for understanding advertisement content and user preferences across multiple dimensions simultaneously.
Interviewer: Your first-author research on document relation extraction represents your independent research contributions. Could you discuss this work and its significance?
Wang Minghui: My first-author paper, "Distilling the Documents for Relation Extraction by Topic Segmentation," published at ICDAR 2021, represents my independent contribution to document understanding. This work addresses a critical challenge in document-level relation extraction—how to identify and focus on relevant content while discarding unnecessary information.
The breakthrough lies in our novel document distillation framework that integrates topic segmentation algorithms to identify and discard irrelevant content for relation extraction tasks. We also developed a topic enhancement module implemented through bidirectional attention mechanisms to improve relation prediction accuracy.
While this research received fewer citations initially, it demonstrates important principles about information filtering and attention mechanisms that are highly relevant to commercial applications. In advertising systems, we face similar challenges of extracting relevant signals from noisy, complex data.
Interviewer: How do you apply your research insights to real-world advertising challenges at ByteDance?
Wang Minghui: At ByteDance, I focus on two major areas: algorithmic optimization of advertising systems and integration of large language models for ad understanding and prediction.
I optimize advertising algorithms from multiple perspectives—features, data flows, model structures, and loss functions. This requires the same systematic approach I developed during my academic research on multi-source knowledge integration. We analyze user signals, content features, and contextual information to improve CTR and CVR predictions.
The second area involves using LLMs to help advertising systems understand advertisements more accurately and fine-tune these models to predict user engagement. This directly builds on my research experience with dialogue generation and knowledge-enhanced systems.
Through these optimization efforts, we've achieved a 10% cumulative increase in TikTok's advertiser value, demonstrating how academic research principles can drive meaningful business outcomes.
Interviewer: Your competitive programming experience is remarkable, including a Kaggle silver medal and multiple national first prizes. How do these experiences complement your research and industry work?
Wang Minghui: My competitive experience has been instrumental in developing practical problem-solving skills. The Kaggle silver medal I earned in the ASHRAE Great Energy Predictor III competition, where I ranked in the top 2%, required advanced feature engineering and machine learning techniques that I regularly apply in my current work.
The competition involved predicting energy usage for over 1,000 buildings using complex multi-modal data. Our solution using LightGBM with sophisticated feature engineering demonstrates the same data integration principles I apply in advertising optimization.
My national first prizes in IoT design and services computing competitions developed my expertise in real-world system integration. The multi-intelligent vehicle SLAM mapping project required coordination between multiple data sources and real-time processing—skills directly applicable to large-scale advertising systems.
Interviewer: How did your educational journey from Harbin Institute of Technology to Peking University prepare you for your current role?
Wang Minghui: My educational journey provided a strong foundation in both theoretical computer science and practical software engineering. At Harbin Institute of Technology, I developed fundamental skills in software engineering, earning consistent merit student recognition and first-prize scholarships.
At Peking University's School of Software & Microelectronics, I specialized in advanced software engineering while focusing on NLP research. The program's emphasis on both academic rigor and practical application perfectly prepared me for bridging research and industry applications.
The recognition I received—including Excellent Graduate awards from both institutions and the Exceptional Award for Academic Innovation at Peking University—reflects my ability to excel in both theoretical research and practical implementation.
Interviewer: What are your current priorities and future research directions?
Wang Minghui: My current focus is advancing the integration of large language models with commercial advertising systems. We're exploring how to leverage the latest developments in foundation models to create more sophisticated understanding of user intent and advertisement relevance.
I'm particularly interested in developing more efficient and interpretable AI systems that can operate at the scale required by platforms like TikTok while maintaining the nuanced understanding capabilities demonstrated in my academic research.
From a research perspective, I'm exploring how principles of knowledge-enhanced dialogue generation can be adapted for commercial recommendation systems. The challenge is maintaining theoretical rigor while meeting the performance, scalability, and interpretability requirements of production systems.
Interviewer: Thank you, Mr. Wang, for sharing your insights. Your work clearly demonstrates how academic excellence can drive meaningful industry innovation.
Wang Minghui: Thank you for this opportunity. I believe we're at an exciting moment where academic NLP research and commercial AI applications are converging to create unprecedented opportunities for impact. My experience shows that rigorous academic training and practical industry experience can be mutually reinforcing, leading to both better research and more effective commercial applications.
|
Home > News >