EMLP 2023 Announces Best Papers: Peking University Collaborates with WeChat AI Team to Secure China's First Best Long Paper Award

The prestigious EMNLP 2023 conference—a leading event in the natural language processing (NLP) field—will be held from December 6 to December 10 in Singapore. This year, the conference saw an impressive 4,909 submissions, with 1,047 papers accepted, resulting in a 21.3% acceptance rate. Organized annually by the Association for Computational Linguistics (ACL), EMNLP is one of the most influential international gatherings in NLP.

The WeChat AI team notably had eight papers selected for presentation, including a collaboration with Peking University that won the Best Long Paper Award. This accolade marks a historic achievement as the first paper from China to receive this honor at EMNLP. Titled “Label Words are Anchors: An Information Flow Perspective for Understanding In-Context Learning,” the award-winning paper explores in-context learning (ICL) mechanisms, revealing that label words in demonstration examples act as "anchors" in the learning process:

1. In shallow computing layers, semantic information is concentrated within the representations of label words.

2. In deeper layers, aggregated information from these label words informs the language model's predictions.

The study proposes a novel “anchor re-weighting” method to improve ICL performance and introduces an analytical framework for diagnosing ICL errors, paving the way for future research.

The authors, part of Professor Sun Xu’s team at Peking University and the WeChat AI team, are led by first author Wang Le'an, a graduate student who excelled in his undergraduate program in intelligent science and technology. This publication marks his first major conference contribution in the NLP field.

This research benefited from extensive collaboration, addressing key research directions, experimental enhancements, and manuscript revisions, showcasing Wang Le'an’s robust mathematical skills and Professor Sun Xu’s insightful guidance.

The WeChat AI team's performance at EMNLP 2023 underscores their expertise in the AI domain. Composed mainly of master’s and Ph.D. students from top Chinese universities, the team maintains a high standard for talent acquisition. In addition to the Best Long Paper Award at EMNLP 2023, they have garnered several awards, including the Best Short Paper Award at RecSys 2023 and the Best Long Paper Award at ACL 2019.

Beyond EMNLP, the WeChat AI team has excelled in various research areas, clinching multiple championships at the Workshop on Machine Translation (WMT) from 2020 to 2022 while competing against leading teams globally. They have also achieved notable success in the Dialogue System Technology Challenge (DSTC) in 2018 and 2019.

Dedicated to merging business with academic research, the WeChat AI team identifies research topics through real-world applications. They advocate for a synergistic relationship between business and research, stating, "By conducting research to find better methods, the results can ultimately be directly applied in business." Their work at this conference has advanced applications in dialogue systems, WeChat Translation, WeChat Reading, and WeChat Input Method, among others.