Recap of Baidu ACL 2020 Paper Sharing Session – Papers 4 to 6

On the "Listen to Experts Talk Papers" Baidu ACL 2020 sharing event, Baidu's Technical Committee Chair Dr. Wu Hua and six NLP engineers hosted a live B‑session where three selected ACL 2020 papers were presented and discussed in depth.

Paper 4: "Exploring Contextual Word‑level Style Relevance for Unsupervised Style Transfer" proposes a new sequence‑to‑sequence generation model that dynamically performs style transfer based on the style relevance of each generated word. The work introduces two main innovations: a hierarchical relevance propagation algorithm that computes word‑level style relevance in the style classifier and uses it as a guiding signal for the generator, and a style‑transfer decoder that takes word‑level relevance and semantic information as inputs, fine‑tuned with multiple loss terms.

Paper 5: "SKEP: Sentiment Knowledge Enhanced Pre‑training for Sentiment Analysis" presents a sentiment‑knowledge‑enhanced language model pre‑training method. Building on general pre‑training, the authors design a sentiment‑oriented masking strategy and a multi‑task learning algorithm that incorporate sentiment words, polarity, and comment‑entity relations, achieving a unified pre‑trained representation for various sentiment analysis tasks.

Paper 6: "Towards Conversational Recommendation over Multi‑Type Dialogs" defines a new task of conversational recommendation within multi‑type dialogs. The goal is for a bot to proactively steer a conversation from non‑recommendation modes (e.g., QA) to recommendation mode, then use collected user interests and real‑time feedback across multiple turns to fulfill the recommendation objective.

The recap concludes by inviting the audience to follow future NLP sharing live streams, highlighting Baidu NLP's mission to "understand language, empower intelligence, change the world," and describing its ongoing development of core NLP technologies and platforms.

About the organizer: DataFunTalk focuses on big data and artificial intelligence technology applications, having hosted over 100 offline salons, forums, and conferences across major Chinese cities since 2017, and producing more than 300 original articles with millions of reads.