CAREER: Structured Output Models of Recommendations, Activities, and Behavior

职业：建议、活动和行为的结构化输出模型

• 批准号: 1750063
• 负责人: Julian McAuley
• 金额: $ 55万
• 依托单位: University of California-San Diego
• 依托单位国家: 美国
• 项目类别: Continuing Grant
• 财政年份: 2018
• 资助国家: 美国
• 起止时间: 2018-08-01 至 2023-07-31
• 项目状态: 已结题
• 来源: https://www.nsf.gov/awardsearch/showAward?AWD_ID=1750063&HistoricalAwards=false
• 关键词: CAREER; Structured; Output; Models; Recommendations

This project will investigate predictive models of human behavior that are capable of estimating rich, structured outputs. Such predictive models underlie many of the most important computing systems in science and industry, ranging from e-Commerce to personalized healthcare. While existing models (i.e., "recommender systems") typcially focus on simple predictions (a user's next click or star rating, a patient's next symptom, etc.), this project shall develop models capable of generating outputs in the form of text, images, and sequences. These new modalities of predictive modeling will allow personalized recommender systems to be adapted to answer complex questions, predict nuanced reactions, and even to design new content that elicits a certain reaction.The project's technical approach combines ideas from personalized recommender systems with newly-emerging techniques for generative modeling. Ideas from recommender systems can be used to handle issues like personalization, subjectivity, or other variance that arises due to differences between individuals; ideas from generative modeling allow complex outputs (text, images, sequences) to be generated. This technical contribution can be viewed either as a new form of recommender system capable of handling more complex queries, or alternately as a new suite of generative modeling approaches that can account for variance between individuals. This project will have impact to applications where complex, high-dimensional data meets issues of personalization and subjectivity. Specific examples to be investigated include online activity traces, e-Commerce, and personalized health.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.

该项目将研究人类行为的预测模型，这些模型能够估计丰富的结构化输出。这种预测模型是科学和工业中许多最重要的计算系统的基础，从电子商务到个性化医疗保健。虽然现有的模型（即，“推荐系统”）通常集中于简单的预测（用户的下一次点击或星星评级、患者的下一个症状等），本项目将开发能够以文本、图像和序列形式生成输出的模型。这些预测建模的新模式将使个性化推荐系统能够适应回答复杂的问题，预测细微的反应，甚至设计新的内容，消除某种反应。该项目的技术方法将个性化推荐系统的思想与新兴的生成建模技术相结合。来自推荐系统的想法可以用来处理个性化、主观性或其他由于个体之间的差异而产生的差异等问题;来自生成建模的想法可以生成复杂的输出（文本、图像、序列）。这种技术贡献可以被看作是一种新形式的推荐系统，能够处理更复杂的查询，或者作为一套新的生成建模方法，可以考虑个人之间的差异。该项目将对复杂的高维数据满足个性化和主观性问题的应用程序产生影响。该奖项反映了NSF的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

项目成果

BERT Learns to Teach: Knowledge Distillation with Meta Learning

• DOI: 10.18653/v1/2022.acl-long.485
• 发表时间: 2021-06
• 作者: Wangchunshu Zhou;Canwen Xu;Julian McAuley

Knowledge-Grounded Self-Rationalization via Extractive and Natural Language Explanations

• 发表时间: 2021-06
• 作者: Bodhisattwa Prasad Majumder;Oana-Maria Camburu;Thomas Lukasiewicz;Julian McAuley

InforMask: Unsupervised Informative Masking for Language Model Pretraining

• DOI: 10.48550/arxiv.2210.11771
• 发表时间: 2022-10
• 期刊: Comput. Syst. Sci. Eng.
• 作者: Nafis Sadeq;Canwen Xu;Julian McAuley

Controlling Bias Exposure for Fair Interpretable Predictions

控制偏差暴露以实现公平可解释的预测

• 发表时间: 2022
• 期刊: Findings of EMNLP
• 作者: Zexue He, Yu Wang

LaPraDoR: Unsupervised Pretrained Dense Retriever for Zero-Shot Text Retrieval

• DOI: 10.48550/arxiv.2203.06169
• 发表时间: 2022-03
• 作者: Canwen Xu;Daya Guo;Nan Duan;Julian McAuley