A5089522357- Topic Modeling
- Speech and dialogue systems
- AI in Service Interactions
- Misinformation and Its Impacts
- Natural Language Processing Techniques
- Sentiment Analysis and Opinion Mining
- Multi-Agent Systems and Negotiation
- Privacy-Preserving Technologies in Data
- Media Influence and Health
- Mental Health via Writing
- Digital Mental Health Interventions
- Ethics and Social Impacts of AI
- Speech Recognition and Synthesis
- Educational Technology and Assessment
- Anomaly Detection Techniques and Applications
- Expert finding and Q&A systems
- Multimodal Machine Learning Applications
- Social Robot Interaction and HRI
- Recommender Systems and Techniques
- Cell Image Analysis Techniques
- Hate Speech and Cyberbullying Detection
- Access Control and Trust
- Text and Document Classification Technologies
- Yersinia bacterium, plague, ectoparasites research
- Engineering Education and Pedagogy
Northeastern University
2025
Stanford University
2024
Columbia University
2021-2024
Singapore University of Technology and Design
2024
Carnegie Mellon University
2023
Google (United States)
2022
Virginia Tech
2022
University of California, Davis
2018-2021
University of California, Los Angeles
2000-2020
University of California, Berkeley
2018-2020
Developing intelligent persuasive conversational agents to change people’s opinions and actions for social good is the frontier in advancing ethical development of automated dialogue systems. To do so, first step understand intricate organization strategic disclosures appeals employed human persuasion conversations. We designed an online task where one participant was asked persuade other donate a specific charity. collected large dataset with 1,017 dialogues annotated emerging strategies...
Despite much progress in training artificial intelligence (AI) systems to imitate human language, building agents that use language communicate intentionally with humans interactive environments remains a major challenge. We introduce Cicero, the first AI agent achieve human-level performance Diplomacy, strategy game involving both cooperation and competition emphasizes natural negotiation tactical coordination between seven players. Cicero integrates model planning reinforcement learning...
In recommendation dialogs, humans commonly disclose their preference and make recommendations in a friendly manner. However, this is challenge when developing sociable dialog system, due to the lack of dataset annotated with such strategies. Therefore, we present INSPIRED, new 1,001 human-human dialogs for movie measures successful recommendations. To better understand how communication, design an annotation scheme related strategies based on social science theories annotate these dialogs....
End-to-end learning framework is useful for building dialog systems its simplicity in training and efficiency model updating. However, current end-to-end approaches only consider user semantic inputs under-utilize other information. Therefore, we propose to include sentiment obtained through multimodal information (acoustic, dialogic textual), the make more user-adaptive effective. We incorporated both supervised reinforcement settings. In settings, adding reduced length improved task...
With the increasing applications of language models, it has become crucial to protect these models from leaking private information. Previous work attempted tackle this challenge by training RNN-based with differential privacy guarantees.However, applying classical leads poor model performance as underlying notion is over-pessimistic and provides undifferentiated protection for all tokens in data. Given that information natural sparse (for example, bulk an email might not carry personally...
Using chatbots to make recommendations is increasingly popular. The design of recommendation has mainly been taking an information-centric approach by focusing on the recommended content per se. Limited attention how social connection and relational strategies, such as self-disclosure from a chatbot, may influence users' perception acceptance recommendation. In this work, we designed, implemented, evaluated chatbot capable performing three different levels self-disclosure: factual...
Weiyan Shi, Kun Qian, Xuewei Wang, Zhou Yu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.
Weiyan Shi, Tiancheng Zhao, Zhou Yu. Proceedings of the 2019 Conference North American Chapter Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers). 2019.
Inducing a meaningful structural representation from one or set of dialogues is crucial but challenging task in computational linguistics. Advancement made this area critical for dialogue system design and discourse analysis. It can also be extended to solve grammatical inference. In work, we propose incorporate structured attention layers into Variational Recurrent Neural Network (VRNN) model with discrete latent states learn structure an unsupervised fashion. Compared vanilla VRNN, enables...
Intelligent conversational agents, or chatbots, can take on various identities and are increasingly engaging in more human-centered conversations with persuasive goals. However, little is known about how inquiry strategies influence the conversation's effectiveness. We conducted an online study involving 790 participants to be persuaded by a chatbot for charity donation. designed two four factorial experiment (two strategies) where were randomly assigned different conditions. Findings showed...
Protecting large language models from privacy leakage is becoming increasingly crucial with their wide adoption in real-world products. Yet applying *differential privacy* (DP), a canonical notion provable guarantees for machine learning models, to those remains challenging due the trade-off between model utility and loss. Utilizing fact that sensitive information data tends be sparse, Shi et al. (2021) formalized DP extension called *Selective Differential Privacy* (SDP) protect only tokens...
Persuasion is important in numerous situations like healthy habit promotion, and emotional support. As AI gets more involved our daily life, it becomes critical to study how they can persuade humans persuasive are. In this talk, I will cover (1) build such systems that persuade, negotiate, cooperate with other the game of Diplomacy. (2) also discuss perceive specialized systems. This validates necessity California's Autobot Law proposes guidance regulate (3) these become powerful, safety...
Independent evaluation and red teaming are critical for identifying the risks posed by generative AI systems. However, terms of service enforcement strategies used prominent companies to deter model misuse have disincentives on good faith safety evaluations. This causes some researchers fear that conducting such research or releasing their findings will result in account suspensions legal reprisal. Although offer researcher access programs, they an inadequate substitute independent access,...
End-to-end task-oriented dialog models have achieved promising performance on collaborative tasks where users willingly coordinate with the system to complete a given task. While in non-collaborative settings, for example, negotiation and persuasion, systems do not share common goal. As result, compared collaborate tasks, people use social content build rapport trust these settings order advance their goals. To handle content, we introduce hierarchical intent annotation scheme, which can be...
Most traditional AI safety research has approached models as machines and centered on algorithm-focused attacks developed by security experts. As large language (LLMs) become increasingly common competent, non-expert users can also impose risks during daily interactions. This paper introduces a new perspective to jailbreak LLMs human-like communicators, explore this overlooked intersection between everyday interaction safety. Specifically, we study how persuade them. First, propose...
Persuasion dialogue system reflects the machine's ability to make strategic moves beyond verbal communication, and therefore differentiates itself from task-oriented or open-domain dialogues has its own unique values. However, repetition inconsistency problems still persist in response generation could substantially impact user experience impede persuasion outcome. Besides, although reinforcement learning (RL) approaches have achieved big success tasks such as games, it requires a...
Developing intelligent persuasive conversational agents to change people's opinions and actions for social good is the frontier in advancing ethical development of automated dialogue systems. To do so, first step understand intricate organization strategic disclosures appeals employed human persuasion conversations. We designed an online task where one participant was asked persuade other donate a specific charity. collected large dataset with 1,017 dialogues annotated emerging strategies...
End-to-end task-oriented dialog models have achieved promising performance on collaborative tasks where users willingly coordinate with the system to complete a given task. While in non-collaborative settings, for example, negotiation and persuasion, systems do not share common goal. As result, compared collaborate tasks, people use social content build rapport trust these settings order advance their goals. To handle content, we introduce hierarchical intent annotation scheme, which can be...
Mixed-initiative dialogue tasks involve repeated exchanges of information and conversational control. Conversational agents gain control by generating responses that follow particular intents or strategies, prescribed a policy planner. The standard approach has been fine-tuning pre-trained language models to perform generation conditioned on these intents. However, supervised are limited the cost quality data annotation.We instead prompt large as drop-in replacement conditional generation....
Learning a shared dialog structure from set of task-oriented dialogs is an important challenge in computational linguistics. The learned can shed light on how to analyze human dialogs, and more importantly contribute the design evaluation systems. We propose extract structures using modified VRNN model with discrete latent vectors. Different existing HMM-based models, our based variational-autoencoder (VAE). Such able capture dynamics beyond surface forms language. find that qualitatively,...
As language models (LMs) are widely utilized in personalized communication scenarios (e.g., sending emails, writing social media posts) and endowed with a certain level of agency, ensuring they act accordance the contextual privacy norms becomes increasingly critical. However, quantifying norm awareness LMs emerging risk LM-mediated is challenging due to (1) long-tailed nature privacy-sensitive cases, (2) lack evaluation approaches that capture realistic application scenarios. To address...