Papers by Lucas Dixon
Detecting Hallucination and Coverage Errors in Retrieval Augmented Generation for Controversial Topics (2024.lrec-main)
Copied to clipboard
Tyler A. Chang, Katrin Tomanek, Jessica Hoffmann, Nithum Thain, Erin MacMurray van Liemt, Kathleen Meier-Hellstern, Lucas Dixon
| Challenge: | a growing audience of users is engaging with LLM-driven chatbots. |
| Approach: | They propose a strategy to handle controversial topics in LLM-based chatbots based on Wikipedia’s Neutral Point of View principle. |
| Outcome: | The proposed methods detect errors in the tuned LLM responses even when no training data is available. |
Conversations Gone Awry: Detecting Early Signs of Conversational Failure (P18-1)
Copied to clipboard
Justine Zhang, Jonathan Chang, Cristian Danescu-Niculescu-Mizil, Lucas Dixon, Yiqing Hua, Dario Taraborelli, Nithum Thain
| Challenge: | Prior work focused on characterizing and detecting content exhibiting antisocial online behavior. |
| Approach: | They propose a task of predicting from the very start of a conversation whether it will get out of hand. |
| Outcome: | The proposed framework can detect early warning signs of antisocial behavior in online conversations. |
Improving Neutral Point-of-View Generation with Data- and Parameter-Efficient RL (2025.emnlp-main)
Copied to clipboard
Jessica Hoffmann, Christiane Ahlheim, Zac Yu, Aria Walfrand, Jarvis Jin, Marie Tano, Ahmad Beirami, Erin MacMurray van Liemt, Nithum Thain, Hakim Sidahmed, Lucas Dixon
| Challenge: | Parameter-efficient reinforcement learning (PE-RL) is a highly effective training regime to improve large language models’ ability to answer queries on sensitive topics with a Neutral Point of View (NPOV). |
| Approach: | They propose to use parameter-efficient reinforcement learning to train large language models to answer queries with a Neutral Point of View (NPOV) This is compared to the strongest baseline, LoRA finetuning, SFT and RLHF. |
| Outcome: | The proposed training regime improves on NPOV quality and scores higher on features identified by linguists as key to separating good answers from the best answers. |
Civil Rephrases Of Toxic Texts With Self-Supervised Transformers (2021.eacl-main)
Copied to clipboard
| Challenge: | et al., 2018a): a poor phrasing may make the conversation go awry. |
| Approach: | They propose a model that can help suggest rephrasings of toxic comments in a more civil manner. |
| Outcome: | The proposed model generates sentences that are more fluent and better at preserving the initial content compared to earlier systems and human evaluation. |
WikiConv: A Corpus of the Complete Conversational History of a Large Online Collaborative Community (D18-1)
Copied to clipboard
Yiqing Hua, Cristian Danescu-Niculescu-Mizil, Dario Taraborelli, Nithum Thain, Jeffery Sorensen, Lucas Dixon
| Challenge: | Compared to large-scale collections of conversations from social media, Wikipedia talk pages only capture a subset of all discussions and only accounts for the final form of each conversation. |
| Approach: | They propose to reconstruct a corpus that encompasses the complete history of conversations between Wikipedia contributors. |
| Outcome: | The proposed corpus extracts high quality data in both Chinese and English. |
To Mask or to Mirror: Human-AI Alignment in Collective Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly used to model and augment collective decision-making. |
| Approach: | They propose a framework for assessing collective alignment using the Lost at Sea social psychology task. |
| Outcome: | The proposed framework compares LLMs with human-AI alignment on the Lost at Sea social psychology task. |
Toxicity Detection: Does Context Really Matter? (2020.acl-main)
Copied to clipboard
| Challenge: | Existing ‘toxicity’ detection datasets and models ignore the context of the posts, implicitly assuming that comments may be judged independently. |
| Approach: | They limit the notion of context to the previous post in the thread and the discussion title and focus on how it affects human judgement. |
| Outcome: | The proposed model can amplify or mitigate perceived toxicity of posts and a small but significant subset of manually labeled posts end up having the opposite toxicity labels if the annotators are not provided with context. |
Towards Agile Text Classifiers for Everyone (2023.findings-emnlp)
Copied to clipboard
Maximilian Mozes, Jessica Hoffmann, Katrin Tomanek, Muhamed Kouate, Nithum Thain, Ann Yuan, Tolga Bolukbasi, Lucas Dixon
| Challenge: | Text-based safety classifiers are widely used for content moderation and increasingly to tune generative language model behavior. |
| Approach: | They propose to use small, targeted datasets to train safety classifiers using small, iterative datasets that can be quickly developed for a particular policy. |
| Outcome: | The proposed method can be quickly developed for a specific policy with a labeled dataset of as few as 80 examples. |