Yejin Choi · page 2
Papers
- Learning to Write with Cooperative Discriminators (2018 · 303 citations)
RNN language models write fluent sentences, but over whole paragraphs their output tends to be bland, repetitive and occasionally self-contradictory. - TreeTalk: Composition and Compression of Trees for Image Descriptions (2014 · 290 citations)
The paper addresses automatic image captioning. Instead of limiting descriptions to what vision detectors can recognize, it draws on the expressive language people write about photos online. Given a query image, the system retrieves visually similar photos from the collection of captioned images assembled by Ordonez… - Surface Form Competition: Why the Highest Probability Answer Isn't Always Right (2021 · 277 citations)
Zero-shot multiple-choice evaluation often scores each option by the probability a language model gives it. - SocialIQA: Commonsense Reasoning about Social Interactions (2019 · 259 citations)
Pretrained language models do well on many commonsense tasks but still struggle with social reasoning: working out why people act as they do, how they feel, and what they will do next. - MAUVE: Measuring the Gap Between Neural Text and Human Text using Divergence Frontiers (2021 · 251 citations)
Judging how close machine-generated text is to human writing in open-ended generation is hard. The distributions involved are discrete and extremely high-dimensional, and they can only be reached through samples. The authors introduce MAUVE, a measure that compares a model's text distribution Q (language model plus… - Reframing Instructional Prompts to GPTk's Language (2022 · 242 citations)
Instructions written by non-experts for language models are often long and abstract, and models such as GPT3 follow them poorly. - WildChat: 1M ChatGPT Interaction Logs in the Wild (2024 · 239 citations)
Researchers have little public data on how people actually use chatbots. Most instruction-tuning data is either proprietary or written by experts, and expert-written data is usually single-turn and does not look like real usage. To fill this gap, the authors ran free chat services on Hugging Face Spaces, one backed… - COMET-ATOMIC 2020: On Symbolic and Neural Commonsense Knowledge Graphs (2020 · 239 citations)
The paper argues that hand-built commonsense knowledge graphs can never cover everything an NLP system will meet. - Maieutic Prompting: Logically Consistent Reasoning with Recursive Explanations (2022 · 219 citations)
Large language models often reason inconsistently, and prompting them to write explanations before answering does not fully help, since those explanations can be wrong or contradict the final answer. - SODA: Million-scale Dialogue Distillation with Social Commonsense Contextualization (2023 · 202 citations)
Open-domain social dialogue research has been held back by a lack of data. Everyday conversations are rarely recorded, private messages raise legal and privacy problems, and crowdsourced corpora are small and narrowly themed. The authors address this by generating dialogues synthetically with CO 3, a three-stage… - Generative Data Augmentation for Commonsense Reasoning (2020 · 198 citations)
High-performing commonsense reasoning models depend on large crowdsourced training sets. These are expensive to build and often contain annotation artifacts that models exploit. The authors propose G-DAUG^c, which expands small training sets with synthetic multiple-choice items produced by finetuned pretrained… - Neural Metaphor Detection in Context (2018 · 198 citations)
The paper addresses metaphor detection: deciding whether a word in a sentence is used figuratively. Earlier systems often reduced context to subject–verb–object triples or relied on shallow unigram features. The authors show that the full sentence gives the model more useful evidence. They test two setups: tagging… - Challenges in Automated Debiasing for Toxic Language Detection (2021 · 191 citations)
Toxic language classifiers often learn shortcuts that harm minority groups. They flag tweets simply for mentioning minority identities, for containing profanity, or for using features of African American English (AAE). This paper asks whether debiasing techniques developed for NLU tasks such as NLI can fix these… - RealTime QA: What's the Answer Right Now? (2022 · 181 citations)
Most open-domain QA benchmarks assume an answer stays fixed regardless of when the question is asked. RealTime QA drops that assumption. It is a living benchmark that publishes fresh questions about current events every week and scores systems on them before the next batch appears. Questions come from the weekly news… - Is Reinforcement Learning (Not) for Natural Language Processing: Benchmarks, Baselines, and Building Blocks for Natural Language Policy Optimization (2022 · 174 citations)
The paper asks whether reinforcement learning is a practical way to align pre-trained language models with human preferences. - Generating Sequences by Learning to Self-Correct (2022 · 172 citations)
Language models usually produce output in one pass and treat it as final. Even when the output almost works, for example a program with one wrong line or a sentence missing a keyword, it gets discarded rather than repaired. Welleck et al. propose SELF-CORRECTION, which splits generation into two parts. A base… - Quark: Controllable Text Generation with Reinforced Unlearning (2022 · 170 citations)
Large pretrained language models often produce text that users do not want, such as toxic language, heavy repetition, or the wrong sentiment. - WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs (2024 · 165 citations)
Open safety classifiers for LLMs, such as Llama-Guard2, do reasonably well on plainly worded harmful requests. - Is GPT-3 Text Indistinguishable from Human Text? Scarecrow: A Framework for Scrutinizing Machine Text (2022 · 152 citations)
Recent work suggested that ordinary crowd workers can no longer reliably tell GPT-3 output from human writing. - ProsocialDialog: A Prosocial Backbone for Conversational Agents (2022 · 152 citations)
Open-domain chatbots such as GPT-3, BlenderBot and OPT tend to go along with harmful or unethical user statements, or to dodge them with generic deflections. - The Unlocking Spell on Base LLMs: Rethinking Alignment via In-Context Learning (2023 · 150 citations)
Alignment tuning through SFT and RLHF is usually seen as the step that turns a pretrained base LLM into a usable assistant. - Magpie: Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing (2024 · 148 citations)
Open-weight chat models such as Llama-3-Instruct are released without the instruction data used to align them. - Dynamic Entity Representations in Neural Language Models (2017 · 144 citations)
Most neural language models do not represent entities beyond the words used to mention them, so they cannot follow who or what a long text is about. - A Roadmap to Pluralistic Alignment (2024 · 143 citations)
This ICML 2024 position paper argues that aligning AI to an "average" human preference ignores real disagreement among people. Using LLMs as a testbed, the authors define three ways a model can be pluralistic: - Symbolic Chain-of-Thought Distillation: Small Models Can Also “Think” Step-by-Step (2023 · 140 citations)
Chain-of-thought (CoT) prompting helps very large language models, but the gains have mostly appeared only at scales above roughly 50–60B parameters.
This page in Markdown → https://md.research.elamur.ai/free/en/authors/yejin-choi/page/2/