Papers by Danqi Chen
Adapting Language Models to Compress Contexts (2023.emnlp-main)
Copied to clipboard
| Challenge: | Transformer-based language models have a finite context window and expensive computational cost of processing long text documents. |
| Approach: | They propose to adapt pre-trained LMs into AutoCompressors to compress text into summary vectors . authors propose to use summary vector to speed up inference over long contexts based on a finite context window . |
| Outcome: | The proposed model can compress long contexts into summary vectors, which are accessible as soft prompts. |
Structured Pruning Learns Compact and Accurate Models (2022.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models have high costs in terms of storage, memory, and computation time. |
| Approach: | They propose a task-specific structured pruning method CoFi which provides highly parallelizable subnetworks and matches distillation methods in both accuracy and latency. |
| Outcome: | The proposed method matches the distillation methods in accuracy and latency without resorting to unlabeled data. |
Prompting ELECTRA: Few-Shot Learning with Discriminative Pre-Trained Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Pre-trained masked language models perform few-shot learning, but discriminative models like ELECTRA do not fit into the paradigm. |
| Approach: | They propose to use ELECTRA to train pre-trained models to score originality of target options without introducing new parameters. |
| Outcome: | The proposed model outperforms masked language models in a wide range of tasks without adding new parameters. |
Single-dataset Experts for Multi-dataset Question Answering (2021.emnlp-main)
Copied to clipboard
| Challenge: | Prior work has focused on training one network on multiple datasets to build a model that performs well on all of the training datasets and generalizes and transfers better to new datasets. |
| Approach: | They combine multiple reading comprehension datasets to build a multi-dataset question answering model with an ensemble of single-data set experts. |
| Outcome: | The proposed model outperforms baseline models in in-distribution accuracy and generalization and transfer performance. |
Learning Dense Representations of Phrases at Scale (2021.acl-long)
Copied to clipboard
| Challenge: | Existing phrase retrieval models rely on sparse representations and still underperform retriever-reader approaches. |
| Approach: | They propose a method to learn phrase representations from reading comprehension tasks using negative sampling methods. |
| Outcome: | The proposed model improves over previous models by 15%-25% absolute accuracy and matches the performance of state-of-the-art retrieval models. |
SpanBERT: Improving Pre-training by Representing and Predicting Spans (2020.tacl-1)
Copied to clipboard
| Challenge: | Pre-training methods like BERT mask individual words or subword units, but many tasks involve reasoning about relationships between two or more spans of text. |
| Approach: | They propose a pre-training method that masks contiguous random spans instead of random tokens to train the span boundary representations to predict the entire content of the masked span. |
| Outcome: | The proposed method outperforms BERT and its better-tuned baselines on span selection tasks and on coreference resolution tasks. |
Should You Mask 15% in Masked Language Modeling? (2023.eacl-main)
Copied to clipboard
| Challenge: | Masked language models (MLMs) traditionally mask 15% of tokens due to the belief that more masking would leave insufficient context to learn good representations. |
| Approach: | They revisit the 15% masking rate of MLMs to examine the role of masking in linguistic training. |
| Outcome: | The proposed masking rate outperforms BERT-large size models on GLUE and SQUAD while maintaining 95% accuracy. |
Simple Entity-Centric Questions Challenge Dense Retrievers (2021.emnlp-main)
Copied to clipboard
| Challenge: | Open-domain question answering has exploded in popularity due to the success of dense retrieval models. |
| Approach: | They construct a set of simple, entity-rich questions based on facts from Wikidata and test their models against supervised datasets. |
| Outcome: | The proposed model outperforms sparse retrieval methods on open-domain question answering datasets by a large margin. |
MQuAKE: Assessing Knowledge Editing in Language Models via Multi-Hop Questions (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for retraining from scratch are limited and only work on the recall of edited facts. |
| Approach: | They propose a benchmark method that allows users to ask multi-hop questions to assess whether edited models correctly answer questions where the answer should change as an entailed consequence of edited facts. |
| Outcome: | The proposed method outperforms existing models and scales well with LLMs (up to 175B) it is based on a memory-based approach that stores all edited facts externally while prompting the language model iteratively to generate answers consistent with the edited facts. |
Retrieval-based Language Models and Applications (2023.acl-tutorials)
Copied to clipboard
| Challenge: | In this tutorial, we will provide a comprehensive overview of retrieval-based language models. |
| Approach: | This tutorial will provide a comprehensive overview of recent advances in retrieval-based language models. |
| Outcome: | This tutorial will provide a comprehensive overview of recent advances in retrieval-based language models. |
Poisoning Retrieval Corpora by Injecting Adversarial Passages (2023.emnlp-main)
Copied to clipboard
| Challenge: | Dense retrievers have outperformed traditional lexical methods in a range of information retrieval tasks, but to what extent can they be safely deployed in real-world applications? |
| Approach: | They propose a method where a malicious user injects a small number of adversarial passages into a retrieval corpus to maximize similarity with a set of training queries. |
| Outcome: | The proposed attack fools retrieval systems into returning top results for queries not seen by the attacker. |
MRQA 2019 Shared Task: Evaluating Generalization in Reading Comprehension (D19-58)
Copied to clipboard
| Challenge: | MRQA datasets have been used to benchmark progress in general-purpose language understanding. |
| Approach: | They propose to combine 18 question answering datasets into one shared task to evaluate their generalization capabilities. |
| Outcome: | The best system achieved an average F1 score of 72.5 on the 12 held-out datasets, 10.7 absolute points higher than baseline based on BERT. |
Query-Focused Retrieval Heads Improve Long-Context Reasoning and Re-ranking (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent work has identified retrieval heads as a subset of attention heads responsible for retrieving salient information in long-context language models. |
| Approach: | They introduce a retrieval head that uses attention scores to enhance retrieval from long context . they use QRRetriever to select the most relevant parts with the highest retrieval scores . |
| Outcome: | The proposed retrieval heads outperform other retrieval-based retrieval retrievers on BEIR benchmarks. |
A Discrete Hard EM Approach for Weakly Supervised Question Answering (D19-1)
Copied to clipboard
| Challenge: | Existing work on question answering tasks only provide weak supervision for how the answer should be computed . weak supervision is attractive because it is relatively easy to gather, allowing for large datasets . but weak supervision complicates learning because there are many different spurious ways to derive the correct answer. |
| Approach: | They propose a method to convert question answering tasks into discrete latent variable learning problems with a precomputed set of possible solutions that contains one correct option. |
| Outcome: | The proposed approach outperforms previous methods on six QA tasks and achieves state-of-the-art on five of them. |
What In-Context Learning “Learns” In-Context: Disentangling Task Recognition and Task Learning (2023.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) can perform in-context learning (ICL) with only a few demonstrations, but its mechanisms are not well-understood. |
| Approach: | They characterize two ways in which LLMs leverage demonstrations to solve tasks with a few demonstrations. |
| Outcome: | The proposed model achieves non-trivial performance with only TR, and TR does not improve with larger models or more demonstrations. |
How to Train Long-Context Language Models (Effectively) (2025.acl-long)
Copied to clipboard
| Challenge: | a new study shows that language models can process extremely long contexts with minimal training. |
| Approach: | They use supervised fine-tuning and continued training to evaluate a language model's long-context capabilities. |
| Outcome: | The proposed model outperforms Llama-3.1-8B-Instruct on most long-context tasks . the model can process 512K tokens, one of the longest context windows of LMs . |
Open-Domain Question Answering (2020.acl-tutorials)
Copied to clipboard
| Challenge: | tutorial provides a comprehensive overview of cutting-edge research in open-domain question answering (QA) |
| Approach: | tutorial provides a comprehensive overview of cutting-edge research in open-domain question answering . focus will shift to cutting- edge models proposed for open- domain QA . |
| Outcome: | The tutorial will cover cutting-edge research in open-domain question answering (QA) it will cover two-stage retriever-reader approaches, dense retriever and end-to-end training, and retriever free methods . |
Privacy Implications of Retrieval-Based Language Models (2023.emnlp-main)
Copied to clipboard
| Challenge: | a study of retrieval-based language models shows improved interpretability, factuality, and adaptability compared to parametric counterparts . kNN-LMs are more susceptible to leaking private information from their private datastore than parametric models . |
| Approach: | They present the first study of privacy risks in retrieval-based language models . they aim to strike a balance between utility and privacy in domains where privacy is of concern . |
| Outcome: | The proposed methods improve interpretability, factuality, and adaptability compared to parametric models . the study finds that kNN-LMs are more susceptible to leaking private data than parametric ones . |
Measuring Inductive Biases of In-Context Learning with Underspecified Demonstrations (2023.acl-long)
Copied to clipboard
| Challenge: | In-context learning is an important paradigm for adapting large language models to new tasks . but the generalization behavior of ICL remains poorly understood . |
| Approach: | They characterize the feature biases of large language models by constructing underspecified demonstrations . they find that LLMs exhibit clear feature bias, and they evaluate interventions . |
| Outcome: | The proposed model prefers the "default" task features over distractor features more often than the base model. |
Dense Passage Retrieval for Open-Domain Question Answering (2020.emnlp-main)
Copied to clipboard
Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, Wen-tau Yih
| Challenge: | Open-domain question answering relies on efficient passage retrieval to select candidate contexts. |
| Approach: | They propose a dual-encoder framework that can be implemented to retrieve passages from a small number of questions and passages. |
| Outcome: | The proposed system outperforms a strong Lucene-BM25 system in top-20 passage retrieval accuracy on multiple open-domain QA benchmarks. |
TextHide: Tackling Data Privacy in Language Understanding Tasks (2020.findings-emnlp)
Copied to clipboard
| Challenge: | Unsolved privacy challenges in distributed or federated learning are a challenge for many domains including Natural Language Processing. |
| Approach: | They propose a federated learning framework that adds an encryption step to prevent an eavesdropping attacker from recovering private text data. |
| Outcome: | The proposed model can effectively defend against attacks on shared gradients or representations and the averaged accuracy reduction is only 1.9%. |
Phrase Retrieval Learns Passage Retrieval, Too (2021.emnlp-main)
Copied to clipboard
| Challenge: | Dense retrieval methods have shown great promise over sparse methods in a range of NLP problems. |
| Approach: | They propose to use dense phrase retrieval to learn coarse-level retrieval including passages . they show phrase retrievals can be fine-tuned for more coarse-grained retrieval units . |
| Outcome: | The proposed method improves passage retrieval accuracy and QA performance with fewer passages. |
Non-Parametric Few-Shot Learning for Word Sense Disambiguation (2021.naacl-main)
Copied to clipboard
| Challenge: | Word sense disambiguation (WSD) is a problem in natural language processing . 84% of annotated words have less than 10 examples in the long-tail distribution . |
| Approach: | They propose a non-parametric few-shot learning approach to mitigate word sense disambiguation . they use a metric space to compute distances among the senses of a given word . |
| Outcome: | The proposed method achieves a 75.1 F1 score on the unified evaluation benchmark. |
Representing Rule-based Chatbots with Transformers (2025.naacl-long)
Copied to clipboard
| Challenge: | Existing work on how Transformers can solve synthetic tasks has not explored how to extend this to a conversational setting. |
| Approach: | They propose to use ELIZA as a framework for formal mechanistic analysis of Transformers . they propose to model local pattern matching and long-term dialogue state tracking . |
| Outcome: | The proposed model can be extended to model key aspects of conversation, the authors show . their model favors an induction head mechanism over a more precise copying mechanism . |
Don’t Prompt, Search! Mining-based Zero-Shot Learning with Language Models (2022.emnlp-main)
Copied to clipboard
| Challenge: | Recent work has obtained strong zero-shot results by prompting language models. |
| Approach: | They propose a mining-based approach that uses regular expressions to mine labeled examples from unlabeled corpora and fine tune a pretrained model. |
| Outcome: | The proposed method outperforms prompting on a wide range of tasks when using comparable templates. |
Ditch the Gold Standard: Re-evaluating Conversational Question Answering (2022.acl-long)
Copied to clipboard
| Challenge: | Existing conversational question answering systems provide natural-language answers to users in information-seeking conversations. |
| Approach: | They conduct the first large-scale human evaluation of state-of-the-art conversational question answering systems . they propose a question rewriting mechanism based on predicted history which better correlates with human judgments . |
| Outcome: | The proposed question rewriting mechanism better correlates with human judgments. |
Training Language Models with Memory Augmentation (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for training memory-augmented language models only introduce mem-ories at testing time or represent them using a separately trained encoder. |
| Approach: | They propose a training approach that directly takes in-batch examples as accessible memory and new methods for memory construction and data batching that are used for adapting to different sets of memories at testing time. |
| Outcome: | The proposed approach reduces perplexity from 18.70 to 15.37 on multiple language modeling and machine translation benchmarks. |
Generating Natural Language Proofs with Verifier-Guided Search (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing stepwise methods struggle to generate valid proof steps based on the hypothesis . instead, they generate invalid steps . |
| Approach: | They propose a stepwise method which generates relevant steps conditioning on the hypothesis. |
| Outcome: | The proposed method improves correctness of predicted proofs from 27.7% to 33.3% on EntailmentBank and RuleTaker. |
A Frustratingly Easy Approach for Entity and Relation Extraction (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing work on end-to-end relation extraction models combine two tasks: named entity recognition and relation extraction. |
| Approach: | They propose a pipelined approach for entity and relation extraction that uses two independent encoders to construct the relation model. |
| Outcome: | The proposed approach achieves an 8.16 speedup with a slight reduction in accuracy on standard benchmarks. |
Can Rationalization Improve Robustness? (2022.naacl-main)
Copied to clipboard
| Challenge: | Existing models that generate rationales before making predictions can ignore noise or adversarially added text by simply masking it out of the generated rationale. |
| Approach: | They propose to use a 'rationalizethen-predict' framework to generate subsets of input to generate rationales and then make predictions using them. |
| Outcome: | The proposed models improve robustness over AddText attacks while struggling in certain scenarios. |
Enabling Large Language Models to Generate Text with Citations (2023.emnlp-main)
Copied to clipboard
| Challenge: | Existing work relies on commercial search engines and human evaluation, making it difficult to reproduce and compare different modeling approaches. |
| Approach: | They propose a new generation paradigm that requires large language models to provide citations to one or a few text passages for any statement they generate. |
| Outcome: | The proposed model improves factual correctness and verifiability of large language models by providing citations to a set of questions and retrieval corpora and generating answers with citation. |
MABEL: Attenuating Gender Bias using Textual Entailment Data (2022.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for mitigating gender bias in language models are insufficient or inconsistent. |
| Approach: | They propose a method for attenuating gender bias using entailment labels . they use a contrastive learning objective on counterfactually augmented enanglement pairs . |
| Outcome: | The proposed method outperforms previous task-agnostic debiasing approaches on intrinsic and extrinsic metrics and preserves task performance after fine-tuning on downstream tasks. |
LitSearch: A Retrieval Benchmark for Scientific Literature Search (2024.emnlp-main)
Copied to clipboard
| Challenge: | Literature search questions pose significant challenges for modern retrieval systems . a lack of domain expertise and reasoning through lengthy papers is a challenge . |
| Approach: | They propose a retrieval benchmark for literature search queries using inline citations from papers and questions about recently published papers. |
| Outcome: | The proposed retrieval benchmarks outperform state-of-the-art retrieval models and reranking pipelines. |
Finding Dataset Shortcuts with Grammar Induction (2022.emnlp-main)
Copied to clipboard
| Challenge: | Prior work on shortcut detection focused on enumerating features like unigrams or bigrams . prior work relied on post-hoc models that reveal qualitative patterns without a clear statistical interpretation . |
| Approach: | They propose to use probabilistic grammars to characterize and discover shortcuts in NLP datasets using context-free grammars and synchronous context- free grammars. |
| Outcome: | The proposed grammars reveal interesting shortcut features in a number of datasets, including simple and high-level features, and automatically identify groups of test examples on which conventional classifiers fail. |
Optimizing Test-Time Query Representations for Dense Retrieval (2023.findings-acl)
Copied to clipboard
| Challenge: | Recent developments of dense retrieval rely on quality representations of queries and contexts from pre-trained query and context encoders. |
| Approach: | They propose a test-time optimization of query representations that provides fine-grained pseudo labels over retrieval results. |
| Outcome: | The proposed algorithm improves open-domain question answering accuracy and direct re-ranking by up to 2.0% while running 1.3–2.4x faster with an efficient implementation. |
Factual Probing Is [MASK]: Learning vs. Learning to Recall (2021.naacl-main)
Copied to clipboard
| Challenge: | Existing methods for factual probing can interpret the model’s prediction accuracy as a lower bound on the amount of factual information it encodes. |
| Approach: | They propose a method which directly optimizes in continuous embedding space and can predict an additional 6.4% of facts in the LAMA benchmark. |
| Outcome: | The proposed method outperforms the best previous prompt method by 6.4% on the LAMA benchmark. |
Making Pre-trained Language Models Better Few-shot Learners (2021.acl-long)
Copied to clipboard
| Challenge: | Recent studies show that the GPT-3 model can perform few-shots on language understanding tasks with a natural-language prompt and a few task demonstrations. |
| Approach: | They propose a technique for fine-tuning language models using a few examples . they propose LM-BFF, which uses prompt-based fine-uning and a pipeline for automating prompt generation . |
| Outcome: | The proposed approach outperforms standard fine-tuning procedures on a range of NLP tasks. |
Training Trajectories of Language Models Across Scales (2023.acl-long)
Copied to clipboard
Mengzhou Xia, Mikel Artetxe, Chunting Zhou, Xi Victoria Lin, Ramakanth Pasunuru, Danqi Chen, Luke Zettlemoyer, Veselin Stoyanov
| Challenge: | Scaling up language models has led to unprecedented performance gains, but little is understood about how the training dynamics change as models get larger. |
| Approach: | They analyze the training checkpoints of different-sized OPT models on next-token prediction, sequence-level generation and downstream tasks. |
| Outcome: | The results show that language models of different sizes learn more during training . small models halt at hallucinations, larger ones learn to assign lower probabilities . |
SimCSE: Simple Contrastive Learning of Sentence Embeddings (2021.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for learning universal sentence embeddings are based on unsupervised approaches with only dropout as noise. |
| Approach: | They propose an unsupervised approach that takes an input sentence and predicts itself in a contrastive objective with only standard dropout used as noise. |
| Outcome: | The proposed framework performs on par with previous supervised approaches and can produce superior sentence embeddings from unlabeled or labeled data. |
C-STS: Conditional Semantic Textual Similarity (2023.emnlp-main)
Copied to clipboard
Ameet Deshpande, Carlos Jimenez, Howard Chen, Vishvak Murahari, Victoria Graf, Tanmay Rajpurohit, Ashwin Kalyan, Danqi Chen, Karthik Narasimhan
| Challenge: | Semantic textual similarity (STS) is a cornerstone task in natural language processing, but it is inherently ambiguous. |
| Approach: | They propose a task called conditional STS which measures similarity conditioned on an aspect elucidated in natural language. |
| Outcome: | The proposed task reduces subjectivity and ambiguity and enables fine-grained similarity evaluation using diverse conditions. |
Long-Context Language Modeling with Parallel Context Encoding (2024.acl-long)
Copied to clipboard
| Challenge: | Existing long-context models degenerate with retrieved contexts. |
| Approach: | They propose a framework that can be applied to existing decoder-only LLMs for context expansion. |
| Outcome: | The proposed framework can be applied to any existing decoder-only LLMs for context expansion. |