Papers by Kyunghyun Cho
Meta-Learning for Low-Resource Neural Machine Translation (D18-1)
Copied to clipboard
| Challenge: | In this paper, we propose to extend the recently introduced model-agnostic meta-learning algorithm for low-resource neural machine translation (NMT). |
| Approach: | They propose to extend the recently introduced meta-learning algorithm for low-resource neural machine translation (NMT) they frame low-Resource translation as a meta- learning problem where we learn to adapt to low-REsource languages based on multilingual high-resourced language tasks. |
| Outcome: | The proposed meta-learning algorithm outperforms the multilingual, transfer learning based approach and can train a competitive NMT system with only a fraction of training examples. |
Translation between Molecules and Natural Language (2022.emnlp-main)
Copied to clipboard
| Challenge: | MolT5 pretrains models on unlabeled natural language text and molecule strings . bringing a new drug to market can cost over a billion dollars and take over ten years . |
| Approach: | They propose a self-supervised learning framework for pretraining models on unlabeled natural language text and molecule strings. |
| Outcome: | The proposed framework pretrains models on unlabeled natural language text and molecule strings, and it generates high quality outputs. |
Show Your Work with Confidence: Confidence Bands for Tuning Curves (2024.naacl-long)
Copied to clipboard
| Challenge: | a rush to scale up has left us with large, costly language models and little understanding of how different designs compare. |
| Approach: | They propose a method to construct valid confidence bands for tuning curves . they validated their method with ablations and analyze the effect of sample size . |
| Outcome: | The proposed method shows that bootstrap confidence bands do not approximate their target confidence. |
Generating Diverse Translations with Sentence Codes (P19-1)
Copied to clipboard
| Challenge: | Existing methods to generate diverse translations use different sentence structures . Xu et al., 2018: generating multiple valid translations with high diversity is difficult . |
| Approach: | They propose to use sentence codes to condition the sentence generation to obtain diverse translations . they propose to sample multiple candidates, each of which conditioned on a unique code . |
| Outcome: | The proposed method generates paraphrase translations with drastically different structures . the proposed method can be easily adopted to existing translation systems . |
AdapterHub: A Framework for Adapting Transformers (2020.emnlp-demos)
Copied to clipboard
Jonas Pfeiffer, Andreas Rücklé, Clifton Poth, Aishwarya Kamath, Ivan Vulić, Sebastian Ruder, Kyunghyun Cho, Iryna Gurevych
| Challenge: | AdapterHub framework enables dynamic “stiching-in” of pre-trained adapters for different tasks and languages. |
| Approach: | They propose a framework that allows dynamic "stiching-in" of pre-trained adapters for different tasks and languages. |
| Outcome: | The proposed framework allows dynamic “stiching-in” of pre-trained adapters for different tasks and languages. |
The Future is not One-dimensional: Complex Event Schema Induction by Graph Modeling for Event Prediction (2021.emnlp-main)
Copied to clipboard
| Challenge: | Event schemas encode knowledge of stereotypical structures of events and their connections . previous work on event schema induction focuses on atomic events or linear temporal sequences . |
| Approach: | They propose a Temporal Complex Event Schema: a graph-based schema representation that encompasses events, arguments, temporal connections and argument relations. |
| Outcome: | The proposed model outperforms existing models on HITS@1 by 17.8%. |
Comparing Test Sets with Item Response Theory (2021.acl-long)
Copied to clipboard
Clara Vania, Phu Mon Htut, William Huang, Dhara Mungra, Richard Yuanzhe Pang, Jason Phang, Haokun Liu, Kyunghyun Cho, Samuel R. Bowman
| Challenge: | Recent results from large pretrained models show that many datasets are saturated and unlikely to detect further progress. |
| Approach: | They evaluate 29 datasets using predictions from 18 pretrained Transformer models on individual test examples. |
| Outcome: | The proposed datasets are saturated and unlikely to detect future improvements. |
Dynamic Meta-Embeddings for Improved Sentence Representations (D18-1)
Copied to clipboard
| Challenge: | A sprawling literature has emerged about what word embeddings are most useful for which tasks . word embed-ding is a technique that can be used to learn word-level meaning representations for a variety of tasks. |
| Approach: | They propose a method for supervised learning of embedding ensembles that leads to state-of-the-art performance on a variety of tasks. |
| Outcome: | The proposed method leads to state-of-the-art performance on a variety of tasks. |
Iterative Refinement in the Continuous Space for Non-Autoregressive Neural Machine Translation (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing non-autoregressive inference procedures that refine in token space often require computational overhead. |
| Approach: | They propose an efficient inference procedure that iteratively refines translation purely in the continuous space using a latent variable instead of the latent variables. |
| Outcome: | The proposed procedure is twice as efficient and more effective than the existing EM-like inference procedure. |
A Systematic Characterization of Sampling Algorithms for Open-ended Language Generation (2020.aacl-main)
Copied to clipboard
| Challenge: | Existing sampling algorithms for auto-regressive language models have similar performance . entropy reduction, order preservation and slope preservation are common properties of existing methods . |
| Approach: | They investigate the quality-diversity trade-off between ancestral sampling algorithms for auto-regressive language models. |
| Outcome: | The proposed methods have similar performance to existing methods for open-ended language generation. |
Following Length Constraints in Instructions (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing instruction following models fail to follow length constraints in their evaluations. |
| Approach: | They propose to train models that can be controlled at inference time with instructions containing desired length constraints. |
| Outcome: | The proposed models outperform standard instruction following models in length instructed evaluations. |
Intriguing Effect of the Correlation Prior on ICD-9 Code Assignment (2023.acl-srw)
Copied to clipboard
| Challenge: | The Ninth Revision of the International Classification of Diseases (ICD-9) is a standardized coding system used worldwide to classify and code diseases, injuries, and other health conditions. |
| Approach: | They evaluate the usefulness of correlation bias and suggest it could improve ICD-9 code assignment in some cases. |
| Outcome: | The proposed model improves on classes that are more imbalanced and less correlated with other codes, but the effect on individual class can be negative or positive. |
Grammar Induction with Neural Language Models: An Unusual Replication (D18-1)
Copied to clipboard
| Challenge: | Recent work on latent tree learning attempts to develop models with parse-valued latent variables and train them on non-parsing tasks. |
| Approach: | They propose a model with parse-valued latent variables and a strong latent tree learning result on constituency parsing. |
| Outcome: | The proposed model outperforms all baselines and performs competitively with symbolic grammar induction systems. |
Training a Ranking Function for Open-Domain Question Answering (N18-4)
Copied to clipboard
| Challenge: | Recent advances in machine reading have inspired researchers to combine Information Retrieval with machine reading to tackle open-domain QA. |
| Approach: | They propose two neural network rankers that assign scores to different passages based on their likelihood of containing the answer to a given question. |
| Outcome: | The proposed models achieve human level performance in open-domain QA compared to reading comprehension-style QA because it is difficult to retrieve the pieces of paragraphs that contain the answer to the question. |
First Tragedy, then Parse: History Repeats Itself in the New Era of Large Language Models (2024.naacl-long)
Copied to clipboard
| Challenge: | a new system trained on well over a trillion words smashes the state of the art by a margin previously thought impossible. |
| Approach: | They argue that disparities in scale are transient and researchers can work to reduce them . they argue that data, rather than hardware, is still a bottleneck for many applications . |
| Outcome: | a new system trained on well over a trillion words smashes the state of the art by a margin previously thought impossible. |
System-Level Natural Language Feedback (2024.eacl-long)
Copied to clipboard
| Challenge: | Existing studies on NL feedback focus on instance-level approaches to refine specific examples, but we present a framework for system-level use of NL. |
| Approach: | They propose a framework for system-level use of natural language feedback . they use feedback to formalize system-design decisions in a human-in-the-loop-process . |
| Outcome: | The proposed framework improves search query and dialog response generation and human written instance-level feedback brings further gains over GPT-3.5 written feedback. |
Don’t Say That! Making Inconsistent Dialogue Unlikely with Unlikelihood Training (2020.acl-main)
Copied to clipboard
| Challenge: | Unlikelihood is a technique developed for removal of repetition in language model completions . it allows for a model to be generalized to solve a number of problems . |
| Approach: | They extend the unlikelihood objective to generate generations that contain repetitions . they show that such an objective can be used to improve logical consistency . |
| Outcome: | The proposed approach can be applied to a number of dialogue tasks. |
Finding Generalizable Evidence by Learning to Convince Q&A Models (D19-1)
Copied to clipboard
| Challenge: | a system that finds the strongest supporting evidence for a given answer is proposed . a study using passage-based question-answering (QA) shows that agents select evidence that generalizes . |
| Approach: | They propose a system that finds the strongest supporting evidence for a given answer . they use passage-based question-answering (QA) as a testbed to train evidence agents . |
| Outcome: | The proposed system improves QA in a robust manner by using agent-selected evidence. |
Neural Unsupervised Parsing Beyond English (D19-61)
Copied to clipboard
| Challenge: | Unsupervised parsing is a task that can be learned without substantial prior knowledge. |
| Approach: | They train an unsupervised model for Arabic, Chinese, English, and German to learn syntactic structure from unlabeled text. |
| Outcome: | The PRPN architecture outperforms trivial baselines and acquires at least some parsing ability for all languages. |
Connecting the Dots: Event Graph Schema Induction with Path Language Modeling (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods to automate event extraction focus on uncertainty, re-occurring events and multiple hypotheses. |
| Approach: | They propose a new Event Graph Schema where two event types are connected through multiple paths involving entities that fill important roles in a coherent story. |
| Outcome: | The proposed model is highly effective at inducing salient and coherent schemas. |
A Stable and Effective Learning Strategy for Trainable Greedy Decoding (D18-1)
Copied to clipboard
| Challenge: | Existing methods for decoding text using beam search are expensive and require reinforcement learning. |
| Approach: | They propose a method that allows us to reap the full benefits of beam search with no additional computational cost. |
| Outcome: | The proposed method outperforms greedy decoding and beam search on machine translation tasks with minimal computational cost. |
Towards Realistic Practices In Low-Resource Natural Language Processing: The Development Set (D19-1)
Copied to clipboard
| Challenge: | Using development sets for low-resource training is often more effective . however, some studies show that early stopping can overestimate performance . |
| Approach: | They find that early stopping on a development set is more effective than using all available data for training. |
| Outcome: | The proposed model overestimates accuracy over languages and tasks by 1.4% compared to a more realistic set of training epochs. |
The Geometry of Prompting: Unveiling Distinct Mechanisms of Task Adaptation in Language Models (2025.findings-naacl)
Copied to clipboard
| Challenge: | Recent studies have focused on the inputoutput behavior of LMs, leaving the internal mechanisms behind ICL largely unexplored. |
| Approach: | They investigate how different prompting methods modify internal representations in pre-trained language models. |
| Outcome: | The proposed model can be programmed with natural language to perform a wide array of tasks without expensive fine-tuning. |
On the Effect of Pretraining Corpora on In-context Learning by a Large-scale Language Model (2022.naacl-main)
Copied to clipboard
Seongjin Shin, Sang-Woo Lee, Hwijeen Ahn, Sungdong Kim, HyoungSeok Kim, Boseop Kim, Kyunghyun Cho, Gichang Lee, Woomyoung Park, Jung-Woo Ha, Nako Sung
| Challenge: | Recent studies on large-scale in-context language models have reported successful in-const zero- and few-shot learning ability. |
| Approach: | They investigate the effects of the pretraining corpus on in-context learning in a Korean-centric model. |
| Outcome: | The study shows that pretraining corpus size does not determine in-context learning ability . the findings suggest that in-constext learning is not always competitive . |
AdapterFusion: Non-Destructive Task Composition for Transfer Learning (2021.eacl-main)
Copied to clipboard
| Challenge: | Existing methods for incorporating knowledge from multiple tasks suffer from catastrophic forgetting and difficulties in dataset balancing. |
| Approach: | They propose an algorithm that extracts and combine adapters in a knowledge composition step. |
| Outcome: | The proposed class outperforms traditional methods such as full fine-tuning and multi-task learning on 16 diverse NLU tasks. |
Analyzing the Forgetting Problem in Pretrain-Finetuning of Open-domain Dialogue Response Models (2021.eacl-main)
Copied to clipboard
| Challenge: | a large-scale unsupervised pretraining has been shown to greatly boost the performance of natural language processing models. |
| Approach: | They propose an intuitive finetuning strategy to regularize the finetune process . they propose a mix-review strategy to alleviate the forgetting problem . |
| Outcome: | The proposed strategy regularizes the finetuning process, and the forgetting problem is alleviated . the proposed strategy also improves the performance of the resulting model . |
Scaling Laws Are Unreliable for Downstream Tasks: A Reality Check (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Downstream scaling laws aim to predict task performance at larger scales from the model’s performance at smaller scales. |
| Approach: | They conduct a meta-analysis of existing data on downstream scaling laws and find that predictable scaling only occurs in a minority of cases: 39% of the time. |
| Outcome: | The proposed scaling laws only occur in a minority of cases, and seemingly benign changes to the experimental setting can completely change the scaling behavior. |
Conditional Word Embedding and Hypothesis Testing via Bayes-by-Backprop (D18-1)
Copied to clipboard
| Challenge: | Whether word's meaning varies across contexts has become a major focus of research in recent years. |
| Approach: | They propose a word embedding model that incorporates document covariates to estimate conditional word embeds. |
| Outcome: | The proposed model estimates word embedding distributions based on document covariates . if word embeds are statistically significant, hypothesis tests can be performed . |
HUE: Pretrained Model and Dataset for Understanding Hanja Documents of Ancient Korea (2022.findings-naacl)
Copied to clipboard
| Challenge: | Historical records in Korea before the 20th century were primarily written in Hanja, an extinct language based on Chinese characters. |
| Approach: | They present a dataset that includes tasks for attribution, topic classification, named entity recognition and summary retrieval for classical Hanja documents. |
| Outcome: | The proposed models improve on the Annals of the Joseon Dynasty and Diaries of the Royal Secretariats datasets. |
Zero-Shot Transfer Learning for Event Extraction (P18-1)
Copied to clipboard
| Challenge: | Existing supervised event extraction methods rely on manual annotations and features specific to each event type. |
| Approach: | They propose a framework that maps event mentions to a specific type in an event ontology . they use existing annotations to extract event types from unstructured text data . |
| Outcome: | The proposed framework can be applied to new unseen event types without manual annotations. |
Consistency of a Recurrent Language Model With Respect to Incomplete Decoding (2020.emnlp-main)
Copied to clipboard
| Challenge: | Neural sequence models trained with maximum likelihood have been shown to exhibit issues such as length bias and degenerate repetition. |
| Approach: | They propose to use a recurrent language model to address inconsistency in decoding algorithms that are inconsistent despite the fact that recursive language models are trained to produce sequences of finite length. |
| Outcome: | The proposed methods prevent inconsistency in the proposed models. |
Leveraging Implicit Feedback from Deployment Data in Dialogue (2024.eacl-short)
Copied to clipboard
| Challenge: | Xu et al., 2023) and Bai ed., 2019) use crowdworkers to collect signals from natural dialogue episodes. |
| Approach: | They use the publicly released BlenderBot deployment data to extract signals from conversations to implicitly measure the quality of a machine-generated utterance. |
| Outcome: | The proposed model improves over baseline models, but some proxy signals can lead to undesirable generations. |
Asking and Answering Questions to Evaluate the Factual Consistency of Summaries (2020.acl-main)
Copied to clipboard
| Challenge: | Existing automatic evaluation metrics for summarization are insensitive to factual inconsistencies. |
| Approach: | They propose an automatic evaluation protocol that detects factual inconsistencies in a model-generated summary. |
| Outcome: | QAGS has higher correlations with human judgments of factual consistency than other evaluation metrics. |
Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (D18-1)
Copied to clipboard
| Challenge: | Despite its success, neural autoregressive modeling has its weakness in decoding, i.e., finding the most likely sequence. |
| Approach: | They propose a conditional non-autoregressive neural sequence model based on iterative refinement based upon latent variable models and conditional denoising autoencoders. |
| Outcome: | The proposed model significantly speeds up decoding while maintaining the generation quality comparable to the autoregressive counterpart. |
On the Blind Spots of Model-Based Evaluation Metrics for Text Generation (2023.acl-long)
Copied to clipboard
| Challenge: | Existing methods for text generation evaluation metrics are lacking in robustness analysis. |
| Approach: | They propose to use stress tests to test for errors in text generation evaluation metrics . they find that BERTScore is confused by truncation errors in summarization . |
| Outcome: | The proposed stress tests show that they are insensitive to errors in open-ended generation, translation, and summarization. |
DEEP: DEnoising Entity Pre-training for Neural Machine Translation (2022.acl-long)
Copied to clipboard
| Challenge: | Earlier named entity translation methods focus on phonetic transliteration, which ignores the sentence context for translation. |
| Approach: | They propose a DEnoising Entity Pre-training method that leverages monolingual data and a knowledge base to improve named entity translation accuracy within sentences. |
| Outcome: | The proposed method improves on three language pairs and denoising auto-encoding baselines. |
SSMBA: Self-Supervised Manifold Based Data Augmentation for Improving Out-of-Domain Robustness (2020.emnlp-main)
Copied to clipboard
| Challenge: | Data augmentation is a common method used to improve out-of-domain (OOD) generalization. |
| Approach: | They propose a data augmentation method that uses corruption and reconstruction functions to move randomly on a manifold to generate training examples. |
| Outcome: | The proposed method outperforms existing methods and baseline models on both in-domain and OOD data and achieves gains of 0.8% on OOD Amazon reviews, 1.8% accuracy on OOO MNLI, and 1.4 BLEU on in- domain IWSLT14 German-English. |
Dialogue Natural Language Inference (P19-1)
Copied to clipboard
| Challenge: | Consistency is a long standing issue faced by dialogue models. |
| Approach: | They propose to frame the consistency of dialogue agents as natural language inference and create a new natural language dataset called Dialogue NLI. |
| Outcome: | The proposed model can improve the consistency of a dialogue model with human evaluation and automatic metrics on a suite of evaluation sets designed to measure the model’s consistency. |
Emergent Linguistic Phenomena in Multi-Agent Communication Games (D19-1)
Copied to clipboard
| Challenge: | a recent study examines the behavior of linguistic agents in a community-level setting . a linguistic continuum emerges where neighboring languages are more mutually intelligible than farther removed ones . |
| Approach: | They propose a multi-agent communication framework for studying linguistic phenomena at the community level. |
| Outcome: | The proposed framework can reproduce complex linguistic behavior observed in natural language . it can be used to study interactions between perceptually-enabled agents . |
Language Models as Causal Effect Generators (2025.emnlp-main)
Copied to clipboard
| Challenge: | Using sequence-driven structural causal models (SD-SCMs) we characterize how SD-SCAMs enables sampling from observational, interventional, and counterfactual distributions according to the desired causal structure. |
| Approach: | They propose a sequence-driven structural causal model that uses language models to parameterize a structural causal system based on a user-specified DAG. |
| Outcome: | The proposed method outperforms state-of-the-art methods and can underpin auditing of language models for (un)desirable causal effects, such as misinformation or discrimination. |
Countering Language Drift via Visual Grounding (D19-1)
Copied to clipboard
| Challenge: | Emergent multi-agent communication protocols are different from natural language . a long-standing goal of artificial intelligence research is to develop agents that can cooperate with other agents . |
| Approach: | They propose to use syntactic and semantic constraints to improve communication . they propose to combine these constraints with auxiliary training constraints to reduce language drift . |
| Outcome: | a new study shows that pre-trained agents retain English syntax while learning to convey intended meaning . the proposed training constraints can be used to mitigate language drift . |
Unsupervised Question Decomposition for Question Answering (2020.emnlp-main)
Copied to clipboard
| Challenge: | Existing QA systems struggle to answer complex questions because information is scattered in different places. |
| Approach: | They propose an unsupervised algorithm that decomposes hard questions into simpler sub-questions . they propose an algorithm that can be used to generate a final answer from millions of questions . |
| Outcome: | The proposed algorithm decomposes hard questions into simpler sub-questions that existing QA systems can answer. |
Translating Hanja Historical Documents to Contemporary Korean and English (2022.findings-emnlp)
Copied to clipboard
| Challenge: | The Annals of Joseon Dynasty contain the daily records of the Kings of Joseont, the 500-year kingdom preceding the modern nation of Korea. |
| Approach: | They propose a neural machine translation model that translates historical documents written in Hanja to more easily understandable Korean and to English. |
| Outcome: | The proposed model outperforms baseline models in terms of BLEU scores for both contemporary Korean and English translations. |
Characterizing and addressing the issue of oversmoothing in neural autoregressive sequence modeling (2022.aacl-main)
Copied to clipboard
| Challenge: | Neural autoregressive sequence models assign high probability to unreasonably short sequences . authors propose to minimize oversmoothing rate during training . |
| Approach: | They propose to minimize the oversmoothing rate during training by tuning the regularization strength. |
| Outcome: | The proposed regularization can control the oversmoothing rate and improve decoding performance. |
Multi-lingual Common Semantic Space Construction via Cluster-consistent Word Embedding (D18-1)
Copied to clipboard
| Challenge: | a new approach to multilingual word embedding is needed to achieve this goal . a multilingual common semantic space is a language-agnostic semantic continuous space . |
| Approach: | They propose a multilingual common semantic space where words from multiple languages are mapped into a shared space so that resources and knowledge can be shared across languages. |
| Outcome: | The proposed approach achieves 14.6% absolute F-score gain over state-of-the-art methods on cross-lingual direct transfer. |
Length-Adaptive Transformer: Train Once with Length Drop, Use Anytime with Search (2021.acl-long)
Copied to clipboard
| Challenge: | Pre-trained language models have achieved notable improvements in various NLP tasks. |
| Approach: | They propose a Length-Adaptive Transformer that can be used for various inference scenarios after one-shot training. |
| Outcome: | The proposed model can be used for various inference scenarios after one-shot training. |
Making the Most Out of the Limited Context Length: Predictive Power Varies with Clinical Note Type and Note Section (2023.acl-srw)
Copied to clipboard
| Challenge: | Clinical notes have a long time span over multiple long documents. |
| Approach: | They propose a framework to analyze clinical notes with high predictive power . they propose to combine different types of notes to improve performance . |
| Outcome: | The proposed framework could be used to extract information from clinical notes . it shows that the sample size can be optimized for large contexts . |
Improving Conversational Question Answering Systems after Deployment using Feedback-Weighted Learning (2020.coling-main)
Copied to clipboard
| Challenge: | In most applications, users are not able to provide the correct answer to the system, but they are able provide binary (correct, incorrect) feedback. |
| Approach: | They propose feedback-weighted learning based on importance sampling to improve upon an initial supervised system using binary user feedback. |
| Outcome: | The proposed method improves on an initial supervised system, getting close to a fully-supervised system that has access to the same labeled examples in in-domain experiments (QuAC) and matching in out-of-domain experiment (DoQA). |
Improved Zero-shot Neural Machine Translation via Ignoring Spurious Correlations (P19-1)
Copied to clipboard
| Challenge: | Existing approaches to train a multilingual NMT model for low-resource languages are lacking in terms of number of supervised examples. |
| Approach: | They propose to use decoder pre-training and back-translation to solve the degeneracy problem by analyzing spurious correlations between source and decoded sentences. |
| Outcome: | The proposed methods show significant improvement over the pivot-based approach on three challenging multilingual datasets. |