Kyunghyun Cho · page 2
Papers
- Zero-Shot Transfer Learning for Event Extraction (2018 · 273 citations)
Supervised event extractors depend on hand-labeled examples for each event type, so covering a new type normally requires fresh annotation. - BERT has a Mouth, and It Must Speak: BERT as a Markov Random Field Language Model (2019 · 256 citations)
BERT is pretrained to fill in masked tokens using context on both sides, so it is not obvious how to use it as an ordinary language model that scores or generates text. - Translation between Molecules and Natural Language (2022 · 238 citations)
The paper frames conversion between chemical structures and English as two new tasks. Molecule captioning produces a textual description from a SMILES string. Text-guided de novo molecule generation does the reverse, producing a SMILES string that fits a given description. The main obstacle is that paired… - Learning to Understand Phrases by Embedding the Dictionary (2016 · 236 citations)
Learning vector representations for whole phrases is hard, mainly because there is no obvious gold-standard target to train or evaluate them against. - Neural Machine Translation with Byte-Level Subwords (2019 · 228 citations)
Most NMT vocabularies are built from characters. Languages with very large character inventories, such as Japanese and Chinese, and noisy text therefore spend many vocabulary slots on rare symbols. Feeding raw bytes (256 symbols) avoids this but makes sequences too long to be practical. Wang, Cho and Gu (Facebook AI… - Unsupervised Question Decomposition for Question Answering (2020 · 222 citations)
Multi-hop questions are hard for QA systems because the evidence they need is spread across several passages, while simpler single-hop questions are handled well. - Don't Say That! Making Inconsistent Dialogue Unlikely with Unlikelihood Training (2020 · 217 citations)
Generative dialogue models trained only with maximum likelihood show several weaknesses. They copy too much from the context, repeat n-grams inside a single reply, overuse frequent tokens, and can contradict what was said earlier. The authors extend the unlikelihood loss of Welleck et al. (2019a) into a general… - Learning to Parse and Translate Improves Neural Machine Translation (2017 · 217 citations)
Most neural machine translation (NMT) systems use no linguistic knowledge. Earlier attempts to add syntax mostly targeted the source side and needed external parsers at test time. Eriguchi, Tsuruoka and Cho address this with a multi-task model that injects target-side syntax only during training. They propose… - Nematus: a Toolkit for Neural Machine Translation (2017 · 206 citations)
Nematus is an open-source toolkit for Neural Machine Translation (NMT) designed for strong translation quality, ease of use, and easy extension. - Improved Zero-shot Neural Machine Translation via Ignoring Spurious Correlations (2019 · 162 citations)
Multilingual NMT systems can translate between language pairs never seen together in training (zero-shot translation), but this ability is fragile. - SSMBA: Self-Supervised Manifold Based Data Augmentation for Improving Out-of-Domain Robustness (2020 · 161 citations)
Classifiers and sequence models trained on one domain often perform poorly on out-of-domain (OOD) inputs. For images, simple transformations like rotation keep new examples realistic, but no comparable meaning-preserving operations exist for text. The authors propose SSMBA, a task-agnostic augmentation scheme based… - Iterative Reasoning Preference Optimization (2024 · 153 citations)
Iterative preference-optimization methods such as Iterative DPO, Self-Rewarding LLMs and SPIN work well for general instruction following. - Insertion-based Decoding with Automatically Inferred Generation Order (2019 · 125 citations)
Standard autoregressive decoders write output strictly from left to right, a convention that may not suit every task. - Importance of Search and Evaluation Strategies in Neural Dialogue Modeling (2019 · 121 citations)
Neural dialogue research has mostly focused on architectures and training objectives. Decoding, meaning how a response is actually generated from a trained model, has received far less attention. This paper isolates the effect of the search algorithm. The authors keep a single attention-based recurrent… - Finding Generalizable Evidence by Learning to Convince Q&A Models (2019 · 43 citations)
The paper asks whether the passage sentences that best support an answer can be found automatically, and whether that evidence is useful beyond the model that produced it. - First Result on Arabic Neural Machine Translation (2016 · 40 citations)
This short 2016 paper reports the first fully neural machine translation results for Arabic, in both directions (Ar→En and En→Ar). - Jump to better conclusions: SCAN both left and right (2018 · 38 citations)
This BlackboxNLP 2018 paper re-examines SCAN, the command-to-action benchmark from Lake and Baroni (2018) that tests compositional generalization in recurrent sequence-to-sequence models. - Two Failures of Self-Consistency in the Multi-Step Reasoning of LLMs (2023 · 25 citations)
Evaluations of LLMs usually check whether answers are correct, not whether a model's answers agree with each other. - Following Length Constraints in Instructions (2025 · 22 citations)
Instruction-tuned LLMs and their evaluators tend to favor longer answers. Many prompts do not say how long the answer should be, which makes evaluation ambiguous and lets training exploit this length bias. Yuan et al. (Meta FAIR / NYU) suggest stating the desired length explicitly in the prompt as a word limit. They…
This page in Markdown → https://md.research.elamur.ai/free/en/authors/kyunghyun-cho/page/2/