Papers by Jun Suzuki

Improving Neural Machine Translation by Incorporating Hierarchical Subword Features (C18-1)

Machine Translation

Challenge:	Recent machine translation algorithms rely on parallel corpora, but only some resource-rich language pairs can benefit from them.
Approach:	They construct a parallel corpus for English-Japanese, which has 8.7 million sentence pairs . they use a web crawler to automatically align parallel sentences in the corpus .
Outcome:	The proposed corpus includes a broader range of domains and can be trained with a pre-trained model.

Copied to clipboard

STEP: Staged Parameter-Efficient Pre-training for Large Language Models (2025.naacl-short)

Challenge:	Using subwords, we find that the appropriate subword units for the three layers differ depending on the model . incorporating hierarchical subword features improves BLEU scores on the IWSLT evaluation datasets.
Approach:	They propose a method that expresses a word by combining "subwords" they propose to incorporate hierarchical subword features into a single embedding layer .
Outcome:	The proposed method improves BLEU scores on the IWSLT evaluation datasets.

Copied to clipboard

Kazuki Yano, Takumi Ito, Jun Suzuki

Filtering Noisy Dialogue Corpora by Connectivity and Content Relatedness (2020.emnlp-main)

Challenge:	Recent LLM development trends involve pre-training models with a vast number of parameters on massive datasets.
Approach:	They propose a method that integrates parameter-efficient tuning techniques with model growth to reduce memory requirements while maintaining equivalent performance.
Outcome:	The proposed method reduces memory requirements by 53.9% while maintaining equivalent performance to vanilla pre-trained models on downstream tasks.

Copied to clipboard

Reina Akama, Sho Yokoi, Jun Suzuki, Kentaro Inui

NTT Neural Machine Translation Systems at WAT 2019 (D19-52)

Challenge:	Large-scale dialogue datasets contain a non-negligible number of unacceptable utterance pairs . previous studies have identified such flaws and reported that the corpus is noisy .
Approach:	They propose a method for scoring the quality of utterance pairs based on their connectivity and relatedness.
Outcome:	The proposed method has a good correlation with human judgment of dialogue quality and is applied to training data filtered by the proposed method.

Copied to clipboard

Transductive Learning of Neural Language Models for Syntactic and Semantic Analysis (D19-1)

Multilinguality

Challenge:	We submitted two systems for scientific paper subtask and timely disclosure subtask . we evaluated the usefulness of incorporating external data from a wide variety of web pages to improve the translation quality.
Approach:	They describe two different translation tasks submitted to WAT 2019 . they submitted scientific paper subtasks and timely disclosure subtask .
Outcome:	The proposed system performed better on scientific paper and timely disclosure subtasks.

Copied to clipboard

Hiroki Ouchi, Jun Suzuki, Kentaro Inui

Detecting Response Generation Not Requiring Factual Judgment (2024.naacl-srw)

Challenge:	despite its practical advantages, transductive learning is underexplored in natural language processing . despite the simplicity of the technique, it is understudied in natural languages .
Approach:	They conduct an empirical study of transductive learning for neural models . they fine-tune language models on an unlabeled test set to obtain test-set-specific word representations.
Outcome:	The proposed method improves state-of-the-art neural models in syntactic and semantic tasks.

Copied to clipboard

Ryohei Kamei, Daiki Shiono, Reina Akama, Jun Suzuki

An Empirical Study of Incorporating Pseudo Data into Grammatical Error Correction (D19-1)

Challenge:	Large language models (LLMs) have undergone considerable development and can solve various natural language processing tasks.
Approach:	They aimed to achieve both attractiveness and factuality in a dialogue response by crowdsourcing a dataset and performing classification tasks on several models.
Outcome:	The proposed model with the highest classification accuracy could yield about 88% accurate classification results.

Copied to clipboard

Shun Kiyono, Jun Suzuki, Masato Mita, Tomoya Mizumoto, Kentaro Inui

Langsmith: An Interactive Academic Text Revision System (2020.emnlp-demos)

Challenge:	incorporating pseudo data in the training of grammatical error correction models has been a key factor in improving performance of such models.
Approach:	They investigate the choice of how pseudo data should be generated or used in a grammatical error correction model and show that the results are state-of-the-art.
Outcome:	The proposed method achieves state-of-the-art on the CoNLL-2014 test set and the official test set of the BEA-2019 shared task without making any modifications to the model architecture.

Copied to clipboard

Takumi Ito, Tatsuki Kuribayashi, Masatoshi Hidaka, Jun Suzuki, Kentaro Inui

Challenge:	Currently, diversity and inclusion initiatives in the academic community are encouraged . however, writing papers in English can be a daunting task .
Approach:	They propose a system that helps non-native English speakers to write papers in English . the system can suggest fluent, academic-style sentences based on their rough, incomplete phrases or sentences .
Outcome:	The proposed system can help non-native English speakers write papers in English . the system can suggest fluent, academic-style sentences based on their rough sentences .

An Empirical Study of Building a Strong Baseline for Constituency Parsing (P18-2)

Copied to clipboard

Jun Suzuki, Sho Takase, Hidetaka Kamigaito, Makoto Morishita, Masaaki Nagata

The Impact of Integration Step on Integrated Gradients (2024.eacl-srw)

Challenge:	Sequence-to-sequence models have been used for natural language generation tasks such as machine translation and summarization.
Approach:	They propose to build a strong baseline based on general purpose sequence-to-sequence models for constituency parsing.
Outcome:	The proposed model outperforms existing models in natural language generation tasks without any explicit task-specific knowledge or architecture of constituent parsing.

Copied to clipboard

Masahiro Makino, Yuya Asazuma, Shota Sasaki, Jun Suzuki

Instance-Based Learning of Span Representations: A Case Study through Named Entity Recognition (2020.acl-main)

Challenge:	Integrated gradients (IG) are a powerful tool for explaining the internal structure of a language model.
Approach:	They propose to customize the step count for each instance to minimize the error.
Outcome:	The optimal number of steps to maintain minimal error varies from instance to instance.

Copied to clipboard

Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi, Sho Yokoi, Tatsuki Kuribayashi, Ryuto Konno, Kentaro Inui

Responsible & Trustworthy NLP

Challenge:	Recent neural networks can induce good span feature representations and achieve high performance in structured prediction tasks.
Approach:	They propose an instance-based learning method that learns similarities between spans . they aim to build models that have high interpretability without sacrificing performance .
Outcome:	The proposed method improves interpretability without sacrificing performance.

A Challenging Multimodal Video Summary: Simultaneously Extracting and Generating Keyframe-Caption Pairs from Video (2023.emnlp-main)

Copied to clipboard

Keito Kudo, Haruki Nagasawa, Jun Suzuki, Nobuyuki Shimizu

Target-Guided Open-Domain Conversation Planning (2022.coling-1)

Challenge:	Existing methods to summarize video content have only considered video and image data, and the trend towards multimodal video summarization is changing.
Approach:	They propose a multimodal video summarization task setting and a dataset to train and evaluate the task.
Outcome:	The proposed task is useful as a practical application and presents a highly challenging problem worthy of study.

Copied to clipboard

Yosuke Kishinami, Reina Akama, Shiki Sato, Ryoko Tokuhisa, Jun Suzuki, Kentaro Inui

An Empirical Study of Span Representations in Argumentation Structure Parsing (P19-1)

Challenge:	Existing studies on goal-oriented conversational tasks lack planning . prior studies on this topic have focused on generating proactive behavior in agents .
Approach:	They propose a task to evaluate whether neural conversational agents have goal-oriented conversation planning abilities.
Outcome:	The proposed task evaluates whether neural conversational agents have goal-oriented conversation planning abilities.

Copied to clipboard

Tatsuki Kuribayashi, Hiroki Ouchi, Naoya Inoue, Paul Reisert, Toshinori Miyoshi, Jun Suzuki, Kentaro Inui

Representation Learning

Challenge:	Argumentation structure parsing (ASP) is a task of identifying argumentation structures in argumentative text.
Approach:	They propose to exploit neural network-based span representations for ASP to improve performance . they also propose task-dependent extensions for a parser that can be used to parse arguments .
Outcome:	The proposed model outperforms neural network-based approaches for argumentation structure parsing (ASP) it also provides some challenging types of instances to be parsed.

Select and Attend: Towards Controllable Content Selection in Text Generation (D19-1)

Copied to clipboard

Xiaoyu Shen, Jun Suzuki, Kentaro Inui, Hui Su, Dietrich Klakow, Satoshi Sekine

STEP: Staged Parameter-Efficient Pre-training for Large Language Models (2024.acl-srw)

Challenge:	Recent neural network models conflate content selection and surface realization into a black-box architecture, resulting in content to be described in text cannot be explicitly controlled.
Approach:	They propose to decouple content selection from the decoder to allow finer-grained control over the generation.
Outcome:	The proposed model can be trained end-to-end without human annotations and achieves promising results in data-totext and headline generation tasks.

Copied to clipboard

Kazuki Yano, Takumi Ito, Jun Suzuki

Encoder-Decoder Models Can Benefit from Pre-trained Masked Language Models in Grammatical Error Correction (2020.acl-main)

Challenge:	Existing methods for reducing computational costs during pre-training have been studied, but they often degrade performance under fair conditions.
Approach:	They propose a method that combines parameter-efficient tuning and staged training to reduce memory requirements while maintaining comparable performance.
Outcome:	The proposed method reduces memory requirements by 40.4% while maintaining comparable performance.

Copied to clipboard

Masahiro Kaneko, Masato Mita, Shun Kiyono, Jun Suzuki, Kentaro Inui

Evaluating Dialogue Generation Systems via Response Selection (2020.acl-main)

Challenge:	Existing methods for incorporating a masked language model into an EncDec model have potential drawbacks when applied to GEC.
Approach:	They propose to incorporate a pre-trained masked language model (MLM) into an encoder-decoder model for grammatical error correction.
Outcome:	The proposed method achieves state-of-the-art on BEA-2019 and CoNLL-2014 benchmarks.

Copied to clipboard

Shiki Sato, Reina Akama, Hiroki Ouchi, Jun Suzuki, Kentaro Inui

Predicting Fine-tuned Performance on Larger Datasets Before Creating Them (2025.coling-industry)

Challenge:	Existing automatic evaluation metrics for open-domain dialogue systems correlate poorly with human evaluation.
Approach:	They propose to construct response selection test sets with well-chosen false candidates to evaluate response generation systems via response selection.
Outcome:	The proposed method correlates with human evaluation better than widely used metrics such as BLEU.

Copied to clipboard

Toshiki Kuramoto, Jun Suzuki

A Large Collection of Model-generated Contradictory Responses for Consistency-aware Dialogue Systems (2024.findings-acl)

Challenge:	a method to estimate the performance of pretrained models fine-tuned with a larger dataset is proposed . a recent study found that fine-timing PMs with the small amount of data does not always result in ideal performance.
Approach:	They propose a method to estimate the performance of pretrained models fine-tuned with a larger dataset from the result with fewer epochs.
Outcome:	The proposed method can help resource-limited companies develop machine-learning models . it shows that when a model is fine-tuned with a larger dataset, its classification performance increases .

Copied to clipboard

Shiki Sato, Reina Akama, Jun Suzuki, Kentaro Inui

Preventing Critical Scoring Errors in Short Answer Scoring with Confidence Estimation (2020.acl-srw)

Challenge:	Recent large-scale neural response generation models (RGMs) have made significant progress but still struggle to generate semantically appropriate responses.
Approach:	They build a large dataset of model-generated contradictions for the first time and analyze the results to gain valuable insights into their characteristics.
Outcome:	The proposed dataset significantly improves the performance of data-driven contradiction suppression methods.

Copied to clipboard

Hiroaki Funayama, Shota Sasaki, Yuichiroh Matsubayashi, Tomoya Mizumoto, Jun Suzuki, Masato Mita, Kentaro Inui

Challenge:	Recent Short Answer Scoring systems use Quadratic Weighted Kappa (QWK) but it is unsatisfactory when measuring their effectiveness in actual usage.
Approach:	They propose a task formulation of Short Answer Scoring (SAS) that matches actual usage and extracts as many scoring predictions that are not critical scoring errors (CSEs).
Outcome:	The proposed system predicts scores with zero critical scoring errors (CSEs) for 50% of test data at maximum by filtering out low-reliability predictions on the basis of a certain confidence estimation.

Can Input Attributions Explain Inductive Reasoning in In-Context Learning? (2025.findings-acl)

Copied to clipboard

Mengyu Ye, Tatsuki Kuribayashi, Goro Kobayashi, Jun Suzuki

Responsible & Trustworthy NLP

Challenge:	interpreting the internal process of neural models has long been a challenge . despite rapid progress, there are still questions bridging the IA and MI eras .
Approach:	They propose to use input attribution methods to interpret in-context learning . they find that a certain simple IA method works best in large models .
Outcome:	The proposed method is the best for interpreting LLM-based ICL, but the larger the model, the harder it is to interpret it.

Enhancing Persuasive Dialogue Agents by Synthesizing Cross‐Disciplinary Communication Strategies (2025.emnlp-industry)

Copied to clipboard

Shinnosuke Nozue, Yuto Nakano, Yotaro Watanabe, Meguru Takasaki, Shoji Moriya, Reina Akama, Jun Suzuki

Context-aware Neural Machine Translation with Mini-batch Embedding (2021.eacl-main)

Challenge:	Current approaches to developing persuasive dialogue agents rely on predefined persuasive strategies that fail to capture the complexity of real-world interactions.
Approach:	They propose a framework for designing persuasive dialogue agents that draws on proven strategies from social psychology, behavioral economics, and communication theory.
Outcome:	The proposed framework demonstrated significant improvement in the persuasion success rate and generalizability of the datasets.

Copied to clipboard

Makoto Morishita, Jun Suzuki, Tomoharu Iwata, Masaaki Nagata

Representation Learning

Challenge:	Existing models that translate sentences in sentence-by-sentence manner do not consider the inter-sentent context in a language translation model.
Approach:	They propose to use mini-batch embedding to represent the features of sentences in a mini-Batch by choosing sentences from the same document and then embedd it into the model.
Outcome:	The proposed method outperforms baseline models and improves writing style or terminology to fit the document’s context.

Pruning Multilingual Large Language Models for Multilingual Inference (2024.findings-emnlp)

Copied to clipboard

Hwichan Kim, Jun Suzuki, Tosho Hirasawa, Mamoru Komachi

Direct Output Connection for a High-Rank Language Model (D18-1)

Challenge:	Multilingual large language models (MLLMs) demonstrate better zeroshot learning performance in non-English languages compared to large language model trained on English-dominant data.
Approach:	They propose a pruning approach to prune large language models using bilingual sentence pairs from English and other languages to enhance their performance in non-English language.
Outcome:	The proposed pruning strategy enhances the MLLMs’ performance in non-English language.

Copied to clipboard

Sho Takase, Jun Suzuki, Masaaki Nagata

PheMT: A Phenomenon-wise Dataset for Machine Translation Robustness on User-Generated Contents (2020.coling-main)

Challenge:	Neural network language models have played a central role in recent natural language processing advances.
Approach:	They propose a recurrent neural network language model that combines probability distributions computed from a final RNN layer and also middle layers.
Outcome:	The proposed language model achieves the best score on the Penn Treebank and WikiText-2 benchmark datasets.

Copied to clipboard

Ryo Fujii, Masato Mita, Kaori Abe, Kazuaki Hanawa, Makoto Morishita, Jun Suzuki, Kentaro Inui

Single Model Ensemble using Pseudo-Tags and Distinct Vectors (2020.acl-main)

Challenge:	Existing studies suggest that Neural Machine Translation still struggles with certain kinds of input with considerable noise, such as User-Generated Contents (UGC) on the Internet.
Approach:	They propose to evaluate the robustness of Neural Machine Translation models against specific linguistic phenomena in Japanese-English translation.
Outcome:	The proposed model can handle user-generated content (UGC) on the Internet, but it is difficult to translate clean inputs.

Copied to clipboard

Ryosuke Kuwabara, Jun Suzuki, Hideki Nakayama

Challenge:	Existing methods for model ensembles require time, memory, and management effort to perform tasks.
Approach:	They propose a method that replicates the effects of a model ensemble with a single model.
Outcome:	The proposed method emulates or outperforms a traditional model ensemble with 1/K-times fewer parameters on text classification and sequence labeling tasks.

Investigating the Effectiveness of Multiple Expert Models Collaboration (2023.findings-emnlp)

Copied to clipboard

Ikumi Ito, Takumi Ito, Jun Suzuki, Kentaro Inui

Machine Translation

Challenge:	Using multi-domain MT, we compare the performance of a single model with a multi-expert model in a fair condition.
Approach:	They propose to combine a multi-domain machine translation model with a aggregation strategy to investigate their results.
Outcome:	The proposed approach outperforms the current multi-domain model and aggregation methods in a fair condition on multiple domain datasets.

TimeMachine-bench: A Benchmark for Evaluating Model Capabilities in Repository-Level Migration Tasks (2026.eacl-long)

Copied to clipboard

Ryo Fujii, Makoto Morishita, Kazuki Yano, Jun Suzuki

Challenge:	Automated software engineering is a critical task of software engineers.
Approach:	They propose a benchmark to evaluate software migration in real-world Python projects.
Outcome:	The proposed benchmark consists of GitHub repositories whose tests fail in response to dependency updates.

Language Models as an Alternative Evaluator of Word Order Hypotheses: A Case Study in Japanese (2020.acl-main)

Copied to clipboard

Tatsuki Kuribayashi, Takumi Ito, Jun Suzuki, Kentaro Inui

B2T Connection: Serving Stability and Performance in Deep Transformers (2023.findings-acl)

Challenge:	a method using neural language models (LMs) for analyzing the word order of language is currently lacking.
Approach:	They propose a method using neural language models to analyze the word order in Japanese . they test whether there is a parallel between LMs and human word order preference .
Outcome:	The proposed method is validated by comparing it with other linguistic studies.

Copied to clipboard

Sho Takase, Shun Kiyono, Sosuke Kobayashi, Jun Suzuki

A Self-Refinement Strategy for Noise Reduction in Grammatical Error Correction (2020.findings-emnlp)

Challenge:	Existing methods to prevent the vanishing gradient problem in deep neural networks are not effective.
Approach:	They propose a method that can equip both higher stability and effective training by a simple modification from Post-LN.
Outcome:	The proposed method outperforms Pre-LN and Post-Ln on a wide range of tasks.

Copied to clipboard

Masato Mita, Shun Kiyono, Masahiro Kaneko, Jun Suzuki, Kentaro Inui

Hunt for Buried Treasures: Extracting Unclaimed Embodiments from Patent Specifications (2023.acl-industry)

Challenge:	Existing approaches for grammatical error correction (GEC) rely on supervised learning with manually created datasets.
Approach:	They propose to denoise GEC datasets by leveraging prediction consistency of existing models.
Outcome:	The proposed method outperforms baseline methods on CoNLL-2014, JFLEG, and BEA-2019 benchmarks.

Copied to clipboard

Chikara Hashimoto, Gautam Kumar, Shuichiro Hashimoto, Jun Suzuki

Information Extraction & Text Mining

Challenge:	Unclaimed embodiment extraction is performed manually and little research has been conducted on its automation.
Approach:	They propose a task of unclaimed embodiment extraction and a dataset for the task . they use a natural language inference task to extract unclaimed inventions .
Outcome:	The proposed task requires performing natural language inference on patent specifications.

Scene-Text Aware Image and Text Retrieval with Dual-Encoder (2022.acl-srw)

Copied to clipboard

Shumpei Miyawaki, Taku Hasegawa, Kyosuke Nishida, Takuma Kato, Jun Suzuki

Multimodality

Challenge:	Existing studies on image and text retrieval using a dual-encoder model have not shown their effectiveness for fast inferences.
Approach:	They propose a dual-encoder model that connects vision and language in the same semantic space and integrates scene-text and visual information into a model.
Outcome:	The proposed model can interpret scene-text and surrounding visual information better than cross-encoder models.

Instance-Based Neural Dependency Parsing (2021.tacl-1)

Copied to clipboard

Hiroki Ouchi, Jun Suzuki, Sosuke Kobayashi, Sho Yokoi, Tatsuki Kuribayashi, Masashi Yoshikawa, Kentaro Inui

Pointwise HSIC: A Linear-Time Kernelized Co-occurrence Norm for Sparse Linguistic Expressions (D18-1)

Challenge:	Existing models that use instance-based inference for dependency parsing are difficult to understand for humans.
Approach:	They develop neural models that adopt an interpretable inference process for dependency parsing.
Outcome:	The proposed models achieve competitive accuracy with standard neural models and have plausibility of instance-based explanations.

Copied to clipboard

Sho Yokoi, Sosuke Kobayashi, Kenji Fukumizu, Jun Suzuki, Kentaro Inui

Challenge:	Empirically, PHSIC is learned thousands of times faster than an RNN-based PMI while outperforming PMI in accuracy.
Approach:	They propose a new kernel-based co-occurrence measure that can be applied to sparse linguistic expressions with a very short learning time.
Outcome:	The proposed measure can be applied to sparse linguistic expressions with a very short learning time, and is called the pointwise HSIC.

Subword-based Compact Reconstruction of Word Embeddings (N19-1)

Copied to clipboard

Shota Sasaki, Jun Suzuki, Kentaro Inui

TEASPN: Framework and Protocol for Integrated Writing Assistance Environments (D19-3)

Challenge:	Existing word-based word embeddings are based on subword information and memory-shared embeddables.
Approach:	They propose a method for reconstructing pre-trained word embeddings using subword information using memory-shared embedds and a variant of the key-value-query self-attention mechanism.
Outcome:	The proposed method can imitate well-trained word embeddings in a small fixed space while preventing quality degradation across several linguistic benchmark datasets.

Copied to clipboard

Masato Hagiwara, Takumi Ito, Tatsuki Kuribayashi, Jun Suzuki, Kentaro Inui

Assessing Step-by-Step Reasoning against Lexical Negation: A Case Study on Syllogism (2023.emnlp-main)

Challenge:	TEASPN is an open-source protocol for integrated writing assistance environments . authors propose that developers and researchers can integrate the latest developments in natural language processing with low cost.
Approach:	They propose a protocol and framework for integrating writing aids with writing software.
Outcome:	The proposed protocol standardizes the way writing software communicates with servers that implement such technologies, allowing developers and researchers to integrate the latest developments in natural language processing (NLP) with low cost.

Copied to clipboard

Mengyu Ye, Tatsuki Kuribayashi, Jun Suzuki, Goro Kobayashi, Hiroaki Funayama

Robustness in NLP

Challenge:	Large language models (LLMs) take advantage of step-by-step reasoning instructions . negation is a core linguistic phenomenon that is difficult to process .
Approach:	They examine the step-by-step reasoning ability of large language models with a focus on negation . negation is a core linguistic phenomenon that is difficult to process .
Outcome:	The proposed models perform better when using chain-of-thought prompting . the results highlight unique limitations in each LLM family .

Word Rotator’s Distance (2020.emnlp-main)

Copied to clipboard

Sho Yokoi, Ryo Takahashi, Reina Akama, Jun Suzuki, Kentaro Inui

Representation Learning

Challenge:	Existing approaches to measure textual similarity are inconsistent with the word alignment and are empirically inferior to the simple cosine similarity between general-purpose sentence vectors.
Approach:	They propose to decouple word vectors into their norm and direction and then grow the norm and directions of word vector.
Outcome:	The proposed methods outperform alignment-based approaches on several benchmarks and strong baselines on the semantic textual similarity task.

Embeddings of Label Components for Sequence Labeling: A Case Study of Fine-grained Named Entity Recognition (2020.acl-srw)

Copied to clipboard

Takuma Kato, Kaori Abe, Hiroki Ouchi, Shumpei Miyawaki, Jun Suzuki, Kentaro Inui

Domain Adaptation of Machine Translation with Crowdworkers (2022.emnlp-industry)

Challenge:	In general, the labels used in sequence labeling consist of different types of elements.
Approach:	They propose to integrate label component information as embeddings into sequence labeling models.
Outcome:	The proposed method improves on English and Japanese fine-grained named entity recognition on low-frequency labels.

Copied to clipboard

Evaluating Model Alignment with Human Perception: A Study on Shitsukan in LLMs and LVLMs (2025.coling-main)

Challenge:	Existing methods for machine translation work well when no in-domain data are available.
Approach:	They propose a framework that efficiently and effectively collects parallel sentences in a target domain from the web with the help of crowdworkers.
Outcome:	The proposed method improves the BLEU scores by an average of +7.8 points compared to a general-purpose translation model.

Copied to clipboard

Daiki Shiono, Ana Brassard, Yukiko Ishizuki, Jun Suzuki

JParaCrawl v3.0: A Large-scale English-Japanese Parallel Corpus (2022.lrec-1)

Challenge:	This work examines the alignment of large language models and large vision-language models with human perception.
Approach:	They use a dataset of shitsukan terms elicited from individuals in response to object images to evaluate their understanding of the Japanese concept of shitukan.
Outcome:	The proposed models demonstrated mixed accuracy across benchmark tasks, with limited overlap between model- and human-generated terms.

Copied to clipboard

Makoto Morishita, Katsuki Chousa, Jun Suzuki, Masaaki Nagata

Machine Translation

Challenge:	Existing parallel corpora for English-Japanese are limited, limiting the accuracy of machine translation models.
Approach:	They propose a web-based English-Japanese parallel corpus with 21 million unique sentence pairs . this is more than twice as many as the previous corpus JParaCrawl v2.0 .
Outcome:	The proposed corpus boosts the accuracy of machine translation models on various domains.

Bipartite-play Dialogue Collection for Practical Automatic Evaluation of Dialogue Systems (2022.aacl-srw)

Copied to clipboard

Shiki Sato, Yosuke Kishinami, Hiroaki Sugiyama, Reina Akama, Ryoko Tokuhisa, Jun Suzuki

Natural Language Interfaces

Challenge:	Existing methods for dialogue system evaluation are inefficient and time-consuming.
Approach:	They propose a dialogue collection method for automating dialogue system evaluation using bipartite-play method . authors propose constructing a better automatic evaluation method which is reproducible and low cost .
Outcome:	The proposed method correlates strongly with human subjectivity and human evaluation.

SHAPE: Shifted Absolute Position Embedding for Transformers (2021.emnlp-main)

Copied to clipboard

Shun Kiyono, Sosuke Kobayashi, Jun Suzuki, Kentaro Inui

Effective Adversarial Regularization for Neural Machine Translation (P19-1)

Challenge:	Existing position representations suffer from a lack of generalization to test data with unseen lengths or high computational cost.
Approach:	They propose to achieve shift invariance by randomly shifting absolute positions during training by a SHAPE algorithm that is empirically comparable to its counterpart.
Outcome:	The proposed method outperforms existing representations on sequence-to-sequence tasks due to extrapolation, i.e., the ability to generalize to sequences that are longer than those observed during training.

Copied to clipboard

Motoki Sato, Jun Suzuki, Shun Kiyono

Robustness in NLP

Challenge:	Existing (small) perturbations that induce a critical prediction error in machine learning models are often referred to as adversarial examples.
Approach:	They propose to use adversarial perturbations to regularize text classification tasks by adding adversarials to a typical NMT model structure.
Outcome:	The proposed method significantly improves performance of NMT models, such as LSTM-based and Transformer-based models.

Suppressing Final Layer Hidden State Jumps in Transformer Pretraining (2026.findings-eacl)

Copied to clipboard

Keigo Shibata, Kazuki Yano, Ryosuke Takahashi, Jaesung Lee, Wataru Ikeda, Jun Suzuki