Yusuke Miyao

Hong Chen, Duc Minh Vo, Hiroya Takamura et al.

Existing automatic story evaluation methods place a premium on story lexical level coherence, deviating from human preference. We go beyond this limitation by considering a novel \textbf{Story} \textbf{E}valuation method that mimics human preference when judging a story, namely \textbf{StoryER}, which consists of three sub-tasks: \textbf{R}anking, \textbf{R}ating and \textbf{R}easoning. Given either a machine-generated or a human-written story, StoryER requires the machine to output 1) a preference score that corresponds to human preference, 2) specific ratings and their corresponding confidences and 3) comments for various aspects (e.g., opening, character-shaping). To support these tasks, we introduce a well-annotated dataset comprising (i) 100k ranked story pairs; and (ii) a set of 46k ratings and comments on various aspects of the story. We finetune Longformer-Encoder-Decoder (LED) on the collected dataset, with the encoder responsible for preference score and aspect prediction and the decoder for comment generation. Our comprehensive experiments result in a competitive benchmark for each task, showing the high correlation to human preference. In addition, we have witnessed the joint learning of the preference scores, the aspect ratings, and the comments brings gain in each single task. Our dataset and benchmarks are publicly available to advance the research of story evaluation tasks.\footnote{Dataset and pre-trained model demo are available at anonymous website \url{http://storytelling-lab.com/eval} and \url{https://github.com/sairin1202/StoryER}}

LLM-jp, Akiko Aizawa, Eiji Aramaki et al.

This paper introduces LLM-jp, a cross-organizational project for the research and development of Japanese large language models (LLMs). LLM-jp aims to develop open-source and strong Japanese LLMs, and as of this writing, more than 1,500 participants from academia and industry are working together for this purpose. This paper presents the background of the establishment of LLM-jp, summaries of its activities, and technical reports on the LLMs developed by LLM-jp. For the latest activities, visit https://llm-jp.nii.ac.jp/en/.

Shunsuke Kando, Yusuke Miyao, Jason Naradowsky et al.

Traditional spoken language processing involves cascading an automatic speech recognition (ASR) system into text processing models. In contrast, "textless" methods process speech representations without ASR systems, enabling the direct use of acoustic speech features. Although their effectiveness is shown in capturing acoustic features, it is unclear in capturing lexical knowledge. This paper proposes a textless method for dependency parsing, examining its effectiveness and limitations. Our proposed method predicts a dependency tree from a speech signal without transcribing, representing the tree as a labeled sequence. scading method outperforms the textless method in overall parsing accuracy, the latter excels in instances with important acoustic features. Our findings highlight the importance of fusing word-level representations and sentence-level prosody for enhanced parsing performance. The code and models are made publicly available: https://github.com/mynlp/SpeechParser.

Erica K. Shimomoto, Edison Marrese-Taylor, Hiroya Takamura et al.

This paper explores the task of Temporal Video Grounding (TVG) where, given an untrimmed video and a natural language sentence query, the goal is to recognize and determine temporal boundaries of action instances in the video described by the query. Recent works tackled this task by improving query inputs with large pre-trained language models (PLM) at the cost of more expensive training. However, the effects of this integration are unclear, as these works also propose improvements in the visual inputs. Therefore, this paper studies the effects of PLMs in TVG and assesses the applicability of parameter-efficient training with NLP adapters. We couple popular PLMs with a selection of existing approaches and test different adapters to reduce the impact of the additional parameters. Our results on three challenging datasets show that, without changing the visual inputs, TVG models greatly benefited from the PLM integration and fine-tuning, stressing the importance of sentence query representation in this task. Furthermore, NLP adapters were an effective alternative to full fine-tuning, even though they were not tailored to our task, allowing PLM integration in larger TVG models and delivering results comparable to SOTA models. Finally, our results shed light on which adapters work best in different scenarios.

Qiang Zhang, Jason Naradowsky, Yusuke Miyao

We introduce the task of implicit offensive text detection in dialogues, where a statement may have either an offensive or non-offensive interpretation, depending on the listener and context. We argue that reasoning is crucial for understanding this broader class of offensive utterances and release SLIGHT, a dataset to support research on this task. Experiments using the data show that state-of-the-art methods of offense detection perform poorly when asked to detect implicitly offensive statements, achieving only ${\sim} 11\%$ accuracy. In contrast to existing offensive text detection datasets, SLIGHT features human-annotated chains of reasoning which describe the mental process by which an offensive interpretation can be reached from each ambiguous statement. We explore the potential for a multi-hop reasoning approach by utilizing existing entailment models to score the probability of these chains and show that even naive reasoning models can yield improved performance in most situations. Furthermore, analysis of the chains provides insight into the human interpretation process and emphasizes the importance of incorporating additional commonsense knowledge.

Shunsuke Kando, Hiroshi Noji, Yusuke Miyao

Incorporating stronger syntactic biases into neural language models (LMs) is a long-standing goal, but research in this area often focuses on modeling English text, where constituent treebanks are readily available. Extending constituent tree-based LMs to the multilingual setting, where dependency treebanks are more common, is possible via dependency-to-constituency conversion methods. However, this raises the question of which tree formats are best for learning the model, and for which languages. We investigate this question by training recurrent neural network grammars (RNNGs) using various conversion methods, and evaluating them empirically in a multilingual setting. We examine the effect on LM performance across nine conversion methods and five languages through seven types of syntactic tests. On average, the performance of our best model represents a 19 \% increase in accuracy over the worst choice across all languages. Our best model shows the advantage over sequential/overparameterized LMs, suggesting the positive effect of syntax injection in a multilingual setting. Our experiments highlight the importance of choosing the right tree formalism, and provide insights into making an informed decision.

Kyosuke Takami, Yuka Tateisi, Satoshi Sekine et al.

Authentic school examinations provide a high-validity test bed for evaluating multimodal large language models (MLLMs), yet benchmarks grounded in Japanese K-12 assessments remain scarce. We present a multimodal dataset constructed from Japan's National Assessment of Academic Ability, comprising officially released middle-school items in Science, Mathematics, and Japanese Language. Unlike existing benchmarks based on synthetic or curated data, our dataset preserves real exam layouts, diagrams, and Japanese educational text, together with nationwide aggregated student response distributions (N $\approx$ 900{,}000). These features enable direct comparison between human and model performance under a unified evaluation framework. We benchmark recent multimodal LLMs using exact-match accuracy and character-level F1 for open-ended responses, observing substantial variation across subjects and strong sensitivity to visual reasoning demands. Human evaluation and LLM-as-judge analyses further assess the reliability of automatic scoring. Our dataset establishes a reproducible, human-grounded benchmark for multimodal educational reasoning and supports future research on evaluation, feedback generation, and explainable AI in authentic assessment contexts. Our dataset is available at: https://github.com/KyosukeTakami/gakucho-benchmark

Qiang Zhang, Jason Naradowsky, Yusuke Miyao

Knowing how to end and resume conversations over time is a natural part of communication, allowing for discussions to span weeks, months, or years. The duration of gaps between conversations dictates which topics are relevant and which questions to ask, and dialogue systems which do not explicitly model time may generate responses that are unnatural. In this work we explore the idea of making dialogue models aware of time, and present GapChat, a multi-session dialogue dataset in which the time between each session varies. While the dataset is constructed in real-time, progress on events in speakers' lives is simulated in order to create realistic dialogues occurring across a long timespan. We expose time information to the model and compare different representations of time and event progress. In human evaluation we show that time-aware models perform better in metrics that judge the relevance of the chosen topics and the information gained from the conversation.

Qiang Zhang, Jason Naradowsky, Yusuke Miyao

When engaging in conversations, dialogue agents in a virtual simulation environment may exhibit their own emotional states that are unrelated to the immediate conversational context, a phenomenon known as self-emotion. This study explores how such self-emotion affects the agents' behaviors in dialogue strategies and decision-making within a large language model (LLM)-driven simulation framework. In a dialogue strategy prediction experiment, we analyze the dialogue strategy choices employed by agents both with and without self-emotion, comparing them to those of humans. The results show that incorporating self-emotion helps agents exhibit more human-like dialogue strategies. In an independent experiment comparing the performance of models fine-tuned on GPT-4 generated dialogue datasets, we demonstrate that self-emotion can lead to better overall naturalness and humanness. Finally, in a virtual simulation environment where agents have discussions on multiple topics, we show that self-emotion of agents can significantly influence the decision-making process of the agents, leading to approximately a 50% change in decisions.

Chung-Chi Chen, Hiroya Takamura, Ichiro Kobayashi et al.

This paper investigates the role of expert-designed hint in enhancing sentiment analysis on financial social media posts. We explore the capability of large language models (LLMs) to empathize with writer perspectives and analyze sentiments. Our findings reveal that expert-designed hint, i.e., pointing out the importance of numbers, significantly improve performances across various LLMs, particularly in cases requiring perspective-taking skills. Further analysis on tweets containing different types of numerical data demonstrates that the inclusion of expert-designed hint leads to notable improvements in sentiment analysis performance, especially for tweets with monetary-related numbers. Our findings contribute to the ongoing discussion on the applicability of Theory of Mind in NLP and open new avenues for improving sentiment analysis in financial domains through the strategic use of expert knowledge.

Chung-Chi Chen, Hen-Hsen Huang, Hsin-Hsi Chen et al.

In the era of rapid Internet and social media platform development, individuals readily share their viewpoints online. The overwhelming quantity of these posts renders comprehensive analysis impractical. This necessitates an efficient recommendation system to filter and present significant, relevant opinions. Our research introduces a dual-pronged argument mining technique to improve recommendation system effectiveness, considering both professional and amateur investor perspectives. Our first strategy involves using the discrepancy between target and closing prices as an opinion indicator. The second strategy applies argument mining principles to score investors' opinions, subsequently ranking them by these scores. Experimental results confirm the effectiveness of our approach, demonstrating its ability to identify opinions with higher profit potential. Beyond profitability, our research extends to risk analysis, examining the relationship between recommended opinions and investor behaviors. This offers a holistic view of potential outcomes following the adoption of these recommended opinions.

Yuto Harada, Yusuke Yamauchi, Yusuke Oda et al.

Supervised fine-tuning (SFT) is a critical step in aligning large language models (LLMs) with human instructions and values, yet many aspects of SFT remain poorly understood. We trained a wide range of base models on a variety of datasets including code generation, mathematical reasoning, and general-domain tasks, resulting in 1,000+ SFT models under controlled conditions. We then identified the dataset properties that matter most and examined the layer-wise modifications introduced by SFT. Our findings reveal that some training-task synergies persist across all models while others vary substantially, emphasizing the importance of model-specific strategies. Moreover, we demonstrate that perplexity consistently predicts SFT effectiveness, often surpassing superficial similarity between the training data and the benchmark, and that mid-layer weight changes correlate most strongly with performance gains. We release these 1,000+ SFT models and benchmark results to accelerate further research. All resources are available at https://github.com/llm-jp/massive-sft.

Bowen Chen, Namgi Han, Yusuke Miyao

Memorization is a fundamental component of intelligence for both humans and LLMs. However, while LLM performance scales rapidly, our understanding of memorization lags. Due to limited access to the pre-training data of LLMs, most previous studies focus on a single model series, leading to isolated observations among series, making it unclear which findings are general or specific. In this study, we collect multiple model series (Pythia, OpenLLaMa, StarCoder, OLMo1/2/3) and analyze their shared or unique memorization behavior at both the statistical and internal levels, connecting individual observations while showing new findings. At the statistical level, we reveal that the memorization rate scales log-linearly with model size, and memorized sequences can be further compressed. Further analysis demonstrated a shared frequency and domain distribution pattern for memorized sequences. However, different models also show individual features under the above observations. At the internal level, we find that LLMs can remove certain injected perturbations, while memorized sequences are more sensitive. By decoding middle layers and attention head ablation, we revealed the general decoding process and shared important heads for memorization. However, the distribution of those important heads differs between families, showing a unique family-level feature. Through bridging various experiments and revealing new findings, this study paves the way for a universal and fundamental understanding of memorization in LLM.

Taisei Katô, Yusuke Miyao

We examine the abilities of intrinsic bias metrics of static word embeddings to predict whether Natural Language Processing (NLP) systems exhibit biased behavior. A word embedding is one of the fundamental NLP technologies that represents the meanings of words through real vectors, and problematically, it also learns social biases such as stereotypes. An intrinsic bias metric measures bias by examining a characteristic of vectors, while an extrinsic bias metric checks whether an NLP system trained with a word embedding is biased. A previous study found that a common intrinsic bias metric usually does not correlate with extrinsic bias metrics. However, the intrinsic and extrinsic bias metrics did not measure the same bias in most cases, which makes us question whether the lack of correlation is genuine. In this paper, we extract characteristic words from datasets of extrinsic bias metrics and analyze correlations with intrinsic bias metrics with those words to ensure both metrics measure the same bias. We observed moderate to high correlations with some extrinsic bias metrics but little to no correlations with the others. This result suggests that intrinsic bias metrics can predict biased behavior in particular settings but not in others. Experiment codes are available at GitHub.

Bowen Chen, Namgi Han, Yusuke Miyao

Large Language Models (LLMs), trained on massive corpora with billions of parameters, show unprecedented performance in various fields. Though surprised by their excellent performances, researchers also noticed some special behaviors of those LLMs. One of those behaviors is memorization, in which LLMs can generate the same content used to train them. Though previous research has discussed memorization, the memorization of LLMs still lacks explanation, especially the cause of memorization and the dynamics of generating them. In this research, we comprehensively discussed memorization from various perspectives and extended the discussion scope to not only just the memorized content but also less and unmemorized content. Through various studies, we found that: (1) Through experiments, we revealed the relation of memorization between model size, continuation size, and context size. Further, we showed how unmemorized sentences transition to memorized sentences. (2) Through embedding analysis, we showed the distribution and decoding dynamics across model size in embedding space for sentences with different memorization scores. The n-gram statistics analysis presents d (3) An analysis over n-gram and entropy decoding dynamics discovered a boundary effect when the model starts to generate memorized sentences or unmemorized sentences. (4)We trained a Transformer model to predict the memorization of different models, showing that it is possible to predict memorizations by context.

Tatsuro Inaba, Go Kamoda, Kentaro Inui et al.

This study explores how bilingual language models develop complex internal representations. We employ sparse autoencoders to analyze internal representations of bilingual language models with a focus on the effects of training steps, layers, and model sizes. Our analysis shows that language models first learn languages separately, and then gradually form bilingual alignments, particularly in the mid layers. We also found that this bilingual tendency is stronger in larger models. Building on these findings, we demonstrate the critical role of bilingual representations in model performance by employing a novel method that integrates decomposed representations from a fully trained model into a mid-training model. Our results provide insights into how language models acquire bilingual capabilities.

Mutsumi Sasaki, Kouta Nakayama, Yusuke Miyao et al.

When introducing Large Language Models (LLMs) into industrial applications, such as healthcare and education, the risk of generating harmful content becomes a significant challenge. While existing machine unlearning methods can erase specific harmful knowledge and expressions, diverse harmful content makes comprehensive removal difficult. In this study, instead of individually listing targets for forgetting, we propose Exclusive Unlearning (EU), which aims for broad harm removal by extensively forgetting everything except for the knowledge and expressions we wish to retain. We demonstrate that through Exclusive Unlearning, it is possible to obtain a model that ensures safety against a wide range of inputs, including jailbreaks, while maintaining the ability to respond to diverse instructions related to specific domains such as medicine and mathematics.

Shunsuke Kando, Yusuke Miyao, Shinnosuke Takamichi

The purpose of speech tokenization is to transform a speech signal into a sequence of discrete representations, serving as the foundation for speech language models (SLMs). While speech tokenization has many options, their effect on the performance of SLMs remains unclear. This paper investigates two key aspects of speech tokenization: the segmentation width and the cluster size of discrete units. First, we segment speech signals into fixed/variable widths and pooled representations. We then train K-means models in multiple cluster sizes. Through the evaluation on zero-shot spoken language understanding benchmarks, we find the positive effect of moderately coarse segmentation and bigger cluster size. Notably, among the best-performing models, the most efficient one achieves a 50% reduction in training data and a 70% decrease in training runtime. Our analysis highlights the importance of combining multiple tokens to enhance fine-grained spoken language understanding.

Junjie Chen, Xiangheng He, Danushka Bollegala et al.

Unsupervised constituency parsing focuses on identifying word sequences that form a syntactic unit (i.e., constituents) in target sentences. Linguists identify the constituent by evaluating a set of Predicate-Argument Structure (PAS) equivalent sentences where we find the constituent appears more frequently than non-constituents (i.e., the constituent corresponds to a frequent word sequence within the sentence set). However, such frequency information is unavailable in previous parsing methods that identify the constituent by observing sentences with diverse PAS. In this study, we empirically show that constituents correspond to frequent word sequences in the PAS-equivalent sentence set. We propose a frequency-based parser span-overlap that (1) computes the span-overlap score as the word sequence's frequency in the PAS-equivalent sentence set and (2) identifies the constituent structure by finding a constituent tree with the maximum span-overlap score. The parser achieves state-of-the-art level parsing accuracy, outperforming existing unsupervised parsers in eight out of ten languages. Additionally, we discover a multilingual phenomenon: participant-denoting constituents tend to have higher span-overlap scores than equal-length event-denoting constituents, meaning that the former tend to appear more frequently in the PAS-equivalent sentence set than the latter. The phenomenon indicates a statistical difference between the two constituent types, laying the foundation for future labeled unsupervised parsing research.

Bowen Chen, Namgi Han, Yusuke Miyao

The lack of data transparency in Large Language Models (LLMs) has highlighted the importance of Membership Inference Attack (MIA), which differentiates trained (member) and untrained (non-member) data. Though it shows success in previous studies, recent research reported a near-random performance in different settings, highlighting a significant performance inconsistency. We assume that a single setting doesn't represent the distribution of the vast corpora, causing members and non-members with different distributions to be sampled and causing inconsistency. In this study, instead of a single setting, we statistically revisit MIA methods from various settings with thousands of experiments for each MIA method, along with study in text feature, embedding, threshold decision, and decoding dynamics of members and non-members. We found that (1) MIA performance improves with model size and varies with domains, while most methods do not statistically outperform baselines, (2) Though MIA performance is generally low, a notable amount of differentiable member and non-member outliers exists and vary across MIA methods, (3) Deciding a threshold to separate members and non-members is an overlooked challenge, (4) Text dissimilarity and long text benefit MIA performance, (5) Differentiable or not is reflected in the LLM embedding, (6) Member and non-members show different decoding dynamics.

Bolei Ma, Yusuke Miyao

Do Large Language Models (LLMs) genuinely grasp the compositional semantics of events, or do they rely on surface-level probabilistic heuristics? We investigate the Imperfective Paradox, a logical phenomenon where the past progressive aspect entails event realization for activities (e.g., running $\to$ ran) but not for accomplishments (e.g., building $\nrightarrow$ built). We introduce ImperfectiveNLI, a diagnostic dataset designed to probe this distinction across diverse semantic classes. Evaluating state-of-the-art open-weight models, we uncover a pervasive Teleological Bias: models systematically hallucinate completion for goal-oriented events, often overriding explicit textual negation. Representational analyses show that while internal embeddings often distinguish process from result, inference decisions are dominated by strong priors about goal attainment. We further find that prompting-based interventions reduce hallucinated completions but also increase incorrect rejections of valid entailments. Our findings suggest that current LLMs lack structural aspectual awareness, operating as predictive narrative engines rather than faithful logical reasoners.

Zhihan Cao, Fumihito Nishino, Hiroaki Yamada et al.

We introduce JBE-QA, a Japanese Bar Exam Question-Answering dataset to evaluate large language models' legal knowledge. Derived from the multiple-choice (tanto-shiki) section of the Japanese bar exam (2015-2024), JBE-QA provides the first comprehensive benchmark for Japanese legal-domain evaluation of LLMs. It covers the Civil Code, the Penal Code, and the Constitution, extending beyond the Civil Code focus of prior Japanese resources. Each question is decomposed into independent true/false judgments with structured contextual fields. The dataset contains 3,464 items with balanced labels. We evaluate 26 LLMs, including proprietary, open-weight, Japanese-specialised, and reasoning models. Our results show that proprietary models with reasoning enabled perform best, and the Constitution questions are generally easier than the Civil Code or the Penal Code questions.

Yurie Koga, Shunsuke Kando, Yusuke Miyao

This paper investigates whether the Critical Period (CP) effects in human language acquisition are observed in self-supervised speech models (S3Ms). CP effects refer to greater difficulty in acquiring a second language (L2) with delayed L2 exposure onset, and greater retention of their first language (L1) with delayed L1 exposure offset. While previous work has studied these effects using textual language models, their presence in speech models remains underexplored despite the central role of spoken language in human language acquisition. We train S3Ms with varying L2 training onsets and L1 training offsets on child-directed speech and evaluate their phone discrimination performance. We find that S3Ms do not exhibit clear evidence of either CP effects in terms of phonological acquisition. Notably, models with delayed L2 exposure onset tend to perform better on L2 and delayed L1 exposure offset leads to L1 forgetting.

Romain Harang, Jason Naradowsky, Yaswitha Gujju et al.

Large Language Models (LLMs) exhibit emergent capabilities in structured domains, suggesting they may implicitly internalize high-fidelity representations of world models. While probing techniques have shown promising signs of this in scientific and game-based settings, they rely on model-specific internal activations, which limit interpretability and generalizability. In this work, we propose a model-agnostic, state-based evaluation framework using chess as a benchmark to assess whether LLMs preserve the semantics of structured environments. Our method analyzes the downstream legal move distributions (state affordances) to estimate semantic fidelity between predicted and actual game states. This approach offers a more meaningful evaluation than conventional string-based metrics by aligning more closely with the strategic and rule-governed nature of chess. Experimental results demonstrate that our metrics capture deficiencies in state-tracking, highlighting limitations of LLMs in maintaining coherent internal models over long sequences. Our framework provides a robust tool for evaluating structured reasoning in LLMs without requiring internal model access, and generalizes to a wide class of symbolic environments.

Daichi Kato, Ryo Ueda, Yusuke Miyao

If humans understood language by randomly selecting parsing actions, it might have been necessary to construct a robust symbolic system capable of being interpreted under any hierarchical structure. However, human parsing strategies do not seem to follow such a random pattern. Why is that the case? In fact, a previous study on emergent communication using models with hierarchical biases have reported that agents adopting random parsing strategies$\unicode{x2013}$ones that deviate significantly from human language comprehension$\unicode{x2013}$can achieve high communication accuracy. In this study, we investigate this issue by making two simple and natural modifications to the experimental setup: (I) we use more complex inputs that have hierarchical structures, such that random parsing makes semantic interpretation more difficult, and (II) we incorporate a surprisal-related term, which is known to influence the order of words and characters in natural language, into the objective function. With these changes, we evaluate whether agents employing random parsing strategies still maintain high communication accuracy.

Ha-Thanh Nguyen, Chaoran Liu, Qianying Liu et al.

We present BIS Reasoning 1.0, the first large-scale Japanese dataset of syllogistic reasoning problems explicitly designed to evaluate belief-inconsistent reasoning in large language models (LLMs). Unlike prior datasets such as NeuBAROCO and JFLD, which focus on general or belief-aligned reasoning, BIS Reasoning 1.0 introduces logically valid yet belief-inconsistent syllogisms to uncover reasoning biases in LLMs trained on human-aligned corpora. We benchmark state-of-the-art models - including GPT models, Claude models, and leading Japanese LLMs - revealing significant variance in performance, with GPT-4o achieving 79.54% accuracy. Our analysis identifies critical weaknesses in current LLMs when handling logically valid but belief-conflicting inputs. These findings have important implications for deploying LLMs in high-stakes domains such as law, healthcare, and scientific literature, where truth must override intuitive belief to ensure integrity and safety.

Kunhang Li, Jason Naradowsky, Yansong Feng et al.

We explore the human motion knowledge of Large Language Models (LLMs) through 3D avatar control. Given a motion instruction, we prompt LLMs to first generate a high-level movement plan with consecutive steps (High-level Planning), then specify body part positions in each step (Low-level Planning), which we linearly interpolate into avatar animations. Using 20 representative motion instructions that cover fundamental movements and balance body part usage, we conduct comprehensive evaluations, including human and automatic scoring of both high-level movement plans and generated animations, as well as automatic comparison with oracle positions in low-level planning. Our findings show that LLMs are strong at interpreting high-level body movements but struggle with precise body part positioning. While decomposing motion queries into atomic components improves planning, LLMs face challenges in multi-step movements involving high-degree-of-freedom body parts. Furthermore, LLMs provide reasonable approximations for general spatial descriptions, but fall short in handling precise spatial specifications. Notably, LLMs demonstrate promise in conceptualizing creative motions and distinguishing culturally specific motion patterns.

Shintaro Ozaki, Tatsuya Hiraoka, Hiroto Otake et al.

Large Language Models (LLMs) are known to process information using a proficient internal language consistently, referred to as latent language, which may differ from the input or output languages. However, how the discrepancy between the latent language and the input and output language affects downstream task performance remains largely unexplored. While many studies research the latent language of LLMs, few address its importance in influencing task performance. In our study, we hypothesize that thinking in latent language consistently enhances downstream task performance. To validate this, our work varies the input prompt languages across multiple downstream tasks and analyzes the correlation between consistency in latent language and task performance. We create datasets consisting of questions from diverse domains such as translation and geo-culture, which are influenced by the choice of latent language. Experimental results across multiple LLMs on translation and geo-culture tasks, which are sensitive to the choice of language, indicate that maintaining consistency in latent language is not always necessary for optimal downstream task performance. This is because these models adapt their internal representations near the final layers to match the target language, reducing the impact of consistency on overall performance.

Pratik Sutar, Jason Naradowsky, Yusuke Miyao

Natural language is commonly used to describe instrument timbre, such as a "warm" or "heavy" sound. As these descriptors are based on human perception, there can be disagreement over which acoustic features correspond to a given adjective. In this work, we pursue a data-driven approach to further our understanding of such adjectives in the context of guitar tone. Our main contribution is a dataset of timbre adjectives, constructed by processing single clips of instrument audio to produce varied timbres through adjustments in EQ and effects such as distortion. Adjective annotations are obtained for each clip by crowdsourcing experts to complete a pairwise comparison and a labeling task. We examine the dataset and reveal correlations between adjective ratings and highlight instances where the data contradicts prevailing theories on spectral features and timbral adjectives, suggesting a need for a more nuanced, data-driven understanding of timbre.

Qiang Zhang, Jason Naradowsky, Yusuke Miyao

Existing dialogue models may encounter scenarios which are not well-represented in the training data, and as a result generate responses that are unnatural, inappropriate, or unhelpful. We propose the "Ask an Expert" framework in which the model is trained with access to an "expert" which it can consult at each turn. Advice is solicited via a structured dialogue with the expert, and the model is optimized to selectively utilize (or ignore) it given the context and dialogue history. In this work the expert takes the form of an LLM. We evaluate this framework in a mental health support domain, where the structure of the expert conversation is outlined by pre-specified prompts which reflect a reasoning strategy taught to practitioners in the field. Blenderbot models utilizing "Ask an Expert" show quality improvements across all expert sizes, including those with fewer parameters than the dialogue model itself. Our best model provides a $\sim 10\%$ improvement over baselines, approaching human-level scores on "engingingness" and "helpfulness" metrics.

Taiga Ishii, Ryo Ueda, Yusuke Miyao

A unique feature of Recurrent Neural Networks (RNNs) is that it incrementally processes input sequences. In this research, we aim to uncover the inherent generalization properties, i.e., inductive bias, of RNNs with respect to how frequently RNNs switch the outputs through time steps in the sequence classification task, which we call output sequence frequency. Previous work analyzed inductive bias by training models with a few synthetic data and comparing the model's generalization with candidate generalization patterns. However, when examining the output sequence frequency, previous methods cannot be directly applied since enumerating candidate patterns is computationally difficult for longer sequences. To this end, we propose to directly calculate the output sequence frequency for each model by regarding the outputs of the model as discrete-time signals and applying frequency domain analysis. Experimental results showed that Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) have an inductive bias towards lower-frequency patterns, while Elman RNN tends to learn patterns in which the output changes at high frequencies. We also found that the inductive bias of LSTM and GRU varies with the number of layers and the size of hidden layers.

Koki Washio, Yusuke Miyao

Open-domain code generation is a challenging problem because the set of functions and classes that we use are frequently changed and extended in programming communities. We consider the challenge of code generation for unknown libraries without additional training. In this paper, we explore a framework of code generation that can refer to relevant API documentations like human programmers to handle unknown libraries. As a first step of this direction, we implement a model that can extract relevant code signatures from API documentations based on a natural language intent and copy primitives from the extracted signatures. Moreover, to evaluate code generation for unknown libraries and our framework, we extend an existing dataset of open-domain code generation and resplit it so that the evaluation data consist of only examples using the libraries that do not appear in the training data. Experiments on our new split show that baseline encoder-decoder models cannot generate code using primitives of unknown libraries as expected. In contrast, our model outperforms the baseline on the new split and can properly generate unknown primitives when extracted code signatures are noiseless.

Fei Cheng, Yusuke Miyao

Extracting event time from news articles is a challenging but attractive task. In contrast to the most existing pair-wised temporal link annotation, Reimers et al.(2016) proposed to annotate the time anchor (a.k.a. the exact time) of each event. Their work represents time anchors with discrete representations of Single-Day/Multi-Day and Certain/Uncertain. This increases the complexity of modeling the temporal relations between two time anchors, which cannot be categorized into the relations of Allen's interval algebra (Allen, 1990). In this paper, we propose an effective method to decompose such complex temporal relations into sub-level relations by introducing a unified quadruple representation for both Single-Day/Multi-Day and Certain/Uncertain time anchors. The temporal relation classifiers are trained in a multi-label classification manner. The system structure of our approach is much simpler than the existing decision tree model (Reimers et al., 2018), which is composed by a dozen of node classifiers. Another contribution of this work is to construct a larger event time corpus (256 news documents) with a reasonable Inter-Annotator Agreement (IAA), for the purpose of overcoming the data shortage of the existing event time corpus (36 news documents). The empirical results show our approach outperforms the state-of-the-art decision tree model and the increase of data size obtained a significant improvement of performance.

Akiko Aizawa, Frederic Bergeron, Junjie Chen et al.

The global pandemic of COVID-19 has made the public pay close attention to related news, covering various domains, such as sanitation, treatment, and effects on education. Meanwhile, the COVID-19 condition is very different among the countries (e.g., policies and development of the epidemic), and thus citizens would be interested in news in foreign countries. We build a system for worldwide COVID-19 information aggregation containing reliable articles from 10 regions in 7 languages sorted by topics. Our reliable COVID-19 related website dataset collected through crowdsourcing ensures the quality of the articles. A neural machine translation module translates articles in other languages into Japanese and English. A BERT-based topic-classifier trained on our article-topic pair dataset helps users find their interested information efficiently by putting articles into different categories.

Hayate Iso, Yui Uehara, Tatsuya Ishigaki et al.

We propose a data-to-text generation model with two modules, one for tracking and the other for text generation. Our tracking module selects and keeps track of salient information and memorizes which record has been mentioned. Our generation module generates a summary conditioned on the state of tracking module. Our model is considered to simulate the human-like writing process that gradually selects the information by determining the intermediate variables while writing the summary. In addition, we also explore the effectiveness of the writer information for generation. Experimental results show that our model outperforms existing models in all evaluation metrics even without writer information. Incorporating writer information further improves the performance, contributing to content planning and surface realization.

Yang Gao, Steffen Eger, Ilia Kuznetsov et al.

Peer review is a core element of the scientific process, particularly in conference-centered fields such as ML and NLP. However, only few studies have evaluated its properties empirically. Aiming to fill this gap, we present a corpus that contains over 4k reviews and 1.2k author responses from ACL-2018. We quantitatively and qualitatively assess the corpus. This includes a pilot study on paper weaknesses given by reviewers and on quality of author responses. We then focus on the role of the rebuttal phase, and propose a novel task to predict after-rebuttal (i.e., final) scores from initial reviews and author responses. Although author responses do have a marginal (and statistically significant) influence on the final scores, especially for borderline papers, our results suggest that a reviewer's final score is largely determined by her initial score and the distance to the other reviewers' initial scores. In this context, we discuss the conformity bias inherent to peer reviewing, a bias that has largely been overlooked in previous research. We hope our analyses will help better assess the usefulness of the rebuttal phase in NLP conferences.

Sang Phan, Gustav Eje Henter, Yusuke Miyao et al.

Captioning models are typically trained using the cross-entropy loss. However, their performance is evaluated on other metrics designed to better correlate with human assessments. Recently, it has been shown that reinforcement learning (RL) can directly optimize these metrics in tasks such as captioning. However, this is computationally costly and requires specifying a baseline reward at each step to make training converge. We propose a fast approach to optimize one's objective of interest through the REINFORCE algorithm. First we show that, by replacing model samples with ground-truth sentences, RL training can be seen as a form of weighted cross-entropy loss, giving a fast, RL-based pre-training algorithm. Second, we propose to use the consensus among ground-truth captions of the same video as the baseline reward. This can be computed very efficiently. We call the complete proposal Consensus-based Sequence Training (CST). Applied to the MSRVTT video captioning benchmark, our proposals train significantly faster than comparable methods and establish a new state-of-the-art on the task, improving the CIDEr score from 47.3 to 54.2.