Cheng-Han Chiang

Cheng-Han Chiang, Yung-Sung Chuang, James Glass et al.

Existing sentence textual similarity benchmark datasets only use a single number to summarize how similar the sentence encoder's decision is to humans'. However, it is unclear what kind of sentence pairs a sentence encoder (SE) would consider similar. Moreover, existing SE benchmarks mainly consider sentence pairs with low lexical overlap, so it is unclear how the SEs behave when two sentences have high lexical overlap. We introduce a high-quality SE diagnostic dataset, HEROS. HEROS is constructed by transforming an original sentence into a new sentence based on certain rules to form a \textit{minimal pair}, and the minimal pair has high lexical overlaps. The rules include replacing a word with a synonym, an antonym, a typo, a random word, and converting the original sentence into its negation. Different rules yield different subsets of HEROS. By systematically comparing the performance of over 60 supervised and unsupervised SEs on HEROS, we reveal that most unsupervised sentence encoders are insensitive to negation. We find the datasets used to train the SE are the main determinants of what kind of sentence pairs an SE considers similar. We also show that even if two SEs have similar performance on STS benchmarks, they can have very different behavior on HEROS. Our result reveals the blind spot of traditional STS benchmarks when evaluating SEs.

Cheng-Han Chiang, Hung-yi Lee

Explanation methods in Interpretable NLP often explain the model's decision by extracting evidence (rationale) from the input texts supporting the decision. Benchmark datasets for rationales have been released to evaluate how good the rationale is. The ground truth rationales in these datasets are often human annotations obtained via crowd-sourced websites. Valuable as these datasets are, the details on how those human annotations are obtained are often not clearly specified. We conduct comprehensive controlled experiments using crowd-sourced websites on two widely used datasets in Interpretable NLP to understand how those unsaid details can affect the annotation results. Specifically, we compare the annotation results obtained from recruiting workers satisfying different levels of qualification. We also provide high-quality workers with different instructions for completing the same underlying tasks. Our results reveal that the annotation quality is highly subject to the workers' qualification, and workers can be guided to provide certain annotations by the instructions. We further show that specific explanation methods perform better when evaluated using the ground truth rationales obtained by particular instructions. Based on these observations, we highlight the importance of providing complete details of the annotation process and call for careful interpretation of any experiment results obtained using those annotations.

Cheng-Han Chiang, Hung-yi Lee

In this paper, we study the differences and commonalities between statistically out-of-distribution (OOD) samples and adversarial (Adv) samples, both of which hurting a text classification model's performance. We conduct analyses to compare the two types of anomalies (OOD and Adv samples) with the in-distribution (ID) ones from three aspects: the input features, the hidden representations in each layer of the model, and the output probability distributions of the classifier. We find that OOD samples expose their aberration starting from the first layer, while the abnormalities of Adv samples do not emerge until the deeper layers of the model. We also illustrate that the models' output probabilities for Adv samples tend to be more unconfident. Based on our observations, we propose a simple method to separate ID, OOD, and Adv samples using the hidden representations and output probabilities of the model. On multiple combinations of ID, OOD datasets, and Adv attacks, our proposed method shows exceptional results on distinguishing ID, OOD, and Adv samples.

Yu-Xiang Lin, Cheng-Han Chiang, Hung-yi Lee

In this paper, we show that when spoken language models (SLMs) are instructed to speak in a specific speaking style at the beginning of a multi-turn conversation, they cannot maintain the required speaking styles after several turns of interaction; we refer to this as the style amnesia of SLMs. We focus on paralinguistic speaking styles, including emotion, accent, volume, and speaking speed. We evaluate three proprietary and two open-source SLMs, demonstrating that none of these models can maintain a consistent speaking style when instructed to do so. We further show that when SLMs are asked to recall the style instruction in later turns, they can recall the style instruction, but they fail to express it throughout the conversation. We also show that explicitly asking the model to recall the style instruction can partially mitigate style amnesia. In addition, we examine various prompting strategies and find that SLMs struggle to follow the required style when the instruction is placed in system messages rather than user messages, which contradicts the intended function of system prompts.

Cheng-Han Chiang, Wei-Chih Chen, Chun-Yi Kuan et al.

Using large language models (LLMs) for automatic evaluation has become an important evaluation method in NLP research. However, it is unclear whether these LLM-based evaluators can be applied in real-world classrooms to assess student assignments. This empirical report shares how we use GPT-4 as an automatic assignment evaluator in a university course with 1,028 students. Based on student responses, we find that LLM-based assignment evaluators are generally acceptable to students when students have free access to these LLM-based evaluators. However, students also noted that the LLM sometimes fails to adhere to the evaluation instructions. Additionally, we observe that students can easily manipulate the LLM-based evaluator to output specific strings, allowing them to achieve high scores without meeting the assignment rubric. Based on student feedback and our experience, we provide several recommendations for integrating LLM-based evaluators into future classrooms. Our observation also highlights potential directions for improving LLM-based evaluators, including their instruction-following ability and vulnerability to prompt hacking.

Cheng-Han Chiang, Hung-yi Lee

Using large language models (LLMs) to evaluate text quality has recently gained popularity. Some prior works explore the idea of using LLMs for evaluation, while they differ in some details of the evaluation process. In this paper, we analyze LLM evaluation (Chiang and Lee, 2023) and G-Eval (Liu et al., 2023), and we discuss how those details in the evaluation process change how well the ratings given by LLMs correlate with human ratings. We find that the auto Chain-of-Thought (CoT) used in G-Eval does not always make G-Eval more aligned with human ratings. We also show that forcing the LLM to output only a numeric rating, as in G-Eval, is suboptimal. Last, we reveal that asking the LLM to explain its own ratings consistently improves the correlation between the ChatGPT and human ratings and pushes state-of-the-art (SoTA) correlations on two meta-evaluation datasets.

Cheng-Han Chiang, Hung-yi Lee

In this paper, we explore the following question: Are synonym substitution attacks really synonym substitution attacks (SSAs)? We approach this question by examining how SSAs replace words in the original sentence and show that there are still unresolved obstacles that make current SSAs generate invalid adversarial samples. We reveal that four widely used word substitution methods generate a large fraction of invalid substitution words that are ungrammatical or do not preserve the original sentence's semantics. Next, we show that the semantic and grammatical constraints used in SSAs for detecting invalid word replacements are highly insufficient in detecting invalid adversarial samples.

Cheng-Han Chiang, Hung-yi Lee

Long-form generations from large language models (LLMs) contain a mix of factual and non-factual claims, making evaluating factuality difficult. Prior works evaluate the factuality of a long paragraph by decomposing it into multiple facts, verifying those facts independently, and aggregating the results. Such methods assume that combining factual claims forms a factual paragraph. The above assumption can be violated: we show that strong open-source models like Llama-chat can generate paragraphs that contain verifiable facts, but the facts are combined into a non-factual paragraph due to entity ambiguity. We further reveal that existing factuality metrics, including FActScore and citation recall, cannot properly evaluate these non-factual paragraphs and overestimate their factuality. To address this, we introduce an enhanced metric, D-FActScore, specifically designed for content with ambiguous entities. We evaluate the D-FActScores of people biographies generated by retrieval-augmented LLMs. We show that D-FActScore can better assess the factuality of paragraphs with entity ambiguity than FActScore. We also find that four widely used open-source LLMs tend to mix information of distinct entities to form non-factual paragraphs, making their D-FActScore much lower than FActScore by over 10%.

Ke-Han Lu, Szu-Wei Fu, Chao-Han Huck Yang et al.

Large language models (LLMs) have been widely used as knowledge backbones of Large Audio Language Models (LALMs), yet how much auditory knowledge they encode through text-only pre-training and how this affects downstream performance remains unclear. We study this gap by comparing different LLMs under two text-only and one audio-grounded setting: (1) direct probing on AKB-2000, a curated benchmark testing the breadth and depth of auditory knowledge; (2) cascade evaluation, where LLMs reason over text descriptions from an audio captioner; and (3) audio-grounded evaluation, where each LLM is fine-tuned into a Large Audio Language Model (LALM) with an audio encoder. Our findings reveal that auditory knowledge varies substantially across families, and text-only results are strongly correlated with audio performance. Our work provides empirical grounding for a comprehensive understanding of LLMs in audio research.

Hsiao-Ying Huang, Cheng-Han Chiang, Hung-yi Lee

Interleaved spoken language models (SLMs) alternately generate text and speech tokens, but decoding at full transformer depth for every step becomes costly, especially due to long speech sequences. We propose SPAR-K, a modality-aware early exit framework designed to accelerate interleaved SLM inference while preserving perceptual quality. SPAR-K introduces a speech alternating-depth schedule: most speech positions exit at a fixed intermediate layer, while periodic full-depth "refresh" steps mitigate distribution shift due to early exit. We evaluate our framework using Step-Audio-2-mini and GLM-4-Voice across four datasets spanning reasoning, factual QA, and dialogue tasks, measuring performance in terms of ASR transcription accuracy and perceptual quality. Experimental results demonstrate that SPAR-K largely preserves question-answering accuracy with a maximum accuracy drop of 0.82\% while reducing average speech decoding depth by up to 11\% on Step-Audio-2-mini and 5\% on GLM-4-Voice, both with negligible changes in MOS and WER and no auxiliary computation overhead. We further demonstrate that confidence-based early exit strategies, widely used in text LLMs, are suboptimal for SLMs, highlighting that the unique statistical nature of speech tokens necessitates a specialized early exit design.

Cheng-Han Chiang, Hung-yi Lee

Large language models (LLMs) can solve problems step-by-step. While this chain-of-thought (CoT) reasoning boosts LLMs' performance, it is unclear if LLMs \textit{know} when to use CoT and whether those CoT are always necessary to answer the question. This paper shows that LLMs tend to generate redundant calculations and reasoning on a manually constructed math QA dataset, GSM8K-Zero. GSM8K-Zero is constructed such that the questions can be answered without any calculations, but LLMs, including Llama-2 models and Claude-2, tend to generate lengthy and unnecessary calculations to answer the questions. We also conduct experiments to explain why LLMs generate redundant calculations and reasonings. GSM8K-Zero is publicly available at https://github.com/d223302/Over-Reasoning-of-LLMs and https://huggingface.co/datasets/dcml0714/GSM8K-Zero.

Cheng-Han Chiang, Sung-Feng Huang, Hung-yi Lee

While behaviors of pretrained language models (LMs) have been thoroughly examined, what happened during pretraining is rarely studied. We thus investigate the developmental process from a set of randomly initialized parameters to a totipotent language model, which we refer to as the embryology of a pretrained language model. Our results show that ALBERT learns to reconstruct and predict tokens of different parts of speech (POS) in different learning speeds during pretraining. We also find that linguistic knowledge and world knowledge do not generally improve as pretraining proceeds, nor do downstream tasks' performance. These findings suggest that knowledge of a pretrained model varies during pretraining, and having more pretrain steps does not necessarily provide a model with more comprehensive knowledge. We will provide source codes and pretrained models to reproduce our results at https://github.com/d223302/albert-embryology.

Guan-Ting Lin, Cheng-Han Chiang, Hung-yi Lee

In spoken dialogue, even if two current turns are the same sentence, their responses might still differ when they are spoken in different styles. The spoken styles, containing paralinguistic and prosodic information, mark the most significant difference between text and speech modality. When using text-only LLMs to model spoken dialogue, text-only LLMs cannot give different responses based on the speaking style of the current turn. In this paper, we focus on enabling LLMs to listen to the speaking styles and respond properly. Our goal is to teach the LLM that "even if the sentences are identical if they are spoken in different styles, their corresponding responses might be different". Since there is no suitable dataset for achieving this goal, we collect a speech-to-speech dataset, StyleTalk, with the following desired characteristics: when two current speeches have the same content but are spoken in different styles, their responses will be different. To teach LLMs to understand and respond properly to the speaking styles, we propose the Spoken-LLM framework that can model the linguistic content and the speaking styles. We train Spoken-LLM using the StyleTalk dataset and devise a two-stage training pipeline to help the Spoken-LLM better learn the speaking styles. Based on extensive experiments, we show that Spoken-LLM outperforms text-only baselines and prior speech LLMs methods.

Ke-Han Lu, Zhehuai Chen, Szu-Wei Fu et al. · mit

We introduce DeSTA2.5-Audio, a general-purpose Large Audio Language Model (LALM) designed for robust auditory perception and instruction-following, without requiring task-specific audio instruction-tuning. Recent LALMs typically augment Large Language Models (LLMs) with auditory capabilities by training on large-scale, manually curated or LLM-synthesized audio-instruction datasets. However, these approaches have often suffered from the catastrophic forgetting of the LLM's original language abilities. To address this, we revisit the data construction pipeline and propose DeSTA, a self-generated cross-modal alignment strategy in which the backbone LLM generates its own training targets. This approach preserves the LLM's native language proficiency while establishing effective audio-text alignment, thereby enabling zero-shot generalization without task-specific tuning. Using DeSTA, we construct DeSTA-AQA5M, a large-scale, task-agnostic dataset containing 5 million training samples derived from 7,000 hours of audio spanning 50 diverse datasets, including speech, environmental sounds, and music. DeSTA2.5-Audio achieves state-of-the-art or competitive performance across a wide range of audio-language benchmarks, including Dynamic-SUPERB, MMAU, SAKURA, Speech-IFEval, and VoiceBench. Comprehensive comparative studies demonstrate that our self-generated strategy outperforms widely adopted data construction and training strategies in both auditory perception and instruction-following capabilities. Our findings underscore the importance of carefully designed data construction in LALM development and offer practical insights for building robust, general-purpose LALMs.

Cheng-Han Chiang, Hung-yi Lee, Michal Lukasik

The LLM-as-a-judge paradigm uses large language models (LLMs) for automated text evaluation, where a numerical assessment is assigned by an LLM to the input text following scoring rubrics. Existing methods for LLM-as-a-judge use cross-entropy (CE) loss for fine-tuning, which neglects the numeric nature of score prediction. Recent work addresses numerical prediction limitations of LLM fine-tuning through regression-aware fine-tuning, which, however, does not consider chain-of-thought (CoT) reasoning for score prediction. In this paper, we introduce TRACT (Two-stage Regression-Aware fine-tuning with CoT), a method combining CoT reasoning with regression-aware training. TRACT consists of two stages: first, seed LLM is fine-tuned to generate CoTs, which serve as supervision for the second stage fine-tuning. The training objective of TRACT combines the CE loss for learning the CoT reasoning capabilities, and the regression-aware loss for the score prediction. Experiments across four LLM-as-a-judge datasets and two LLMs show that TRACT significantly outperforms existing methods. Extensive ablation studies validate the importance of each component in TRACT.

Cheng-Han Chiang, Xiaofei Wang, Linjie Li et al. · microsoft-research

Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before responding. In contrast, humans typically engage in complex mental reasoning internally, enabling them to communicate ideas clearly and concisely. Thus, integrating an unspoken thought process into SLMs is highly desirable. While naively generating a complete chain-of-thought (CoT) reasoning before starting to talk can enable thinking for SLMs, this induces additional latency for the speech response, as the CoT reasoning can be arbitrarily long. To solve this issue, we propose Stitch, a novel generation method that alternates between the generation of unspoken reasoning chunks and spoken response chunks. Since the audio duration of a chunk of spoken response is much longer than the time to generate the tokens in a chunk of spoken response, we use the remaining free time to generate the unspoken reasoning tokens. When a chunk of audio is played to the user, the model continues to generate the next unspoken reasoning chunk, achieving simultaneous thinking and talking. Remarkably, Stitch matches the latency of baselines that cannot generate unspoken CoT by design while outperforming those baselines by 15% on math reasoning datasets; Stitch also performs equally well on non-reasoning datasets as those baseline models. Some animations and demonstrations are on the project page: https://d223302.github.io/STITCH.

Tzu-wen Hsu, Ke-Han Lu, Cheng-Han Chiang et al.

Large Audio-Language Models (LALMs) can take audio and text as the inputs and answer questions about the audio. While prior LALMs have shown strong performance on standard benchmarks, there has been alarming evidence that LALMs can hallucinate what is presented in the audio. To mitigate the hallucination of LALMs, we introduce Audio-Aware Decoding (AAD), a lightweight inference-time strategy that uses contrastive decoding to compare the token prediction logits with and without the audio context. By contrastive decoding, AAD promotes the tokens whose probability increases when the audio is present. We conduct our experiment on object hallucination datasets with three LALMs and show that AAD improves the F1 score by 0.046 to 0.428. We also show that AAD can improve the accuracy on general audio QA datasets like Clotho-AQA by 5.4% to 10.3%. We conduct thorough ablation studies to understand the effectiveness of each component in AAD.

Liang-Hsuan Tseng, En-Pei Hu, Cheng-Han Chiang et al.

Unsupervised automatic speech recognition (ASR) aims to learn the mapping between the speech signal and its corresponding textual transcription without the supervision of paired speech-text data. A word/phoneme in the speech signal is represented by a segment of speech signal with variable length and unknown boundary, and this segmental structure makes learning the mapping between speech and text challenging, especially without paired data. In this paper, we propose REBORN,Reinforcement-Learned Boundary Segmentation with Iterative Training for Unsupervised ASR. REBORN alternates between (1) training a segmentation model that predicts the boundaries of the segmental structures in speech signals and (2) training the phoneme prediction model, whose input is the speech feature segmented by the segmentation model, to predict a phoneme transcription. Since supervised data for training the segmentation model is not available, we use reinforcement learning to train the segmentation model to favor segmentations that yield phoneme sequence predictions with a lower perplexity. We conduct extensive experiments and find that under the same setting, REBORN outperforms all prior unsupervised ASR models on LibriSpeech, TIMIT, and five non-English languages in Multilingual LibriSpeech. We comprehensively analyze why the boundaries learned by REBORN improve the unsupervised ASR performance.

Cheng-Han Chiang, Xiaofei Wang, Linjie Li et al.

Current large language models (LLMs) and spoken language models (SLMs) begin thinking and taking actions only after the user has finished their turn. This prevents the model from interacting during the user's turn and can lead to high response latency while it waits to think. Consequently, thinking after receiving the full input is not suitable for speech-to-speech interaction, where real-time, low-latency exchange is important. We address this by noting that humans naturally "think while listening." In this paper, we propose SHANKS, a general inference framework that enables SLMs to generate unspoken chain-of-thought reasoning while listening to the user input. SHANKS streams the input speech in fixed-duration chunks and, as soon as a chunk is received, generates unspoken reasoning based on all previous speech and reasoning, while the user continues speaking. SHANKS uses this unspoken reasoning to decide whether to interrupt the user and to make tool calls to complete the task. We demonstrate that SHANKS enhances real-time user-SLM interaction in two scenarios: (1) when the user is presenting a step-by-step solution to a math problem, SHANKS can listen, reason, and interrupt when the user makes a mistake, achieving 37.1% higher interruption accuracy than a baseline that interrupts without thinking; and (2) in a tool-augmented dialogue, SHANKS can complete 56.9% of the tool calls before the user finishes their turn. Overall, SHANKS moves toward models that keep thinking throughout the conversation, not only after a turn ends. Animated illustrations of Shanks can be found at https://d223302.github.io/SHANKS/

Cheng-Han Chiang, Hung-yi Lee

Human evaluation is indispensable and inevitable for assessing the quality of texts generated by machine learning models or written by humans. However, human evaluation is very difficult to reproduce and its quality is notoriously unstable, hindering fair comparisons among different natural language processing (NLP) models and algorithms. Recently, large language models (LLMs) have demonstrated exceptional performance on unseen tasks when only the task instructions are provided. In this paper, we explore if such an ability of the LLMs can be used as an alternative to human evaluation. We present the LLMs with the exact same instructions, samples to be evaluated, and questions used to conduct human evaluation, and then ask the LLMs to generate responses to those questions; we dub this LLM evaluation. We use human evaluation and LLM evaluation to evaluate the texts in two NLP tasks: open-ended story generation and adversarial attacks. We show that the result of LLM evaluation is consistent with the results obtained by expert human evaluation: the texts rated higher by human experts are also rated higher by the LLMs. We also find that the results of LLM evaluation are stable over different formatting of the task instructions and the sampling algorithm used to generate the answer. We are the first to show the potential of using LLMs to assess the quality of texts and discuss the limitations and ethical considerations of LLM evaluation.

Cheng-Han Chiang, Hung-yi Lee

Pre-training language models (LMs) on large-scale unlabeled text data makes the model much easier to achieve exceptional downstream performance than their counterparts directly trained on the downstream tasks. In this work, we study what specific traits in the pre-training data, other than the semantics, make a pre-trained LM superior to their counterparts trained from scratch on downstream tasks. We propose to use artificially constructed datasets as the pre-training data to exclude the effect of semantics, and further control what characteristics the pre-training corpora have. By fine-tuning the pre-trained models on GLUE benchmark, we can learn how beneficial it is to transfer the knowledge from the model trained on the dataset possessing that specific trait. We define and discuss three different characteristics in the artificial dataset: 1) matching the token's uni-gram or bi-gram distribution between pre-training and downstream fine-tuning, 2) the presence of the explicit dependencies among the tokens in a sequence, 3) the length of the implicit dependencies among the tokens in a sequence. Our experiments show that the explicit dependencies in the sequences of the pre-training data are critical to the downstream performance. Our results also reveal that models achieve better downstream performance when pre-trained on a dataset with a longer range of implicit dependencies. Based on our analysis, we find that models pre-trained with artificial datasets are prone to learn spurious correlation in downstream tasks. Our work reveals that even if the LMs are not pre-trained on natural language, they still gain transferability on certain human language downstream tasks once the LMs learn to model the token dependencies in the sequences. This result helps us understand the exceptional transferability of pre-trained LMs.

Cheng-Han Chiang, Hung-yi Lee

In this paper, we study how the intrinsic nature of pre-training data contributes to the fine-tuned downstream performance. To this end, we pre-train different transformer-based masked language models on several corpora with certain features, and we fine-tune those language models on GLUE benchmarks. We find that models pre-trained on unstructured data beat those trained directly from scratch on downstream tasks. Our results also show that pre-training on structured data does not always make the model acquire ability that can be transferred to natural language downstream tasks. To our great astonishment, we uncover that pre-training on certain non-human language data gives GLUE performance close to performance pre-trained on another non-English language.