Large language models have become a vital component in modern NLP, achieving state of the art performance in a variety of tasks. However, they are often inefficient for real-world deployment due to their expensive inference costs. Knowledge distillation is a promising technique to improve their efficiency while retaining most of their effectiveness. In this paper, we reproduce, compare and analyze several representative methods for task-agnostic (general-purpose) distillation of Transformer language models. Our target of study includes Output Distribution (OD) transfer, Hidden State (HS) transfer with various layer mapping strategies, and Multi-Head Attention (MHA) transfer based on MiniLMv2. Through our extensive experiments, we study the effectiveness of each method for various student architectures in both monolingual (English) and multilingual settings. Overall, we show that MHA transfer based on MiniLMv2 is generally the best option for distillation and explain the potential reasons behind its success. Moreover, we show that HS transfer remains as a competitive baseline, especially under a sophisticated layer mapping strategy, while OD transfer consistently lags behind other approaches. Findings from this study helped us deploy efficient yet effective student models for latency-critical applications.
Transferring the knowledge of large language models (LLMs) is a promising technique to incorporate linguistic knowledge into end-to-end automatic speech recognition (ASR) systems. However, existing works only transfer a single representation of LLM (e.g. the last layer of pretrained BERT), while the representation of a text is inherently non-unique and can be obtained variously from different layers, contexts and models. In this work, we explore a wide range of techniques to obtain and transfer multiple representations of LLMs into a transducer-based ASR system. While being conceptually simple, we show that transferring multiple representations of LLMs can be an effective alternative to transferring only a single representation.
Large pre-trained language models have achieved state-of-the-art results on a variety of downstream tasks. Knowledge Distillation (KD) of a smaller student model addresses their inefficiency, allowing for deployment in resource-constraint environments. KD however remains ineffective, as the student is manually selected from a set of existing options already pre-trained on large corpora, a sub-optimal choice within the space of all possible student architectures. This paper proposes KD-NAS, the use of Neural Architecture Search (NAS) guided by the Knowledge Distillation process to find the optimal student model for distillation from a teacher, for a given natural language task. In each episode of the search process, a NAS controller predicts a reward based on a combination of accuracy on the downstream task and latency of inference. The top candidate architectures are then distilled from the teacher on a small proxy set. Finally the architecture(s) with the highest reward is selected, and distilled on the full downstream task training set. When distilling on the MNLI task, our KD-NAS model produces a 2 point improvement in accuracy on GLUE tasks with equivalent GPU latency with respect to a hand-crafted student architecture available in the literature. Using Knowledge Distillation, this model also achieves a 1.4x speedup in GPU Latency (3.2x speedup on CPU) with respect to a BERT-Base Teacher, while maintaining 97% performance on GLUE Tasks (without CoLA). We also obtain an architecture with equivalent performance as the hand-crafted student model on the GLUE benchmark, but with a 15% speedup in GPU latency (20% speedup in CPU latency) and 0.8 times the number of parameters
The sentence is a fundamental unit in many NLP applications. Sentence segmentation is widely used as the first preprocessing task, where an input text is split into consecutive sentences considering the end of the sentence (EOS) as their boundaries. This task formulation relies on a strong assumption that the input text consists only of sentences, or what we call the sentential units (SUs). However, real-world texts often contain non-sentential units (NSUs) such as metadata, sentence fragments, nonlinguistic markers, etc. which are unreasonable or undesirable to be treated as a part of an SU. To tackle this issue, we formulate a novel task of sentence identification, where the goal is to identify SUs while excluding NSUs in a given text. To conduct sentence identification, we propose a simple yet effective method which combines the beginning of the sentence (BOS) and EOS labels to determine the most probable SUs and NSUs based on dynamic programming. To evaluate this task, we design an automatic, language-independent procedure to convert the Universal Dependencies corpora into sentence identification benchmarks. Finally, our experiments on the sentence identification task demonstrate that our proposed method generally outperforms sentence segmentation baselines which only utilize EOS labels.
Off-policy evaluation (OPE) aims to accurately evaluate the performance of counterfactual policies using only offline logged data. Although many estimators have been developed, there is no single estimator that dominates the others, because the estimators' accuracy can vary greatly depending on a given OPE task such as the evaluation policy, number of actions, and noise level. Thus, the data-driven estimator selection problem is becoming increasingly important and can have a significant impact on the accuracy of OPE. However, identifying the most accurate estimator using only the logged data is quite challenging because the ground-truth estimation accuracy of estimators is generally unavailable. This paper studies this challenging problem of estimator selection for OPE for the first time. In particular, we enable an estimator selection that is adaptive to a given OPE task, by appropriately subsampling available logged data and constructing pseudo policies useful for the underlying estimator selection task. Comprehensive experiments on both synthetic and real-world company data demonstrate that the proposed procedure substantially improves the estimator selection compared to a non-adaptive heuristic.
Large-scale language models (LLMs) such as GPT-2, BERT and RoBERTa have been successfully applied to ASR N-best rescoring. However, whether or how they can benefit competitive, near state-of-the-art ASR systems remains unexplored. In this study, we incorporate LLM rescoring into one of the most competitive ASR baselines: the Conformer-Transducer model. We demonstrate that consistent improvement is achieved by the LLM's bidirectionality, pretraining, in-domain finetuning and context augmentation. Furthermore, our lexical analysis sheds light on how each of these components may be contributing to the ASR performance.
Off-policy evaluation (OPE) is the method that attempts to estimate the performance of decision making policies using historical data generated by different policies without conducting costly online A/B tests. Accurate OPE is essential in domains such as healthcare, marketing or recommender systems to avoid deploying poor performing policies, as such policies may hart human lives or destroy the user experience. Thus, many OPE methods with theoretical backgrounds have been proposed. One emerging challenge with this trend is that a suitable estimator can be different for each application setting. It is often unknown for practitioners which estimator to use for their specific applications and purposes. To find out a suitable estimator among many candidates, we use a data-driven estimator selection procedure for off-policy policy performance estimators as a practical solution. As proof of concept, we use our procedure to select the best estimator to evaluate coupon treatment policies on a real-world online content delivery service. In the experiment, we first observe that a suitable estimator might change with different definitions of the outcome variable, and thus the accurate estimator selection is critical in real-world applications of OPE. Then, we demonstrate that, by utilizing the estimator selection procedure, we can easily find out suitable estimators for each purpose.
Off-policy Evaluation (OPE), or offline evaluation in general, evaluates the performance of hypothetical policies leveraging only offline log data. It is particularly useful in applications where the online interaction involves high stakes and expensive setting such as precision medicine and recommender systems. Since many OPE estimators have been proposed and some of them have hyperparameters to be tuned, there is an emerging challenge for practitioners to select and tune OPE estimators for their specific application. Unfortunately, identifying a reliable estimator from results reported in research papers is often difficult because the current experimental procedure evaluates and compares the estimators' performance on a narrow set of hyperparameters and evaluation policies. Therefore, it is difficult to know which estimator is safe and reliable to use. In this work, we develop Interpretable Evaluation for Offline Evaluation (IEOE), an experimental procedure to evaluate OPE estimators' robustness to changes in hyperparameters and/or evaluation policies in an interpretable manner. Then, using the IEOE procedure, we perform extensive evaluation of a wide variety of existing estimators on Open Bandit Dataset, a large-scale public real-world dataset for OPE. We demonstrate that our procedure can evaluate the estimators' robustness to the hyperparamter choice, helping us avoid using unsafe estimators. Finally, we apply IEOE to real-world e-commerce platform data and demonstrate how to use our protocol in practice.
Common grounding is the process of creating and maintaining mutual understandings, which is a critical aspect of sophisticated human communication. While various task settings have been proposed in existing literature, they mostly focus on creating common ground under static context and ignore the aspect of maintaining them overtime under dynamic context. In this work, we propose a novel task setting to study the ability of both creating and maintaining common ground in dynamic environments. Based on our minimal task formulation, we collected a large-scale dataset of 5,617 dialogues to enable fine-grained evaluation and analysis of various dialogue systems. Through our dataset analyses, we highlight novel challenges introduced in our setting, such as the usage of complex spatio-temporal expressions to create and maintain common ground. Finally, we conduct extensive experiments to assess the capabilities of our baseline dialogue system and discuss future prospects of our research.
Recent models achieve promising results in visually grounded dialogues. However, existing datasets often contain undesirable biases and lack sophisticated linguistic analyses, which make it difficult to understand how well current models recognize their precise linguistic structures. To address this problem, we make two design choices: first, we focus on OneCommon Corpus \citep{udagawa2019natural,udagawa2020annotated}, a simple yet challenging common grounding dataset which contains minimal bias by design. Second, we analyze their linguistic structures based on \textit{spatial expressions} and provide comprehensive and reliable annotation for 600 dialogues. We show that our annotation captures important linguistic structures including predicate-argument structure, modification and ellipsis. In our experiments, we assess the model's understanding of these structures through reference resolution. We demonstrate that our annotation can reveal both the strengths and weaknesses of baseline models in essential levels of detail. Overall, we propose a novel framework and resource for investigating fine-grained language understanding in visually grounded dialogues.