ACL 2023 Papers

Unsupervised speech recognition ({pasted macro ‘ASRU’}/) is the problem of learning automatic speech recognition (ASR) systems from unpaired speech-only and text-only corpora. While various algorithms exist to solve this problem, a theoretical framework is missing to study their properties and address such issues as sensitivity to hyperparameters and training instability. In this paper, we proposed a general theoretical framework to study the properties of {pasted macro ‘ASRU’}/ systems based on random matrix theory and the theory of neural tangent kernels. Such a framework allows us to prove various learnability conditions and sample complexity bounds of {pasted macro ‘ASRU’}/. Extensive {pasted macro ‘ASRU’}/ experiments on synthetic languages with three classes of transition graphs provide strong empirical evidence for our theory (code available at https://github.com/cactuswiththoughts/UnsupASRTheory.gitcactuswiththoughts/UnsupASRTheory.git).

Recent data-driven conversational models are able to return fluent, consistent, and informative responses to many kinds of requests and utterances in task-oriented scenarios. However, these responses are typically limited to just the immediate local topic instead of being wider-ranging and proactively taking the conversation further, for example making suggestions to help customers achieve their goals. This inadequacy reflects a lack of understanding of the interlocutor’s situation and implicit goal. To address the problem, we introduce a task of proactive response selection based on situational information. We present a manually-curated dataset of 1.7k English conversation examples that include situational background information plus for each conversation a set of responses, only some of which are acceptable in the situation. A responsive and informed conversation system should select the appropriate responses and avoid inappropriate ones; doing so demonstrates the ability to adequately understand the initiating request and situation. Our benchmark experiments show that this is not an easy task even for strong neural models, offering opportunities for future research.

Modern Natural Language Generation (NLG) models come with massive computational and storage requirements. In this work, we study the potential of compressing them, which is crucial for real-world applications serving millions of users. We focus on Knowledge Distillation (KD) techniques, in which a small student model learns to imitate a large teacher model, allowing to transfer knowledge from the teacher to the student. In contrast to much of the previous work, our goal is to optimize the model for a specific NLG task and a specific dataset. Typically in real-world applications, in addition to labeled data there is abundant unlabeled task-specific data, which is crucial for attaining high compression rates via KD. In this work, we conduct a systematic study of task-specific KD techniques for various NLG tasks under realistic assumptions. We discuss the special characteristics of NLG distillation and particularly the exposure bias problem. Following, we derive a family of Pseudo-Target (PT) augmentation methods, substantially extending prior work on sequence-level KD. We propose the Joint-Teaching method, which applies word-level KD to multiple PTs generated by both the teacher and the student. Finally, we validate our findings in an extreme setup with no labeled examples using GPT-4 as the teacher. Our study provides practical model design observations and demonstrates the effectiveness of PT training for task-specific KD in NLG.

The development of large language models (LLMs) such as ChatGPT has brought a lot of attention recently. However, their evaluation in the benchmark academic datasets remains under-explored due to the difficulty of evaluating the generative outputs produced by this model against the ground truth. In this paper, we aim to present a thorough evaluation of ChatGPT’s performance on diverse academic datasets, covering tasks like question-answering, text summarization, code generation, commonsense reasoning, mathematical problem-solving, machine translation, bias detection, and ethical considerations. Specifically, we evaluate ChatGPT across 140 tasks and analyze 255K responses it generates in these datasets. This makes our work the largest evaluation of ChatGPT in NLP benchmarks. In short, our study aims to validate the strengths and weaknesses of ChatGPT in various tasks and provide insights for future research using LLMs. We also report a new emergent ability to follow multi-query instructions that we mostly found in ChatGPT and other instruction-tuned models. Our extensive evaluation shows that even though ChatGPT is capable of performing a wide variety of tasks, and may obtain impressive performance in several benchmark datasets, it is still far from achieving the ability to reliably solve many challenging tasks. By providing a thorough assessment of ChatGPT’s performance across diverse NLP tasks, this paper sets the stage for a targeted deployment of ChatGPT-like LLMs in real-world applications.

Training grounded response generation models often requires a large collection of grounded dialogues. However, it is costly to build such dialogues. In this paper, we present a synthetic data generation framework (SynDG) for grounded dialogues. The generation process utilizes large pre-trained language models and freely available knowledge data (e.g., Wikipedia pages, persona profiles, etc.). The key idea of designing SynDG is to consider dialogue flow and coherence in the generation process. Specifically, given knowledge data, we first heuristically determine a dialogue flow, which is a series of knowledge pieces. Then, we employ T5 to incrementally turn the dialogue flow into a dialogue. To ensure coherence of both the dialogue flow and the synthetic dialogue, we design a two-level filtering strategy, at the flow-level and the utterance-level respectively. Experiments on two public benchmarks show that the synthetic grounded dialogue data produced by our framework is able to significantly boost model performance in both full training data and low-resource scenarios.

Zero pronouns (ZPs) are frequently omitted in pro-drop languages (e.g. Chinese, Hungarian, and Hindi), but should be recalled in non-pro-drop languages (e.g. English). This phenomenon has been studied extensively in machine translation (MT), as it poses a significant challenge for MT systems due to the difficulty in determining the correct antecedent for the pronoun. This survey paper highlights the major works that have been undertaken in zero pronoun translation (ZPT) after the neural revolution so that researchers can recognize the current state and future directions of this field. We provide an organization of the literature based on evolution, dataset, method, and evaluation. In addition, we compare and analyze competing models and evaluation metrics on different benchmarks. We uncover a number of insightful findings such as: 1) ZPT is in line with the development trend of large language model; 2) data limitation causes learning bias in languages and domains; 3) performance improvements are often reported on single benchmarks, but advanced methods are still far from real-world use; 4) general-purpose metrics are not reliable on nuances and complexities of ZPT, emphasizing the necessity of targeted metrics; 5) apart from commonly-cited errors, ZPs will cause risks of gender bias.

The ability to understand a user’s underlying needs is critical for conversational systems, especially with limited input from users in a conversation. Thus, in such a domain, Asking Clarification Questions (ACQs) to reveal users’ true intent from their queries or utterances arise as an essential task. However, it is noticeable that a key limitation of the existing ACQs studies is their incomparability, from inconsistent use of data, distinct experimental setups and evaluation strategies. Therefore, in this paper, to assist the development of ACQs techniques, we comprehensively analyse the current ACQs research status, which offers a detailed comparison of publicly available datasets, and discusses the applied evaluation metrics, joined with benchmarks for multiple ACQs-related tasks. In particular, given a thorough analysis of the ACQs task, we discuss a number of corresponding research directions for the investigation of ACQs as well as the development of conversational systems.

Mathematical reasoning is a fundamental aspect of human intelligence and is applicable in various fields, including science, engineering, finance, and everyday life. The development of artificial intelligence (AI) systems capable of solving math problems and proving theorems in language has garnered significant interest in the fields of machine learning and natural language processing. For example, mathematics serves as a testbed for aspects of reasoning that are challenging for powerful deep learning models, driving new algorithmic and modeling advances. On the other hand, recent advances in large-scale neural language models have opened up new benchmarks and opportunities to use deep learning for mathematical reasoning. In this survey paper, we review the key tasks, datasets, and methods at the intersection of mathematical reasoning and deep learning over the past decade. We also evaluate existing benchmarks and methods, and discuss future research directions in this domain.

Open domain question answering (ODQA) is a longstanding task aimed at answering factual questions from a large knowledge corpus without any explicit evidence in natural language processing (NLP). Recent works have predominantly focused on improving the answering accuracy and have achieved promising progress. However, higher accuracy often requires more memory consumption and inference latency, which might not necessarily be efficient enough for direct deployment in the real world. Thus, a trade-off between accuracy, memory consumption and processing speed is pursued. In this paper, we will survey recent advancements in the efficiency of ODQA models and conclude core techniques for achieving efficiency. Additionally, we will provide a quantitative analysis of memory cost, query speed, accuracy, and overall performance comparison. Our goal is to keep scholars informed of the latest advancements and open challenges in ODQA efficiency research and contribute to the further development of ODQA efficiency.

Hybrid retrievers can take advantage of both sparse and dense retrievers. Previous hybrid retrievers leverage indexing-heavy dense retrievers. In this work, we study “Is it possible to reduce the indexing memory of hybrid retrievers without sacrificing performance”? Driven by this question, we leverage an indexing-efficient dense retriever (i.e. DrBoost) and introduce a LITE retriever that further reduces the memory of DrBoost. LITE is jointly trained on contrastive learning and knowledge distillation from DrBoost. Then, we integrate BM25, a sparse retriever, with either LITE or DrBoost to form light hybrid retrievers. Our Hybrid-LITE retriever saves 13× memory while maintaining 98.0% performance of the hybrid retriever of BM25 and DPR. In addition, we study the generalization capacity of our light hybrid retrievers on out-of-domain dataset and a set of adversarial attacks datasets. Experiments showcase that light hybrid retrievers achieve better generalization performance than individual sparse and dense retrievers. Nevertheless, our analysis shows that there is a large room to improve the robustness of retrievers, suggesting a new research direction.

Distillation from Weak Teacher (DWT) is a method of transferring knowledge from a smaller, weaker teacher model to a larger student model to improve its performance. Previous studies have shown that DWT can be effective in the vision domain and natural language processing (NLP) pre-training stage. Specifically, DWT shows promise in practical scenarios, such as enhancing new generation or larger models using pre-trained yet older or smaller models and lacking a resource budget. However, the optimal conditions for using DWT have yet to be fully investigated in NLP pre-training. Therefore, this study examines three key factors to optimize DWT, distinct from those used in the vision domain or traditional knowledge distillation. These factors are:(i) the impact of teacher model quality on DWT effectiveness, (ii) guidelines for adjusting the weighting value for DWT loss, and (iii) the impact of parameter remapping as a student model initialization technique for DWT.

We present a pipeline for a statistical stylometric exploration of a hypothesized partition of a text. Given a parameterization of the text, our pipeline: (1) detects literary features yielding the optimal overlap between the hypothesized and unsupervised partitions, (2) performs a hypothesis-testing analysis to quantify the statistical significance of the optimal overlap, while conserving implicit correlations between units of text that are more likely to be grouped, and (3) extracts and quantifies the importance of features most responsible for the classification, estimates their statistical stability and cluster-wise abundance. We apply our pipeline to the first two books in the Bible, where one stylistic component stands out in the eyes of biblical scholars, namely, the Priestly component. We identify and explore statistically significant stylistic differences between the Priestly and non-Priestly components.

Large language models trained on code have shown great potential to increase productivity of software developers. Several execution-based benchmarks have been proposed to evaluate functional correctness of model-generated code on simple programming problems. Nevertheless, it is expensive to perform the same evaluation on complex real-world projects considering the execution cost. On the other hand, static analysis tools such as linters, which can detect errors without running the program, haven’t been well explored for evaluating code generation models. In this work, we propose a static evaluation framework to quantify static errors in Python code completions, by leveraging Abstract Syntax Trees. Compared with execution-based evaluation, our method is not only more efficient, but also applicable to code in the wild. For experiments, we collect code context from open source repos to generate one million function bodies using public models. Our static analysis reveals that Undefined Name and Unused Variable are the most common errors among others made by language models. Through extensive studies, we also show the impact of sampling temperature, model size, and context on static errors in code completions.

Recently, high-performing code generation systems based on large language models have surfaced. They are trained on massive corpora containing much more natural text than actual executable computer code. This work shows that current code generation systems exhibit undesired biases inherited from their large language model backbones, which can reduce the quality of the generated code under specific circumstances. To investigate the effect, we propose the “block of influence” concept, which enables a modular decomposition and analysis of the coding challenges. We introduce an automated intervention mechanism reminiscent of adversarial testing that exposes undesired biases through the failure modes of the models under test. Finally, we demonstrate how our framework can be used as a data transformation technique during fine-tuning, acting as a mitigation strategy for these biases.

Social media is one of the most highly sought resources for analyzing characteristics of the language by its users. In particular, many researchers utilized various linguistic features of mental health problems from social media. However, existing approaches to detecting mental disorders face critical challenges, such as the scarcity of high-quality data or the trade-off between addressing the complexity of models and presenting interpretable results grounded in expert domain knowledge. To address these challenges, we design a simple but flexible model that preserves domain-based interpretability. We propose a novel approach that captures the semantic meanings directly from the text and compares them to symptom-related descriptions. Experimental results demonstrate that our model outperforms relevant baselines on various mental disorder detection tasks. Our detailed analysis shows that the proposed model is effective at leveraging domain knowledge, transferable to other mental disorders, and providing interpretable detection results.

In recent years, large language models (LLMs) have achieved strong performance on benchmark tasks, especially in zero or few-shot settings. However, these benchmarks often do not adequately address the challenges posed in the real-world, such as that of hierarchical classification. In order to address this challenge, we propose refactoring conventional tasks on hierarchical datasets into a more indicative long-tail prediction task. We observe LLMs are more prone to failure in these cases. To address these limitations, we propose the use of entailment-contradiction prediction in conjunction with LLMs, which allows for strong performance in a strict zero-shot setting. Importantly, our method does not require any parameter updates, a resource-intensive process and achieves strong performance across multiple datasets.

Zero-shot prompt-based learning has made much progress in sentiment analysis, and considerable effort has been dedicated to designing high-performing prompt templates. However, two problems exist; First, large language models are often biased to their pre-training data, leading to poor performance in prompt templates that models have rarely seen. Second, in order to adapt to different domains, re-designing prompt templates is usually required, which is time-consuming and inefficient. To remedy both shortcomings, we propose a simple yet strong data construction method to de-bias a given prompt template, yielding a large performance improvement in sentiment analysis tasks across different domains, pre-trained language models, and prompt templates. Also, we demonstrate the advantage of using domain-agnostic generic responses over the in-domain ground-truth data.

A triple speech translation data comprises speech, transcription, and translation. In the end-to-end paradigm, text machine translation (MT) usually plays the role of a teacher model for the speech translation (ST) via knowledge distillation. Parameter sharing with the teacher is often adopted to construct the ST model architecture, however, the two modalities are independently fed and trained via different losses. This situation does not match ST’s properties across two modalities and also limits the upper bound of the performance. Inspired by the works of video Transformer, we propose a simple unified cross-modal ST method, which concatenates speech and text as the input, and builds a teacher that can utilize both cross-modal information simultaneously. Experimental results show that in our unified ST framework, models can effectively utilize the auxiliary information from speech and text, and achieve compelling results on MuST-C datasets.

Extractive summarization focuses on extracting salient sentences from the source document and incorporating them in the summary without changing their wording or structure. The naive approach for extractive summarization is sentence classification, which makes independent binary decisions for each sentence, resulting in the model cannot detect the dependencies between sentences in the summary. Recent approaches introduce an autoregressive decoder to detect redundancy relationship between sentences by step-by-step sentence selection, but bring train-inference gap. To address these issues, we formulate extractive summarization as a salient sentence set recognition task. To solve the sentence set recognition task, we propose a set prediction network (SetSum), which sets up a fixed set of learnable queries to extract the entire sentence set of the summary, while capturing the dependencies between them.Different from previous methods with an auto-regressive decoder, we employ a non-autoregressive decoder to predict the sentences within the summary in parallel during both the training and inference process, which eliminates the train-inference gap. Experimental results on both single-document and multi-document extracted summary datasets show that our approach outperforms previous state-of-the-art models.