EMNLP 2023 Papers

As one of the most exciting features of large language models (LLMs), in-context learning is a mixed blessing. While it allows users to fast-prototype a task solver with only a few training examples, the performance is generally sensitive to various configurations of the prompt such as the choice or order of the training examples. In this paper, we for the first time theoretically and empirically identify that such a paradox is mainly due to the label shift of the in-context model to the data distribution, in which LLMs shift the label marginal $p(y)$ while having a good label conditional $p(x|y)$. With this understanding, we can simply calibrate the in-context predictive distribution by adjusting the label marginal, which is estimated via Monte-Carlo sampling over the in-context model, i.e., generation of LLMs. We call our approach as generative calibration. We conduct exhaustive experiments with 12 text classification tasks and 12 LLMs scaling from 774M to 33B, generally find that the proposed method greatly and consistently outperforms the ICL as well as state-of-the-art calibration methods, by up to 27% absolute in macro-F1. Meanwhile, the proposed method is also stable under different prompt configurations.

Large Language Model (LLM) has demonstrated significant ability in various Natural Language Processing tasks. However, their effectiveness is highly dependent on the phrasing of the task prompt, leading to research on automatic prompt optimization using labeled task data. We reveal that these prompt optimization techniques are vulnerable to distribution shifts such as subpopulation shifts, which are common for LLMs in real-world scenarios such as customer reviews analysis. In this light, we propose a new problem of robust prompt optimization for LLMs against distribution shifts, which requires the prompt optimized over the labeled source group can simultaneously generalize to an unlabeled target group. To solve this problem, we propose Generalized Prompt Optimization framework , which incorporates the unlabeled data from the target group into prompt optimization. Extensive experimental results demonstrate the effectiveness of the proposed framework with significant performance improvement on the target group and comparable performance on the source group.

Large language models (LLMs) have great potential for synthetic data generation. This work shows that useful data can be synthetically generated even for tasks that cannot be solved directly by LLMs: for problems with structured outputs, it is possible to prompt an LLM to perform the task in the reverse direction, by generating plausible input text for a target output structure. Leveraging this asymmetry in task difficulty makes it possible to produce large-scale, high-quality data for complex tasks. We demonstrate the effectiveness of this approach on closed information extraction, where collecting ground-truth data is challenging, and no satisfactory dataset exists to date. We synthetically generate a dataset of 1.8M data points, establish its superior quality compared to existing datasets in a human evaluation, and use it to finetune small models (220M and 770M parameters), termed SynthIE, that outperform the prior state of the art (with equal model size) by a substantial margin of 57 absolute points in micro-F1 and 79 points in macro-F1. Code, data, and models are available at anonymous.

Few-shot relation extraction involves identifying the type of relationship between two specific entities within a text, using a limited number of annotated samples. A variety of solutions to this problem have emerged by applying meta-learning and neural graph techniques which typically necessitate a training process for adaptation. Recently, the strategy of in-context learning has been demonstrating notable results without the need of training. Few studies have already utilized in-context learning for zero-shot information extraction. Unfortunately, the evidence for inference is either not considered or implicitly modeled during the construction of chain-of-thought prompts. In this paper, we propose a novel approach for few-shot relation extraction using large language models, named CoT-ER, chain-of-thought with explicit evidence reasoning. In particular, CoT-ER first induces large language models to generate evidences using task-specific and concept-level knowledge. Then these evidences are explicitly incorporated into chain-of-thought prompting for relation extraction. Experimental results demonstrate that our CoT-ER approach (with 0% training data) achieves competitive performance compared to the fully-supervised (with 100% training data) state-of-the-art approach on the FewRel1.0 and FewRel2.0 datasets.

To alleviate the data scarcity problem in End-to-end speech translation (ST), pre-training on data for speech recognition and machine translation is considered as an important technique. However, the modality gap between speech and text prevents the ST model from efficiently inheriting knowledge from the pre-trained models. In this work, we propose AdaTranS for end-to-end ST. It adapts the speech features with a new shrinking mechanism to mitigate the length mismatch between speech and text features by predicting word boundaries. Experiments on the MUST-C dataset demonstrate that AdaTranS achieves better performance than the other shrinking-based methods, with higher inference speed and lower memory usage. Further experiments also show that AdaTranS can be equipped with additional alignment losses to further improve performance.

We introduce a simple yet efficient sentence-level attack on black-box toxicity detector models. By adding several positive words or sentences to the end of a hateful message, we are able to change the prediction of a neural network and pass the toxicity detection system check. This approach is shown to be working on seven languages from three different language families. We also describe the defence mechanism against the aforementioned attack and discuss its limitations.

Incorporating language-specific (LS) modules or Mixture-of-Experts (MoE) are proven methods to boost performance in multilingual model performance, but the scalability of these approaches to hundreds of languages or experts tends to be hard to manage. We present Language-specific Matrix Synthesis (LMS), a novel method that addresses the issue. LMS utilizes parameter-efficient and lightweight modules, reducing the number of parameters while outperforming existing methods, e.g., +1.73 BLEU over Switch Transformer on OPUS-100 multilingual translation. Additionally, we introduce Fuse Distillation (FD) to condense multilingual knowledge from multiple LS modules into a single shared module, improving model inference and storage efficiency. Our approach demonstrates superior scalability and performance compared to state-of-the-art methods.

Increased focus on the computational efficiency of systems in natural language processing has motivated the design of efficient model architectures and improvements to underlying hardware accelerators. However, the resulting increases in computational throughput and reductions in floating point operations have not directly translated to improvements in wall-clock inference latency. We demonstrate that these discrepancies can be largely attributed to bottlenecks introduced by deep learning frameworks. We denote this phenomena as the framework tax, and observe that the disparity is growing as hardware speed increases over time. In this work, we examine this phenomena through a series of case studies analyzing the effects of model design decisions, framework paradigms, and hardware platforms on total model latency. Based on our findings, we provide actionable recommendations to researchers and practitioners aimed at narrowing the gap between efficient NLP model research and practice.

Psychology research has long explored aspects of human personality like extroversion, agreeableness and emotional stability, three of the personality traits that make up the 'Big Five'. Categorizations like the 'Big Five' are commonly used to assess and diagnose personality types. In this work, we explore whether text generated from large language models exhibits consistency in it's perceived 'Big Five' personality traits. For example, is a language model such as GPT2 likely to respond in a consistent way if asked to go out to a party? We also show that when exposed to different types of contexts (such as personality descriptions, or answers to diagnostic questions about personality traits), language models such as BERT and GPT2 consistently identify and mirror personality markers in those contexts. This behavior illustrates an ability to be manipulated in a predictable way (with correlations up to 0.84 between intended and realized changes in personality traits), and frames them as tools for controlling personas in applications such as dialog systems. We contribute two data-sets of personality descriptions of humans subjects.

This paper presents the first few-shot LLM-based chatbot that almost never hallucinates and has high conversationality and low latency. WikiChat is grounded on the English Wikipedia, the largest curated free-text corpus. WikiChat generates a response from an LLM, retains only the grounded facts, and combines them with additional information it retrieves from the corpus to form factual and engaging responses. We distill WikiChat based on GPT-4 into a 7B-parameter LLaMA model with minimal loss of quality, to significantly improve its latency, cost and privacy, and facilitate research and deployment. Using a novel hybrid human-and-LLM evaluation methodology, we show that our best system achieves 97.3% factual accuracy in simulated conversations. It significantly outperforms all retrieval-based and LLM-based baselines, and by 3.9%, 38.6% and 51.0% on head, tail and recent knowledge compared to GPT-4. Compared to previous state-of-the-art retrieval-based chatbots, WikiChat is also significantly more informative and engaging, just like an LLM. WikiChat achieves 97.9% factual accuracy in conversations with human users about recent topics, 55.0% better than GPT-4, while receiving significantly higher user ratings and more favorable comments.

A particularly successful class of approaches for few-shot learning combines language models with prompts - hand-crafted task descriptions that complement data samples. However, designing prompts by hand for each task commonly requires domain knowledge and substantial guesswork. We observe, in the context of classification tasks, that instruction finetuned language models are remarkably robust towards some dimensions of a prompt's design. We subsequently propose a simple method to eliminate the need for handcrafted prompts, named AuT-Few. This approach consists of (i) a prompt retrieval module that selects suitable task instructions from the instruction-tuning knowledge base, and (ii) the generation of two distinct, semantically meaningful, class descriptions and a selection mechanism via cross-validation. Over 12 datasets, spanning 8 classification tasks, we show that AuT-Few outperforms current state-of-the-art few-shot learning methods. Moreover, AuT-Few is the best ranking method across datasets on the RAFT few-shot benchmark. Notably, these results are achieved without task-specific handcrafted prompts on unseen tasks.

Text-to-SQL benchmarks play a crucial role in evaluating the progress made in the field and the ranking of different models. However, accurately matching a model-generated SQL query to a reference SQL query in a benchmark fails for various reasons, such as underspecified natural language queries, inherent assumptions in both model-generated and reference queries, and the non-deterministic nature of SQL output under certain conditions. In this paper, we conduct an extensive study of several prominent cross-domain text-to-SQL benchmarks and re-evaluate some of the top-performing models within these benchmarks, by both manually evaluating the SQL queries and rewriting them in equivalent expressions. Our evaluation reveals that attaining a perfect performance on these benchmarks is unfeasible due to the multiple interpretations that can be derived from the provided samples. Furthermore, we find that the true performance of the models is underestimated and their relative performance changes after a re-evaluation. Most notably, our evaluation reveals a surprising discovery: a recent GPT4-based model surpasses the gold standard reference queries in the Spider benchmark in our human evaluation. This finding highlights the importance of interpreting benchmark evaluations cautiously, while also acknowledging the critical role of additional independent evaluations in driving advancements in the field.

Recent work in Natural Language Processing and Computer Vision has been using textual information – e.g., entity names and descriptions – available in knowledge graphs to ground neural models to high-quality structured data. However, when it comes to non-English languages, the quantity and quality of textual information are comparatively scarce. To address this issue, we introduce the novel task of automatic Knowledge Graph Completion (KGE) and perform a thorough investigation on bridging the gap in both the quantity and quality of textual information between English and non-English languages. More specifically, we: i) bring to light the problem of increasing multilingual coverage and precision of entity names and descriptions in Wikidata; ii) demonstrate that state-of-the-art methods, namely, Machine Translation (MT), Web Search (WS), and Large Language Models (LLMs), struggle with this task; iii) present M-NTA, a novel unsupervised approach that combines MT, WS, and LLMs to generate high-quality textual information; and, iv) study the impact of increasing multilingual coverage and precision of non-English textual information in Entity Linking, Knowledge Graph Completion, and Question Answering. As part of our effort towards better multilingual knowledge graphs, we also introduce WikiKGE-10, the first human-curated benchmark to evaluate KGE approaches in 10 languages across 7 language families.

We investigate the task of out-of-domain (OOD) text classification with the aim of extending a classification model, trained on multiple source domains, to an unseen target domain. Recent studies have shown that learning invariant representations can enhance the performance of OOD generalization. However, the inherent disparity in data distribution across different domains poses challenges for achieving effective invariance learning. This study addresses this issue by employing memory augmentations. Specifically, we augment the original feature space using key-value memory and employ a meta-learning-based approach to enhance the quality of the invariant representations. Experimental results on sentiment analysis and natural language inference tasks show the effectiveness of memory-based method for invariance learning, leading to state-of-the-art performance on six datasets.

Post-training quantization (PTQ) of transformer language models faces significant challenges due to the existence of detrimental outliers in activations. We observe that these outliers are concentrated in specific channels and are asymmetric across channels. To address this issue, we propose the Outlier Suppression+ (OS+) framework, which contains the channel-wise shifting for asymmetry and channel-wise scaling for concentration. We show that these operations can be seamlessly migrated into subsequent modules while maintaining equivalence. Second, we propose a fast and stable scheme to calculate effective shifting and scaling values. The channel-wise shifting aligns the center of each channel for removal of outlier asymmetry. The channel-wise scaling quantitatively evaluates changes brought by migration and quantization for better quantization burden balance. We validate our OS+ under both standard and fine-grained quantization settings with models including BERT, OPT, BLOOM, BLOOMZ, and LLaMA. Comprehensive results across various tasks demonstrate the superiority of our approach. Especially, with standard quantization, OS+ can achieve near-floating-point performance on both small models and large language models on 8-bit and 6-bit. Besides, we establish a new state-of-the-art for 4-bit BERT with 15.5\% improvement. Our code is available at \url{https://github.com/ModelTC/Outlier_Suppression_Plus}.

Recently, many companies have been providing the capabilities of large language models as services. These Language-Model-as-a-Service (LMaaS) offerings support a variety of user tasks through in-context learning from prompts, which include instructions and demonstrations of the task. However, for users, manually crafting prompts or running automatic prompt tuning methods themselves can be demanding. Despite these challenges, LMaaS providers do not offer automatic prompt engineering methods as part of their services. One of the major obstacles to deploying them on an LMaaS is the heavy computational costs associated with automatic prompt engineering methods. These methods are typically designed to iterate through tens of thousands of examples, which impose unaffordable overheads for LMaaS providers. In this paper, we introduce MetaL-Prompt, a novel lightweight automatic prompt generation method for LMaaS. MetaL-Prompt meta-trains a prompt generation model (PGM) to enable robust learning by the language model from the contexts created by the generated prompts (i.e., in-context learning). Thanks to our meta-learning approach, a PGM can generate prompts for unseen tasks without requiring additional training for those specific tasks. Furthermore, the PGM can generate prompts with a single forward pass, significantly reducing computational costs compared to previous methods. We evaluate MetaL-Prompt on a range of unseen tasks and find that it improves performance by up to 19.4\% in terms of mean F1 score on QA datasets compared to the state-of-the-art baseline P-tuning, with limited computational cost.

Large language models (LLMs) such as ChatGPT can produce coherent, cohesive, relevant, and fluent answers for various natural language processing (NLP) tasks. Taking document-level machine translation (MT) as a testbed, this paper provides an in-depth evaluation of LLMs' ability on discourse modeling. The study focuses on three aspects: 1) Effects of Context-Aware Prompts, where we investigate the impact of different prompts on document-level translation quality and discourse phenomena; 2) Comparison of Translation Models, where we compare the translation performance of ChatGPT with commercial MT systems and advanced document-level MT methods; 3) Analysis of Discourse Modelling Abilities, where we further probe discourse knowledge encoded in LLMs and shed light on impacts of training techniques on discourse modeling. By evaluating on a number of benchmarks, we surprisingly find that LLMs have demonstrated superior performance and show potential to become a new paradigm for document-level translation: 1) leveraging their powerful long-text modeling capabilities, GPT-3.5 and GPT-4 outperform commercial MT systems in terms of human evaluation; 2) GPT-4 demonstrates a stronger ability for probing linguistic knowledge than GPT-3.5. This work highlights the challenges and opportunities of LLMs for MT, which we hope can inspire the future design and evaluation of LLMs (We release our data and annotations at https://github.com/longyuewangdcu/Document-MT-LLM).

The vital role of analogical reasoning in human cognition allows us to grasp novel concepts by linking them with familiar ones through shared relational structures. Despite the attention previous research has given to word analogies, this work suggests that Large Language Models (LLMs) often overlook the structures that underpin these analogies, raising questions about the efficacy of word analogies as a measure of analogical reasoning skills akin to human cognition. In response to this, our paper introduces a task of analogical structure abduction, grounded in cognitive psychology, designed to abduce structures that form an analogy between two systems. In support of this task, we establish a benchmark called SCAR, containing 400 scientific analogies from 13 distinct fields, tailored for evaluating analogical reasoning with structure abduction. The empirical evidence underlines the continued challenges faced by LLMs, including ChatGPT and GPT-4, in mastering this task, signifying the need for future exploration to enhance their abilities.

Text-video based multimodal event extraction refers to identifying event information from the given text-video pairs. Existing methods predominantly utilize video appearance features (VAF) and text sequence features (TSF) as input information. Some of them employ contrastive learning to align VAF with the event types extracted from TSF. However, they disregard the motion representations in videos and the optimization of contrastive objective could be misguided by the background noise from RGB frames. We observe that the same event triggers correspond to similar motion trajectories, which are hardly affected by the background noise. Moviated by this, we propose a Three Stream Multimodal Event Extraction framework (TSEE) that simultaneously utilizes the features of text sequence and video appearance, as well as the motion representations to enhance the event extraction capacity. Firstly, we extract the optical flow features (OFF) as motion representations from videos to incorporate with VAF and TSF. Then we introduce a Multi-level Event Contrastive Learning module to align the embedding space between OFF and event triggers, as well as between event triggers and types. Finally, a Dual Querying Text module is proposed to enhance the interaction between modalities. Experimental results show that TSEE outperforms the state-of-the-art methods, which demonstrates its superiority.