EMNLP 2023 Papers

We evaluate a range of recent LLMs on English creative writing, a challenging and complex task that requires imagination, coherence, and style. We use a difficult, open-ended scenario chosen to avoid training data reuse: an epic narration of a single combat between Ignatius J. Reilly, the protagonist of the Pulitzer Prize-winning novel A Confederacy of Dunces (1980), and a pterodactyl, a prehistoric flying reptile. We ask several LLMs and humans to write such a story and conduct a human evalution involving various criteria such as fluency, coherence, originality, humor, and style. Our results show that some state-of-the-art commercial LLMs match or slightly outperform our writers in most dimensions; whereas open-source LLMs lag behind. Humans retain an edge in creativity, while humor shows a binary divide between LLMs that can handle it comparably to humans and those that fail at it. We discuss the implications and limitations of our study and suggest directions for future research.

We present 1-Pager the first system that answers a question and retrieves evidence using a single Transformer-based model and decoding process. 1-Pager incrementally partitions the retrieval corpus using constrained decoding to select a document and answer string, and we show that this is competitive with comparable retrieve-and-read alternatives according to both retrieval and answer accuracy metrics. 1-Pager also outperforms the equivalent `closed-book' question answering model, by grounding predictions in an evidence corpus. While 1-Pager is not yet on-par with more expensive systems that read many more documents before generating an answer, we argue that it provides an important step toward attributed generation by folding retrieval into the sequence-to-sequence paradigm that is currently dominant in NLP. We also show that the search paths used to partition the corpus are easy to read and understand, paving a way forward for interpretable neural retrieval.

Large language models (LLMs) encode parametric knowledge about world facts and have shown remarkable performance in knowledge-driven NLP tasks. However, their reliance on parametric knowledge may cause them to overlook contextual cues, leading to incorrect predictions in context-sensitive NLP tasks (e.g., knowledge acquisition tasks). In this paper, we seek to assess and enhance LLMs' contextual faithfulness in two aspects: knowledge conflict and prediction with abstention. We demonstrate that LLMs' faithfulness can be significantly improved using carefully designed prompting strategies. In particular, we identify opinion-based prompts and counterfactual demonstrations as the most effective methods. Opinion-based prompts reframe the context as a narrator's statement and inquire about the narrator's opinions, while counterfactual demonstrations use instances containing false facts to improve faithfulness in knowledge conflict situations. Neither technique requires additional training. We conduct experiments on three datasets of two standard NLP tasks, machine reading comprehension and relation extraction, and the results demonstrate significant improvement in faithfulness to contexts. Code and data are released at https://github.com/wzhouad/context-faithful-llm.

In this paper, we present a novel approach for distilling math word problem solving capabilities from large language models (LLMs) into smaller, more efficient student models. Our approach is designed to consider the student model's weaknesses and foster a tailored learning experience by generating targeted exercises aligned with educational science principles, such as knowledge tracing and personalized learning. Concretely, we let GPT-3 be a math tutor and run two steps iteratively: 1) assessing the student model's current learning status on a GPT-generated exercise book, and 2) improving the student model by training it with tailored exercise samples generated by GPT-3. Experimental results reveal that our approach outperforms LLMs (e.g., GPT-3 and PaLM) in accuracy across three distinct benchmarks while employing significantly fewer parameters. Furthermore, we provide a comprehensive analysis of the various components within our methodology to substantiate their efficacy.

Theory of mind (ToM) evaluations currently focus on testing models using passive narratives that inherently lack interactivity. We introduce FANToM, a new benchmark designed to stress-test ToM within information-asymmetric conversational contexts via question answering. Our benchmark draws upon important theoretical requisites from psychology and necessary empirical considerations when evaluating large language models (LLMs). In particular, we formulate multiple types of questions that demand the same underlying reasoning to identify illusory or false sense of ToM capabilities in LLMs. We show that FANToM is challenging for state-of-the-art LLMs, which perform significantly worse than humans even with chain-of-thought reasoning or fine-tuning.

The recent success of general-domain large language models (LLMs) has significantly changed the natural language processing paradigm towards a unified foundation model across domains and applications. In this paper, we focus on assessing the performance of GPT-4, the most capable LLM so far, on the text-based applications for radiology reports, comparing against state-of-the-art (SOTA) radiology-specific models. Exploring various prompting strategies, we evaluated GPT-4 on a diverse range of common radiology tasks and we found GPT-4 either outperforms or is on par with current SOTA radiology models. With zero-shot prompting, GPT-4 already obtains substantial gains ($\approx$ 10\% absolute improvement) over radiology models in temporal sentence similarity classification (accuracy) and natural language inference ($F_1$). For tasks that require learning dataset-specific style or schema (e.g. findings summarisation), GPT-4 improves with example-based prompting and matches supervised SOTA. Our extensive error analysis with a board-certified radiologist shows GPT-4 has a sufficient level of radiology knowledge with only occasional errors in complex context that require nuanced domain knowledge. For findings summarisation, GPT-4 outputs are found to be overall comparable with existing manually-written impressions.

Efficient inference has become crucial for hyper-scale AI models, including large language models, as their parameter count continues to increase for enhanced performance. This necessity holds true regardless of the computing environment, whether it be mobile devices or cloud servers. Quantization emerges as a solution to alleviate the computational burden during inference. By representing models with a reduced bit-width, quantization minimizes the frequency of DRAM access while fully exploiting the parallelism of operations through a dense matrix format. Consequently, quantized models achieve low end-to-end latency and optimize resource utilization by addressing both memory and computing bottlenecks. In this paper, we propose a straightforward post-training quantization scheme, called \textsc{Z-Fold}, that fully utilizes the feature of the Transformer structure widely employed in large language models.

Biomedical entity linking (BioEL) is the process of connecting entities referenced in documents to entries in biomedical databases such as the Unified Medical Language System (UMLS) or Medical Subject Headings (MeSH). The study objective was to comprehensively evaluate nine recent state-of-the-art biomedical entity linking models under a unified framework. We compare these models along axes of (1) accuracy, (2) speed, (3) ease of use, (4) generalization, and (5) adaptability to new ontologies and datasets. We additionally quantify the impact of various preprocessing choices such as abbreviation detection. Systematic evaluation reveals several notable gaps in current methods. In particular, current methods struggle to correctly link genes and proteins and often have difficulty effectively incorporating context into linking decisions. To expedite future development and baseline testing, we release our unified evaluation framework and all included models on GitHub at https://github.com/davidkartchner/biomedical-entity-linking

Continual learning (CL) aims to constantly learn new knowledge over time while avoiding catastrophic forgetting on old tasks. We focus on continual text classification under the class-incremental setting. Recent CL studies have identified the severe performance decrease on analogous classes as a key factor for catastrophic forgetting. In this paper, through an in-depth exploration of the representation learning process in CL, we discover that the compression effect of the information bottleneck leads to confusion on analogous classes. To enable the model learn more sufficient representations, we propose a novel replay-based continual text classification method, InfoCL. Our approach utilizes fast-slow and current-past contrastive learning to perform mutual information maximization and better recover the previously learned representations. In addition, InfoCL incorporates an adversarial memory augmentation strategy to alleviate the overfitting problem of replay. Experimental results demonstrate that InfoCL effectively mitigates forgetting and achieves state-of-the-art performance on three text classification tasks.

Generating paragraph captions for untrimmed videos without event annotations is challenging, especially when aiming to enhance precision and minimize repetition at the same time. To address this challenge, we propose a module called Sparse Frame Grouping (SFG). It dynamically groups event information with the help of action information for the entire video and excludes redundant frames within pre-defined clips. To enhance the performance, an Intra Contrastive Learning technique is designed to align the SFG module with the core event content in the paragraph, and an Inter Contrastive Learning technique is employed to learn action-guided context with reduced static noise simultaneously. Extensive experiments are conducted on two benchmark datasets (ActivityNet Captions and YouCook2). Results demonstrate that SFG outperforms the state-of-the-art methods on all metrics.

In many domains of argumentation, people’s arguments are driven by so-called attitude roots, i.e., underlying beliefs and world views, and their corresponding attitude themes. Given the strength of these latent drivers of arguments, recent work in psychology suggests that instead of directly countering surface-level reasoning (e.g., falsifying the premises), one should follow an argumentation style inspired by the Jiu-Jitsu ``soft'' combat system: first, identify an arguer's attitude roots and themes, and then choose a prototypical rebuttal that is aligned with those drivers instead of trying to invalidate those. In this work, we are the first to explore Jiu-Jitsu argumentation for peer reviews by proposing the novel task of attitude and theme-guided rebuttal generation. To this end, we enrich an existing dataset for discourse structure in peer reviews with attitude roots, attitude themes, and canonical rebuttals. To facilitate this process, we recast established annotation concepts from the domain of peer reviews (e.g., aspects a review sentence is relating to) and train domain-specific models. We then propose strong rebuttal generation strategies, which we benchmark on our novel dataset for the task of end-to-end attitude and theme-guided rebuttal generation and two subtasks.

Binary code analysis has immense importance in the research domain of software security. Today, software is very often compiled for various Instruction Set Architectures (ISAs). As a result, cross-architecture binary code analysis has become an emerging problem. Recently, deep learning-based binary analysis has shown promising success. It is widely known that training a deep learning model requires a massive amount of data. However, for some low-resource ISAs, an adequate amount of data is hard to find, preventing deep learning from being widely adopted for binary analysis. To overcome the data scarcity problem and facilitate cross-architecture binary code analysis, we propose to apply the ideas and techniques in Neural Machine Translation (NMT) to binary code analysis. Our insight is that a binary, after disassembly, is represented in some assembly language. Given a binary in a low-resource ISA, we translate it to a binary in a high-resource ISA (e.g., x86). Then we can use a model that has been trained on the high-resource ISA to test the translated binary. We have implemented the model called UNSUPERBINTRANS, and conducted experiments to evaluate its performance. Specifically, we conducted two downstream tasks, including code similarity detection and vulnerability discovery. In both tasks, we achieved high accuracies.

We address the task of evidence retrieval for long document question answering, which involves locating relevant paragraphs within a document to answer a question. We aim to assess the applicability of large language models (LLMs) in the task of zero-shot long document evidence retrieval, owing to their unprecedented performance across various NLP tasks. However, currently the LLMs can consume limited context lengths as input, thus providing document chunks as inputs might overlook the global context while missing out on capturing the inter-segment dependencies. Moreover, directly feeding the large input sets can incur significant computational costs, particularly when processing the entire document (and potentially incurring monetary expenses with enterprise APIs like OpenAI's GPT variants). To address these challenges, we propose a suite of techniques that exploit the discourse structure commonly found in documents. By utilizing this structure, we create a condensed representation of the document, enabling a more comprehensive understanding and analysis of relationships between different parts. We retain $99.6$% of the best zero-shot approach's performance, while processing only $26$% of the total tokens used by the best approach in the information seeking evidence retrieval setup. We also show how our approach can be combined with *self-ask* reasoning agent to achieve best zero-shot performance in complex multi-hop question answering, just $\approx 4$% short of zero-shot performance using gold evidence.

Does inverse scaling only occur as a function of model size, or can it also occur over the course of training? We carry out an exploratory study investigating whether the performance of language models on specific tasks can decrease (while general performance remains high) during training on the language modeling task. We find 8 tasks on which Pythia 12B (Biderman et al., 2023) shows decreased performance over the course of training. Five of these tasks (TruthfulQA-MC1, TruthfulQA-MC2, Hindsight Neglect, Memo Trap, and Pattern Match Suppression) additionally show a consistent relationship whereby larger language models show a greater decrease in performance the more they are trained, despite showing standard (positive) scaling overall. This highlights the importance of testing performance at all relevant benchmarks any time models are trained on additional data, even if their overall performance improves.

Knowing the language of an input text/audio is a necessary first step for using almost every NLP tool such as taggers, parsers, or translation systems. Language identification is a well-studied problem, sometimes even considered solved; in reality, due to lack of data and computational challenges, current systems cannot accurately identify most of the world's 7000 languages. To tackle this bottleneck, we first compile a corpus, MCS-350, of 50K multilingual and parallel children's stories in 350+ languages. MCS-350 can serve as a benchmark for language identification of short texts and for 1400+ new translation directions in low-resource Indian and African languages. Second, we propose a novel misprediction-resolution hierarchical model, LIMIT, for language identification that reduces error by 55% (from 0.71 to 0.32) on our compiled children's stories dataset and by 40% (from 0.23 to 0.14) on the FLORES-200 benchmark. Our method can expand language identification coverage into low-resource languages by relying solely on systemic misprediction patterns, bypassing the need to retrain large models from scratch.

Despite the significant progress in developing named entity recognition models, scaling to novel-emerging types still remains challenging in real-world scenarios. Continual learning and zero-shot learning approaches have been explored to handle novel-emerging types with less human supervision, but they have not been as successfully adopted as supervised approaches. Meanwhile, humans possess a much larger vocabulary size than these approaches and have the ability to learn the alignment between entities and concepts effortlessly through natural supervision. In this paper, we consider a more realistic and challenging setting called open-vocabulary named entity recognition (OVNER) to imitate human-level ability. OVNER aims to recognize entities in novel types by their textual names or descriptions. Specifically, we formulate OVNER as a semantic matching task and propose a novel and scalable two-stage method called Context-Type SemAntiC Alignment and FusiOn (CACAO). In the pre-training stage, we adopt Dual-Encoder for context-type semantic alignment and pre-train Dual-Encoder on 80M context-type pairs which are easily accessible through natural supervision. In the fine-tuning stage, we use Cross-Encoder for context-type semantic fusion and fine-tune Cross-Encoder on base types with human supervision. Experimental results show that our method outperforms the previous state-of-the-art methods on three challenging OVNER benchmarks by 9.7%, 9.5%, and 1.8% F1-score of novel types. Moreover, CACAO also demonstrates its flexible transfer ability in cross-domain NER.

Fine-tuning contextualized representations learned by pre-trained language models remains a prevalent practice in NLP. However, fine-tuning can lead to representation degradation (also known as representation collapse), which may result in instability, sub-optimal performance, and weak generalization. In this paper, we propose Representation Projection Invariance (REPINA), a novel regularization method to maintain the information content of representation and reduce representation collapse during fine-tuning by discouraging undesirable changes in the representations. We study the empirical behavior of the proposed regularization in comparison to 5 comparable baselines across 13 language understanding tasks (GLUE benchmark and six additional datasets). When evaluating in-domain performance, REPINA consistently outperforms other baselines on most tasks (10 out of 13). Additionally, REPINA improves out-of-distribution performance. We also demonstrate its effectiveness in few-shot settings and robustness to label perturbation. As a by-product, we extend previous studies of representation collapse and propose several metrics to quantify it. Our empirical findings show that our approach is significantly more effective at mitigating representation collapse.

Federated learning (FL) enables multiple participants to collaboratively train machine learning models using decentralized data sources, alleviating privacy concerns that arise from directly sharing local data. However, the lack of model privacy protection in FL becomes an unneglectable challenge, especially when people want to federally finetune models based on a proprietary large language model. In this study, we propose a novel FL training approach that accomplishes information exchange among participants via tunable soft prompts. These soft prompts, updated and transmitted between the server and clients, assume the role of the global model parameters and serve as messengers to deliver useful knowledge from the local data and global model. As the global model itself is not required to be shared and the local training is conducted based on an auxiliary model with fewer parameters than the global model, the proposed approach provides protection for the global model while reducing communication and computation costs in FL. Extensive experiments show the effectiveness of the proposed approach compared to several baselines. We have released the source code at https://github.com/alibaba/FederatedScope/tree/fedsp/federatedscope/nlp/fedsp.

Collecting high-quality labeled data for model training is notoriously time-consuming and labor-intensive for various NLP tasks. While copious solutions, such as active learning for small language models (SLMs) and prevalent in-context learning in the era of large language models (LLMs), have been proposed and alleviate the labeling burden to some extent, their performances are still subject to human intervention. It is still underexplored how to reduce the annotation cost in the LLMs era. To bridge this, we revolutionize traditional active learning and propose an innovative collaborative learning framework FreeAL to interactively distill and filter the task-specific knowledge from LLMs. During collaborative training, an LLM serves as an active annotator inculcating its coarse-grained knowledge, while a downstream SLM is incurred as a student to filter out high-quality in-context samples to feedback LLM for the subsequent label refinery. Extensive experiments on eight benchmark datasets demonstrate that FreeAL largely enhances the zero-shot performances for both SLM and LLM without any human supervision.