EMNLP 2024 Papers

Low-resource domains, characterized by scarce data and annotations, present significant challenges for language and visual understanding tasks, with the latter much under-explored in the literature. Recent advancements in Vision-Language Models (VLM) have shown promising results in high-resource domains but fall short in low-resource concepts that are under-represented (e.g. only a handful of images per category) in the pre-training set. We tackle the challenging task of zero-shot low-resource image classification from a novel perspective. By leveraging a retrieval-based strategy, we achieve this in a training-free fashion. Specifically, our method, named CoRE (Combination of Retrieval Enrichment), enriches the representation of both query images and class prototypes by retrieving relevant textual information from large web-crawled databases. This retrieval-based enrichment significantly boosts classification performance by incorporating the broader contextual information relevant to the specific class. We validate our method on a newly established benchmark covering diverse low-resource domains, including medical imaging, rare plants, and circuits. Our experiments demonstrate that CoRE outperforms existing state-of-the-art methods that rely on synthetic data generation and model fine-tuning.

In the Minecraft Collaborative Building Task, two players collaborate: an Architect (A) provides instructions to a Builder (B) to assemble a specified structure using 3D blocks. In this work, we investigate the use of large language models (LLMs) to predict the sequence of actions taken by the Builder. Leveraging LLMs’ in-context learning abilities, we use few-shot prompting techniques, that significantly improve performance over baseline methods. Additionally, we present a detailed analysis of the gaps in performance for future work.

Despite Large Language Models (LLMs) have performed impressively in various Natural Language Processing (NLP) tasks, their inherent hallucination phenomena severely challenge their credibility in complex reasoning. Combining explainable Knowledge Graphs (KGs) with LLMs is a promising path to address this issue. However, structured KGs are difficult to utilize, and how to make LLMs understand and incorporate them is a challenging topic. We thereby reorganize a more efficient structure of KGs, while designing the KG-related instruction tuning and continual pre-training strategies to enable LLMs to learn and internalize this form of representation effectively. Moreover, we construct subgraphs to further enhance the retrieval capabilities of KGs via CoT reasoning. Extensive experiments on two KGQA datasets demonstrate that our model achieves convincing performance compared to strong baselines.

The rapid introduction of new brand names into everyday language poses a unique challenge for e-commerce spelling correction services, which must distinguish genuine misspellings from novel brand names that use unconventional spelling. We seek to address this challenge via Retrieval Augmented Generation (RAG). On this approach, product names are retrieved from a catalog and incorporated into the context used by a large language model (LLM) that has been fine-tuned to do contextual spelling correction. Through quantitative evaluation and qualitative error analyses, we find improvements in spelling correction utilizing the RAG framework beyond a stand-alone LLM. We also demonstrate the value of additional finetuning of the LLM to incorporate retrieved context.

Retrieval Augmented Generation (RAG) has been a powerful tool for Large Language Models (LLMs) to efficiently process overly lengthy contexts. However, recent LLMs like Gemini-1.5 and GPT-4 show exceptional capabilities to understand long contexts directly. We conduct a comprehensive comparison between RAG and long-context (LC) LLMs, aiming to leverage the strengths of both. We benchmark RAG and LC across various public datasets using three latest LLMs. Results reveal that when resourced sufficiently, LC consistently outperforms RAG in terms of average performance. However, RAG’s significantly lower cost remains a distinct advantage. Based on this observation, we propose Self-Route, a simple yet effective method that routes queries to RAG or LC based on model self-reflection. Self-Route significantly reduces the computation cost while maintaining a comparable performance to LC. Our findings provide a guideline for long-context applications of LLMs using RAG and LC.

Learning from human feedback via proxy reward modeling has been studied to align Large Language Models (LLMs) with human values. However, achieving reliable training through that proxy reward model (RM) is not a trivial problem, and its behavior remained as a black-box. In this paper, we study the role of proxy rewards in the LLM alignment via ‘reverse reward engineering’ by composing interpretable features as a white-box reward function. We aim to replicate the ground truth (gold) reward signal by achieving a monotonic relationship between the proxy and gold reward signals after training the model using the proxy reward in reinforcement learning (RL). Our findings indicate that successfully emulating the gold reward requires generating responses that are relevant with enough length to open-ended questions, while also ensuring response consistency in closed-ended questions. Furthermore, resulting models optimizing our devised white-box reward show competitive performances with strong open-source RMs in alignment benchmarks. We highlight its potential usage as a simple but strong reward baseline for the LLM alignment, not requiring explicit human feedback dataset and RM training.

Large Language Models (LLMs) have exhibited exceptional performance across diverse domains. However, recent studies reveal that LLMs are plagued by the “reversal curse”. Most existing methods rely on aggressive sample permutation and pay little attention to delving into the underlying reasons for this issue, resulting in only partial mitigation. In this paper, inspired by human knowledge reversal, we investigate and quantify the individual influence of three potential reasons on the reversal curse: 1) knowledge clarity, 2) entity correlation modeling, and 3) pairwise relationship reasoning capability. Motivated by the analysis of these reasons, we propose a novel **P**airwise entity **O**rder- and **R**elationship-**E**nhanced (**PORE**) data strategy, which facilitates bidirectional entity correlation modeling and pairwise relationship reasoning to overcome the reversal curse. Specifically, PORE augments the samples with entity order-reversal and semantically preserved question-answer pairs, enhancing the encoding of entity correlations in both directions. PORE also employs entity-interleaved pairwise relationship data, which elevates the model’s capability for relationship reasoning. Additionally, to improve the recall of reverse relationships, we leverage knowledge clarity to construct high-clarity data for PORE. Extensive experimental results on available and two newly assembled datasets demonstrate the effectiveness and generalization of our method in both data-sufficient and -constrained situations.

In-context learning (ICL) has demonstrated excellent performance across various downstream NLP tasks, especially when synergized with powerful large language models (LLMs). Existing studies evaluate ICL methods primarily based on downstream task performance. This evaluation protocol overlooks the significant cost associated with the demonstration configuration process, i.e., tuning the demonstration as the ICL prompt. However, in this work, we point out that the evaluation protocol leads to unfair comparisons and potentially biased evaluation, because we surprisingly find the correlation between the configuration costs and task performance. Then we call for a two-dimensional evaluation paradigm that considers both of these aspects, facilitating a fairer comparison.Finally, based on our empirical finding that the optimized demonstration on one language model generalizes across language models of different sizes, we introduce a simple yet efficient strategy that can be applied to any ICL method as a plugin, yielding a better trade-off between the two dimensions according to the proposed evaluation paradigm.

Recent advancements in State Space Models (SSMs) have attracted significant interest, particularly in models optimized for parallel training and handling long-range dependencies. Architectures like Mamba have scaled to billions of parameters with selective SSM. To facilitate broader applications using Mamba, exploring its efficiency is crucial. While token reduction techniques offer a straightforward post-training strategy, we find that applying existing methods directly to SSMs leads to substantial performance drops. Through insightful analysis, we identify the reasons for this failure and the limitations of current techniques. In response, we propose a tailored, unified post-training token reduction method for SSMs. Our approach integrates token importance and similarity, thus taking advantage of both pruning and merging, to devise a fine-grained intra-layer token reduction strategy. Extensive experiments show that our method improves the average accuracy by 5.7% to 13.1% on six benchmarks with Mamba-2 compared to existing methods, while significantly reducing computational demands and memory requirements.

This work suggests fundamentally rethinking the current practice of pruning large language models (LLMs). The way it is done is by divide and conquer: split the model into submodels, sequentially prune them, and reconstruct predictions of the dense counterparts on small calibration data one at a time; the final model is obtained simply by putting the resulting sparse submodels together. While this approach enables pruning under memory constraints, it generates high reconstruction errors. In this work, we first present an array of reconstruction techniques that can significantly reduce this error by more than 90%. Unwittingly, however, we discover that minimizing reconstruction error is not always ideal and can overfit the given calibration data, resulting in rather increased language perplexity and poor performance at downstream tasks. We find out that a strategy of self-generating calibration data can mitigate this trade-off between reconstruction and generalization, suggesting new directions in the presence of both benefits and pitfalls of reconstruction for pruning LLMs.

This study addresses the challenges of assessing and enhancing social-pragmatic inference in large language models (LLMs). We first highlight the inadequacy of current accuracy-based multiple choice question answering (MCQA) formats in assessing social-pragmatic reasoning, and propose the direct evaluation of models’ free-form responses as measure, which correlates better with human judgment. Furthermore, we explore methods to improve pragmatic abilities in LLMs, advocating for preference optimization (PO) over supervised finetuning (SFT), given the absence of a definitive “gold” answer in social contexts. Our results show that preferential tuning consistently outperforms SFT across pragmatic phenomena and offers a near-free launch in pragmatic abilities without compromising general capabilities. Lastly, we examine the internal structure of LLMs, revealing that the significant boost in pragmatic reasoning is tied to deeper layer representations, analogous to human high-level thinking. Our experiments span a variety of pragmatic and social reasoning datasets, as well as an image referential game requiring a multimodal theory of mind (ToM). With our refined paradigms for evaluating and enhancing pragmatic inference, this paper offers key insights into building more socially aware language models.

Machine *unlearning* refers to methods for deleting information about specific training instances from a trained machine learning model. This enables models to delete user information and comply with privacy regulations. While retraining the model from scratch on the training set excluding the instances to be “*forgotten*” would result in a desired unlearned model, owing to the size of datasets and models, it is infeasible. Hence, unlearning algorithms have been developed, where the goal is to obtain an unlearned model that behaves as closely as possible to the retrained model. Consequently, evaluating an unlearning method involves - (i) randomly selecting a *forget* set (i.e., the training instances to be unlearned), (ii) obtaining an unlearned and a retrained model, and (iii) comparing the performance of the unlearned and the retrained model on the test and forget set. However, when the forget set is randomly selected, the unlearned model is almost often similar to the original (i.e., prior to unlearning) model. Hence, it is unclear if the model did really unlearn or simply copied the weights from the original model. For a more robust evaluation, we instead propose to consider training instances with significant influence on the trained model. When such influential instances are considered in the forget set, we observe that the unlearned model deviates significantly from the retrained model. Such deviations are also observed when the size of the forget set is increased. Lastly, choice of dataset for evaluation could also lead to misleading interpretation of results.

Large Language Models (LLMs) have demonstrated impressive capabilities in understanding and generating codes. Due to these capabilities, many recent methods are proposed to automatically refine the codes with LLMs. However, we should rethink that the refined codes (from LLMs and even humans) are not always more efficient than their original versions. On the other hand, running two different versions of codes and comparing them every time is not ideal and time-consuming. Therefore, in this work, we propose a novel method based on the code language model that is trained to judge the efficiency between two different codes (generated across humans and machines) by either classifying the superior one or predicting the relative improvement. We validate our method on multiple programming languages with multiple refinement steps, demonstrating that the proposed method can effectively distinguish between more and less efficient versions of code.

As the rapidly advancing domain of natural language processing (NLP), large language models (LLMs) have emerged as powerful tools for interpreting human commands and generating text across various tasks. Nonetheless, the resilience of LLMs to handle text containing inherent errors, stemming from human interactions and collaborative systems, has not been thoroughly explored. Our study investigates the resilience of LLMs against five common types of disruptions including 1) ASR (Automatic Speech Recognition) errors, 2) OCR (Optical Character Recognition) errors, 3) grammatical mistakes, 4) typographical errors, and 5) distractive content. We aim to investigate how these models react by deliberately embedding these errors into instructions. Our findings reveal that while some LLMs show a degree of resistance to certain types of noise, their overall performance significantly suffers. This emphasizes the importance of further investigation into enhancing model resilience. In response to the observed decline in performance, our study also evaluates a “re-pass” strategy, designed to purify the instructions of noise before the LLMs process them. Our analysis indicates that correcting noisy instructions, particularly for open-source LLMs, presents significant challenges.

We tackle societal bias in image-text datasets by removing spurious correlations between protected groups and image attributes. Traditional methods only target labeled attributes, ignoring biases from unlabeled ones. Using text-guided inpainting models, our approach ensures protected group independence from all attributes and mitigates inpainting biases through data filtering. Evaluations on multi-label image classification and image captioning tasks show our method effectively reduces bias without compromising performance across various models. Specifically, we achieve an average societal bias reduction of 46.1% in leakage-based bias metrics for multi-label classification and 74.8% for image captioning.

In this paper, we investigate the capability of Large Language Models (LLMs) to represent texts in multilingual contexts. Our findings show that sentence representations derived from LLMs exhibit a high degree of isomorphism across languages.This existing isomorphism can facilitate representational alignments in zero-shot and few-shot settings.Specifically, by applying a contrastive objective at the representation level with only a small number of translation pairs (e.g., 100), we substantially improve models’ performance on Semantic Textual Similarity (STS) tasks across languages. This representation-level approach proves to be more efficient and effective for semantic alignment than continued pretraining or instruction tuning. Interestingly, we also observe substantial STS improvements within individual languages, even without a monolingual objective specifically designed for this purpose.

Entity tracking is essential for complex reasoning. To perform in-context entity tracking, language models (LMs) must bind an entity to its attribute (e.g., bind a container to its content) to recall attribute for a given entity. For example, given a context mentioning “The coffee is in Box Z, the stone is in Box M, the map is in Box H”, to infer “Box Z contains the coffee” later, LMs must bind “Box Z” to “coffee”. To explain the binding behaviour of LMs, existing research introduces a Binding ID mechanism and states that LMs use a abstract concept called Binding ID (BI) to internally mark entity-attribute pairs. However, they have not directly captured the BI information from entity activations. In this work, we provide a novel view of the Binding ID mechanism by localizing the BI information. Specifically, we discover that there exists a low-rank subspace in the hidden state (or activation) of LMs, that primarily encodes BIs. To identify this subspace, we take principle component analysis as our first attempt and it is empirically proven to be effective. Moreover, we also discover that when editing representations along directions in the subspace, LMs tend to bind a given entity to other attributes accordingly. For example, by patching activations along the BI encoding direction we can make the LM to infer “Box Z contains the stone” and “Box Z contains the map”.

With the strong representational capabilities of pre-trained language models, dependency parsing in resource-rich languages has seen significant advancements. However, the parsing accuracy drops sharply when the model is transferred to low-resource language due to distribution shifts. To alleviate this issue, we propose a representation alignment and adversarial model to filter out useful knowledge from rich-resource language and ignore useless ones. Our proposed model consists of two components, i.e., an alignment network in the input layer for selecting useful language-specific features and an adversarial network in the encoder layer for augmenting the language-invariant contextualized features. Experiments on the benchmark datasets show that our proposed model outperforms RoBERTa-enhanced strong baseline models by 1.37 LAS and 1.34 UAS. Detailed analysis shows that both alignment and adversarial networks are equally important in alleviating the distribution shifts problem and can complement each other. In addition, the comparative experiments demonstrate that both the alignment and adversarial networks can substantially facilitate extracting and utilizing relevant target language features, thereby increasing the adaptation capability of our proposed model.

Recent developments in the quality and accessibility of large language models have precipitated a surge in user-facing tools for content generation. Motivated by a necessity for human quality control of these systems, we introduce ReportGPT: a pipeline framework for verifiable human-in-the-loop table-to-text generation. ReportGPT is based on a domain specific language, which acts as a proof mechanism for generating verifiable commentary. This allows users to quickly check the relevancy and factuality of model outputs. User selections then become few-shot examples for improving the performance of the pipeline. We configure 3 approaches to our pipeline, and find that usage of language models in ReportGPT’s components trade off precision for more insightful downstream commentary. Furthermore, ReportGPT learns from human feedback in real-time, needing only a few samples to improve performance.