EMNLP 2024 Papers

Can human reading comprehension be assessed from eye movements in reading? In this work, we address this longstanding question using large-scale eyetracking data. We focus on a cardinal and largely unaddressed variant of this question: predicting reading comprehension of a single participant for a single question from their eye movements over a single paragraph. We tackle this task using a battery of recent models from the literature, and three new multimodal language models. We evaluate the models in two different reading regimes: ordinary reading and information seeking, and examine their generalization to new textual items, new participants, and the combination of both. The evaluations suggest that the task is highly challenging, and highlight the importance of benchmarking against a strong text-only baseline. While in some cases eye movements provide improvements over such a baseline, they tend to be small. This could be due to limitations of current modelling approaches, limitations of the data, or because eye movement behavior does not sufficiently pertain to fine-grained aspects of reading comprehension processes. Our study provides an infrastructure for making further progress on this question.

Solidarity is a crucial concept to understand social relations in societies. In this study, we investigate the frequency of (anti-)solidarity towards women and migrants in German parliamentary debates between 1867 and 2022. Using 2,864 manually annotated text snippets, we evaluate large language models (LLMs) like Llama 3, GPT-3.5, and GPT-4. We find that GPT-4 outperforms other models, approaching human annotation accuracy. Using GPT-4, we automatically annotate 18,300 further instances and find that solidarity with migrants outweighs anti-solidarity but that frequencies and solidarity types shift over time. Most importantly, group-based notions of (anti-)solidarity fade in favor of compassionate solidarity, focusing on the vulnerability of migrant groups, and exchange-based anti-solidarity, focusing on the lack of (economic) contribution. This study highlights the interplay of historical events, socio-economic needs, and political ideologies in shaping migration discourse and social cohesion.

This paper investigates the impact of different Byte Pair Encoding (BPE) configurations, specifically, merge operations on neural machine translation (NMT) performance for the Filipino-Cebuano language pair across various text domains. Results demonstrate that smaller BPE configurations, notably 2k, 5k, and 8k consistently yield higher BLEU scores, indicating improved translation quality through finer tokenization granularity. Conversely, larger BPE configurations and the absence of BPE result in lower BLEU scores, suggesting a decline in translation quality due to coarser tokenization. Additionally, these findings help us understand how the size of the model and how finely we break down words affect the quality of translations. This knowledge will be useful for improving translation systems, especially for languages that don’t have many parallel texts available for training.

Large Language Models (LLMs) are increasingly relied upon to evaluate text outputs of other LLMs, thereby influencing leaderboards and development decisions. However, concerns persist over the accuracy of these assessments and the potential for misleading conclusions. In this work, we investigate the effectiveness of LLMs as evaluators for text generation tasks. We propose FBI, a novel framework designed to examine the proficiency of Evaluator LLMs in assessing four critical abilities in other LLMs: factual accuracy, instruction following, coherence in long-form writing, and reasoning proficiency. By introducing targeted perturbations in answers generated by LLMs, that clearly impact one of these key capabilities, we test whether an Evaluator LLM can detect these quality drops. By creating a total of 2400 perturbed answers covering 22 perturbation categories, we conduct a comprehensive study using different evaluation strategies on five prominent LLMs commonly used as evaluators in the literature. Our findings reveal significant shortcomings in current Evaluator LLMs, which failed to identify quality drops in over 50% of cases on average. Single-answer and pairwise evaluations demonstrated notable limitations, whereas reference-based evaluations showed comparatively better performance. These results underscore the unreliable nature of current Evaluator LLMs and advocate for cautious implementation in practical applications.

There is a variety of multimodal data pertinent to public companies, spanning from accounting statements, macroeconomic statistics, earnings conference calls, and financial reports. These diverse modalities capture the state of firms from a variety of different perspectives but requires complex interactions to reconcile in the formation of accurate financial predictions. The commonality between these different modalities is that they all represent a time series, typically observed for a particular firm at a quarterly horizon, providing the ability to model trends and variations of company data over time. However, the time series of these diverse modalities contains varying temporal and cross-channel patterns that are challenging to model without the appropriate inductive biases. In this work, we design a novel multimodal time series prediction task that includes numerical financial results, macroeconomic states, and long financial documents to predict next quarter’s company earnings relative to analyst expectations. We explore a variety of approaches for this novel setting, establish strong unimodal baselines, and propose a multimodal model that exhibits state-of-the-art performance on this unique task. We demonstrate that each modality contains unique information and that the best performing model requires careful fusion of the different modalities in a multi-stage training approach. To better understand model behavior, we conduct a variety of probing experiments, reveal insights into the value of different modalities, and demonstrate the practical utility of our proposed method in a simulated trading setting.

We introduce FinDVer, a comprehensive benchmark specifically designed to evaluate the explainable claim verification capabilities of LLMs in the context of understanding and analyzing long, hybrid-content financial documents. FinDVer contains 4,000 expert-annotated examples across four subsets, each focusing on a type of scenario that frequently arises in real-world financial domains. We assess a broad spectrum of 25 LLMs under long-context and RAG settings. Our results show that even the current best-performing system (i.e., GPT-4o) significantly lags behind human experts. Our detailed findings and insights highlight the strengths and limitations of existing LLMs in this new task. We believe FinDVer can serve as a valuable benchmark for evaluating LLM capabilities in claim verification over complex, expert-domain documents.

Reinforcement learning from human feedback (RLHF) plays a crucial role in aligning language models with human preferences. While the significance of dataset quality is generally recognized, explicit investigations into its impact within the RLHF framework, to our knowledge, have been limited. This paper addresses the issue of text quality within the preference dataset by focusing on direct preference optimization (DPO), an increasingly adopted reward-model-free RLHF method. We confirm that text quality significantly influences the performance of models optimized with DPO more than those optimized with reward-model-based RLHF. Building on this new insight, we propose an extension of DPO, termed filtered direct preference optimization (fDPO). fDPO uses a trained reward model to monitor the quality of texts within the preference dataset during DPO training. Samples of lower quality are discarded based on comparisons with texts generated by the model being optimized, resulting in a more accurate dataset. Experimental results demonstrate that fDPO enhances the final model performance. Our code is available at https://github.com/CyberAgentAILab/filtered-dpo.

As machine learning models continue to swiftly advance, calibrating their performance has become a major concern prior to practical and widespread implementation. Most existing calibration methods often negatively impact model accuracy due to the lack of diversity of validation data, resulting in reduced generalizability. To address this, we propose a calibration method that incorporates synthetic data without compromising accuracy. We derive the expected calibration error (ECE) bound using the Probably Approximately Correct (PAC) learning framework. Large language models (LLMs), known for their ability to mimic real data and generate text with mixed class labels, are utilized as a synthetic data generation strategy to lower the ECE bound and improve model accuracy on real test data. Additionally, we propose data generation mechanisms for efficient calibration. Testing our method on four different natural language processing tasks, we observed an average up to 34% increase in accuracy and 33% decrease in ECE.

While fine-tuning of pre-trained language models generally helps to overcome the lack of labelled training samples, it also displays model performance instability. This instability mainly originates from randomness in initialisation or data shuffling. To address this, researchers either modify the training process or augment the available samples, which typically results in increased computational costs. We propose a new mitigation strategy, called **Delayed Ensemble with Noisy Interpolation (DENI)**, that leverages the strengths of ensembling, noise regularisation and model interpolation, while retaining computational efficiency. We compare DENI with 9 representative mitigation strategies across 3 models, 4 tuning strategies and 7 text classification datasets. We show that: 1) DENI outperforms the best performing mitigation strategy (Ensemble), while using only a fraction of its cost; 2) the mitigation strategies are beneficial for parameter-efficient fine-tuning (PEFT) methods, outperforming full fine-tuning in specific cases; and 3) combining DENI with data augmentation often leads to even more effective instability mitigation.

Large Language Models (LLMs) have shown impressive capabilities, yet they still struggle with math reasoning. In this work, we propose CoT-Influx, a novel approach that pushes the boundary of few-shot Chain-of-Thoughts (CoT) learning to improve LLM mathematical reasoning. Motivated by the observation that adding more concise CoT examples in the prompt can improve LLM reasoning performance, CoT-Influx employs a coarse-to-fine pruner to maximize the input of effective and concise CoT examples. The pruner first selects as many crucial CoT examples as possible and then prunes unimportant tokens to fit the context window. As a result, by enabling more CoT examples with double the context window size in tokens, CoT-Influx significantly outperforms various prompting baselines across various LLMs (LLaMA2-7B, 13B, 70B) and 5 math datasets, achieving up to 4.55% absolute improvements. Remarkably, without any fine-tuning, LLaMA2-70B with CoT-Influx surpasses GPT-3.5 and a wide range of larger LLMs (PaLM, Minerva 540B, etc.) on the GSM8K. CoT-Influx is a plug-and-play module for LLMs, adaptable in various scenarios. It’s compatible with advanced reasoning prompting techniques, such as self-consistency, and supports different long-context LLMs, including Mistral-7B-v0.3-32K and Yi-6B-200K.

Large language models (LLMs) have become the preferred solution for many natural language processing tasks. In low-resource environments such as specialized domains, their few-shot capabilities are expected to deliver high performance. Named Entity Recognition (NER) is a critical task in information extraction that is not covered in recent LLM benchmarks. There is a need for better understanding the performance of LLMs for NER in a variety of settings including languages other than English. This study aims to evaluate generative LLMs, employed through prompt engineering, for few-shot clinical NER. We compare 13 auto-regressive models using prompting and 16 masked models using fine-tuning on 14 NER datasets covering English, French and Spanish. While prompt-based auto-regressive models achieve competitive F1 for general NER, they are outperformed within the clinical domain by lighter biLSTM-CRF taggers based on masked models. Additionally, masked models exhibit lower environmental impact compared to auto-regressive models. Findings are consistent across the three languages studied, which suggests that LLM prompting is not yet suited for NER production in the clinical domain.

A supervised ranking model, despite its effectiveness over traditional approaches, usually involves complex processing - typically multiple stages of task-specific pre-training and fine-tuning. This has motivated researchers to explore simpler pipelines leveraging large language models (LLMs) that can work in a zero-shot manner. However, since zero-shot inference does not make use of a training set of pairs of queries and their relevant documents, its performance is mostly worse than that of supervised models, which are trained on such example pairs. Motivated by the existing findings that training examples generally improve zero-shot performance, in our work, we explore if this also applies to ranking models. More specifically, given a query and a pair of documents, the preference prediction task is improved by augmenting examples of preferences for similar queries from a training set. Our proposed pairwise few-shot ranker demonstrates consistent improvements over the zero-shot baseline on both in-domain (TREC DL) and out-domain (BEIR subset) retrieval benchmarks.

In this paper, we propose a modified version of the MedQA-USMLE dataset, named MEDQA-OPEN, which contains open-ended medical questions without options to mimic clinical scenarios, along with clinician-approved reasoned answers. Additionally, we implement a prompt driven by Chain of Thought (CoT) reasoning, CLINICR, to mirror the prospective process of incremental reasoning, reaching a correct response to medical questions. We empirically demonstrate how CLINICR outperforms the state-of-the-art 5-shot CoT-based prompt (Liévin et al., 2022). We also present an approach that mirrors real-life clinical practice by first exploring multiple differential diagnoses through MCQ-CLINICR and subsequently narrowing down to a final diagnosis using MCQ-ELIMINATIVE. Finally, emphasizing the importance of response verification in medical settings, we utilize a reward model mechanism, replacing the elimination process performed by MCQ-ELIMINATIVE.

Large language models (LLMs) excel in generating coherent text, but they often struggle with context awareness, leading to inaccuracies in tasks requiring faithful adherence to provided information. We introduce FastMem, a novel method designed to enhance instruction fine-tuned LLMs’ context awareness through fast memorization of the prompt. FastMem maximizes the likelihood of the prompt before inference by updating only the last Feed-Forward Network (FFN) module. This targeted approach ensures efficient optimization without overfitting, significantly improving the model’s ability to comprehend and accurately follow the context. Our experiments demonstrate substantial gains in reading comprehension, text summarization and adherence to output structures. For instance, FastMem improves the accuracy of Llama 3-8B-Inst on the NQ-SWAP dataset from 59.1% to 71.6%, and reduces the output structure failure rate of Qwen 1.5-4B-Chat from 34.9% to 25.5%. Extensive experimental results highlight FastMem’s potential to offer a robust solution to enhance the reliability and accuracy of LLMs in various applications. Our code is available at: https://github.com/IAAR-Shanghai/FastMem.

In this study, we aim to explore Multitask Speech Language Model (SpeechLM) efficient inference via token reduction. Unlike other modalities such as vision or text, speech has unique temporal dependencies, making previous efficient inference works on other modalities not directly applicable. Furthermore, methods for efficient SpeechLM inference on long sequence and sparse signals remain largely unexplored. In this work, we propose FastAdaSP, a weighted token merging framework specifically designed for various speech-related tasks to improve the trade-off between efficiency and performance. Experimental results on WavLLM and Qwen-Audio show that our method achieves the state-of-the-art (SOTA) efficiency-performance trade-off compared with other baseline methods. Specifically, FastAdaSP achieved 7x memory efficiency and 1.83x decoding throughput without any degradation on tasks like Emotion Recognition (ER) and Spoken Question Answering (SQA).

The training of automatic speech recognition (ASR) with little to no supervised data remains an open question. In this work, we demonstrate that streaming Transformer-Transducer (TT) models can be trained from scratch in consumer and accessible GPUs in their entirety with pseudo-labeled (PL) speech from foundational speech models (FSM). This allows training a robust ASR model just in one stage and does not require large data and computational budget compared to the two-step scenario with pre-training and fine-tuning. We perform a comprehensive ablation on different aspects of PL-based streaming TT models such as the impact of (1) shallow fusion of n-gram LMs, (2) contextual biasing with named entities, (3) chunk-wise decoding for low-latency streaming applications, and (4) TT overall performance as the function of the FSM size. Our results demonstrate that TT can be trained from scratch without supervised data, even with very noisy PLs. We validate the proposed framework on 6 languages from CommonVoice and propose multiple heuristics to filter out hallucinated PLs.

The deployment of large language models (LLMs) is often constrained by memory bandwidth, where the primary bottleneck is the cost of transferring model parameters from the GPU’s global memory to its registers. When coupled with custom kernels that fuse the dequantization and matmul operations, weight-only quantization can thus enable faster inference by reducing the amount of memory movement. However, developing high-performance kernels for weight-quantized LLMs presents substantial challenges, especially when the weights are compressed to non-evenly-divisible bit widths (e.g., 3 bits) with non-uniform, lookup table (LUT) quantization. This paper describes FLUTE, a flexible lookup table engine for LUT-quantized LLMs, which uses offline restructuring of the quantized weight matrix to minimize bit manipulations associated with unpacking, and vectorization and duplication of the lookup table to mitigate shared memory bandwidth constraints. At batch sizes < 32 and quantization group size of 128 (typical in LLM inference), the FLUTE kernel can be 2-4x faster than existing GEMM kernels. As an application of FLUTE, we explore a simple extension to lookup table-based NormalFloat quantization and apply it to quantize LLaMA3 to various configurations, obtaining competitive quantization performance against strong baselines while obtaining an end-to-end throughput increase of 1.5 to 2 times.

Parameter efficient finetuning methods like low-rank adaptation (LoRA) aim to reduce the computational costs of finetuning pretrained Language Models (LMs). Enabled by these low-rank settings, we propose an even more efficient optimization strategy: Fast Forward, a simple and effective approach to accelerate large segments of SGD training. In a Fast Forward stage, we repeat the most recent optimizer step until the loss stops improving on a tiny validation set. By alternating between regular optimization steps and Fast Forward stages, Fast Forward provides up to an 87% reduction in FLOPs over standard SGD with Adam. We validate Fast Forward by finetuning various models on different tasks and demonstrate that it speeds up training without compromising model performance. Additionally, we analyze when and how to apply Fast Forward.

Full-parameter fine-tuning is computationally prohibitive for large language models (LLMs), making parameter-efficient fine-tuning (PEFT) methods like low-rank adaptation (LoRA) increasingly popular. However, LoRA and its existing variants introduce significant latency in multi-tenant settings, hindering their applications in the industry. To address this issue, we propose the Fantastic LoRA (FanLoRA) framework, which consists of four steps: (a) adding LoRA modules to all the Transformer linear weights and fine-tuning on a large-scale instruction tuning dataset. (b) The importance of each module is then assessed using a novel importance scoring method. (c) only the most critical modules per layer are retained, resulting in the FanLoRA setting. (d) The FanLoRA setting is applied to fine-tune various downstream tasks. Our extensive experiments demonstrate that: (a) FanLoRA outperforms existing PEFT baselines across a wide collection of tasks with comparable tunable parameters. (b) FanLoRA significantly reduces the inference latency of LoRA, making it valuable for further broadening the applications of LLMs in the industry.