EMNLP 2024 Papers

We present a new approach called MeritOpt based on the Personalized Federated Learning algorithm MeritFed that can be applied to Natural Language Tasks with heterogeneous data. We evaluate it on the Low-Resource Machine Translation task, using the datasets of South East Asian and Finno-Ugric languages. In addition to its effectiveness, MeritOpt is also highly interpretable, as it can be applied to track the impact of each language used for training. Our analysis reveals that target dataset size affects weight distribution across auxiliary languages, that unrelated languages do not interfere with the training, and auxiliary optimizer parameters have minimal impact. Our approach is easy to apply with a few lines of code, and we provide scripts for reproducing the experiments (https://github.com/VityaVitalich/MeritOpt).

The recent integration of chemistry with natural language processing (NLP) has advanced drug discovery. Molecule representation in language models (LMs) is crucial in enhancing chemical understanding. We propose Augmented Molecular Retrieval (AMORE), a flexible zero-shot framework for assessment of Chemistry LMs of different natures: trained solely on molecules for chemical tasks and on a combined corpus of natural language texts and string-based structures. The framework relies on molecule augmentations that preserve an underlying chemical, such as kekulization and cycle replacements. We evaluate encoder-only and generative LMs by calculating a metric based on the similarity score between distributed representations of molecules and their augmentations. Our experiments on ChEBI-20 and QM9 benchmarks show that these models exhibit significantly lower scores than graph-based molecular models trained without language modeling objectives. Additionally, our results on the molecule captioning task for cross-domain models, MolT5 and Text+Chem T5, demonstrate that the lower the representation-based evaluation metrics, the lower the classical text generation metrics like ROUGE and METEOR.

We present LoCoVQA, a dynamic benchmark generator for evaluating long-context reasoning in vision language models (VLMs). LoCoVQA augments test examples for mathematical reasoning, VQA, and character recognition tasks with increasingly long visual contexts composed of both in-distribution and out-of-distribution distractor images.Across these tasks, a diverse set of VLMs rapidly lose performance as the visual context length grows, often exhibiting a striking logarithmic decay trend. This test assesses how well VLMs can ignore irrelevant information when answering queries—a task that is quite easy for language models (LMs) in the text domain—demonstrating that current state-of-the-art VLMs lack this essential capability for many long-context applications.

When asked to summarize articles or answer questions given a passage, large language models (LLMs) can hallucinate details and respond with unsubstantiated answers that are inaccurate with respect to the input context. This paper describes a simple approach for detecting such **contextual hallucinations**. We hypothesize that contextual hallucinations are related to the extent to which an LLM attends to information in the provided context versus its own generations. Based on this intuition, we propose a simple hallucination detection model whose input features are given by the ratio of attention weights on the context versus newly generated tokens (for each attention head). We find that a linear classifier based on these _lookback ratio_ features is as effective as a richer detector that utilizes the entire hidden states of an LLM or a text-based entailment model. The lookback ratio-based detector—**Lookback Lens**—is found to transfer across tasks and even models, allowing a detector that is trained on a 7B model to be applied (without retraining) to a larger 13B model. We further apply this detector to mitigate contextual hallucinations, and find that a simple classifier-guided decoding approach is able to reduce the amount of hallucination, for example by 9.6% in the XSum summarization task.

Large language model-based steganography encodes hidden messages into model-generated tokens. The key tradeoff is between how much hidden information can be introduced and how much the model can be perturbed. To address this tradeoff, we show how to adapt strategies previously used for LLM watermarking to encode large amounts of information. We tackle the practical (but difficult) setting where we do not have access to the full model when trying to recover the hidden information. Theoretically, we study the fundamental limits in how much steganographic information can be inserted into LLM-created outputs. We provide practical encoding schemes and present experimental results showing that our proposed strategies are nearly optimal.

The quality of training data is crucial for enhancing the long-text capabilities of foundation models. Despite existing efforts to refine data quality through heuristic rules and evaluations based on data diversity and difficulty, there’s a lack of systematic approaches specifically tailored for assessing long texts. Addressing this gap, our work systematically measures the quality of long texts by evaluating three fundamental linguistic dimensions: coherence, cohesion, and complexity. Drawing inspiration from the aforementioned three dimensions, we introduce a suite of metrics designed to evaluate the quality of long texts, encompassing both statistical and pre-trained language model-based ones. Leveraging these metrics, we present LongWanjuan, a bilingual dataset specifically tailored to enhance the training of language models for long-text tasks with over 160B tokens. In LongWanjuan, we categorize long texts into holistic, aggregated, and chaotic types, enabling a detailed analysis of long-text quality. Furthermore, we devise a data mixture recipe that strategically balances different types of long texts within LongWanjuan, leading to significant improvements in model performance on long-text tasks.

Long-Context Question Answering (LCQA), a challenging task, aims to reason over long-context documents to yield accurate answers to questions. Existing long-context Large Language Models (LLMs) for LCQA often struggle with the “lost in the middle” issue. Retrieval-Augmented Generation (RAG) mitigates this issue by providing external factual evidence. However, its chunking strategy disrupts the global long-context information, and its low-quality retrieval in long contexts hinders LLMs from identifying effective factual details due to substantial noise. To this end, we propose LongRAG, a general, dual-perspective, and robust LLM-based RAG system paradigm for LCQA to enhance RAG’s understanding of complex long-context knowledge (i.e., global information and factual details). We design LongRAG as a plug-and-play paradigm, facilitating adaptation to various domains and LLMs. Extensive experiments on three multi-hop datasets demonstrate that LongRAG significantly outperforms long-context LLMs (up by 6.94%), advanced RAG (up by 6.16%), and Vanilla RAG (up by 17.25%). Furthermore, we conduct quantitative ablation studies and multi-dimensional analyses, highlighting the effectiveness of the system’s components and fine-tuning strategies.Data and code are available at [https://github.com/QingFei1/LongRAG](https://github.com/QingFei1/LongRAG).

Large language models (LLMs) have achieved impressive performance in numerous domains but often struggle to process lengthy inputs effectively and efficiently due to limited length generalization and attention’s quadratic computational demands. Many sought to mitigate this by restricting the attention window within the pre-trained length. However, these methods introduce new issues such as ignoring the middle context and requiring additional training. To address these problems, we propose LongHeads, a training-free framework that enhances LLM’s long context ability by unlocking multi-head attention’s untapped potential. Instead of allowing each head to attend to the full sentence, which struggles with generalizing to longer sequences due to out-of-distribution (OOD) issues, we allow each head to process in-distribution length by selecting and attending to important context chunks. To this end, we propose a chunk selection strategy that relies on the inherent correlation between the query and the key representations, efficiently distributing context chunks to different heads. In this way, each head ensures it can effectively process attended tokens within the trained length, while different heads in different layers can collectively process longer contexts. LongHeads works efficiently and fits seamlessly with many LLMs that use relative positional encoding. LongHeads achieves 100% accuracy at the 128k length on passkey retrieval task, verifying LongHeads’ efficacy in extending the usable context window for existing models.

Current long-context benchmarks primarily focus on retrieval-based tests, requiring Large Language Models (LLMs) to locate specific information within extensive input contexts, such as the needle-in-a-haystack (NIAH) benchmark. Long-context generation refers to the ability of a language model to generate coherent and contextually accurate text that spans across lengthy passages or documents. While recent studies show strong performance on NIAH and other retrieval-based long-context benchmarks, there is a significant lack of benchmarks for evaluating long-context generation capabilities. To bridge this gap and offer a comprehensive assessment, we introduce a synthetic benchmark, LongGenBench, which allows for flexible configurations of customized generation context lengths. LongGenBench advances beyond traditional benchmarks by redesigning the format of questions and necessitating that LLMs respond with a single, cohesive long-context answer. Upon extensive evaluation using LongGenBench, we observe that: (1) both API accessed and open source models exhibit performance degradation in long-context generation scenarios, ranging from 1.2% to 47.1%; (2) different series of LLMs exhibit varying trends of performance degradation, with the Gemini-1.5-Flash model showing the least degradation among API accessed models, and the Qwen2 series exhibiting the least degradation in LongGenBench among open source models.

Instruction tuning enables language models to more effectively generalize and better follow user intent. However, obtaining instruction data is costly and challenging. Prior work employs methods such as expensive human annotation, crowd-sourced datasets with alignment issues, and generating noisy examples via LLMs. We introduce the LongForm-C dataset, which is created by reverse instructions. We generate instructions via LLMs for human-written corpus examples using reverse instructions. First we select a diverse set of human-written documents from corpora such as C4 and Wikipedia; then we generate instructions for these documents via LLMs. This approach provides a cheaper and cleaner instruction-tuning dataset with natural output and one suitable for long text generation. Our models outperform 10x larger language models without instruction tuning on tasks such as story/recipe generation and long-form question answering. Moreover, LongForm models outperform prior instruction-tuned models such as FLAN-T5 and Alpaca by a large margin, and improve language understanding capabilities further. We publicly release our data and models: [Anonymized-URL].

Embedding models play a pivotal role in modern NLP applications such as document retrieval. However, existing embedding models are limited to encoding short documents of typically 512 tokens, restrained from application scenarios requiring long inputs. This paper explores context window extension of existing embedding models, pushing their input length to a maximum of 32,768. We begin by evaluating the performance of existing embedding models using our newly constructed LongEmbed benchmark, which includes two synthetic and four real-world tasks, featuring documents of varying lengths and dispersed target information. The benchmarking results highlight huge opportunities for enhancement in current models. Via comprehensive experiments, we demonstrate that training-free context window extension strategies can effectively increase the input length of these models by several folds. Moreover, comparison of models using Absolute Position Encoding (APE) and Rotary Position Encoding (RoPE) reveals the superiority of RoPE-based embedding models in context window extension, offering empirical guidance for future models. Our benchmark, code and trained models will be released to advance the research in long context embedding models.

Extending large language models to effectively handle long contexts requires instruction fine-tuning on input sequences of similar length. To address this, we present LongAlign—a recipe of the instruction data, training, and evaluation for long context alignment. First, we construct a long instruction-following dataset using Self-Instruct. To ensure the data diversity, it covers a broad range of tasks from various long context sources. Second, we adopt the packing and sorted batching strategies to speed up supervised fine-tuning on data with varied length distributions. Additionally, we develop a loss weighting method to balance the contribution to the loss across different sequences during packing training. Third, we introduce the LongBench-Chat benchmark for evaluating instruction-following capabilities on queries of 10k-100k in length. Experiments show that LongAlign outperforms existing recipes for LLMs in long context tasks by up to 30%, while also maintaining their proficiency in handling short, generic tasks.

As the demand for processing extended textual data grows, the ability to handle long-range dependencies and maintain computational efficiency is more critical than ever. One of the key issues for long-sequence modeling using attention-based model is the mismatch between the limited-range modeling power of full attention and the long-range token dependency in the input sequence. In this work, we propose to scale up the attention receptive field by tensorizing long input sequences into compact tensor representations followed by attention on each transformed dimension. The resulting Tensorized Attention can be adopted as efficient transformer backbones to extend input context length with improved memory and time efficiency. We show that the proposed attention tensorization encodes token dependencies as a multi-hop attention process, and is equivalent to Kronecker decomposition of full attention. Extensive experiments show that tensorized attention can be used to adapt pretrained LLMs with improved efficiency. Notably, using customized Triton kernels, tensorization enables Llama-8B training under 32,768 context length and can steadily extrapolate to 128k length during inference with 11 times speedup (compared to full attention with FlashAttention-2).

Pre-trained Language Models (PLMs) have achieved significant success in text classification. However, they still face challenges with hard samples, which refer to instances where the model exhibits diminished confidence in distinguishing new samples. Existing research has addressed related issues, but often overlooks the semantic information inherent in the labels, treating them merely as one-hot vectors. In this paper, we propose Logits Reranking via Semantic Labels (LRSL), a model-agnostic post-processing method that leverages label semantics and auto detection of hard samples to improve classification accuracy. LRSL automatically identifies hard samples, which are then jointly processed by MLP-based and Similarity-based approaches. Applied only during inference, LRSL operates solely on classification logits, reranking them based on semantic similarities without interfering with the model’s training process. The experiments demonstrate the effectiveness of our method, showing significant improvements across different PLMs. Our codes are publicly available at https://github.com/SIGSDSscau/LRSL.

Document-level relation extraction (DocRE) aims to identify relationships between entities within a document. Due to the vast number of entity pairs, fully annotating all fact triplets is challenging, resulting in datasets with numerous false negative samples. Recently, self-training-based methods have been introduced to address this issue. However, these methods are purely black-box and sub-symbolic, making them difficult to interpret and prone to overlooking symbolic interdependencies between relations.To remedy this deficiency, our insight is that symbolic knowledge, such as logical rules, can be used as diagnostic tools to identify conflicts between pseudo-labels. By resolving these conflicts through logical diagnoses, we can correct erroneous pseudo-labels, thus enhancing the training of neural models.To achieve this, we propose **LogicST**, a neural-logic self-training framework that iteratively resolves conflicts and constructs the minimal diagnostic set for updating models. Extensive experiments demonstrate that LogicST significantly improves performance and outperforms previous state-of-the-art methods. For instance, LogicST achieves an increase of **7.94%** in F1 score compared to CAST (Tan et al., 2023a) on the DocRED benchmark (Yao et al., 2019). Additionally, LogicST is more time-efficient than its self-training counterparts, requiring only **10%** of the training time of CAST.

We introduce LogicAsker, a novel approach for evaluating and enhancing the logical reasoning capabilities of large language models (LLMs) such as ChatGPT and GPT-4. Despite LLMs’ prowess in tasks like writing assistance, code generation, and machine translation, assessing their ability to reason has been challenging. Traditional evaluations often prioritize accuracy on downstream tasks over direct assessments of reasoning processes. LogicAsker addresses this gap by employing a set of atomic reasoning skills grounded in propositional and predicate logic to systematically examine and improve the reasoning prowess of LLMs. Our methodology reveals significant gaps in LLMs’ learning of logical rules, with identified reasoning failures ranging from 29% to 90% across different models. Moreover, we leverage these findings to construct targeted demonstration examples and fine-tune data, notably enhancing logical reasoning in models like GPT-4o by up to 5%. To our knowledge, this is the first effort to utilize test case outcomes to effectively refine LLMs’ formal reasoning capabilities. We make our code, data, and results publicly available(https://github.com/yxwan123/LogicAsker) to facilitate further research and replication of our findings.

To explain social phenomena and identify systematic biases, much research in computational social science focuses on comparative text analyses. These studies often rely on coarse corpus-level statistics or local word-level analyses, mainly in English. We introduce the InfoGap method—an efficient and reliable approach to locating information gaps and inconsistencies in articles at the fact level, across languages. We evaluate InfoGap by analyzing LGBT people’s portrayals, across 2.7K biography pages on English, Russian, and French Wikipedias. We find large discrepancies in factual coverage across the languages. Moreover, our analysis reveals that biographical facts carrying negative connotations are more likely to be highlighted in Russian Wikipedia. Crucially, InfoGap both facilitates large scale analyses, and pinpoints local document- and fact-level information gaps, laying a new foundation for targeted and nuanced comparative language analysis at scale.

NLP research on aligning lexical representation spaces to one another has so far focused on aligning language spaces in their entirety. However, cognitive science has long focused on a local perspective, investigating whether translation equivalents truly share the same meaning or the extent that cultural and regional influences result in meaning variations. With recent technological advances and the increasing amounts of available data, the longstanding question of cross-lingual lexical alignment can now be approached in a more data-driven manner. However, developing metrics for the task requires some methodology for comparing metric efficacy. We address this gap and present a methodology for analyzing both synthetic validations and a novel naturalistic validation using lexical gaps in the kinship domain.We further propose new metrics, hitherto unexplored on this task, based on contextualized embeddings. Our analysis spans 16 diverse languages, demonstrating that there is substantial room for improvement with the use of newer language models. Our research paves the way for more accurate and nuanced cross-lingual lexical alignment methodologies and evaluation.

Text generation, a component in applications such as dialogue systems, relies heavily on decoding algorithms that sample strings from a language model distribution. Traditional methods like top-k and top-𝜋 decoding locally normalise the model’s output, which can significantly distort the original distribution. In this paper, we investigate the effects of such distortions by introducing globally-normalised versions of these decoding methods. Further, we propose an independent Metropolis-Hastings (IMH) algorithm to approximate sampling from these globally-normalised distributions without explicitly computing them. Our empirical analyses compare the performance of local and global decoding across two algorithms (top-k and top-𝜋) with various hyperparameters, using the Pythia language models. Results show that in most configuration, global decoding performs worse than the local decoding versions of the same algorithms, despite preserving the distribution’s integrity. Our results thus suggest that distortion might be an important feature of local decoding algorithms.