NAACL 2024 论文列表

We introduce SOLAR 10.7B, a large language model (LLM) with 10.7 billion parameters, demonstrating superior performance in various natural language processing (NLP) tasks. Inspired by recent efforts to efficiently up-scale LLMs, we present a method for scaling LLMs called depth up-scaling (DUS), which encompasses depthwise scaling and continued pretraining. In contrast to other LLM up-scaling methods that use mixture-of-experts, DUS does not require complex changes to train and inference efficiently. We show experimentally that DUS is simple yet effective in scaling up high-performance LLMs from small ones. Building on the DUS model, we additionally present SOLAR 10.7B-Instruct, a variant fine-tuned for instruction-following capabilities, surpassing Mixtral-8x7B-Instruct. SOLAR 10.7B is publicly available under the Apache 2.0 license, promoting broad access and application in the LLM field.

Despite the recent advances in artificial intelligence, building social intelligence remains a challenge.Among social signals, laughter is one of the distinctive expressions that occurs during social interactions between humans.In this work, we tackle a new challenge for machines to understand the rationale behind laughter in video, Video Laugh Reasoning.We introduce this new task to explain why people laugh in a particular video and a dataset for this task.Our proposed dataset, SMILE, comprises video clips and language descriptions of why people laugh. We propose a baseline by leveraging the reasoning capacity of large language models (LLMs) with textual video representation. Experiments show that our baseline can generate plausible explanations for laughter. We further investigate the scalability of our baseline by probing other video understanding tasks and in-the-wild videos. We release our dataset, code, and model checkpoints on https://github.com/postech-ami/SMILE-Dataset.

Speculative decoding has emerged as a prominent alternative to autoregressive decoding for expediting inference in large language models (LLMs). However, prevailing assumptions often focus solely on latency reduction, neglecting the computational expenses. In this paper, we present Speculate Less, validate More (SLiM), a speculative decoding enhancement to reduce the speculation set while validating more effective tokens. SLiM is designed to mitigate LLMs’ computation costs associated with the token verification by introducing hypothesis reduction based on a fast posterior estimation. It consistently surpasses counterparts lacking cost reduction across a spectrum from CPU to GPU. Our evaluation with diverse conversational datasets shows that SLiM can achieve a substantial 70% reduction in FLOPs while generating more effective predictions on top of prior arts.

Reference-based metrics that operate at the sentence-level typically outperform quality estimation metrics, which have access only to the source and system output.This is unsurprising, since references resolve ambiguities that may be present in the source.In this paper, we investigate whether additional source context can effectively substitute for a reference.We present a metric named SLIDE (SLIding Document Evaluator), which operates on blocks of sentences. SLIDE leverages a moving window that slides over each document in the test set, feeding each chunk of sentences into an unmodified, off-the-shelf quality estimation model.We find that SLIDE obtains significantly higher pairwise system accuracy than its sentence-level baseline, in some cases even eliminating the gap with reference-base metrics.This suggests that source context may provide the same information as a human reference in disambiguating source ambiguities. This finding is especially pertinent for reference-free document-level evaluation, wherein SLIDE could provide higher-quality pairwise system assessments while only requiring document boundary annotations.

Contrastive learning, which utilizes positive pairs and in-batch negatives to optimize the loss objective, has been proven to be an effective method for learning sentence embeddings. However, we argue that the previous methods of constructing positive pairs only through dropout perturbation or entailment relation are limited. Since there is more sentence knowable information (SKI) to be mined, such as sentence external knowledge, semantic analysis, and grammatical description. In this work, we first hand-craft a simple and effective prompt template that is able to obtain the knowable information of input sentences from LLMs (e.g., LLaMA). Then we combine the original sentence and its knowable information to form a positive pair for contrastive learning. We evaluate our method on standard semantic textual similarity (STS) tasks. Experimental results show that our unsupervised and supervised models using BERTbase achieve an average of 78.65% and 82.45% Spearman’s correlation respectively, a 2.40% and 0.88% improvement compared to SimCSE. Our model outperforms the previous state-of-the-art model PromptBERT in both unsupervised and supervised settings and specifically yields a new state-of-the-art performance in supervised setting.

Knowledge base question generation (KBQG) aims to generate natural language questions from a set of triplet facts extracted from KB. Existing methods have significantly boosted the performance of KBQG via pre-trained language models (PLMs) thanks to the richly endowed semantic knowledge. With the advance of pre-training techniques, large language models (LLMs) (e.g., GPT-3.5) undoubtedly possess much more semantic knowledge. Therefore, how to effectively organize and exploit the abundant knowledge for KBQG becomes the focus of our study. In this work, we propose SGSH — a simple and effective framework to Stimulate GPT-3.5 with Skeleton Heuristics to enhance KBQG. The framework incorporates “skeleton heuristics”, which provides more fine-grained guidance associated with each input to stimulate LLMs to generate optimal questions, encompassing essential elements like the question phrase and the auxiliary verb.More specifically, we devise an automatic data construction strategy leveraging ChatGPT to construct a skeleton training dataset, based on which we employ a soft prompting approach to train a BART model dedicated to generating the skeleton associated with each input.Subsequently, skeleton heuristics are encoded into the prompt to incentivize GPT-3.5 to generate desired questions. Extensive experiments demonstrate that SGSH derives the new state-of-the-art performance on the KBQG tasks.

Recently proposed long-form question answering (QA) systems, supported by large language models (LLMs), have shown promising capabilities. Yet, attributing and verifying their generated abstractive answers can be difficult, and automatically evaluating their accuracy remains an ongoing challenge.In this work, we introduce a new QA task for answering multi-answer questions by summarizing multiple diverse sources in a semi-extractive fashion. Specifically, Semi-extractive Multi-source QA (SEMQA) requires models to output a comprehensive answer, while mixing factual quoted spans—copied verbatim from given input sources—and non-factual free-text connectors that glue these spans together into a single cohesive passage. This setting bridges the gap between the outputs of well-grounded but constrained extractive QA systems and more fluent but harder to attribute fully abstractive answers. Particularly, it enables a new mode for language models that leverages their advanced language generation capabilities, while also producing fine in-line attributions by-design that are easy to verify, interpret, and evaluate. To study this task, we create the first dataset of this kind, QuoteSum, with human-written semi-extractive answers to natural and generated questions, and define text-based evaluation metrics. Experimenting with several LLMs in various settings, we find this task to be surprisingly challenging, demonstrating the importance of QuoteSum for developing and studying such consolidation capabilities.

With the increasing risk posed by jailbreak attacks, recent studies have investigated various methods to improve the safety of large language models (LLMs), mainly falling into two strategies: safety training and safeguards. Safety training involves fine-tuning the LLM with adversarial samples, which activate the LLM’s capabilities against jailbreak. However, it is not always effective in countering new attacks and often leads to potential performance degradation. Safeguards, on the other hand, are methods using additional models to filter harmful content from the LLM’s response. Nevertheless, they can only reduce a limited amount of harmful output and introduce extra computational costs. Given the distinct strengths and weaknesses of both, we combine them to balance out their flaws and propose a more effective method called Self-Guard.Specifically, we train the LLM to review its responses for any harmful content and append a [harmful] or [harmless] tag to the end of the response. In this way, Self-Guard possesses the advantages of safety training, leveraging the powerful capabilities of the LLMs themselves to detect harmfulness. Besides that, it gains flexibility like safeguards, making the safety check target the output side, which makes the system less vulnerable to attack updates. Experimental results indicate that our Self-Guard can effectively defend against jailbreak attacks and will not cause LLMs’ performance degradation.

The advent of instruction-tuned large language models (LLMs) has significantly advanced the field of automatic instruction dataset augmentation. However, the method of generating instructions and outputs from inherent knowledge of LLM can unintentionally produce hallucinations — instances of generating factually incorrect or misleading information. To overcome this, we propose SELF-EXPERTISE, automatically generating instruction dataset in the legal domain from a seed dataset. SELF-EXPERTISE extracts knowledge from the outputs of the seed dataset, and generates new instructions, inputs, and outputs. In this way, the proposed method reduces hallucination in automatic instruction augmentation. We trained an SELF-EXPERTISE augmented instruction dataset on the LLaMA-2 7B model to construct Korean legal specialized model, called LxPERT. LxPERT has demonstrated performance surpassing GPT-3.5-turbo in both in-domain and out-of-domain datasets. The SELF-EXPERTISE augmentation pipeline is not only applicable to the legal field but is also expected to be extendable to various domains, potentially advancing domain-specialized LLMs.

Recent advances in named entity recognition (NER) have pushed the boundary of the task to incorporate visual signals, leading to many variants, including multi-modal NER (MNER) or grounded MNER (GMNER). A key challenge to these tasks is that the model should be able to generalize to the entities unseen during the training, and should be able to handle the training samples with noisy annotations.To address this obstacle, we propose SCANNER (Span CANdidate detection and recognition for NER), a model capable of effectively handling all three NER variants.SCANNER is a two-stage structure; we extract entity candidates in the first stage and use it as a query to get knowledge, effectively pulling knowledge from various sources.We can boost our performance by utilizing this entity-centric extracted knowledge to address unseen entities.Furthermore, to tackle the challenges arising from noisy annotations in NER datasets, we introduce a novel self-distillation method, enhancing the robustness and accuracy of our model in processing training data with inherent uncertainties.Our approach demonstrates competitive performance on the NER benchmark and surpasses existing methods on both MNER and GMNER benchmarks.Further analysis shows that the proposed distillation and knowledge utilization methods improve the performance of our model on various benchmarks.

The rapid development of Large Language Models (LLMs) has led to great strides in model capabilities like long-context understanding and reasoning.However, as LLMs are able to process longer contexts, it becomes more challenging to evaluate whether they have acquired certain capabilities, since the length of text (e.g., 200K tokens) they can process far exceeds what humans can reliably assess in a reasonable duration.In this paper, we propose using complex synthetic tasks as a proxy evaluation method, and present S3Eval, a Synthetic, Scalable, Systematic evaluation suite for LLMs evaluation.The synthetic nature of S3Eval provides users full control over the dataset, allowing them to systematically probe LLM capabilities by scaling text length and varying task difficulty across diverse scenarios.The strong correlation between S3Eval and real-world benchmarks demonstrates the soundness of using S3Eval for evaluation of LLMs.S3Eval provides a flexible and infinite long-context data generation method. We have generated a comprehensive dataset called S3Eval-Standard, and experimental results have shown that it poses significant challenges for all existing LLMs.

The complementary potential of Large Language Models (LLM) assumes off-the-shelf LLMs have heterogeneous expertise in a wide range of domains and tasks so that an ensemble of LLMs can achieve consistently better performance. Existing ensemble methods for LLMs mainly focus on reward model ranking of outputs, leading to significant computation overhead. To combat this issue, we revisit the complementary potential of LLMs and further elaborate on it by mining latent expertise with off-the-shelf reward models. We propose ZOOTER, a reward-guided routing method distilling rewards on training queries to train a routing function, which can precisely distribute each query to the LLM with expertise about it. We also integrate a tag-based label enhancement to mitigate noise from uncertainty when using rewards as silver supervision. ZOOTER shows computation efficiency in inference as it only introduces minor computation overhead of a routing function compared with reward model ranking methods. We evaluate ZOOTER on a comprehensive benchmark collection with 26 subsets in different domains and tasks. ZOOTER outperforms the best single model on average and ranks first on 44% of tasks, even surpassing multiple reward model ranking methods.

The growing interest in Large Language Models (LLMs) for specialized applications has revealed a significant challenge: when tailored to specific domains, LLMs tend to experience catastrophic forgetting, compromising their general capabilities and leading to a suboptimal user experience. Additionally, crafting a versatile model for multiple domains simultaneously often results in a decline in overall performance due to confusion between domains. In response to these issues, we present the RolE Prompting Guided Multi-Domain Adaptation (REGA) strategy. This novel approach effectively manages multi-domain LLM adaptation through three key components: 1) Self-Distillation constructs and replays general-domain exemplars to alleviate catastrophic forgetting. 2) Role Prompting assigns a central prompt to the general domain and a unique role prompt to each specific domain to minimize inter-domain confusion during training. 3) Role Integration reuses and integrates a small portion of domain-specific data to the general-domain data, which are trained under the guidance of the central prompt. The central prompt is used for a streamlined inference process, removing the necessity to switch prompts for different domains.Empirical results demonstrate that REGA effectively alleviates catastrophic forgetting and inter-domain confusion. This leads to improved domain-specific performance compared to standard fine-tuned models, while still preserving robust general capabilities.

Pre-trained language models (PLMs) have consistently demonstrated outstanding performance across a diverse spectrum of natural language processing tasks. Nevertheless, despite their success with unseen data, current PLM-based representations often exhibit poor robustness in adversarial settings. In this paper, we introduce RobustSentEmbed, a self-supervised sentence embedding framework designed to improve both generalization and robustness in diverse text representation tasks and against a diverse set of adversarial attacks. Through the generation of high-risk adversarial perturbations and their utilization in a novel objective function, RobustSentEmbed adeptly learns high-quality and robust sentence embeddings. Our experiments confirm the superiority of RobustSentEmbed over state-of-the-art representations. Specifically, Our framework achieves a significant reduction in the success rate of various adversarial attacks, notably reducing the BERTAttack success rate by almost half (from 75.51% to 38.81%). The framework also yields improvements of 1.59% and 0.23% in semantic textual similarity tasks and various transfer tasks, respectively.

We introduce RoDia, the first dataset for Romanian dialect identification from speech. The RoDia dataset includes a varied compilation of speech samples from five distinct regions of Romania, covering both urban and rural environments, totaling 2 hours of manually annotated speech data. Along with our dataset, we introduce a set of competitive models to be used as baselines for future research. The top scoring model achieves a macro F1 score of 59.83% and a micro F1 score of 62.08%, indicating that the task is challenging. We thus believe that RoDia is a valuable resource that will stimulate research aiming to address the challenges of Romanian dialect identification. We release our dataset at https://github.com/codrut2/RoDia.

In this work, we measure the impact of affixal negation on modern English large language models (LLMs). In affixal negation, the negated meaning is expressed through a negative morpheme, which is potentially challenging for LLMs as their tokenizers are often not morphologically plausible. We conduct extensive experiments using LLMs with different subword tokenization methods, which lead to several insights on the interaction between tokenization performance and negation sensitivity. Despite some interesting mismatches between tokenization accuracy and negation detection performance, we show that models can, on the whole, reliably recognize the meaning of affixal negation.

We characterize and study zero-shot abstractive summarization in Large Language Models (LLMs) by measuring position bias, which we propose as a general formulation of the more restrictive lead bias phenomenon studied previously in the literature. Position bias captures the tendency of a model unfairly prioritizing information from certain parts of the input text over others, leading to undesirable behavior. Through numerous experiments on four diverse real-world datasets, we study position bias in multiple LLM models such as GPT 3.5-Turbo, Llama-2, and Dolly-v2, as well as state-of-the-art pretrained encoder-decoder abstractive summarization models such as Pegasus and BART. Our findings lead to novel insights and discussion on performance and position bias of models for zero-shot summarization tasks.

While enabling large language models to implement function calling (known as APIs) can greatly enhance the performance of Large Language Models (LLMs), function calling is still a challenging task due to the complicated relations between different APIs, especially in a context-learning setting without fine-tuning. This paper introduces “Reverse Chain”, a controllable, target-driven approach designed to empower LLMs with the capability to operate external APIs only via prompts. Recognizing that most LLMs have limited tool-use capabilities, Reverse Chain limits LLMs to executing simple tasks, e.g., API Selection and Argument Completion. Furthermore, to manage a controllable multi-function calling, Reverse Chain adopts a generic rule-based on a backward reasoning process. This rule determines when to do API selection or Argument completion. To evaluate the multi-tool-use capability of LLMs, we have released a compositional multi-tool task dataset, available at https://github.com/zhangyingerjelly/reverse-chain. Extensive numerical experiments validate the remarkable proficiency of Reverse Chain in managing multiple API calls.

Human writers often *bookend* their writing with ending sentences that relate back to the beginning sentences in order to compose a satisfying narrative that “closes the loop.” Motivated by this observation, we propose RENarGen, a controllable story-generation paradigm that generates narratives by ensuring the first and last sentences are related and then infilling the middle sentences. Our contributions include an initial exploration of how various methods of bookending from Narratology affect language modeling for stories. Automatic and human evaluations indicate RENarGen produces better stories with more narrative closure than current autoregressive models.