NAACL 2024 Papers

Large multimodal models suffer from multimodal hallucination, where they provide incorrect responses misaligned with the given visual information. Recent works have conjectured that one of the reasons behind multimodal hallucination is due to the vision encoder failing to ground on the image properly. To mitigate this issue, we propose a novel approach that leverages self-feedback as visual cues. Building on this approach, we introduce Volcano, a multimodal self-feedback guided revision model. Volcano generates natural language feedback to its initial response based on the provided visual information and utilizes this feedback to self-revise its initial response. Volcano effectively reduces multimodal hallucination and achieves state-of-the-art on MMHal-Bench, POPE, and GAVIE. It also improves on general multimodal abilities and outperforms previous models on MM-Vet and MMBench. Through qualitative analysis, we show that Volcano’s feedback is properly grounded on the image than the initial response. This indicates that Volcano can provide itself with richer visual information through feedback generation, leading to self-correct hallucinations. We publicly release our model, data, and code at https://github.com/kaistAI/Volcanogithub.com/kaistAI/Volcano

News Image Captioning aims to create captions from news articles and images, emphasizing the connection between textual context and visual elements. Recognizing the significance of human faces in news images and the face-name co-occurrence pattern in existing datasets, we propose a face-naming module for learning better name embeddings. Apart from names, which can be directly linked to an image area (faces), news image captions mostly contain context information that can only be found in the article. We design a retrieval strategy using CLIP to retrieve sentences that are semantically close to the image, mimicking human thought process of linking articles to images. Furthermore, to tackle the problem of the imbalanced proportion of article context and image context in captions, we introduce a simple yet effective method Contrasting with Language Model backbone (CoLaM) to the training pipeline. We conduct extensive experiments to demonstrate the efficacy of our framework. We out-perform the previous state-of-the-art (without external data) by 7.97/5.80 CIDEr scores on GoodNews/NYTimes800k. Our code is available at https://github.com/tingyu215/VACNIC.

Prior study shows that pre-training techniques can boost the performance of visual document understanding (VDU), which typically requires models to gain abilities to perceive and reason both document texts and layouts (e.g., locations of texts and table-cells). To this end, we propose visually guided generative text-layout pre-training, named ViTLP. Given a document image, the model optimizes hierarchical language and layout modeling objectives to generate the interleaved text and layout sequence. In addition, to address the limitation of processing long documents by Transformers, we introduce a straightforward yet effective multi-segment generative pre-training scheme, facilitating ViTLP to process word-intensive documents of any length. ViTLP can function as a native OCR model to localize and recognize texts of document images. Besides, ViTLP can be effectively applied to various downstream VDU tasks. Extensive experiments show that ViTLP achieves competitive performance over existing baselines on benchmark VDU tasks, including information extraction, document classification, and document question answering.

Enabling autonomous language agents to drive application user interfaces (UIs) as humans do can significantly expand the capability of today’s API-based agents. Essential to this vision is the ability of agents to ground natural language commands to on-screen UI elements. Prior UI grounding approaches work by relaying on developer-provided UI metadata (UI trees, such as web DOM, and accessibility labels) to detect on-screen elements. However, such metadata is often unavailable or incomplete. Object detection techniques applied to UI screens remove this dependency, by inferring location and types of UI elements directly from the UI’s visual appearance. The extracted semantics, however, are too limited to directly enable grounding. We overcome the limitations of both approaches by introducing the task of visual UI grounding, which unifies detection and grounding. A model takes as input a UI screenshot and a free-form language expression, and must identify the referenced UI element. We propose a solution to this problem, LVG, which learns UI element detection and grounding using a new technique called layout-guided contrastive learning, where the semantics of individual UI objects are learned also from their visual organization. Due to the scarcity of UI datasets, LVG integrates synthetic data in its training using multi-context learning. LVG outperforms baselines pre-trained on much larger datasets by over 4.9 points in top-1 accuracy, thus demonstrating its effectiveness.

Modern neural language models (LMs) are powerful tools for modeling human sentence production and comprehension, and their internal representations are remarkably well-aligned with representations of language in the human brain. But to achieve these results, LMs must be trained in distinctly un-human-like ways — requiring orders of magnitude more language data than children receive during development, and without perceptual or social context. Do models trained more naturalistically — with grounded supervision — exhibit more humanlike language learning? We investigate this question in the context of word learning, a key sub-task in language acquisition. We train a diverse set of LM architectures, with and without auxiliary visual supervision, on datasets of varying scales. We then evaluate these models’ learning of syntactic categories, lexical relations, semantic features, word similarity, and alignment with human neural representations. We find that visual supervision can indeed improve the efficiency of word learning. However, these improvements are limited: they are present almost exclusively in the low-dataregime, and sometimes canceled out by the inclusion of rich distributional signals from text. The information conveyed by text and images isnot redundant—models mainly driven by visual information yield qualitatively different from those mainly driven by word co-occurrences. However, our results suggest that current multimodal modeling approaches fail to effectively leverage visual information to build human-like word representations from human-scale data.

MultiModal Summarization (MMS) aims to generate a concise summary based on multimodal data like texts and images and has wide application in multimodal fields.Previous works mainly focus on the coarse-level textual and visual features in which the overall features of the image interact with the whole sentence.However, the entities of the input text and the objects of the image may be underutilized, limiting the performance of current MMS models.In this paper, we propose a novel Visual Enhanced Entity-Level Interaction Network (VE-ELIN) to address the problem of underutilization of multimodal inputs at a fine-grained level in two ways.We first design a cross-modal entity interaction module to better fuse the entity information in text and the object information in vision.Then, we design an object-guided visual enhancement module to fully extract the visual features and enhance the focus of the image on the object area.We evaluate VE-ELIN on two MMS datasets and propose new metrics to measure the factual consistency of entities in the output.Finally, experimental results demonstrate that VE-ELIN is effective and outperforms previous methods under both traditional metrics and ours.The source code is available at https://github.com/summoneryhl/VE-ELIN.

This paper presents VisLingInstruct, a novel approach to advancing Multi-Modal Language Models (MMLMs) in zero-shot learning. Current MMLMs show impressive zero-shot abilities in multi-modal tasks, but their performance depends heavily on the quality of instructions. VisLingInstruct tackles this by autonomously evaluating and optimizing instructional texts through In-Context Learning, improving the synergy between visual perception and linguistic expression in MMLMs. Alongside this instructional advancement, we have also optimized the visual feature extraction modules in MMLMs, further augmenting their responsiveness to textual content. Our comprehensive experiments on MMLMs, based on FlanT5 and Vicuna, show that VisLingInstruct significantly improves zero-shot performance in visual multi-modal tasks. Notably, it achieves a 13.1% and 9% increase in accuracy over the prior state-of-the-art on the TextVQA and HatefulMemes datasets. Our main code is available at https://github.com/Zhudongsheng75/VisLingInstruct

As the number of language models has increased, various benchmarks have been suggested to assess the proficiency of the models in natural language understanding. However, there is a lack of such a benchmark in Vietnamese due to the difficulty in accessing natural language processing datasets or the scarcity of task-specific datasets. **ViGLUE**, the proposed dataset collection, is a **Vi**etnamese **G**eneral **L**anguage **U**nderstanding **E**valuation benchmark developed using three methods: translating an existing benchmark, generating new corpora, and collecting available datasets. ViGLUE contains twelve tasks and encompasses over ten areas and subjects, enabling it to evaluate models comprehensively over a broad spectrum of aspects. Baseline models utilizing multilingual language models are also provided for all tasks in the proposed benchmarks. In addition, the study of the available Vietnamese large language models is conducted to explore the language models’ ability in the few-shot learning framework, leading to the exploration of the relationship between specific tasks and the number of shots.

Text classification systems have continuouslyimproved in performance over the years. How-ever, nearly all current SOTA classifiers have asimilar shortcoming, they process text in a hor-izontal manner. Vertically written words willnot be recognized by a classifier. In contrast,humans are easily able to recognize and readwords written both horizontally and vertically.Hence, a human adversary could write problem-atic words vertically and the meaning wouldstill be preserved to other humans. We simulatesuch an attack, VertAttack. VertAttack identifieswhich words a classifier is reliant on and thenrewrites those words vertically. We find thatVertAttack is able to greatly drop the accuracyof 4 different transformer models on 5 datasets.For example, on the SST2 dataset, VertAttackis able to drop RoBERTa’s accuracy from 94 to13%. Furthermore, since VertAttack does notreplace the word, meaning is easily preserved.We verify this via a human study and find thatcrowdworkers are able to correctly label 77%perturbed texts perturbed, compared to 81% ofthe original texts. We believe VertAttack offersa look into how humans might circumvent clas-sifiers in the future and thus inspire a look intomore robust algorithms.

There are several linguistic claims about situations where words are more likely to be used as metaphors.However, few studies have sought to verify such claims with large corpora.This study entails a large-scale, corpus-based analysis of certain existing claims about verb metaphors, by applying metaphor detection to sentences extracted from Common Crawl and using the statistics obtained from the results.The verification results indicate that the direct objects of verbs used as metaphors tend to have lower degrees of concreteness, imageability, and familiarity, and that metaphors are more likely to be used in emotional and subjective sentences.

Value alignment is crucial for the responsible development of Large Language Models (LLMs). However, how to define values in this context remains largely unexplored. Existing work mainly specifies values as risk criteria formulated in the AI community, e.g., fairness and privacy protection, suffering from poor clarity, adaptability and transparency. Leveraging basic values established in humanity and social science that are compatible with values across cultures, this paper introduces a novel value space spanned by multiple basic value dimensions and proposes BaseAlign, a corresponding value alignment paradigm. Applying the representative Schwartz’s Theory of Basic Values as an instantiation, we construct FULCRA, a dataset consisting of 20k (LLM output, value vector) pairs. LLMs’ outputs are mapped into the K-dim value space beyond simple binary labels, by identifying their underlying priorities for these value dimensions. Extensive analysis and experiments on FULCRA: (1) reveal the essential relation between basic values and LLMs’ behaviors, (2) demonstrate that our paradigm with basic values not only covers existing risks but also anticipates the unidentified ones, and (3) manifest BaseAlign’s superiority in alignment performance with less data, paving the way for addressing the above three challenges.

The natural language generation domain has witnessed great success thanks to Transformer models. Although they have achieved state-of-the-art generative quality, they often neglect generative diversity. Prior attempts to tackle this issue suffer from either low model capacity or over-complicated architectures. Some recent methods employ the VAE framework to enhance diversity, but their latent variables fully depend on the input context, restricting exploration of the latent space. In this paper, we introduce VOLTA, a framework that elevates generative diversity by bridging Transformer with VAE via a more effective cross-attention-based connection, departing from conventional embedding concatenation or summation. Additionally, we propose integrating InfoGAN-style latent codes to enable input-independent variability, further diversifying the generation. Moreover, our framework accommodates discrete inputs alongside its existing support for continuous inputs. We perform comprehensive experiments with two types of Transformers on six datasets from three different NLG tasks to show that our approach can significantly improve generative diversity while maintaining generative quality.

The success of Natural Language Understanding (NLU) benchmarks in various languages, such as GLUE for English, CLUE for Chinese, KLUE for Korean, and IndoNLU for Indonesian, has facilitated the evaluation of new NLU models across a wide range of tasks. To establish a standardized set of benchmarks for Vietnamese NLU, we introduce the first Vietnamese Language Understanding Evaluation (VLUE) benchmark. The VLUE benchmark encompasses five datasets covering different NLU tasks, including text classification, span extraction, and natural language understanding. To provide an insightful overview of the current state of Vietnamese NLU, we then evaluate seven state-of-the-art pre-trained models, including both multilingual and Vietnamese monolingual models, on our proposed VLUE benchmark. Furthermore, we present CafeBERT, a new state-of-the-art pre-trained model that achieves superior results across all tasks in the VLUE benchmark. Our model combines the proficiency of a multilingual pre-trained model with Vietnamese linguistic knowledge. CafeBERT is developed based on the XLM-RoBERTa model, with an additional pretraining step utilizing a significant amount of Vietnamese textual data to enhance its adaptation to the Vietnamese language. For the purpose of future research, CafeBERT is made publicly available for research purposes.

Retrieval-augmented generation framework addresses the limitations of large language models by enabling real-time knowledge updates for more accurate answers. An efficient way in the training phase of retrieval-augmented models is attention distillation, which uses attention scores as supervision signals instead of manually annotated query-document pairs. Despite its growing popularity, the detailed mechanisms behind the success of attention distillation remain unexplored, particularly the specific patterns it leverages to benefit training. In this paper, we address this gap by conducting a comprehensive investigation of attention distillation workflow and identifying key factors influencing the learning performance of retrieval-augmented language models. We further propose several insightful indicators for optimizing models’ training methods and avoiding ineffective training.

While Large Language Models (LLMs) have demonstrated exceptional multitasking abilities, fine-tuning these models on downstream, domain-specific datasets is often necessary to yield superior performance on test sets compared to their counterparts without fine-tuning. However, the comprehensive effects of fine-tuning on the LLMs’ generalization ability are not fully understood.This paper delves into the differences between original, unmodified LLMs and their fine-tuned variants. Our primary investigation centers on whether fine-tuning affects the generalization ability intrinsic to LLMs. To elaborate on this, we conduct extensive experiments across five distinct language tasks on various datasets.Our main findings reveal that models fine-tuned on generation and classification tasks exhibit dissimilar behaviors in generalizing to different domains and tasks.Intriguingly, we observe that integrating the in-context learning strategy during fine-tuning on generation tasks can enhance the model’s generalization ability.Through this systematic investigation, we aim to contribute valuable insights into the evolving landscape of fine-tuning practices for LLMs.

Unraveling the intricate details of events in natural language necessitates a subtle understanding of temporal dynamics. Despite the adeptness of Large Language Models (LLMs) in discerning patterns and relationships from data, their inherent comprehension of temporal dynamics remains a formidable challenge. This research meticulously explores these intrinsic challenges within LLMs, with a specific emphasis on evaluating the performance of GPT-3.5 and GPT-4 models in the analysis of temporal data. Employing two distinct prompt types, namely Question Answering (QA) format and Textual Entailment (TE) format, our analysis probes into both implicit and explicit events. The findings underscore noteworthy trends, revealing disparities in the performance of GPT-3.5 and GPT-4. Notably, biases toward specific temporal relationships come to light, with GPT-3.5 demonstrating a preference for “AFTER” in the QA format for both implicit and explicit events, while GPT-4 leans towards “BEFORE”. Furthermore, a consistent pattern surfaces wherein GPT-3.5 tends towards “TRUE”, and GPT-4 exhibits a preference for “FALSE” in the TE format for both implicit and explicit events. This persistent discrepancy between GPT-3.5 and GPT-4 in handling temporal data highlights the intricate nature of inductive bias in LLMs, suggesting that the evolution of these models may not merely mitigate bias but may introduce new layers of complexity.

Recent large language models (LLMs) have shown remarkable performance in aligning generated text with user intentions across various tasks. When it comes to long-form text generation, there has been a growing interest in generation from a discourse coherence perspective.However, existing lexical or semantic metrics such as BLEU, ROUGE, BertScore cannot effectively capture the discourse coherence.The development of discourse-specific automatic evaluation methods for assessing the output of LLMs warrants greater focus and exploration. In this paper, we present a novel automatic metric designed to quantify the discourse divergence between two long-form articles.Extensive experiments on three datasets from representative domains demonstrate that our metric aligns more closely with human preferences and GPT-4 coherence evaluation, outperforming existing evaluation methods.

Parameter-efficient fine-tuning (PEFT) methods are increasingly vital in adapting large-scale pre-trained language models for diverse tasks, offering a balance between adaptability and computational efficiency. They are important in Low-Resource Language (LRL) Neural Machine Translation (NMT) to enhance translation accuracy with minimal resources. However, their practical effectiveness varies significantly across different languages. We conducted comprehensive empirical experiments with varying LRL domains and sizes to evaluate the performance of 8 PEFT methods with in total of 15 architectures using the SacreBLEU score. We showed that 6 PEFT architectures outperform the baseline for both in-domain and out-domain tests and the Houlsby+Inversion adapter has the best performance overall, proving the effectiveness of PEFT methods.

Modular Neural Networks (MNNs) demonstrate various advantages over monolithic models.Existing MNNs are generally explicit: their modular architectures are pre-defined, with individual modules expected to implement distinct functions.Recent works reveal that there exists implicit modularity in standard pre-trained transformers, namely Emergent Modularity.They indicate that such modular structures spontaneously exhibit during the early pre-training phase.Despite the benefits of modularity, most Language Models (LMs) are still treated as monolithic models in the pre-train and fine-tune paradigm, with their emergent modularity locked and underutilized.In this work, focusing on unlocking the emergent modularity in LMs, we showcase that standard LMs could be fine-tuned as their Mixture-of-Expert (MoEs) counterparts without introducing any extra parameters. Such MoEs are derived from emergent modularity and are referred to as Emergent MoEs (EMoE).Our experiments demonstrate that fine-tuning EMoE effectively improves downstream in-domain and out-of-domain generalization compared with vanilla fine-tuning.Our analysis and ablation studies further illustrate that it is robust to various configurations and can scale up to Large Language Models (i.e., Llama2-7B and Llama-30B). Code is available at https://github.com/qiuzh20/EMoE.