EMNLP 2023 Papers

Text is ubiquitous in our visual world, conveying crucial information, such as in documents, websites, and everyday photographs. In this work, we propose UReader, a first exploration of universal OCR-free visually-situated language understanding based on the Multimodal Large Language Model (MLLM). By leveraging the shallow text recognition ability of the MLLM, we only finetuned 1.2% parameters and the training cost is much lower than previous work following domain-specific pretraining and finetuning paradigms. Concretely, UReader is jointly finetuned on a wide range of Visually-situated Language Understanding tasks via a unified instruction format. To enhance the visual text and semantic understanding, we further apply two auxiliary tasks with the same format, namely text reading and key points generation tasks. We design a shape-adaptive cropping module before the encoder-decoder architecture of MLLM to leverage the frozen low-resolution vision encoder for processing high-resolution images. Without downstream finetuning, our single model achieves state-of-the-art ocr-free performance in 8 out of 10 visually-situated language understanding tasks, across 5 domains: documents, tables, charts, natural images, and webpage screenshots. Codes and instruction-tuning datasets will be released.

Reinforcement learning from human feedback serves as a crucial bridge, aligning large language models with human and societal values. This alignment requires a vast corpus of human feedback to learn a reward model, which is subsequently used to finetune language models. However, we have identified that the reward model often finds shortcuts to bypass its intended objectives, misleadingly assuming that humans prefer longer responses. The emergence of length bias often induces the model to favor longer outputs, yet it doesn't equate to an increase in helpful information within these outputs. In this paper, we propose an innovative solution, applying the Product-of-Experts (PoE) technique to separate reward modeling from the influence of sequence length. In our framework, the main expert concentrates on understanding human intents, while the biased expert targets the identification and capture of length bias. To further enhance the learning of bias, we introduce perturbations into the bias-focused expert, disrupting the flow of semantic information. Experimental results validate the effectiveness of our approach, indicating that language model performance is improved, irrespective of sequence length.

Large Language Models (LLMs) demonstrate impressive reasoning ability and the maintenance of world knowledge not only in natural language tasks, but also in some vision-language tasks such as open-domain knowledge-based visual question answering (OK-VQA). As images are invisible to LLMs, researchers convert images to text to engage LLMs into the visual question reasoning procedure. This leads to discrepancies between images and their textual representations presented to LLMs, which consequently impedes final reasoning performance. To fill the information gap and better leverage the reasoning capability, we design a framework that enables LLMs to proactively ask relevant questions to unveil more details in the image, along with filters for refining the generated information. We validate our idea on OK-VQA and A-OKVQA. Our method continuously boosts the performance of baselines methods by an average gain of 2.15\% on OK-VQA, and achieves consistent improvements across different LLMs.

Generative approaches powered by large language models (LLMs) have demonstrated emergent abilities in tasks that require complex reasoning abilities. Yet the generative nature still makes the generated content suffer from hallucinations, thus unsuitable for entity-centric tasks like entity linking (EL) requiring precise entity predictions over a large knowledge base. We present Instructed Generative Entity Linker (INSGENEL), the first approach that enables casual language models to perform entity linking over knowledge bases. Several methods of equipping language models with EL ability were proposed in this work, including (i) a sequence-to-sequence training EL objective with instruction-tuning, (ii) a novel generative EL framework based on a light-weight potential mention retriever that frees the model from heavy and non-parallelizable decoding, achieving 4$\times$ speedup without compromise on linking metrics. INSGENEL outperforms previous generative alternatives with +6.8 F1 points gain on average, also with a huge advantage in training data efficiency and training compute consumption. In addition, our skillfully-engineered in-context learning (ICL) framework for EL still lags behind INSGENEL significantly, reaffirming that the EL task remains a persistent hurdle for general LLMs.

Recent studies have revealed the remarkable cross-lingual capability of multilingual pre-trained language models (mPLMs), even when pre-trained without parallel corpora (mono-mPLMs). Intuitively, semantic alignments may be the reason behind such capability but remain under-explored. In this work, we investigate the alignment properties from the token perspective in mono-mPLMs and find that the alignments correspond to the geometric similarity of embedding space across different languages. Nevertheless, mono-mPLMs tend to damage this geometric similarity at the higher layers due to the lack of cross-lingual interactions, thus limiting their cross-lingual transfer capabilities. To address this issue, we introduce token-level and semantic-level code-switched masked language modeling, employing the self-induced token alignments to explicitly improve cross-lingual interactions over layers of mono-mPLMs without relying on parallel sentences. We evaluate our method on various natural language understanding tasks and unsupervised machine translation tasks. The results demonstrate that our methods outperform the strong baselines and achieve comparable performance with mPLMs trained with parallel corpora.

Linguistic communication is prevalent in Human-Computer Interaction (HCI). Speech (spoken language) serves as a convenient yet potentially ambiguous form due to noise and accents, exposing a gap compared to text. In this study, we investigate the prominent HCI task, Referring Video Object Segmentation (R-VOS), which aims to segment and track objects using linguistic references. While text input is well-investigated, speech input is under-explored. Our objective is to bridge the gap between speech and text, enabling the adaptation of existing text-input R-VOS models to accommodate noisy speech input effectively. Specifically, we propose a method to align the semantic spaces between speech and text by incorporating two key modules: 1) Noise-Aware Semantic Adjustment (NSA) for clear semantics extraction from noisy speech; and 2) Semantic Jitter Suppression (SJS) enabling R-VOS models to tolerate noisy queries. Comprehensive experiments conducted on the challenging AVOS benchmarks reveal that our proposed method outperforms state-of-the-art approaches.

Generative Pre-trained Transformer 4 (GPT-4) demonstrates impressive chain-of-thought reasoning ability. Recent work on self-instruction tuning, such as Alpaca, has focused on enhancing the general proficiency of models. These instructions enable the model to achieve performance comparable to GPT-3.5 on general tasks like open-domain text generation and paraphrasing. However, they fall short of helping the model handle complex reasoning tasks. To bridge the gap, this paper presents LogiCoT, a new instruction-tuning dataset for Logical Chain-of-Thought reasoning with GPT-4. We elaborate on the process of harvesting instructions for prompting GPT-4 to generate chain-of-thought rationales. LogiCoT serves as an instruction set for teaching models of logical reasoning and elicits general reasoning skills.

Neural Machine Translation (NMT) systems encounter a significant challenge when translating a pro-drop ('pronoun-dropping') language (e.g., Chinese) to a non-pro-drop one (e.g., English), since the pro-drop phenomenon demands NMT systems to recover omitted pronouns. This unique and crucial task, however, lacks sufficient datasets for benchmarking. To bridge this gap, we introduce PROSE, a new benchmark featured in diverse pro-drop instances for document-level Chinese-English spoken language translation. Furthermore, we conduct an in-depth investigation of the pro-drop phenomenon in spoken Chinese on this dataset, reconfirming that pro-drop reduces the performance of NMT systems in Chinese-English translation. To alleviate the negative impact introduced by pro-drop, we propose Mention-Aware Semantic Augmentation, a novel approach that leverages the semantic embedding of dropped pronouns to augment training pairs. Results from the experiments on four Chinese-English translation corpora show that our proposed method outperforms existing methods regarding omitted pronoun retrieval and overall translation quality.

Understanding the fundamental concepts and trends in a scientific field is crucial for keeping abreast of its continuous advancement. In this study, we propose a systematic framework for analyzing the evolution of research topics in a scientific field using causal discovery and inference techniques. We define three variables to encompass diverse facets of the evolution of research topics within NLP and utilize a causal discovery algorithm to unveil the causal connections among these variables using observational data. Subsequently, we leverage this structure to measure the intensity of these relationships. By conducting extensive experiments on the ACL Anthology corpus, we demonstrate that our framework effectively uncovers evolutionary trends and the underlying causes for a wide range of NLP research topics. Specifically, we show that tasks and methods are primary drivers of research in NLP, with datasets following, while metrics have minimal impact.

To avoid detection by current NLP monitoring applications, progenitors of hate speech often replace one or more letters in offensive words with homoglyphs, visually similar Unicode characters. Harvesting real-world hate speech containing homoglyphs is challenging due to the vast replacement possibilities. We developed a character substitution scraping method and assembled the Offensive Tweets with Homoglyphs (OTH) Dataset (N=90,788) with more than 1.5 million occurrences of 1,281 non-Latin characters (emojis excluded). In an annotated sample (n=700), 40.14% of the tweets were found to contain hate speech. We assessed the performance of seven transformer-based hate speech detection models and found that they performed poorly in a zero-shot setting (F1 scores between 0.04 and 0.52) but normalizing the data dramatically improved detection (F1 scores between 0.59 and 0.71). Training the models using the annotated data further boosted performance (highest micro-averaged F1 score=0.88, using five-fold cross validation). This study indicates that a dataset containing homoglyphs known and unknown to the scraping script can be collected, and that neural models can be trained to recognize camouflaged real-world hate speech.

In recent years, the injection of factual knowledge has been observed to have a significant positive correlation to the downstream task performance of pre-trained language models. However, existing work neither demonstrates that pre-trained models successfully learn the injected factual knowledge nor proves that there is a causal relation between injected factual knowledge and downstream performance improvements. In this paper, we introduce a counterfactual-based analysis framework to explore the causal effects of factual knowledge injection on the performance of language models within pretrain-finetune paradigm. Instead of directly probing the language model or exhaustively enumerating potential confounding factors, we analyze this issue by perturbing the factual knowledge sources at different scales and comparing the performance of pre-trained language models before and after the perturbation. Surprisingly, throughout our experiments, we find that although the knowledge seems to be successfully injected, the correctness of injected knowledge only has a very limited effect on the models' downstream performance. This finding strongly challenges previous assumptions that the injected factual knowledge is the key for language models to achieve performance improvements on downstream tasks in pretrain-finetune paradigm.

Syntax is a latent hierarchical structure which underpins the robust and compositional nature of human language. In this work, we explore the hypothesis that syntactic dependencies can be represented in language model attention distributions and propose a new method to induce these structures theory-agnostically. Instead of modeling syntactic relations as defined by annotation schemata, we model a more general property implicit in the definition of dependency relations, syntactic substitutability. This property captures the fact that words at either end of a dependency can be substituted with words from the same category. Substitutions can be used to generate a set of syntactically invariant sentences whose representations are then used for parsing. We show that increasing the number of substitutions used improves parsing accuracy on natural data. On long-distance subject-verb agreement constructions, our method achieves 79.5% recall compared to 8.9% using a previous method. Our method also provides improvements when transferred to a different parsing setup, demonstrating that it generalizes.

Recently, aspect-based sentiment analysis (ABSA) models have yielded promising results. However, they are susceptible to learning spurious correlations between certain words of the input text and output labels while modeling the sentiment feature of the aspect. This spurious correlation will potentially undermine the performance of ABSA models. One direct solution for this problem is to make the model see and learn an explanation of sentiment expression rather than certain words. Motivated by this, we exploit explanations for the sentiment polarity of each aspect from large language models (LLMs) to reduce spurious correlations in ABSA. First, we formulate a prompt template that wraps the sentence, an aspect, and the sentiment label. This template is utilized to prompt LLMs to generate an appropriate explanation that states the sentiment cause. Then, we propose two straightforward yet effective methods to leverage the explanation for preventing the learning of spurious correlations. We conducted extensive comparative experiments on five datasets by integrating them with some representative ABSA models. Results show that our methods can achieve performance gains and enhance the performance and generalization ability of ABSA models.

Distantly supervised named entity recognition (DS-NER) aims to locate entity mentions and classify their types with only knowledge bases or gazetteers and unlabeled corpus. However, distant annotations are noisy and degrade the performance of NER models. In this paper, we propose a noise-robust prototype network named MProto for the DS-NER task. Different from previous prototype-based NER methods, MProto represents each entity type with multiple prototypes to characterize the intra-class variance among entity representations. To optimize the classifier, each token should be assigned an appropriate ground-truth prototype and we consider such token-prototype assignment as an optimal transport (OT) problem. Furthermore, to mitigate the noise from incomplete labeling, we propose a novel denoised optimal transport (DOT) algorithm. Specifically, we utilize the assignment result between *Other* class tokens and all prototypes to distinguish unlabeled entity tokens from true negatives. Experiments on several DS-NER benchmarks demonstrate that our MProto achieves state-of-the-art performance. The source code is now available on Github.

Recent progress in Large Language Models (LLMs) has produced models that exhibit remarkable performance across a variety of NLP tasks. However, it remains unclear whether the existing focus of NLP research accurately captures the genuine requirements of human users. This paper provides a comprehensive analysis of the divergence between academic research in NLP and the needs of real-world NLP applications via a large-scale collection of user-GPT conversations. We analyze a large-scale collection of real user queries to GPT. We compare these queries against existing NLP benchmark tasks and identify a significant gap between the tasks that users frequently request from LLMs and the tasks that are commonly studied in academic research. For example, we find that tasks such as "design" and "planning" are prevalent in user interactions but largely neglected or different from traditional NLP benchmarks. We investigate these overlooked tasks, dissect the practical challenges, and provide insights toward a roadmap to make LLMs better aligned with user needs.

Visual text evokes an image in a person's mind, while non-visual text fails to do so. A method to automatically detect visualness in text will enable text-to-image retrieval and generation models to augment text with relevant images. This is particularly challenging with long-form text as text-to-image generation and retrieval models are often triggered for text that is designed to be explicitly visual in nature, whereas long-form text could contain many non-visual sentences. To this end, we curate a dataset of 3,620 English sentences and their visualness scores provided by multiple human annotators. We also propose a fine-tuning strategy that adapts large vision-language models like CLIP by modifying the model's contrastive learning objective to map text identified as non-visual to a common NULL image while matching visual text to their corresponding images in the document. We evaluate the proposed approach on its ability to (i) classify visual and non-visual text accurately, and (ii) attend over words that are identified as visual in psycholinguistic studies. Empirical evaluation indicates that our approach performs better than several heuristics and baseline models for the proposed task. Furthermore, to highlight the importance of modeling the visualness of text, we conduct qualitative analyses of text-to-image generation systems like DALL-E.

Human feedback is increasingly used to steer the behaviours of Large Language Models (LLMs). However, it is unclear how to collect and incorporate feedback in a way that is efficient, effective and unbiased, especially for highly subjective human preferences and values. In this paper, we survey existing approaches for learning from human feedback, drawing on 95 papers primarily from the ACL and arXiv repositories. First, we summarise the past, pre-LLM trends for integrating human feedback into language models. Second, we give an overview of present techniques and practices, as well as the motivations for using feedback; conceptual frameworks for defining values and preferences; and how feedback is collected and from whom. Finally, we encourage a better future of feedback learning in LLMs by raising five unresolved conceptual and practical challenges.

It has been shown that high quality fine-tuning boosts the performance of language models, even if the size of the fine-tuning is small. In this work we show how highly targeted fine-tuning improves the task of hate speech counter-narrative generation in user-generated text, even for very small sizes of training (1722 counter-narratives for English and 355 for Spanish). Providing a small subset of examples focusing on single argumentative strategies, together with the argumentative analysis relevant to that strategy, yields counter-narratives that are as satisfactory as providing the whole set of counter-narratives. We also show that a good base model is required for the fine-tuning to have a positive impact. Indeed, for Spanish, the counter-narratives obtained without fine-tuning are mostly unacceptable, and, while fine-tuning improves their overall quality, the performance still remains quite unsatisfactory.

Topic segmentation, in the context of natural language processing, is the process of finding boundaries in a sequence of sentences that separate groups of adjacent sentences at shifts in semantic meaning. Currently, assessing the quality of a segmentation is done by comparing segmentation boundaries selected by a human or algorithm to those selected by a known good reference. This means that it is not possible to quantify the quality of a segmentation without a human annotator, which can be costly and time consuming. This work seeks to improve assessment of segmentation by proposing a reference-free segmentation quality index (SegReFree). The metric takes advantage of the fact that segmentation at a sentence level generally seeks to identify segment boundaries at semantic boundaries within the text. The proposed metric uses a modified cluster validity metric with semantic embeddings of the sentences to determine the quality of the segmentation. Multiple segmentation data sets are used to compare our proposed metric with existing reference-based segmentation metrics by progressively degrading the reference segmentation while computing all possible metrics; through this process, a strong correlation with existing segmentation metrics is shown. A Python library implementing the metric is released under the GNU General Public License and the repository is available at \url{https://github.com/evan-person/reference_free_segmentation_metric}.