The process of scale calibration in ranking systems involves adjusting the outputs of rankers to correspond with significant qualities like click-through rates or relevance, crucial for mirroring real-world value and thereby boosting the system's effectiveness and reliability. Although there has been research on calibrated ranking losses within learning-to-rank models, the particular issue of adjusting the scale for neural rankers, which excel in handling textual information, has not been thoroughly examined. Neural ranking models are adept at processing text data, yet the application of existing scale calibration techniques to these models poses significant challenges due to their complexity and the intensive training they require, often resulting in suboptimal outcomes. This study delves into the potential of large language models (LLMs) to provide uncertainty measurements for a query and document pair that correlate with the scale-calibrated scores. By employing Monte Carlo sampling to gauge relevance probabilities from LLMs and incorporating natural language explanations (NLEs) to articulate this uncertainty, we carry out comprehensive tests on two major document ranking datasets. Our findings reveal that the approach leveraging NLEs outperforms existing calibration methods under various training scenarios, leading to better calibrated neural rankers.
Pre-trained language models (LMs) are capable of in-context learning (ICL): they can adapt to a task with only a few examples given in the prompt without any parameter update. However, it is unclear where this capability comes from as there is a stark distribution shift between pre-training text and ICL prompts. In this work, we study what patterns of the pre-training data contribute to ICL. We find that LMs' ICL ability depends on $\textit{parallel structures}$ in the pre-training data -- pairs of phrases following similar templates in the same context window. Specifically, we detect parallel structures by checking whether training on one phrase improves prediction of the other, and conduct ablation experiments to study their effect on ICL. We show that removing parallel structures in the pre-training data reduces LMs' ICL accuracy by 51% (vs 2% from random ablation). This drop persists even when excluding common patterns such as n-gram repetitions and long-range dependency, showing the diversity and generality of parallel structures. A closer look at the detected parallel structures indicates that they cover diverse linguistic tasks and span long distances in the data.
Screen user interfaces (UIs) and infographics, sharing similar visual language and design principles, play important roles in human communication and human-machine interaction. We introduce ScreenAI, a vision-language model that specializes in UI and infographics understanding. Our model improves upon the PaLI architecture with the flexible patching strategy of pix2struct and is trained on a unique mixture of datasets. At the heart of this mixture is a novel screen annotation task in which the model has to identify the type and location of UI elements. We use these text annotations to describe screens to Large Language Models and automatically generate question-answering (QA), UI navigation, and summarization training datasets at scale. We run ablation studies to demonstrate the impact of these design choices. At only 5B parameters, ScreenAI achieves new state-of-the-artresults on UI- and infographics-based tasks (Multi-page DocVQA, WebSRC, MoTIF and Widget Captioning), and new best-in-class performance on others (Chart QA, DocVQA, and InfographicVQA) compared to models of similar size. Finally, we release three new datasets: one focused on the screen annotation task and two others focused on question answering.
Large Language Models (LLMs) possess outstanding capabilities in addressing various natural language processing (NLP) tasks. However, the sheer size of these models poses challenges in terms of storage, training and inference due to the inclusion of billions of parameters through layer stacking. While traditional approaches such as model pruning or distillation offer ways for reducing model size, they often come at the expense of performance retention. In our investigation, we systematically explore the approach of reducing the number of layers in LLMs. Surprisingly, we observe that even with fewer layers, LLMs maintain similar or better performance levels, particularly in prompt-based fine-tuning for text classification tasks. Remarkably, in certain cases, models with a single layer outperform their fully layered counterparts. These findings offer valuable insights for future work aimed at mitigating the size constraints of LLMs while preserving their performance, thereby opening avenues for significantly more efficient use of LLMs.
Argument mining (AM) is the process of automatically extracting arguments, their components and/or relations amongst arguments and components from text. As the number of platforms supporting online debate increases, the need for AM becomes ever more urgent, especially in support of downstream tasks. Relation-based AM (RbAM) is a form of AM focusing on identifying agreement (support) and disagreement (attack) relations amongst arguments. RbAM is a challenging classification task, with existing methods failing to perform satisfactorily. In this paper, we show that general-purpose Large Language Models (LLMs), appropriately primed and prompted, can significantly outperform the best performing (RoBERTa-based) baseline. Specifically, we experiment with two open-source LLMs (Llama-2 and Mistral) with ten datasets.
Text Augmentation is an important task for low-resource languages. It helps deal with the problem of data scarcity. A data augmentation strategy is used to deal with the problem of data scarcity. Through the years, much work has been done on data augmentation for the English language. In contrast, very less work has been done on Indian languages. This is contrary to the fact that data augmentation is used to deal with data scarcity. In this work, we focus on implementing techniques like Easy Data Augmentation, Back Translation, Paraphrasing, Text Generation using LLMs, and Text Expansion using LLMs for text classification on different languages. We focus on 6 Indian languages namely: Sindhi, Marathi, Hindi, Gujarati, Telugu, and Sanskrit. According to our knowledge, no such work exists for text augmentation on Indian languages. We carry out binary as well as multi-class text classification to make our results more comparable. We get surprising results as basic data augmentation techniques surpass LLMs.
Bimodal objects, such as the checkerboard pattern used in camera calibration, markers for object tracking, and text on road signs, to name a few, are prevalent in our daily lives and serve as a visual form to embed information that can be easily recognized by vision systems. While binarization from intensity images is crucial for extracting the embedded information in the bimodal objects, few previous works consider the task of binarization of blurry images due to the relative motion between the vision sensor and the environment. The blurry images can result in a loss in the binarization quality and thus degrade the downstream applications where the vision system is in motion. Recently, neuromorphic cameras offer new capabilities for alleviating motion blur, but it is non-trivial to first deblur and then binarize the images in a real-time manner. In this work, we propose an event-based binary reconstruction method that leverages the prior knowledge of the bimodal target's properties to perform inference independently in both event space and image space and merge the results from both domains to generate a sharp binary image. We also develop an efficient integration method to propagate this binary image to high frame rate binary video. Finally, we develop a novel method to naturally fuse events and images for unsupervised threshold identification. The proposed method is evaluated in publicly available and our collected data sequence, and shows the proposed method can outperform the SOTA methods to generate high frame rate binary video in real-time on CPU-only devices.
There is a trilemma in reinforcement learning from human feedback (RLHF): the incompatibility between highly diverse contexts, low labeling cost, and reliable alignment performance. Here we aim to mitigate such incompatibility through the design of dataset information structures during reward modeling, and meanwhile propose new, generalizable methods of analysis that have wider applications, including potentially shedding light on goal misgeneralization. Specifically, we first reexamine the RLHF process and propose a theoretical framework portraying it as an autoencoding process over text distributions. Our framework formalizes the RLHF objective of ensuring distributional consistency between human preference and large language model (LLM) behavior. Based on this framework, we introduce a new method to model generalization in the reward modeling stage of RLHF, the induced Bayesian network (IBN). Drawing from random graph theory and causal analysis, it enables empirically grounded derivation of generalization error bounds, a key improvement over classical methods of generalization analysis. An insight from our analysis is the superiority of the tree-based information structure in reward modeling, compared to chain-based baselines in conventional RLHF methods. We derive that in complex contexts with limited data, the tree-based reward model (RM) induces up to $\Theta(\log n/\log\log n)$ times less variance than chain-based RM where $n$ is the dataset size. As validation, we demonstrate that on three NLP tasks, the tree-based RM achieves 65% win rate on average against chain-based baselines. Looking ahead, we hope to extend the IBN analysis to help understand the phenomenon of goal misgeneralization.
Feature attribution, the ability to localize regions of the input data that are relevant for classification, is an important capability for machine learning models in scientific and biomedical domains. Current methods for feature attribution, which rely on "explaining" the predictions of end-to-end classifiers, suffer from imprecise feature localization and are inadequate for use with small sample sizes and high-dimensional datasets due to computational challenges. We introduce prospector heads, an efficient and interpretable alternative to explanation-based methods for feature attribution that can be applied to any encoder and any data modality. Prospector heads generalize across modalities through experiments on sequences (text), images (pathology), and graphs (protein structures), outperforming baseline attribution methods by up to 49 points in mean localization AUPRC. We also demonstrate how prospector heads enable improved interpretation and discovery of class-specific patterns in the input data. Through their high performance, flexibility, and generalizability, prospectors provide a framework for improving trust and transparency for machine learning models in complex domains.
Advances in language modeling have paved the way for novel human-AI co-writing experiences. This paper explores how varying levels of scaffolding from large language models (LLMs) shape the co-writing process. Employing a within-subjects field experiment with a Latin square design, we asked participants (N=131) to respond to argumentative writing prompts under three randomly sequenced conditions: no AI assistance (control), next-sentence suggestions (low scaffolding), and next-paragraph suggestions (high scaffolding). Our findings reveal a U-shaped impact of scaffolding on writing quality and productivity (words/time). While low scaffolding did not significantly improve writing quality or productivity, high scaffolding led to significant improvements, especially benefiting non-regular writers and less tech-savvy users. No significant cognitive burden was observed while using the scaffolded writing tools, but a moderate decrease in text ownership and satisfaction was noted. Our results have broad implications for the design of AI-powered writing tools, including the need for personalized scaffolding mechanisms.