It is relatively easy to mine a large parallel corpus for any machine learning task, such as speech-to-text or speech-to-speech translation. Although these mined corpora are large in volume, their quality is questionable. This work shows that the simplest filtering technique can trim down these big, noisy datasets to a more manageable, clean dataset. We also show that using this clean dataset can improve the model's performance, as in the case of the multilingual-to-English Speech Translation (ST) model, where, on average, we obtain a 4.65 BLEU score improvement.
Warning: This paper contains examples of offensive language targetting marginalized population. The digitization of historical texts invites researchers to explore the large-scale corpus of historical texts with computational methods. In this study, we present computational text analysis on a relatively understudied topic of how Asian workers are represented in historical newspapers in the United States. We found that the word "coolie" was semantically different in some States (e.g., Massachusetts, Rhode Island, Wyoming, Oklahoma, and Arkansas) with the different discourses around coolie. We also found that then-Confederate newspapers and then-Union newspapers formed distinctive discourses by measuring over-represented words. Newspapers from then-Confederate States associated coolie with slavery-related words. In addition, we found Asians were perceived to be inferior to European immigrants and subjected to the target of racism. This study contributes to supplementing the qualitative analysis of racism in the United States with quantitative discourse analysis.
Previous solutions to knowledge-based visual question answering~(K-VQA) retrieve knowledge from external knowledge bases and use supervised learning to train the K-VQA model. Recently pre-trained LLMs have been used as both a knowledge source and a zero-shot QA model for K-VQA and demonstrated promising results. However, these recent methods do not explicitly show the knowledge needed to answer the questions and thus lack interpretability. Inspired by recent work on knowledge generation from LLMs for text-based QA, in this work we propose and test a similar knowledge-generation-based K-VQA method, which first generates knowledge from an LLM and then incorporates the generated knowledge for K-VQA in a zero-shot manner. We evaluate our method on two K-VQA benchmarks and found that our method performs better than previous zero-shot K-VQA methods and our generated knowledge is generally relevant and helpful.
Feature attribution methods (FAs), such as gradients and attention, are widely employed approaches to derive the importance of all input features to the model predictions. Existing work in natural language processing has mostly focused on developing and testing FAs for encoder-only language models (LMs) in classification tasks. However, it is unknown if it is faithful to use these FAs for decoder-only models on text generation, due to the inherent differences between model architectures and task settings respectively. Moreover, previous work has demonstrated that there is no `one-wins-all' FA across models and tasks. This makes the selection of a FA computationally expensive for large LMs since input importance derivation often requires multiple forward and backward passes including gradient computations that might be prohibitive even with access to large compute. To address these issues, we present a model-agnostic FA for generative LMs called Recursive Attribution Generator (ReAGent). Our method updates the token importance distribution in a recursive manner. For each update, we compute the difference in the probability distribution over the vocabulary for predicting the next token between using the original input and using a modified version where a part of the input is replaced with RoBERTa predictions. Our intuition is that replacing an important token in the context should have resulted in a larger change in the model's confidence in predicting the token than replacing an unimportant token. Our method can be universally applied to any generative LM without accessing internal model weights or additional training and fine-tuning, as most other FAs require. We extensively compare the faithfulness of ReAGent with seven popular FAs across six decoder-only LMs of various sizes. The results show that our method consistently provides more faithful token importance distributions.
Vision-Language Models (VLMs) excel in generating textual responses from visual inputs, yet their versatility raises significant security concerns. This study takes the first step in exposing VLMs' susceptibility to data poisoning attacks that can manipulate responses to innocuous, everyday prompts. We introduce Shadowcast, a stealthy data poisoning attack method where poison samples are visually indistinguishable from benign images with matching texts. Shadowcast demonstrates effectiveness in two attack types. The first is Label Attack, tricking VLMs into misidentifying class labels, such as confusing Donald Trump for Joe Biden. The second is Persuasion Attack, which leverages VLMs' text generation capabilities to craft narratives, such as portraying junk food as health food, through persuasive and seemingly rational descriptions. We show that Shadowcast are highly effective in achieving attacker's intentions using as few as 50 poison samples. Moreover, these poison samples remain effective across various prompts and are transferable across different VLM architectures in the black-box setting. This work reveals how poisoned VLMs can generate convincing yet deceptive misinformation and underscores the importance of data quality for responsible deployments of VLMs. Our code is available at: https://github.com/umd-huang-lab/VLM-Poisoning.
Image captioning aims at generating descriptive and meaningful textual descriptions of images, enabling a broad range of vision-language applications. Prior works have demonstrated that harnessing the power of Contrastive Image Language Pre-training (CLIP) offers a promising approach to achieving zero-shot captioning, eliminating the need for expensive caption annotations. However, the widely observed modality gap in the latent space of CLIP harms the performance of zero-shot captioning by breaking the alignment between paired image-text features. To address this issue, we conduct an analysis on the CLIP latent space which leads to two findings. Firstly, we observe that the CLIP's visual feature of image subregions can achieve closer proximity to the paired caption due to the inherent information loss in text descriptions. In addition, we show that the modality gap between a paired image-text can be empirically modeled as a zero-mean Gaussian distribution. Motivated by the findings, we propose a novel zero-shot image captioning framework with text-only training to reduce the modality gap. In particular, we introduce a subregion feature aggregation to leverage local region information, which produces a compact visual representation for matching text representation. Moreover, we incorporate a noise injection and CLIP reranking strategy to boost captioning performance. We also extend our framework to build a zero-shot VQA pipeline, demonstrating its generality. Through extensive experiments on common captioning and VQA datasets such as MSCOCO, Flickr30k and VQAV2, we show that our method achieves remarkable performance improvements. Code is available at https://github.com/Artanic30/MacCap.
In recent years, language models and deep learning techniques have revolutionized natural language processing tasks, including emotion detection. However, the specific emotion of guilt has received limited attention in this field. In this research, we explore the applicability of three transformer-based language models for detecting guilt in text and compare their performance for general emotion detection and guilt detection. Our proposed model outformed BERT and RoBERTa models by two and one points respectively. Additionally, we analyze the challenges in developing accurate guilt-detection models and evaluate our model's effectiveness in detecting related emotions like "shame" through qualitative analysis of results.
During both pretraining and fine-tuning, Large Language Models (\textbf{LLMs}) are trained on trillions of tokens of text of widely varying quality. Both phases of training typically involve heuristically filtering out ``low-quality'' or \textit{noisy} training samples, yet little is known quantitatively about how the type or intensity of noise affects downstream performance. In this work, we study how noise in chain of thought (\textbf{CoT}) impacts task performance in the highly-controlled setting of algorithmically solvable tasks. First, we develop the Traced Integer (\textbf{TInt}) framework to generate highly customizable noised execution traces for any arithmetic function on lists of integers. We then define two types of noise: \textit{static} noise, a local form of noise which is applied after the CoT trace is computed, and \textit{dynamic} noise, a global form of noise which propagates errors in the trace as it is computed. We then evaluate the test performance of pretrained models both prompted and fine-tuned on noised datasets with varying levels of dataset contamination and intensity. We find fine-tuned models are extremely robust to high levels of static noise but struggle significantly more with lower levels of dynamic noise. In contrast, few-shot prompted models appear more sensitive to even static noise. We conclude with a discussion of how our findings impact noise filtering best-practices, in particular emphasizing the importance of removing samples containing destructive dynamic noise with global errors.
This work delves into the expanding role of large language models (LLMs) in generating artificial data. LLMs are increasingly employed to create a variety of outputs, including annotations, preferences, instruction prompts, simulated dialogues, and free text. As these forms of LLM-generated data often intersect in their application, they exert mutual influence on each other and raise significant concerns about the quality and diversity of the artificial data incorporated into training cycles, leading to an artificial data ecosystem. To the best of our knowledge, this is the first study to aggregate various types of LLM-generated text data, from more tightly constrained data like "task labels" to more lightly constrained "free-form text". We then stress test the quality and implications of LLM-generated artificial data, comparing it with human data across various existing benchmarks. Despite artificial data's capability to match human performance, this paper reveals significant hidden disparities, especially in complex tasks where LLMs often miss the nuanced understanding of intrinsic human-generated content. This study critically examines diverse LLM-generated data and emphasizes the need for ethical practices in data creation and when using LLMs. It highlights the LLMs' shortcomings in replicating human traits and behaviors, underscoring the importance of addressing biases and artifacts produced in LLM-generated content for future research and development. All data and code are available on our project page.
Current IR evaluation is based on relevance judgments, created either manually or automatically, with decisions outsourced to Large Language Models (LLMs). We offer an alternative paradigm, that never relies on relevance judgments in any form. Instead, a text is defined as relevant if it contains information that enables the answering of key questions. We use this idea to design the EXAM Answerability Metric to evaluate information retrieval/generation systems for their ability to provide topically relevant information. We envision the role of a human judge to edit and define an exam question bank that will test for the presence of relevant information in text. We support this step by generating an initial set of exam questions. In the next phase, an LLM-based question answering system will automatically grade system responses by tracking which exam questions are answerable with which system responses. We propose two evaluation measures, the recall-oriented EXAM Cover metric, and the precision-oriented EXAM Qrels metric, the latter which can be implemented with trec_eval. This paradigm not only allows for the expansion of the exam question set post-hoc but also facilitates the ongoing evaluation of future information systems, whether they focus on retrieval, generation, or both.