Confidence estimation of predictions from an End-to-End (E2E) Automatic Speech Recognition (ASR) model benefits ASR's downstream and upstream tasks. Class-probability-based confidence scores do not accurately represent the quality of overconfident ASR predictions. An ancillary Confidence Estimation Model (CEM) calibrates the predictions. State-of-the-art (SOTA) solutions use binary target scores for CEM training. However, the binary labels do not reveal the granular information of predicted words, such as temporal alignment between reference and hypothesis and whether the predicted word is entirely incorrect or contains spelling errors. Addressing this issue, we propose a novel Temporal-Lexeme Similarity (TeLeS) confidence score to train CEM. To address the data imbalance of target scores while training CEM, we use shrinkage loss to focus on hard-to-learn data points and minimise the impact of easily learned data points. We conduct experiments with ASR models trained in three languages, namely Hindi, Tamil, and Kannada, with varying training data sizes. Experiments show that TeLeS generalises well across domains. To demonstrate the applicability of the proposed method, we formulate a TeLeS-based Acquisition (TeLeS-A) function for sampling uncertainty in active learning. We observe a significant reduction in the Word Error Rate (WER) as compared to SOTA methods.
相關內容
Critique ability are crucial in the scalable oversight and self-improvement of Large Language Models (LLMs). While many recent studies explore the critique ability of LLMs to judge and refine flaws in generations, how to comprehensively and reliably measure the critique abilities of LLMs is under-explored. This paper introduces \shortname, a novel benchmark designed to comprehensively and reliably evaluate four key critique ability dimensions of LLMs: feedback, comparison, refinement and meta-feedback. CriticBench encompasses nine diverse tasks, each assessing the LLMs' ability to critique responses at varying levels of quality granularity. Our extensive evaluations of open-source and closed-source LLMs reveal intriguing relationships between the critique ability and tasks, response qualities, and model scales. Datasets, resources and evaluation toolkit for CriticBench will be publicly released at \url{//github.com/open-compass/CriticBench}.
Taming the generation outcome of state of the art Diffusion and Flow-Matching (FM) models without having to re-train a task-specific model unlocks a powerful tool for solving inverse problems, conditional generation, and controlled generation in general. In this work we introduce D-Flow, a simple framework for controlling the generation process by differentiating through the flow, optimizing for the source (noise) point. We motivate this framework by our key observation stating that for Diffusion/FM models trained with Gaussian probability paths, differentiating through the generation process projects gradient on the data manifold, implicitly injecting the prior into the optimization process. We validate our framework on linear and non-linear controlled generation problems including: image and audio inverse problems and conditional molecule generation reaching state of the art performance across all.
Large Language Models (LLMs) such as GPT and Llama have demonstrated significant achievements in summarization tasks but struggle with factual inaccuracies, a critical issue in clinical NLP applications where errors could lead to serious consequences. To counter the high costs and limited availability of expert-annotated data for factual alignment, this study introduces an innovative pipeline that utilizes GPT-3.5 and GPT-4 to generate high-quality feedback aimed at enhancing factual consistency in clinical note summarization. Our research primarily focuses on edit feedback, mirroring the practical scenario in which medical professionals refine AI system outputs without the need for additional annotations. Despite GPT's proven expertise in various clinical NLP tasks, such as the Medical Licensing Examination, there is scant research on its capacity to deliver expert-level edit feedback for improving weaker LMs or LLMs generation quality. This work leverages GPT's advanced capabilities in clinical NLP to offer expert-level edit feedback. Through the use of two distinct alignment algorithms (DPO and SALT) based on GPT edit feedback, our goal is to reduce hallucinations and align closely with medical facts, endeavoring to narrow the divide between AI-generated content and factual accuracy. This highlights the substantial potential of GPT edits in enhancing the alignment of clinical factuality.
The widespread adoption of implicit neural representations, especially Neural Radiance Fields (NeRF), highlights a growing need for editing capabilities in implicit 3D models, essential for tasks like scene post-processing and 3D content creation. Despite previous efforts in NeRF editing, challenges remain due to limitations in editing flexibility and quality. The key issue is developing a neural representation that supports local edits for real-time updates. Current NeRF editing methods, offering pixel-level adjustments or detailed geometry and color modifications, are mostly limited to static scenes. This paper introduces SealD-NeRF, an extension of Seal-3D for pixel-level editing in dynamic settings, specifically targeting the D-NeRF network. It allows for consistent edits across sequences by mapping editing actions to a specific timeframe, freezing the deformation network responsible for dynamic scene representation, and using a teacher-student approach to integrate changes.
Large Language Models (LLMs) face threats from unsafe prompts. Existing methods for detecting unsafe prompts are primarily online moderation APIs or finetuned LLMs. These strategies, however, often require extensive and resource-intensive data collection and training processes. In this study, we propose GradSafe, which effectively detects unsafe prompts by scrutinizing the gradients of safety-critical parameters in LLMs. Our methodology is grounded in a pivotal observation: the gradients of an LLM's loss for unsafe prompts paired with compliance response exhibit similar patterns on certain safety-critical parameters. In contrast, safe prompts lead to markedly different gradient patterns. Building on this observation, GradSafe analyzes the gradients from prompts (paired with compliance responses) to accurately detect unsafe prompts. We show that GradSafe, applied to Llama-2 without further training, outperforms Llama Guard, despite its extensive finetuning with a large dataset, in detecting unsafe prompts. This superior performance is consistent across both zero-shot and adaptation scenarios, as evidenced by our evaluations on the ToxicChat and XSTest. The source code is available at //github.com/xyq7/GradSafe.
We present CFEVER, a Chinese dataset designed for Fact Extraction and VERification. CFEVER comprises 30,012 manually created claims based on content in Chinese Wikipedia. Each claim in CFEVER is labeled as "Supports", "Refutes", or "Not Enough Info" to depict its degree of factualness. Similar to the FEVER dataset, claims in the "Supports" and "Refutes" categories are also annotated with corresponding evidence sentences sourced from single or multiple pages in Chinese Wikipedia. Our labeled dataset holds a Fleiss' kappa value of 0.7934 for five-way inter-annotator agreement. In addition, through the experiments with the state-of-the-art approaches developed on the FEVER dataset and a simple baseline for CFEVER, we demonstrate that our dataset is a new rigorous benchmark for factual extraction and verification, which can be further used for developing automated systems to alleviate human fact-checking efforts. CFEVER is available at //ikmlab.github.io/CFEVER.
Neural Information Retrieval (NIR) has significantly improved upon heuristic-based IR systems. Yet, failures remain frequent, the models used often being unable to retrieve documents relevant to the user's query. We address this challenge by proposing a lightweight abstention mechanism tailored for real-world constraints, with particular emphasis placed on the reranking phase. We introduce a protocol for evaluating abstention strategies in a black-box scenario, demonstrating their efficacy, and propose a simple yet effective data-driven mechanism. We provide open-source code for experiment replication and abstention implementation, fostering wider adoption and application in diverse contexts.
In his 2022 IMS Medallion Lecture delivered at the Joint Statistical Meetings, Prof. Dylan S. Small eloquently advocated for the use of protocols in observational studies. We discuss his proposal and, inspired by his ideas, we develop a protocol for the regression discontinuity design.
This paper introduces the batch-parallel Compressed Packed Memory Array (CPMA), a compressed, dynamic, ordered set data structure based on the Packed Memory Array (PMA). Traditionally, batch-parallel sets are built on pointer-based data structures such as trees because pointer-based structures enable fast parallel unions via pointer manipulation. When compared with cache-optimized trees, PMAs were slower to update but faster to scan. The batch-parallel CPMA overcomes this tradeoff between updates and scans by optimizing for cache-friendliness. On average, the CPMA achieves 3x faster batch-insert throughput and 4x faster range-query throughput compared with compressed PaC-trees, a state-of-the-art batch-parallel set library based on cache-optimized trees. We further evaluate the CPMA compared with compressed PaC-trees and Aspen, a state-of-the-art system, on a real-world application of dynamic-graph processing. The CPMA is on average 1.2x faster on a suite of graph algorithms and 2x faster on batch inserts when compared with compressed PaC-trees. Furthermore, the CPMA is on average 1.3x faster on graph algorithms and 2x faster on batch inserts compared with Aspen.
Adversarial Attacks on Face Recognition (FR) encompass two types: impersonation attacks and evasion attacks. We observe that achieving a successful impersonation attack on FR does not necessarily ensure a successful dodging attack on FR in the black-box setting. Introducing a novel attack method named Pre-training Pruning Restoration Attack (PPR), we aim to enhance the performance of dodging attacks whilst avoiding the degradation of impersonation attacks. Our method employs adversarial example pruning, enabling a portion of adversarial perturbations to be set to zero, while tending to maintain the attack performance. By utilizing adversarial example pruning, we can prune the pre-trained adversarial examples and selectively free up certain adversarial perturbations. Thereafter, we embed adversarial perturbations in the pruned area, which enhances the dodging performance of the adversarial face examples. The effectiveness of our proposed attack method is demonstrated through our experimental results, showcasing its superior performance.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191