分享

This paper empirically investigates the relationship between subword vocabulary size and the performance of large language models (LLMs) to provide insights on how to define the vocabulary size. Experimental results show that larger vocabulary sizes lead to better performance in LLMs. Moreover, we consider a continual training scenario where a pre-trained language model is trained on a different target language. We introduce a simple method to use a new vocabulary instead of the pre-defined one. We show that using the new vocabulary outperforms the model with the vocabulary used in pre-training.

Adversarial datasets should ensure AI robustness that matches human performance. However, as models evolve, datasets can become obsolete. Thus, adversarial datasets should be periodically updated based on their degradation in adversarialness. Given the lack of a standardized metric for measuring adversarialness, we propose AdvScore, a human-grounded evaluation metric. AdvScore assesses a dataset's true adversarialness by capturing models' and humans' varying abilities, while also identifying poor examples. AdvScore then motivates a new dataset creation pipeline for realistic and high-quality adversarial samples, enabling us to collect an adversarial question answering (QA) dataset, AdvQA. We apply AdvScore using 9,347 human responses and ten language model predictions to track the models' improvement over five years (from 2020 to 2024). AdvScore assesses whether adversarial datasets remain suitable for model evaluation, measures model improvements, and provides guidance for better alignment with human capabilities.

Contextualized end-to-end automatic speech recognition has been an active research area, with recent efforts focusing on the implicit learning of contextual phrases based on the final loss objective. However, these approaches ignore the useful contextual knowledge encoded in the intermediate layers. We hypothesize that employing explicit biasing loss as an auxiliary task in the encoder intermediate layers may better align text tokens or audio frames with the desired objectives. Our proposed intermediate biasing loss brings more regularization and contextualization to the network. Our method outperforms a conventional contextual biasing baseline on the LibriSpeech corpus, achieving a relative improvement of 22.5% in biased word error rate (B-WER) and up to 44% compared to the non-contextual baseline with a biasing list size of 100. Moreover, employing RNN-transducer-driven joint decoding further reduces the unbiased word error rate (U-WER), resulting in a more robust network.

Decoder-only language models (LMs) have been successfully adopted for speech-processing tasks including automatic speech recognition (ASR). The LMs have ample expressiveness and perform efficiently. This efficiency is a suitable characteristic for streaming applications of ASR. In this work, we propose to use a decoder-only architecture for blockwise streaming ASR. In our approach, speech features are compressed using CTC output and context embedding using blockwise speech subnetwork, and are sequentially provided as prompts to the decoder. The decoder estimates the output tokens promptly at each block. To this end, we also propose a novel training scheme using random-length prefix prompts to make the model robust to the truncated prompts caused by blockwise processing. An experimental comparison shows that our proposed decoder-only streaming ASR achieves 8% relative word error rate reduction in the LibriSpeech test-other set while being twice as fast as the baseline model.

Understanding the abilities of LLMs to reason about natural language plans, such as instructional text and recipes, is critical to reliably using them in decision-making systems. A fundamental aspect of plans is the temporal order in which their steps needs to be executed, which reflects the underlying causal dependencies between them. We introduce CaT-Bench, a benchmark of Step Order Prediction questions, which test whether a step must necessarily occur before or after another in cooking recipe plans. We use this to evaluate how well frontier LLMs understand causal and temporal dependencies. We find that SOTA LLMs are underwhelming (best zero-shot is only 0.59 in F1), and are biased towards predicting dependence more often, perhaps relying on temporal order of steps as a heuristic. While prompting for explanations and using few-shot examples improve performance, the best F1 result is only 0.73. Further, human evaluation of explanations along with answer correctness show that, on average, humans do not agree with model reasoning. Surprisingly, we also find that explaining after answering leads to better performance than normal chain-of-thought prompting, and LLM answers are not consistent across questions about the same step pairs. Overall, results show that LLMs' ability to detect dependence between steps has significant room for improvement.

Large language model unlearning has garnered increasing attention due to its potential to address security and privacy concerns, leading to extensive research in the field. However, much of this research has concentrated on instance-level unlearning, specifically targeting the removal of predefined instances containing sensitive content. This focus has left a significant gap in the exploration of full entity-level unlearning, which is critical in real-world scenarios such as copyright protection. To this end, we propose a novel task of Entity-level unlearning, which aims to erase entity-related knowledge from the target model completely. To thoroughly investigate this task, we systematically evaluate trending unlearning algorithms, revealing that current methods struggle to achieve effective entity-level unlearning. Then, we further explore the factors that influence the performance of the unlearning algorithms, identifying that knowledge coverage and the size of the forget set play pivotal roles. Notably, our analysis also uncovers that entities introduced through fine-tuning are more vulnerable to unlearning than pre-trained entities. These findings collectively offer valuable insights for advancing entity-level unlearning for LLMs.

As edge-based automatic speech recognition (ASR) technologies become increasingly prevalent for the development of intelligent and personalized assistants, three important challenges must be addressed for these resource-constrained ASR models, i.e., adaptivity, incrementality, and inclusivity. We propose a novel ASR framework, PI-Whisper, in this work and show how it can improve an ASR's recognition capabilities adaptively by identifying different speakers' characteristics in real-time, how such an adaption can be performed incrementally without repetitive retraining, and how it can improve the equity and fairness for diverse speaker groups. More impressively, our proposed PI-Whisper framework attains all of these nice properties while still achieving state-of-the-art accuracy with up to 13.7% reduction of the word error rate (WER) with linear scalability with respect to computing resources.

The goal of text style transfer is to transform the style of texts while preserving their original meaning, often with only a few examples of the target style. Existing style transfer methods generally rely on the few-shot capabilities of large language models or on complex controllable text generation approaches that are inefficient and underperform on fluency metrics. We introduce TinyStyler, a lightweight but effective approach, which leverages a small language model (800M params) and pre-trained authorship embeddings to perform efficient, few-shot text style transfer. We evaluate on the challenging task of authorship style transfer and find TinyStyler outperforms strong approaches such as GPT-4. We also evaluate TinyStyler's ability to perform text attribute style transfer (formal $\leftrightarrow$ informal) with automatic and human evaluations and find that the approach outperforms recent controllable text generation methods. Our model has been made publicly available at https://huggingface.co/tinystyler/tinystyler .

This paper describes our solution for the Diarization of Speaker and Language in Conversational Environments Challenge (Displace 2023). We used a combination of VAD for finding segfments with speech, Resnet architecture based CNN for feature extraction from these segments, and spectral clustering for features clustering. Even though it was not trained with using Hindi, the described algorithm achieves the following metrics: DER 27. 1% and DER 27. 4%, on the development and phase-1 evaluation parts of the dataset, respectively.

Mental disorders pose a global challenge, aggravated by the shortage of qualified mental health professionals. Mental disorder prediction from social media posts by current LLMs is challenging due to the complexities of sequential text data and the limited context length of language models. Current language model-based approaches split a single data instance into multiple chunks to compensate for limited context size. The predictive model is then applied to each chunk individually, and the most voted output is selected as the final prediction. This results in the loss of inter-post dependencies and important time variant information, leading to poor performance. We propose a novel framework which first compresses the large sequence of chronologically ordered social media posts into a series of numbers. We then use this time variant representation for mental disorder classification. We demonstrate the generalization capabilities of our framework by outperforming the current SOTA in three different mental conditions: depression, self-harm, and anorexia, with an absolute improvement of 5% in the F1 score. We investigate the situation where current data instances fall within the context length of language models and present empirical results highlighting the importance of temporal properties of textual data. Furthermore, we utilize the proposed framework for a cross-domain study, exploring commonalities across disorders and the possibility of inter-domain data usage.

Emergency department (ED) overcrowding and the complexity of rapid decision-making in critical care settings pose significant challenges to healthcare systems worldwide. While clinical decision support systems (CDSS) have shown promise, the integration of large language models (LLMs) offers new possibilities for enhancing triage accuracy and clinical decision-making. This study presents an LLM-driven CDSS designed to assist ED physicians and nurses in patient triage, treatment planning, and overall emergency care management. We developed a multi-agent CDSS utilizing Llama-3-70b as the base LLM, orchestrated by CrewAI and Langchain. The system comprises four AI agents emulating key ED roles: Triage Nurse, Emergency Physician, Pharmacist, and ED Coordinator. It incorporates the Korean Triage and Acuity Scale (KTAS) for triage assessment and integrates with the RxNorm API for medication management. The model was evaluated using the Asclepius dataset, with performance assessed by a clinical emergency medicine specialist. The CDSS demonstrated high accuracy in triage decision-making compared to the baseline of a single-agent system. Furthermore, the system exhibited strong performance in critical areas, including primary diagnosis, critical findings identification, disposition decision-making, treatment planning, and resource allocation. Our multi-agent CDSS demonstrates significant potential for supporting comprehensive emergency care management. By leveraging state-of-the-art AI technologies, this system offers a scalable and adaptable tool that could enhance emergency medical care delivery, potentially alleviating ED overcrowding and improving patient outcomes. This work contributes to the growing field of AI applications in emergency medicine and offers a promising direction for future research and clinical implementation.

With significant advancements in Transformers LLMs, NLP has extended its reach into many research fields due to its enhanced capabilities in text generation and user interaction. One field benefiting greatly from these advancements is cybersecurity. In cybersecurity, many parameters that need to be protected and exchanged between senders and receivers are in the form of text and tabular data, making NLP a valuable tool in enhancing the security measures of communication protocols. This survey paper provides a comprehensive analysis of the utilization of Transformers and LLMs in cyber-threat detection systems. The methodology of paper selection and bibliometric analysis is outlined to establish a rigorous framework for evaluating existing research. The fundamentals of Transformers are discussed, including background information on various cyber-attacks and datasets commonly used in this field. The survey explores the application of Transformers in IDSs, focusing on different architectures such as Attention-based models, LLMs like BERT and GPT, CNN/LSTM-Transformer hybrids, emerging approaches like ViTs, among others. Furthermore, it explores the diverse environments and applications where Transformers and LLMs-based IDS have been implemented, including computer networks, IoT devices, critical infrastructure protection, cloud computing, SDN, as well as in autonomous vehicles. The paper also addresses research challenges and future directions in this area, identifying key issues such as interpretability, scalability, and adaptability to evolving threats, and more. Finally, the conclusion summarizes the findings and highlights the significance of Transformers and LLMs in enhancing cyber-threat detection capabilities, while also outlining potential avenues for further research and development.