分享

High-quality automated poetry generation systems are currently only available for a small subset of languages. We introduce a new model for generating poetry in Czech language, based on fine-tuning a pre-trained Large Language Model. We demonstrate that guiding the generation process by explicitly specifying strophe parameters within the poem text strongly improves the effectiveness of the model. We also find that appropriate tokenization is crucial, showing that tokenization methods based on syllables or individual characters instead of subwords prove superior in generating poetic strophes. We further enhance the results by introducing \textit{Forced~generation}, adding explicit specifications of meter and verse parameters at inference time based on the already generated text. We evaluate a range of setups, showing that our proposed approach achieves high accuracies in rhyming and metric aspects of formal quality of the generated poems.

Accurately finding the wrong words in the automatic speech recognition (ASR) hypothesis and recovering them well-founded is the goal of speech error correction. In this paper, we propose a non-autoregressive speech error correction method. A Confidence Module measures the uncertainty of each word of the N-best ASR hypotheses as the reference to find the wrong word position. Besides, the acoustic feature from the ASR encoder is also used to provide the correct pronunciation references. N-best candidates from ASR are aligned using the edit path, to confirm each other and recover some missing character errors. Furthermore, the cross-attention mechanism fuses the information between error correction references and the ASR hypothesis. The experimental results show that both the acoustic and confidence references help with error correction. The proposed system reduces the error rate by 21% compared with the ASR model.

Large Language Models (LLMs) have emerged as a potential solution to assist digital health development with patient education, commonly medication-related enquires. We trained and validated Med-Pal, a medication domain-specific LLM-chatbot fine-tuned with a fine-grained and expert curated dataset from a selection of five light-weighted open-source LLMs of smaller parameter size (7 billion or less) regarding computational constraints and prioritizing operational efficiency. A multi-disciplinary team performed a clinical evaluation of LLMs responses using the SCORE criteria, focusing on safety, accuracy, bias, reproducibility, and ease of understanding. Best performing light-weighted LLM was chosen as Med-Pal for further engineering with guard-railing using adversarial prompting. Med-Pal and existing light-weighted LLMs, including pretrained Biomistral and finetuned Meerkat, were validated on an independent dataset on a broad range of medication-related questions (231 in total), 12 different question types across 14 different medication classes. Mistral-7b emerged as the top performer among selected lightweight LLMs, achieving the highest median score of 14 and 71.9% high-quality responses in accuracy and safety domains, hence chosen as the backbone LLM for Med-Pal. When compared against Biomistral, Med-pal outperformed in generating responses appropriate for patient communication, with significant reductions bias and errors typical of general LLMs. Comparable performance was observed when comparing Med-Pal with Meerkat. Med-Pal showcases the feasibility of developing and employing fine-tuned light-weighted LLMs to enhance digital health communications.

In this research short, we examine the potential of using GPT-4o, a state-of-the-art large language model (LLM) to undertake evidence synthesis and systematic assessment tasks. Traditional workflows for such tasks involve large groups of domain experts who manually review and synthesize vast amounts of literature. The exponential growth of scientific literature and recent advances in LLMs provide an opportunity to complementing these traditional workflows with new age tools. We assess the efficacy of GPT-4o to do these tasks on a sample from the dataset created by the Global Adaptation Mapping Initiative (GAMI) where we check the accuracy of climate change adaptation related feature extraction from the scientific literature across three levels of expertise. Our results indicate that while GPT-4o can achieve high accuracy in low-expertise tasks like geographic location identification, their performance in intermediate and high-expertise tasks, such as stakeholder identification and assessment of depth of the adaptation response, is less reliable. The findings motivate the need for designing assessment workflows that utilize the strengths of models like GPT-4o while also providing refinements to improve their performance on these tasks.

Exploratory data analysis (EDA) is an essential step for analyzing a dataset to derive insights. Several EDA techniques have been explored in the literature. Many of them leverage visualizations through various plots. But it is not easy to interpret them for a non-technical user, and producing appropriate visualizations is also tough when there are a large number of columns. Few other works provide a view of some interesting slices of data but it is still difficult for the user to draw relevant insights from them. Of late, conversational data exploration is gaining a lot of traction among non-technical users. It helps the user to explore the dataset without having deep technical knowledge about the data. Towards this, we propose a system that recommends interesting questions in natural language based on relevant slices of a dataset in a conversational setting. Specifically, given a dataset, we pick a select set of interesting columns and identify interesting slices of such columns and column combinations based on few interestingness measures. We use our own fine-tuned variation of a pre-trained language model(T5) to generate natural language questions in a specific manner. We then slot-fill values in the generated questions and rank them for recommendations. We show the utility of our proposed system in a coversational setting with a collection of real datasets.

We present an efficient method for adapting a monolingual Large Language Model (LLM) to another language, addressing challenges of catastrophic forgetting and tokenizer limitations. We focus this study on adapting Llama 2 to Arabic. Our two-stage approach begins with expanding the vocabulary and training only the embeddings matrix, followed by full model continual pre-training on a bilingual corpus. By continually pre-training on a mix of Arabic and English corpora, the model retains its proficiency in English while acquiring capabilities in Arabic. Our approach results in significant improvements in Arabic and slight enhancements in English, demonstrating cost-effective cross-lingual transfer. We perform ablations on embedding initialization techniques, data mix ratios, and learning rates and release a detailed training recipe. To demonstrate generalizability of this approach we also adapted Llama 3 8B to Arabic and Llama 2 13B to Hindi.

While diffusion models excel at conditional generating high-quality images, prior works in discrete diffusion models were not evaluated on conditional long-text generation. In this work, we address the limitations of prior discrete diffusion models for conditional long-text generation, particularly in long sequence-to-sequence tasks such as abstractive summarization. Despite fast decoding speeds compared to autoregressive methods, previous diffusion models failed on the abstractive summarization task due to the incompatibility between the backbone architectures and the random noising process. To overcome these challenges, we introduce a novel semantic-aware noising process that enables Transformer backbones to handle long sequences effectively. Additionally, we propose CrossMamba, an adaptation of the Mamba model to the encoder-decoder paradigm, which integrates seamlessly with the random absorbing noising process. Our approaches achieve state-of-the-art performance on three benchmark summarization datasets: Gigaword, CNN/DailyMail, and Arxiv, outperforming existing discrete diffusion models on ROUGE metrics as well as possessing much faster speed in inference compared to autoregressive models.

Long-text generation is seemingly ubiquitous in real-world applications of large language models such as generating an email or writing a review. Despite the fundamental importance and prevalence of long-text generation in many practical applications, existing work on personalized generation has focused on the generation of very short text. To overcome these limitations, we study the problem of personalized long-text generation, that is, generating long-text that is personalized for a specific user while being practically useful for the vast majority of real-world applications that naturally require the generation of longer text. In this work, we demonstrate the importance of user-specific personalization for long-text generation tasks and develop the Long-text Language Model Personalization (LongLaMP) Benchmark. LongLaMP provides a comprehensive and diverse evaluation framework for personalized long-text generation. Extensive experiments on LongLaMP for zero-shot and fine-tuned language tasks demonstrate the effectiveness of the proposed benchmark and its utility for developing and evaluating techniques for personalized long-text generation across a wide variety of long-text generation tasks. The results highlight the importance of personalization across a wide variety of long-text generation tasks. Finally, we release the benchmark for others to use for this important problem.

In this work we evaluate the utility of synthetic data for training automatic speech recognition (ASR). We use the ASR training data to train a text-to-speech (TTS) system similar to FastSpeech-2. With this TTS we reproduce the original training data, training ASR systems solely on synthetic data. For ASR, we use three different architectures, attention-based encoder-decoder, hybrid deep neural network hidden Markov model and a Gaussian mixture hidden Markov model, showing the different sensitivity of the models to synthetic data generation. In order to extend previous work, we present a number of ablation studies on the effectiveness of synthetic vs. real training data for ASR. In particular we focus on how the gap between training on synthetic and real data changes by varying the speaker embedding or by scaling the model size. For the latter we show that the TTS models generalize well, even when training scores indicate overfitting.

This article investigates how translation memories (TM) can be created by translators or other language professionals in order to compile domain-specific parallel corpora , which can then be used in different scenarios, such as machine translation training and fine-tuning, TM leveraging, and/or large language model fine-tuning. The article introduces a semi-automatic TM preparation methodology leveraging primarily translation tools used by translators in favor of data quality and control by the translators. This semi-automatic methodology is then used to build a cardiology-based Turkish -> English corpus from bilingual abstracts of Turkish cardiology journals. The resulting corpus called TRENCARD Corpus has approximately 800,000 source words and 50,000 sentences. Using this methodology, translators can build their custom TMs in a reasonable time and use them in their bilingual data requiring tasks.

In this work, we tackle the problem of person search, which is a challenging task consisted of pedestrian detection and person re-identification~(re-ID). Instead of sharing representations in a single joint model, we find that separating detector and re-ID feature extraction yields better performance. In order to extract more representative features for each identity, we segment out the foreground person from the original image patch. We propose a simple yet effective re-ID method, which models foreground person and original image patches individually, and obtains enriched representations from two separate CNN streams. From the experiments on two standard person search benchmarks of CUHK-SYSU and PRW, we achieve mAP of $83.0\%$ and $32.6\%$ respectively, surpassing the state of the art by a large margin (more than 5pp).

Representing speech as discrete tokens provides a framework for transforming speech into a format that closely resembles text, thus enabling the use of speech as an input to the widely successful large language models (LLMs). Currently, while several speech tokenizers have been proposed, there is ambiguity regarding the properties that are desired from a tokenizer for specific downstream tasks and its overall generalizability. Evaluating the performance of tokenizers across different downstream tasks is a computationally intensive effort that poses challenges for scalability. To circumvent this requirement, we present STAB (Speech Tokenizer Assessment Benchmark), a systematic evaluation framework designed to assess speech tokenizers comprehensively and shed light on their inherent characteristics. This framework provides a deeper understanding of the underlying mechanisms of speech tokenization, thereby offering a valuable resource for expediting the advancement of future tokenizer models and enabling comparative analysis using a standardized benchmark. We evaluate the STAB metrics and correlate this with downstream task performance across a range of speech tasks and tokenizer choices.