Typhoon: Thai Large Language Models

Typhoon is a series of Thai large language models (LLMs) developed specifically for the Thai language. This technical report presents challenges and insights in developing Thai LLMs, including data preparation, pretraining, instruction-tuning, and evaluation. As one of the challenges of low-resource languages is the amount of pretraining data, we apply continual training to transfer existing world knowledge from a strong LLM. To evaluate the Thai knowledge encapsulated in each model from the pretraining stage, we develop ThaiExam, a benchmark based on examinations for high-school students and investment professionals in Thailand. In addition, we fine-tune Typhoon to follow Thai instructions, and we evaluate instruction-tuned models on Thai instruction datasets as well as translation, summarization, and question-answering tasks. Experimental results on a suite of Thai benchmarks show that Typhoon outperforms all open-source Thai language models, and its performance is on par with GPT-3.5 in Thai while having only 7 billion parameters and being 2.62 times more efficient in tokenizing Thai text.

Mitigating Word Bias in Zero-shot Prompt-based Classifiers

Prompt-based classifiers are an attractive approach for zero-shot classification. However, the precise choice of the prompt template and label words can largely influence performance, with semantically equivalent settings often showing notable performance difference. This discrepancy can be partly attributed to word biases, where the classifier may be biased towards classes. To address this problem, it is possible to optimise classification thresholds on a labelled data set, however, this mitigates some of the advantages of prompt-based classifiers. This paper instead approaches this problem by examining the expected marginal probabilities of the classes. Here, probabilities are reweighted to have a uniform prior over classes, in an unsupervised fashion. Further, we draw a theoretical connection between the class priors and the language models' word prior, and offer the ability to set a threshold in a zero-resource fashion. We show that matching class priors correlates strongly with the oracle upper bound performance and demonstrate large consistent performance gains for prompt settings over a range of NLP tasks.

Zero-shot NLG evaluation through Pairware Comparisons with LLMs

Evaluating Natural Language Generation (NLG) outputs is crucial but laborious and expensive. While various automatic NLG assessment methods have been proposed, they often are quite task-specific and have to be engineered with a particular domain and attribute in mind. In this work, we propose a robust zero-shot approach to NLG evaluation using pairwise comparative judgment with open-source Large Language Models (LLMs). The motivation for this approach is that even as humans, it is easier to determine which of two options are better, than it is to independently objectively score each option. We use this insight and leverage the emergent abilities of LLMs, where we probe FlanT5 to determine which of two candidate responses is better, rather than assigning absolute scores. Our results demonstrate that comparative assessment is a more effective approach than absolute scoring, enabling smaller open-source LLMs to achieve comparable performance to larger public access APIs. We evaluate systems on both summary evaluation and dialogue response generation, and show that opensource LLMs can lead to good correlations with human scores for a range of different attributes.

Can Generative Large Language Models Perform ASR Error Correction?

ASR error correction continues to serve as an important part of post-processing for speech recognition systems. Traditionally, these models are trained with supervised training using the decoding results of the underlying ASR system and the reference text. This approach is computationally intensive and the model needs to be re-trained when switching the underlying ASR model. Recent years have seen the development of large language models and their ability to perform natural language processing tasks in a zero-shot manner. In this paper, we take ChatGPT as an example to examine its ability to perform ASR error correction in the zero-shot or 1-shot settings. We use the ASR N-best list as model input and propose unconstrained error correction and N-best constrained error correction methods. Results on a Conformer-Transducer model and the pre-trained Whisper model show that we can largely improve the ASR system performance with error correction using the powerful ChatGPT model.

CUED at ProbSum 2023: Hierarchical Ensemble of Summarization Models

In this paper, we consider the challenge of summarizing patients' medical progress notes in a limited data setting. For the Problem List Summarization (shared task 1A) at the BioNLP Workshop 2023, we demonstrate that Clinical-T5 fine-tuned to 765 medical clinic notes outperforms other extractive, abstractive and zero-shot baselines, yielding reasonable baseline systems for medical note summarization. Further, we introduce Hierarchical Ensemble of Summarization Models (HESM), consisting of token-level ensembles of diverse fine-tuned Clinical-T5 models, followed by Minimum Bayes Risk (MBR) decoding. Our HESM approach lead to a considerable summarization performance boost, and when evaluated on held-out challenge data achieved a ROUGE-L of 32.77, which was the best-performing system at the top of the shared task leaderboard.

SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models

Generative Large Language Models (LLMs) such as GPT-3 are capable of generating highly fluent responses to a wide variety of user prompts. However, LLMs are known to hallucinate facts and make non-factual statements which can undermine trust in their output. Existing fact-checking approaches either require access to token-level output probability distribution (which may not be available for systems such as ChatGPT) or external databases that are interfaced via separate, often complex, modules. In this work, we propose "SelfCheckGPT", a simple sampling-based approach that can be used to fact-check black-box models in a zero-resource fashion, i.e. without an external database. SelfCheckGPT leverages the simple idea that if a LLM has knowledge of a given concept, sampled responses are likely to be similar and contain consistent facts. However, for hallucinated facts, stochastically sampled responses are likely to diverge and contradict one another. We investigate this approach by using GPT-3 to generate passages about individuals from the WikiBio dataset, and manually annotate the factuality of the generated passages. We demonstrate that SelfCheckGPT can: i) detect non-factual and factual sentences; and ii) rank passages in terms of factuality. We compare our approach to several existing baselines and show that in sentence hallucination detection, our approach has AUC-PR scores comparable to grey-box methods, while SelfCheckGPT is best at passage factuality assessment.

MQAG: Multiple-choice Question Answering and Generation for Assessing Information Consistency in Summarization

State-of-the-art summarization systems can generate highly fluent summaries. These summaries, however, may contain factual inconsistencies and/or information not present in the source. Hence, an important component of assessing the quality of summaries is to determine whether there is information consistency between the source and the summary. Existing approaches are typically based on lexical matching or representation-based methods. In this work, we introduce an alternative scheme based on standard information-theoretic measures in which the information present in the source and summary is directly compared. We propose a Multiple-choice Question Answering and Generation framework, MQAG, which approximates the information consistency by computing the expected KL-divergence between summary and source answer distributions over automatically generated multiple-choice questions. This approach exploits multiple-choice answer probabilities, as predicted answer distributions can be easily compared. We conduct experiments on four summary evaluation datasets: QAG-CNNDM/XSum, XSum-Faithfulness, Podcast Assessment, and SummEval. Experiments show that MQAG (using models trained on RACE) outperforms existing evaluation methods on the majority of tasks.

Podcast Summary Assessment: A Resource for Evaluating Summary Assessment Methods

Automatic summary assessment is useful for both machine-generated and human-produced summaries. Automatically evaluating the summary text given the document enables, for example, summary generation system development and detection of inappropriate summaries. Summary assessment can be run in a number of modes: ranking summary generation systems; ranking summaries of a particular document; and estimating the quality of a document-summary pair on an absolute scale. Existing datasets with annotation for summary assessment are usually based on news summarization datasets such as CNN/DailyMail or XSum. In this work, we describe a new dataset, the podcast summary assessment corpus, a collection of podcast summaries that were evaluated by human experts at TREC2020. Compared to existing summary assessment data, this dataset has two unique aspects: (i) long-input, speech podcast based, documents; and (ii) an opportunity to detect inappropriate reference summaries in podcast corpus. First, we examine existing assessment methods, including model-free and model-based methods, and provide benchmark results for this long-input summary assessment dataset. Second, with the aim of filtering reference summary-document pairings for training, we apply summary assessment for data selection. The experimental results on these two aspects provide interesting insights on the summary assessment and generation tasks. The podcast summary assessment data is available.

Sparsity and Sentence Structure in Encoder-Decoder Attention of Summarization Systems

Transformer models have achieved state-of-the-art results in a wide range of NLP tasks including summarization. Training and inference using large transformer models can be computationally expensive. Previous work has focused on one important bottleneck, the quadratic self-attention mechanism in the encoder. Modified encoder architectures such as LED or LoBART use local attention patterns to address this problem for summarization. In contrast, this work focuses on the transformer's encoder-decoder attention mechanism. The cost of this attention becomes more significant in inference or training approaches that require model-generated histories. First, we examine the complexity of the encoder-decoder attention. We demonstrate empirically that there is a sparse sentence structure in document summarization that can be exploited by constraining the attention mechanism to a subset of input sentences, whilst maintaining system performance. Second, we propose a modified architecture that selects the subset of sentences to constrain the encoder-decoder attention. Experiments are carried out on abstractive summarization tasks, including CNN/DailyMail, XSum, Spotify Podcast, and arXiv.

Long-Span Summarization via Local Attention and Content Selection

Transformer-based models have achieved state-of-the-art results in a wide range of natural language processing (NLP) tasks including document summarization. Typically these systems are trained by fine-tuning a large pre-trained model to the target task. One issue with these transformer-based models is that they do not scale well in terms of memory and compute requirements as the input length grows. Thus, for long document summarization, it can be challenging to train or fine-tune these models. In this work, we exploit large pre-trained transformer-based models and address long-span dependencies in abstractive summarization using two methods: local self-attention; and explicit content selection. These approaches are compared on a range of network configurations. Experiments are carried out on standard long-span summarization tasks, including Spotify Podcast, arXiv, and PubMed datasets. We demonstrate that by combining these methods, we can achieve state-of-the-art results on all three tasks in the ROUGE scores. Moreover, without a large-scale GPU card, our approach can achieve comparable or better results than existing approaches.