Faithfulness in Natural Language Generation: A Systematic Survey of Analysis, Evaluation and Optimization Methods

Abstract

Natural Language Generation (NLG) has made great progress in recent years due to the development of deep learning techniques such as pre-trained language models. This advancement has resulted in more fluent, coherent and even properties controllable (e.g. stylistic, sentiment, length etc.) generation, naturally leading to development in downstream tasks such as abstractive summarization, dialogue generation, machine translation, and data-to-text generation. However, the faithfulness problem that the generated text usually contains unfaithful or non-factual information has become the biggest challenge, which makes the performance of text generation unsatisfactory for practical applications in many real-world scenarios. Many studies on analysis, evaluation, and optimization methods for faithfulness problems have been proposed for various tasks, but have not been organized, compared and discussed in a combined manner. In this survey, we provide a systematic overview of the research progress on the faithfulness problem of NLG, including problem analysis, evaluation metrics and optimization methods. We organize the evaluation and optimization methods for different tasks into a unified taxonomy to facilitate comparison and learning across tasks. Several research trends are discussed further. NLG Fluency Grammatical Coherence InformativenessDiversity Specificity Redundnacy Controllability Stylistic Attributes Content Faithfulness Consistency Fidelity Factuality Figure 1: Four aspects of the NLG challenge. Faithfulness has become the biggest challenge in modern natural language generation.

Faithful AI in Medicine: A Systematic Review with Large Language Models and Beyond

“…Ground truth summary Despite the promises of medical AI 16 , one major concern that has attracted a lot of attention is its potential risk of generating non-factual or unfaithful information 17 , which is usually referred to as the faithfulness problem 18 . Specifically, generative AI methods can generate contents that are factually inaccurate or biased.…”

“…Faithfulness generally means being loyal to something or some person. In the context of AI or NLP, faithful AI means the algorithm can produce contents that are factually correct, namely staying faithful to facts 18,21 . Generative medical AI systems 11,22,23 learn to map from various types of medical data such as electronic health records (EHRs), medical images or protein sequences, to desired output such as the summarization or explanation of medical scans, radiology reports, and • Extrinsic Error: the generated output cannot be confirmed (either supported or contradicted) by existing knowledge, reference or data.…”

Studying psychosis using Natural Language Generation: A review of emerging opportunities

“…In a similar manner, NLG degeneration with repetitions, non-sequiturs and dead-ends are more frequent in NLG systems that generate less fluent and thus less engaging content, though this reduces the number of unverifiable statements. When adjustments are made to increase fluency and reduce degeneration, as in using DLNs, unverifiable statements (33,43) increase in frequency with more than 70% of certain fluent outputs being falsehoods in some cases (44).…”

Studying Psychosis Using Natural Language Generation: A Review of Emerging Opportunities

“…Similarly, NLG degeneration with repetitions, nonsequiturs and dead-ends are more frequent in NLG systems that generate less fluent and thus less engaging content, though this reduces the number of unverifiable statements. When adjustments are made to increase fluency and reduce degeneration, as in using DLNs, unverifiable statements (37,47) increase in frequency with more than 70% of certain fluent outputs being falsehoods in some cases (48).…”