Improved Natural Language Generation via Loss Truncation

By Daniel Kang, Tatsunori B. Hashimoto. 2020

105 citations recorded by the source corpus.

Neural language models are usually trained to match the distributional properties of large-scale corpora by minimizing the log loss. While straightforward to optimize, this approach forces the model to reproduce all variations in the dataset, including noisy and invalid references (e.g., misannotations and hallucinated facts). Even a small fraction of noisy data can degrade the performance of log loss. As an alternative, prior work has shown that minimizing the distinguishability of generated samples is a principled and robust loss that can handle invalid references. However, distinguishability has not been used in practice due to challenges in optimization and estimation. We propose loss truncation: a simple and scalable procedure which adaptively removes high log loss examples as a way to optimize for distinguishability. Empirically, we demonstrate that loss truncation outperforms existing baselines on distinguishability on a summarization task. Furthermore, we show that samples generated by the loss truncation model have factual accuracy ratings that exceed those of baselines and match human references.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

Citations in the corpus

Survey of Hallucination in Natural Language Generation (2022, methodology): “This is similar to the method proposed by Kang and Hashimoto [61], the latter for full sentences in the summarization task.”
Detecting Hallucinated Content in Conditional Neural Sequence Generation (2020, neutral): “We also compare with the recently proposed loss truncation method (Kang and Hashimoto, 2020) that adaptively removes entire examples with high log loss, which was shown to reduce hallucinations.”
Evaluating Attribution in Dialogue Systems: The BEGIN Benchmark (2021, background): “Because humans often include unattributable information in their utterances, language models trained on those corpora can replicate and perhaps even amplify the prevalence of unattributable responses at test time (Kang and Hashimoto, 2020; Dziri et al., 2022b).”
Trusting Your Evidence: Hallucinate Less with Context-aware Decoding (2023, background): “, 2020) and removing noisy training samples (Kang and Hashimoto, 2020; Goyal and Durrett, 2021).”
Improving Faithfulness in Abstractive Summarization with Contrast Candidate Generation and Selection (2021, background): “light the existence of hallucinations in training data, truncating potentially unfaithful gold summaries during training is an effective strategy (Kang and Hashimoto, 2020; Filippova, 2020).”
Language Model Evaluation Beyond Perplexity (2021, background): “, truncated loss (Kang and Hashimoto, 2020) or ↵entmax loss (Peters et al.”
Text Generation by Learning from Demonstrations (2020, unknown): “Recently, a concurrent work, Kang & Hashimoto (2020), proposed truncated log-loss which both optimizes distinguishability and enjoys efﬁcient optimization.”
Faithful or Extractive? On Mitigating the Faithfulness-Abstractiveness Trade-off in Abstractive Summarization (2021, methodology): “Gigaword reference summaries have substantially less hallucinations than XSum (Kang and Hashimoto, 2020), and WikiHow summaries tend to be of a higher quality since they are written and curated by humans (Koupaee and Wang, 2018; Ladhak et al., 2020).”
Benchmarking Large Language Models for News Summarization (2023, methodology): “As a result, the reference summaries in these datasets are known to have quality issues (Maynez et al., 2020; Kang and Hashimoto, 2020), motivating us to addresses these defects to improve LLM evaluation.”
FactPEGASUS: Factuality-Aware Pre-training and Fine-tuning for Abstractive Summarization (2022, methodology): “Our model lies above the graph, performing better than MLE-baseline, DAE (Goyal and Durrett, 2021), and Loss Truncation (Kang and Hashimoto, 2020). tuning process, we extend our analysis by studying the factuality dynamics, similar to Goyal et al. (2022).”
Hallucinated but Factual! Inspecting the Factuality of Hallucinations in Abstractive Summarization (2021, methodology): “Kang and Hashimoto (2020) propose a “loss truncation” training algorithm that filters out noisy training samples which may lead a model to hallucinate. Zhao et al. (2020) use a verification system to recognize non-factual quantities in summaries and adopt a re-ranking system to reduce the number of ”
On the Origin of Hallucinations in Conversational Models: Is it the Datasets or the Models? (2022, background): “, 2021; Wang and Sennrich, 2020) and summarization (Durmus et al., 2020; Kang and Hashimoto, 2020).”