A Synthetic Data Approach for Domain Generalization of NLI Models (2024)

1 Introduction

Over the past decade, NLI tasks have been critical for benchmarking the representation strengths of our language models. Today, the accuracy on the oft-reported Multi-NLI Williams etal. (2018) (MNLI) dataset has reached above 92%¹¹1https://gluebenchmark.com/leaderboard and supersedes human-level performance Nangia and Bowman (2019). Simultaneously, the practical applications of NLI models has gathered immense attention in fact-checking and source attribution of LLM outputs Honovich etal. (2022); Rashkin etal. (2023). For such downstream tasks, model generalization to new domains and data-distributions is critical. We present a method of synthetic data generation to create a general dataset with varied but balanced distribution of premise lengths, domains of text, and NLI labels (Table 1), and demonstrate improved accuracy with the new data.

MNLI was a first effort to create a multi-domain NLI training dataset with examples from 5 genres spanning fiction, formal texts and conversations, and with single-sentence premise/hypothesis texts. Today, the use of models trained from such datasets has expanded well past routine benchmarking into a variety of practical tasks. In downstream problems involving web-scale LLMs, such as fact-checking of social media text, the domains and texts are clearly more diverse. Yet, as a field, we have not fully explored the distribution-general abilities of our models for downstream semantic tasks beyond NLI itself. There are many other NLI training sets: ANLI Nie etal. (2019) for harder reasoning going beyond stylistic elements, and WANLI Liu etal. (2022) that generates synthetic examples through worker and AI collaboration and which replicate complex reasoning patterns.

To complement these efforts, we explorethe opportunity for synthetic high-quality datasets to adapt NLI models for zero-shot use in downstream applications across new and unseen text domains.

Our generation covers nearly 40 realistic and distinct domains, ranging from reviews, social media comments, to legal texts, also with varying lengths of the premise text. Our technique employs a chain of LLM tasks tuned to generate high quality, creative premise-hypothesis pairs together with a 3-way NLI label (entailment, contradiction, or neutral Williams etal. (2018)). A first step generates domain names, the second produces premises of different lengths in these domains, and the final LLM call produces hypotheses and labels conditioned on each premise. We demonstrate how this approaches creates data with a balanced distribution of domains, labels, and premise lengths.

We fine-tune NLI models on this synthetic data corpus, and present their accuracy on the TRUE factual consistency benchmark Honovich etal. (2022), consisting of 11 tasks unseen by our data and other training sets. We show that our general data-trained models obtain state-of-the-art NLI performance and single-handedly outperform models trained on MNLI, ANLI, or WANLI with around $7\%$ improvement over the best alternative for T5 Raffel etal. (2020) small models. The gap is lower but still around $2\%$ for T5 XXL model size.

Our main contribution is thus that the general synthetic data approach improves the generalization power of NLI models, especially when small models and fast inference is key. We further show that, while in-distribution performance is hard to beat for tasks with in-distribution training data, our synthetic data can still improve in-distribution performance when used to augment the training data for models with sufficient capacity.

2 Related Work

We describe various ways of creating NLI data, and why models should generalize beyond the data.

Human annotation of NLI examples. The major datasetsavailable today were created via costly, time consuming annotation tasks,and significant human effort.Standardly, for datasets such as SNLI Bowman etal. (2015a) and MNLI Williams etal. (2018), the annotation starts with a premise sentencetaken verbatim from different sources. Annotators are asked to write a hypothesis sentence that isentailed, contradicted, or is neutral to the premise. This writing task is complex for humans, and it is well-known that the collected examples often have undesired stylistic artifacts, for example, thehypotheses alone beinghighly predictive of the label Gururangan etal. (2018).Later efforts (ANLI dataset) have focused on improving the quality of examples by including model adversaries into the annotation rounds Nie etal. (2019).

The diversity in genre in these datasets depends on the sources from which the premises are drawn.The SNLI dataset contains image captions Bowman etal. (2015b). The MNLI dataset Williams etal. (2018) has 10 domains from the Open American NationalCorpus (OANC)²²2https://anc.org/,only 5 of which are used for training. The ANLI dataset Nie etal. (2019) contains text from Wikipedia, news, fiction, formal spoken text, and causal or procedural text. In this way, their domain coverage is limited by the chosen corpora and licensing constraints. For example, none of these datasets contain reviews, forum discussions, and science texts, domains which are prevalent and important in applications.

Understandably, different sources and methods ofdata collection produces training examples of a certain style and distribution.The generalization of these NLI sets to fully new settings is an interesting problem Adila and Kang (2022), yetless explored. Our work aims to shed some insights here.

Domain generalization. This problem of training/test mismatch receives less importance during developmentsince models are trained and evaluated on splits of the same dataset. But real world applications of these models cannot assume that the test data matches the training distribution. Models need to be adapted to individual test domains using domain adaptation, or alternatively one could train domain-general models which work well on multiple unseen domains or distributions. We focus on this problem of domain generalization.

There is a large body on work on how training and optimization of models can beadapted for better generalization Muandet etal. (2013); Wang etal. (2021). Another well-knownapproach, especially in computer vision, is to augment the training data to increase its diversity and reduce model overfitting Tobin etal. (2017); Rahman etal. (2019); Zhou etal. (2020). Liu etal. (2020) uses data augmentation to overcome multiple adverserial attacks.In this work, we explore the usefulness of synthetic data generation for the domain-generalization of NLI models.

Synthetic NLI data generation. Today’s LLMs have opened up the possibility of synthetic data to aid many NLP tasks Puri etal. (2020); He etal. (2022); Agrawal etal. (2023); Li etal. (2023).For NLI, synthetic data has been used for different goals:augmenting small training sets and adding in-domainexamples for self-training Vu etal. (2021); He etal. (2022), and increasing the size of harderexamples Liu etal. (2022). Most of these methods prompt LLMs withsequences of premise-hypothesis sentence pairs.The pairs are then labelled by a teacher model.Liu etal. (2022) specialize the generation towards complex linguistic and reasoningpatterns inexisting datasets. We employ synthetic data to improve the diversityand balance of training data along the dimensions of domain, premise length,and label skew.

3 Synthesizing a General-NLI Dataset

We aim to generate NLI examples in different domains, and with premises of varied lengths.³³3Longer hypotheses are not of interest typically. A hypothesis is entailed if hypothesis is true given the premise. Long hypotheses are less likely to contain precise entailment and contradiction relations, with some exceptions such as summaries.

Generating synthetic examples for NLI is in fact a challenging problem. The goal is to produce a pair of texts, which exemplify the reasoning behind different NLI labels. But for the data to be useful, these texts must have creative content and language, and require reasoning. Prior synthetic data generation approaches He etal. (2022); Liu etal. (2022) generate thepremise and hypothesis sentences as a single sequence,followed by annotating the label using a teacher model or human raters. Instead, to achieve maximum control over multiple dimensions: genre, length, and different NLI labels, wegenerate this dataset in two steps: (i) enumerate diverse domains and generate premises in those domains (Section 3.1), (ii) generate hypotheses and labels given the premises (Section 3.2).Figure 1 provides an overall view into the LLM tasks we use for generating our data.

3.1 Generating Premises with Varied Lengths in Multiple Domains

Going beyond distinctions based on the source of a text, it is hard to define the boundaries of a domain in a strict manner. Properties of a text differ along many dimensions: its genre (e.g., news, poetry, or fiction), topic (e.g. politics, science), and the platform or venue for the content, either spoken or written (e.g., reddit, email, image captions, or telephone conversations).We adopt a practical perspective, and consider all these distinctions as the latent features leading to differences between text collections.It is in fact well-known that stylistic variations in NLI datasets impact generalization Belinkov etal. (2019); Adila and Kang (2022).

So we do not start with predefined domains. Rather, we first build a text-generation model which generates triples of domain name, text length, and text in the domain. The resulting texts from different domains are collected into our premise set. We build this model using few-shot prompting of FLAN-PaLM2 L (Unicorn)model Googleand etal. (2023)⁴⁴4Available from https://developers.generativeai.google. using texts from a few seed domains. We draw 18 in-prompt examples of varied lengths from 8 domains (news headlines, news, shopping reviews, wikipedia, movie reviews, place reviews, twitter and reddit post). The example texts are takenfrom public websites with a few edits if needed.⁵⁵5This includes news websites (BBC) for news and news headlines, e-commerce and review websites (eBay and thetechoutlook.com) for shopping reviews, Wikipedia, Google Play for movie reviews, citymaps.uk and top-rated.online websites for place reviews, X (Twitter) and Reddit. We provide the full list of seed examples in Appendix A. We also select these texts to be of different lengths.

Figure 2 shows our prompt.⁶⁶6We note that we also tried prompting without any instruction (just with few-shot examples) and the generated text had similar quality. The length category is set to either short for single sentences and paragraph for longer texts. We sample from this model, with a temperature of 1 to get creative new domains and texts. Ideally, these samples would directly be useful as premises (with domain and length labels). However, these samples were skewed towards certain domains, e.g., certain types of forums, which neither correspond to real word distributions of web-text, nor serve the purpose of a general model. So we first identify new domains generated by the model. We examined the new domains generated in about 1000 samples. Some generated domains were closely related to, or paraphrases of, each other; e.g. both travel forums and US travel forums were generated. Others were rare or noisy. So we manually selected 38 diverse domains including the seed domain names (Table 2).

We then generate balanced samples in these domains, and for the length labels (short, paragraph). We use the same prompt as in Figure 2, but substitute a new domain and length category of interest to it, to generate a text with those properties. This simple text generation model produced high quality and creative texts in different domains. We use these texts as the premises in our data.

ads, blog post, book reviews, casual dialog, chat message, email, essay, fans forum, forum post, google play reviews, government documents, legal, legal document, medical, movie plot, movie reviews, news, news comments, news headlines, phone conversation, place reviews, quora, recipe, reddit comment, reddit title, research paper abstract, scientific article, shopping reviews, song lyrics, sports news, story for kids, student forum, student papers, support forum, travel guides, twitter, wikipedia, youtube comments

3.2 Generating Hypotheses and Labels

We now discuss how we attach hypotheses and labels to our premises to generate complete NLI examples, i.e., (premise, hypothesis, label) triples.

We train LLMs to leverage existing NLI datasets, and learn the task of writing hypotheses for given premises.

Our model conditions on a premise to generate a (hypothesis, label) pair. We generate the label automatically (and accurately, details in next section), and do not need an additional human/teacher labelling step.This model is trained via prompt-tuning of FLAN-PaLM 540B Chowdhery etal. (2023); Chung etal. (2022)on the training split of the MNLI dataset. Figure 3 shows our prompt which has definitions for the three NLI labels similar to the MNLI annotation guidelines Williams etal. (2018).⁷⁷7We also tried another similar instruction for defining the task and labels (different wordings). However, we did not observe meaningful differences with the final prompt shown in the paper. This is probably because the LLM learns the task well after being prompt-tuned on a large set of examples (MNLI).

See Also

Human iNKT cell activation using glycolipids

We used prompt-tuning Lester etal. (2021) for training, in lieu of fine-tuning, for two reasons: a) With prompt-tuning, only a few embeddings are updated (100 in our experiments) leading to efficient training, and b) prompt-tuning provides regularization and avoids memorization of the training set details.⁸⁸8See details of our prompt-tuning running time and hyper-parameters in Appendix B. We note that in our preliminary experiments, we also tried just prompting LLMs (no training) to generate NLI examples; however, we did not obtain high quality examples.

Using the prompt-tuned model, we perform inference once on each of the premises obtained from Section 3.1.⁹⁹9We discard examples if a) the generated text for hypothesis-premise pair is mis-formatted, or b) the generated labels are not among entailment, neutral, and contradiction. Such errors account for less than $1$% of the generated data.We note that large models and regularization were important for creative generation of hypotheses. A T5 XL model (3B parameters) fine-tuned on the same task lead to examples with poor creativity and low utility for training. In many cases, the synthetic hypothesis was a subset of the premise (entailment) or had some simple modifications (e.g., negation) to introduce contradiction.

4 Experiments

We explore the strengths of our general-dataset (GNLI for brevity), by examining the model predictions on data unseen during training.

We compare models trained on our data with those trained on other large NLI training sets. We choose three such sources: the MNLI dataset ($392$K training examples), ANLI ($162$K) with examples that are harder for MNLI trained models, and WANLI ($10$2K), a dataset created by machine-human collaboration.We note that all of these datasets are collected with a similar methodology, i.e., given a premise (and optionally a label), annotators (or LLMs) write a hypothesis. The final label is then manually assigned to the example (if not given as input). In addition, WANLI and GNLI have used MNLI exemplars(few-shot or supervised learning) for data generation. So these datasets would have similar properties in theory.

For all these sources, we trained T5 models Raffel etal. (2020) as a standard test bed, and explore models of different sizes: small (60M), large (770M), and XXL (11B). We trained on the respective training splits, and tune hyper-parameters on the corresponding validation sets. For WANLI which does not contain a validation split, we used the MNLI validation data. We also trained models on the combination of GNLI and other datasets. For these combined models, we tuned hyper-parameters based on the classification accuracy on the development set of GNLI.¹⁰¹⁰10See hyper-parameter details in Appendix C.

We train two classifiers for each model size (e.g., T5 XXL) and training data (e.g., ANLI): a 3-way classifier with all the three labels, and a binary classifier. For the binary case, we convert each NLI dataset into a binary dataset with entailment and non-entailment (neutral and contradiction) labels. We use the binary classifiers and 3-way classifiers for factual consistency evaluation and NLI benchmarks, respectively.

4.1 Performance on Unseen Factual Consistency Benchmarks

We first test different models ondata unseen by all of them. We use the TRUE benchmark, a collection of 11 evaluation datasets that contain human annotations for factual consistency in diverse tasks. The tasks include: A) abstractive summarization: FRANK Pagnoni etal. (2021), SummEval Fabbri etal. (2021), MNBM Maynez etal. (2020), Wang etal. (2020), QAGS-CNNDM Wang etal. (2020), and QAGS-XSum Wang etal. (2020). B) dialogue generation: BEGIN Dziri etal. (2022), Q² Honovich etal. (2021), and DialFact Gupta etal. (2022). C) fact verification: FEVER Thorne etal. (2018), and VitaminC Schuster etal. (2021). D) paraphrase detection: PAWS Zhang etal. (2019). The benchmark standardizes the above datasets by converting all annotations to binary labels corresponding to whether the entire text is factuality consistent w.r.t the grounding text or not.This task is a downstream application of NLI models and importantly, the data in this benchmark was not created using the same protocol as NLI benchmarks.

We train different sizes of T5 models on MNLI, ANLI, WANLI, and GNLI. In addition, we report results on models trained on the mixture of MNLI, ANLI, and WANLI (M + A + W). Table 5 shows the results. Following previous work, we report the Receiver Operating Characteristic Area Under the Curve (ROC AUC) for binary detection of inconsistent examples. GNLI outperforms MNLI on all datasets across model sizes showing that it has much stronger generalization. On average, GNLI outperforms MNLI for T5-small trained models by a $8.58\%$ margin, T5-large trained models by $6.88\%$, and T5-XXL trained models by $2.96\%$.

To the best of our knowledge, the previously reported best NLI model on the TRUE benchmark was T5 XXL trained on ANLI Honovich etal. (2022); Gekhman etal. (2023). GNLI obtains $6.85\%$ improvement on average over the best alternative with a single dataset for T5 small (WANLI), $3.26\%$ for T5 large (ANLI), and $2.03\%$ for T5 XXL (WANLI). Therefore, GNLI obtains a new state-of-the-art result on TRUE, outperforming other models with large margins on average and on almost all of the individual test sets within TRUE. In addition, GNLI alone outperforms the mixture of MNLI, ANLI, and WANLI by a relatively large margin for T5 small and T5 large, and by a small margin for T5 XXL.

We note that Gekhman etal. (2023) have recently proposed a synthetic dataset, called TrueTeacher, for the task of factual consistency detection. They used summarization models to condense CNN/DM articles, and labelled the document-summary pair with FLAN-PaLM 540B Chowdhery etal. (2023) based NLI model. Models trained on the TrueTeacher outperformed ANLI ones. We trained a T5 XXL on the TrueTeacher data and observed better performance compared to GNLI as well ($88.06\%$ vs $85.75\%$ on average). This is expected since TrueTeacher is collected directly for the task of factual consistency detection and makes only the binary distinction.

Train / Eval	MNLI	ANLI	WANLI	GNLI Human
T5 small
MNLI	83.37	31.34	56.52	75.31
ANLI	70.35	48.31	52.70	67.14
WANLI	60.40	36.41	72.60	57.76
GNLI	82.18	33.00	56.56	77.14
MNLI + GNLI	82.66	30.94	55.82	77.76
ANLI + GNLI	72.89	37.94	48.65	69.80
WANLI + GNLI	78.02	34.87	86.69	76.53
T5 large
MNLI	90.83	40.22	63.58	82.24
ANLI	86.87	63.62	63.70	81.22
WANLI	81.10	48.03	92.08	77.14
GNLI	90.61	45.72	65.10	83.67
MNLI + GNLI	91.04	44.94	65.54	82.45
ANLI + GNLI	90.61	63.69	65.03	83.67
WANLI + GNLI	87.80	48.34	96.84	83.88
T5 xxl
MNLI	92.11	55.44	65.92	83.27
ANLI	90.01	73.37	67.04	82.86
WANLI	84.61	60.00	86.46	81.84
GNLI	91.77	57.87	65.43	82.65
MNLI + GNLI	92.13	55.94	67.06	84.08
ANLI + GNLI	91.96	72.94	66.26	84.08
WANLI + GNLI	90.34	60.19	87.90	84.69

5 Conclusion

A decade of increasingly useful NLI benchmarks and datasets have been instrumental in improving LLMs for various tasks. We have presented a new exploration of how the data distribution of each data source still impacts downstream performance on new examples. We proposed a synthetic data approach to mitigate these effects with examples balanced for domain, length and labels. We show that, by drawing on an LLM’s parametric knowledge of a broad range of domains, such synthetic data enables us to both train significantly more domain-general NLI models, and to improve intrinsic NLI model performance on in-domain test data by augmenting in-domain training data.

6 Limitations

We do not release the synthetic general NLI data with our paper. However, our method for generating them can be replicated with access to an LLM, either to directly reproduce our results or to apply our approach to other domains, text lengths, and/or training set sizes of interest. Our process for generation requires multiple LLM tasks which uses more compute than a single stage one. But we note that the data is of high linguistic quality with this method. At the same time, the generated premises could potentially contain fictional information, and should not be used for training models that learn facts from data. We have applied our approach to generalize only one dataset (MNLI), which has examples in English. While we obtain positive results, the results on other datasets, and for other languages remain an empirical question. In addition, we performed experiments with FLAN-PaLM 540B (for prompt-tuning) and FLAN-PaLM2 L (for prompting). However, our method is straightforward and can be easily replicated by other LLMs. We expect comparable LLMs should lead to similar results.

References

• Adila and Kang (2022)Dyah Adila and Dongyeop Kang. 2022.Understanding out-of-distribution: A perspective of data dynamics.In I (Still) Can’t Believe It’s Not Better! Workshop at NeurIPS2021, pages 1–8. PMLR.
• Agrawal etal. (2023)Priyanka Agrawal, Chris Alberti, Fantine Huot, Joshua Maynez, JiMa, SebastianRuder, Kuzman Ganchev, Dipanjan Das, and Mirella Lapata. 2023.Qameleon: Multilingual qa with only 5 examples.Transactions of the Association for Computational Linguistics,11:1754.
• Belinkov etal. (2019)Yonatan Belinkov, Adam Poliak, StuartM. Shieber, BenjaminVan Durme, andAlexanderM. Rush. 2019.Don’t take the premise for granted: Mitigating artifacts in naturallanguage inference.In Proceedings of the 57th Conference of the Association forComputational Linguistics, ACL 2019, Florence, Italy, July 28- August 2,2019, Volume 1: Long Papers, pages 877–891.
• Bowman etal. (2015a)SamuelR. Bowman, Gabor Angeli, Christopher Potts, and ChristopherD. Manning.2015a.A large annotatedcorpus for learning natural language inference.In Proceedings of the 2015 Conference on Empirical Methods inNatural Language Processing, pages 632–642, Lisbon, Portugal. Associationfor Computational Linguistics.
• Bowman etal. (2015b)SamuelR Bowman, Gabor Angeli, Christopher Potts, and ChristopherD Manning.2015b.A large annotated corpus for learning natural language inference.In Conference on Empirical Methods in Natural LanguageProcessing, EMNLP 2015, pages 632–642. Association for ComputationalLinguistics (ACL).
• Chowdhery etal. (2023)Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra,Adam Roberts, Paul Barham, HyungWon Chung, Charles Sutton, SebastianGehrmann, etal. 2023.Palm: Scaling language modeling with pathways.Journal of Machine Learning Research, 24(240):1–113.
• Chung etal. (2022)HyungWon Chung, LeHou, Shayne Longpre, Barret Zoph, YiTay, William Fedus,Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, etal. 2022.Scaling instruction-finetuned language models.arXiv e-prints, pages arXiv–2210.
• Dziri etal. (2022)Nouha Dziri, Hannah Rashkin, Tal Linzen, and David Reitter. 2022.Evaluating attribution in dialogue systems: The begin benchmark.Transactions of the Association for Computational Linguistics,10:1066–1083.
• Fabbri etal. (2021)AlexanderR Fabbri, Wojciech Kryściński, Bryan McCann, Caiming Xiong,Richard Socher, and Dragomir Radev. 2021.Summeval: Re-evaluating summarization evaluation.Transactions of the Association for Computational Linguistics,9:391–409.
• Gekhman etal. (2023)Zorik Gekhman, Jonathan Herzig, Roee Aharoni, Chen Elkind, and Idan Szpektor.2023.Trueteacher: Learning factual consistency evaluation with largelanguage models.arXiv preprint arXiv:2305.11171.
• Googleand etal. (2023)RohanAnil Googleand, AndrewM. Dai, Orhan Firat, Melvin Johnson, DmitryLepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey,Zhifeng Chen, Eric Chu, JonathanH. Clark, LaurentEl Shafey, Yanping Huang,Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, KevinRobinson, Sebastian Ruder, YiTay, Kefan Xiao, Yuanzhong Xu, Yujing Zhang,GustavoHernandez Abrego, Junwhan Ahn, Jacob Austin, Paul Barham, Jan Botha,James Bradbury, Siddhartha Brahma, Kevin Brooks, Michele Catasta, Yong Cheng,Colin Cherry, ChristopherA. Choquette-Choo, Aakanksha Chowdhery, ClémentCrepy, Shachi Dave, Mostafa Dehghani, Sunipa Dev, Jacob Devlin, Mark Díaz,Nan Du, Ethan Dyer, Vlad Feinberg, Fangxiaoyu Feng, Vlad Fienber, MarkusFreitag, Xavier Garcia, Sebastian Gehrmann, Lucas Gonzalez, Guy Gur-Ari,Steven Hand, Hadi Hashemi, LeHou, Joshua Howland, Andrea Hu, Jeffrey Hui,Jeremy Hurwitz, Michael Isard, Abe Ittycheriah, Matthew Jagielski, WenhaoJia, Kathleen Kenealy, Maxim Krikun, Sneha Kudugunta, Chang Lan, KatherineLee, Benjamin Lee, Eric Li, Music Li, Wei Li, YaGuang Li, Jian Li, HyeontaekLim, Hanzhao Lin, Zhongtao Liu, Frederick Liu, Marcello Maggioni, AromaMahendru, Joshua Maynez, Vedant Misra, Maysam Moussalem, Zachary Nado, JohnNham, Eric Ni, Andrew Nystrom, Alicia Parrish, Marie Pellat, Martin Polacek,Alex Polozov, Reiner Pope, Siyuan Qiao, Emily Reif, Bryan Richter, ParkerRiley, AlexCastro Ros, Aurko Roy, Brennan Saeta, Rajkumar Samuel, ReneeShelby, Ambrose Slone, Daniel Smilkov, DavidR. So, Daniel Sohn, SimonTokumine, Dasha Valter, Vijay Vasudevan, Kiran Vodrahalli, Xuezhi Wang,Pidong Wang, Zirui Wang, Tao Wang, John Wieting, Yuhuai Wu, Kelvin Xu, YunhanXu, Linting Xue, Pengcheng Yin, Jiahui Yu, Qiao Zhang, Steven Zheng,CeZheng, Weikang Zhou, Denny Zhou, Slav Petrov, and Yonghui Wu. 2023.Palm 2 technical report.
• Gupta etal. (2022)Prakhar Gupta, Chien-Sheng Wu, Wenhao Liu, and Caiming Xiong. 2022.Dialfact: A benchmark for fact-checking in dialogue.In Proceedings of the 60th Annual Meeting of the Associationfor Computational Linguistics (Volume 1: Long Papers), pages 3785–3801.
• Gururangan etal. (2018)Suchin Gururangan, Swabha Swayamdipta, Omer Levy, Roy Schwartz, Samuel Bowman,and NoahA. Smith. 2018.Annotation artifacts innatural language inference data.In Proceedings of the 2018 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, Volume 2 (Short Papers), pages 107–112, New Orleans,Louisiana. Association for Computational Linguistics.
• He etal. (2022)Xuanli He, Islam Nassar, Jamie Kiros, Gholamreza Haffari, and Mohammad Norouzi.2022.Generate, Annotate, and Learn: NLP with Synthetic Text.Transactions of the Association for Computational Linguistics,10:826–842.
• Honovich etal. (2022)OrHonovich, Roee Aharoni, Jonathan Herzig, Hagai Taitelbaum, Doron Kukliansy,Vered Cohen, Thomas Scialom, Idan Szpektor, Avinatan Hassidim, and YossiMatias. 2022.TRUE:Re-evaluating factual consistency evaluation.In Proceedings of the 2022 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, pages 3905–3920, Seattle, United States. Association forComputational Linguistics.
• Honovich etal. (2021)OrHonovich, Leshem Choshen, Roee Aharoni, Ella Neeman, Idan Szpektor, and OmriAbend. 2021.Q2:: Evaluating factual consistency in knowledge-grounded dialoguesvia question generation and question answering.In Proceedings of the 2021 Conference on Empirical Methods inNatural Language Processing, pages 7856–7870.
• Lester etal. (2021)Brian Lester, Rami Al-Rfou, and Noah Constant. 2021.The power ofscale for parameter-efficient prompt tuning.In Proceedings of the 2021 Conference on Empirical Methods inNatural Language Processing, pages 3045–3059, Online and Punta Cana,Dominican Republic. Association for Computational Linguistics.
• Li etal. (2023)Zhuoyan Li, Hangxiao Zhu, Zhuoran Lu, and Ming Yin. 2023.Synthetic data generation with large language models for textclassification: Potential and limitations.In Proceedings of the 2023 Conference on Empirical Methods inNatural Language Processing, pages 10443–10461.
• Liu etal. (2022)Alisa Liu, Swabha Swayamdipta, NoahA Smith, and Yejin Choi. 2022.Wanli: Worker and ai collaboration for natural language inferencedataset creation.In Findings of the Association for Computational Linguistics:EMNLP 2022, pages 6826–6847.
• Liu etal. (2020)Tianyu Liu, Zheng Xin, Xiaoan Ding, Baobao Chang, and Zhifang Sui. 2020.An empirical study on model-agnostic debiasing strategies for robustnatural language inference.In Proceedings of the 24th Conference on Computational NaturalLanguage Learning, pages 596–608.
• Maynez etal. (2020)Joshua Maynez, Shashi Narayan, Bernd Bohnet, and Ryan McDonald. 2020.On faithfulness and factuality in abstractive summarization.In Proceedings of the 58th Annual Meeting of the Associationfor Computational Linguistics, pages 1906–1919.
• Muandet etal. (2013)Krikamol Muandet, David Balduzzi, and Bernhard Schölkopf. 2013.Domain generalization via invariant feature representation.In Proceedings of the 30th International Conference on MachineLearning, volume28 of Proceedings of Machine Learning Research,pages 10–18. PMLR.
• Nangia and Bowman (2019)Nikita Nangia and Samuel Bowman. 2019.Human vs. muppet: A conservative estimate of human performance on theglue benchmark.In Proceedings of the 57th Annual Meeting of the Associationfor Computational Linguistics, pages 4566–4575.
• Nie etal. (2019)Yixin Nie, Adina Williams, Emily Dinan, Mohit Bansal, Jason Weston, and DouweKiela. 2019.Adversarial nli: A new benchmark for natural language understanding.arXiv preprint arXiv:1910.14599.
• Pagnoni etal. (2021)Artidoro Pagnoni, Vidhisha Balachandran, and Yulia Tsvetkov. 2021.Understanding factuality in abstractive summarization with frank: Abenchmark for factuality metrics.In Proceedings of the 2021 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, pages 4812–4829.
• Puri etal. (2020)Raul Puri, Ryan Spring, Mohammad Shoeybi, Mostofa Patwary, and Bryan Catanzaro.2020.Trainingquestion answering models from synthetic data.In Proceedings of the 2020 Conference on Empirical Methods inNatural Language Processing (EMNLP), pages 5811–5826, Online. Associationfor Computational Linguistics.
• Raffel etal. (2020)Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, MichaelMatena, Yanqi Zhou, Wei Li, and PeterJ Liu. 2020.Exploring the limits of transfer learning with a unified text-to-texttransformer.The Journal of Machine Learning Research, 21(1):5485–5551.
• Rahman etal. (2019)MohammadMahfujur Rahman, Clinton f*ckes, Mahsa Baktashmotlagh, and SridhaSridharan. 2019.Multi-component image translation for deep domain generalization.In 2019 IEEE Winter Conference on Applications of ComputerVision (WACV), pages 579–588.
• Rashkin etal. (2023)Hannah Rashkin, Vitaly Nikolaev, Matthew Lamm, Lora Aroyo, Michael Collins,Dipanjan Das, Slav Petrov, GauravSingh Tomar, Iulia Turc, and David Reitter.2023.Measuring Attribution in Natural Language Generation Models.Computational Linguistics, 49(4):777–840.
• Schuster etal. (2021)Tal Schuster, Adam Fisch, and Regina Barzilay. 2021.Get yourvitamin C! robust fact verification with contrastive evidence.In Proceedings of the 2021 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, pages 624–643, Online. Association for ComputationalLinguistics.
• Thorne etal. (2018)James Thorne, Andreas Vlachos, Oana Cocarascu, Christos Christodoulopoulos, andArpit Mittal. 2018.The fact extraction andVERification (FEVER) shared task.In Proceedings of the First Workshop on Fact Extraction andVERification (FEVER), pages 1–9, Brussels, Belgium. Association forComputational Linguistics.
• Tobin etal. (2017)Josh Tobin, Rachel Fong, Alex Ray, Jonas Schneider, Wojciech Zaremba, andPieter Abbeel. 2017.Domain randomization for transferring deep neural networks fromsimulation to the real world.In 2017 IEEE/RSJ International Conference on Intelligent Robotsand Systems (IROS), pages 23–30.
• Vu etal. (2021)TuVu, Minh-Thang Luong, Quoc Le, Grady Simon, and Mohit Iyyer. 2021.STraTA:Self-training with task augmentation for better few-shot learning.In Proceedings of the 2021 Conference on Empirical Methods inNatural Language Processing, pages 5715–5731, Online and Punta Cana,Dominican Republic. Association for Computational Linguistics.
• Wang etal. (2020)Alex Wang, Kyunghyun Cho, and Mike Lewis. 2020.Asking andanswering questions to evaluate the factual consistency of summaries.In Proceedings of the 58th Annual Meeting of the Associationfor Computational Linguistics, pages 5008–5020, Online. Association forComputational Linguistics.
• Wang etal. (2021)Jindong Wang, Cuiling Lan, Chang Liu, Yidong Ouyang, and Tao Qin. 2021.Generalizing to unseen domains: A survey on domain generalization.In Proceedings of the Thirtieth International Joint Conferenceon Artificial Intelligence, IJCAI-21, pages 4627–4635. InternationalJoint Conferences on Artificial Intelligence Organization.Survey Track.
• Williams etal. (2018)Adina Williams, Nikita Nangia, and Samuel Bowman. 2018.A broad-coverage challenge corpus for sentence understanding throughinference.In Proceedings of the 2018 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, Volume 1 (Long Papers), pages 1112–1122.
• Zhang etal. (2019)Yuan Zhang, Jason Baldridge, and Luheng He. 2019.PAWS: Paraphraseadversaries from word scrambling.In Proceedings of the 2019 Conference of the North AmericanChapter of the Association for Computational Linguistics: Human LanguageTechnologies, Volume 1 (Long and Short Papers), pages 1298–1308,Minneapolis, Minnesota. Association for Computational Linguistics.
• Zhou etal. (2020)Kaiyang Zhou, Yongxin Yang, Timothy Hospedales, and Tao Xiang. 2020.Learning to generate novel domains for domain generalization.In Computer Vision – ECCV 2020, pages 561–578, Cham.Springer International Publishing.

Appendix A Seed Examples for Prompting

We provide the seed examples for prompting FLAN-PaLM2 L (Unicorn) to generate premises in Table 7.

Appendix B Running time and Hyper-parameter Details of Prompt Tuning Experiments

For prompt-tuning of FLAN-PaLM 540B, we used an input length and output length of $512$. We tuned $100$ prompt embeddings and used them during inference. We used a learning rate of $0.3$ and did not use dropout. We trained with a batch size of $16$ for $24,544$ steps that is equivalent to one epoch on the MNLI dataset with $392,702$ training examples. The prompt-tuning took around 110 hours to complete with 256 Cloud TPU v4 chips.

Appendix C Hyper-parameter Details of T5 Fine-tuning Experiments

For training T5 models (both binary and 3-way), we tuned learning rates $\in\{5e-4,1e-4,5e-5\}$ and fine-tuned with batch size of $32$ for $50$K steps. We checkpoint every $1K$ steps for early stopping. We use a dropout rate of $0.1$. We trained with an input length of $512$. During inference for factual consistency evaluation (Section 4.1) and NLI benchmarks (Section 4.2), we used an input length of $1024$ and $512$ respectively.

We report the best selected hyper-parameters for T5 binary and 3-way models in Table 8 and Table 9.