WT5?! Training Text-to-Text Models to Explain their Predictions

By Sharan Narang, Colin Raffel, Katherine Lee, Adam Roberts, Noah Fiedel, Karishma Malkan. 2020

219 citations recorded by the source corpus.

Neural networks have recently achieved human-level performance on various challenging natural language processing (NLP) tasks, but it is notoriously difficult to understand why a neural network produced a particular prediction. In this paper, we leverage the text-to-text framework proposed by Raffel et al.(2019) to train language models to output a natural text explanation alongside their prediction. Crucially, this requires no modifications to the loss function or training and decoding procedures -- we simply train the model to output the explanation after generating the (natural text) prediction. We show that this approach not only obtains state-of-the-art results on explainability benchmarks, but also permits learning from a limited set of labeled explanations and transferring rationalization abilities across datasets. To facilitate reproducibility and future work, we release our code use to train the models.

Subfields: Artificial Intelligence, 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 Natural Language Processing

Citations in the corpus

Chain of Thought Prompting Elicits Reasoning in Large Language Models (2022, neutral): “That line of work typically focuses on natural language inference (Camburu et al., 2018; Yordanov et al., 2021; Bostrom et al., 2021), and produces explanations either simultaneously to or after the ﬁnal prediction (Narang et al., 2020; Majumder et al., 2021; Wiegreffe et al., 2021, 2022).”
mT5: A Massively Multilingual Pre-trained Text-to-Text Transformer (2020, methodology): “…for the “Text-to-Text Transfer Transformer” (T5) model released by Raﬀel et al. (2019) have been used to achieve state-of-the-art results on many benchmarks (Khashabi et al., 2020; Roberts et al., 2020; Kale, 2020; Izacard and Grave, 2020; Nogueira et al., 2020; Narang et al., 2020, etc.).”
Large Language Models Can Self-Improve (2022, neutral): “For example, in the supervised learning setting, a model can be fine-tuned using human-annotated rationales (Zaidan et al., 2007; Ling et al., 2017a; Narang et al., 2020; Camburu et al., 2018; Cobbe et al., 2021; Chung et al., 2022).”
Can language models learn from explanations in context? (2022, methodology): “…(e.g. Wei et al., 2022; Reynolds and McDonell, 2021), as well as a huge variety of broader work beyond prompting on topics like training or tuning with explanations (e.g. Camburu et al., 2018; Ling et al., 2017; Hendricks et al., 2016; Narang et al., 2020; Lertvittayakumjorn and Toni, 2021).”
Reframing Human-AI Collaboration for Generating Free-Text Explanations (2021, methodology): “ECQA was released to address the quality issues of CoS-E (Narang et al., 2020); for completeness, we experiment with both.”
Few-Shot Out-of-Domain Transfer Learning of Natural Language Explanations (2021, methodology): “Meaning CD–ﬁne-tune ﬁne-tune T5 on the child dataset, and then ﬁne-tune on 50 NLEs CD–union ﬁne-tune T5 on the union of the child dataset and 50 NLEs WT5–ﬁne-tune ﬁne-tune T5 on the union of e-SNLI and SNLI, and then ﬁne-tune on the child dataset WT5 ﬁne-tune T5 on the union of e-SNLI, SNLI, and the”
Few-Shot Self-Rationalization with Natural Language Prompts (2021, methodology): “Self-rationalization models (Narang et al., 2020; Wiegreffe et al., 2021) are currently based on T5 for at least two reasons.”
Measuring Association Between Labels and Free-Text Rationales (2020, methodology): “To illustrate this point, we reference Narang et al. (2020): ...”
Rationale-Augmented Ensembles in Language Models (2022, methodology): “An important question for open-domain question answering in the closed-book setting is how to utilize a language model to retrieve relevant knowledge and compose the ﬁnal answer (Roberts et al., 2020).”
Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answering (2022, methodology): “Following [41], we formulate the task to output a natural explanation alongside the predicted answer.”
BERTology Meets Biology: Interpreting Attention in Protein Language Models (2020, methodology): “Other approaches include using language models for generating explanations [53, 59].”
e-ViL: A Dataset and Benchmark for Natural Language Explanations in Vision-Language Tasks (2021, background): “First adoptions of NLEs have been in image classification [23] and were further extended to self-driving cars [26], visual question-answering [37], and natural language processing [11; 39; 9; 6; 27; 35; 28; 12].”
Knowledge-Grounded Self-Rationalization via Extractive and Natural Language Explanations (2021, methodology): “In the human evaluation sample for e-SNLI, we found that 73% of NLEs from RE X C are longer (at least by a token) compared to NLEs from WT5.”
Learning to Rationalize for Nonmonotonic Reasoning with Distant Supervision (2020, methodology): “2020); (ii) joint: generating the rationale and the label jointly (Narang et al. 2020); and (iii) post-hoc: generating a rationale given the input and the gold or predicted label.”
Leakage-Adjusted Simulatability: Can Models Generate Non-Trivial Explanations of Their Behavior in Natural Language? (2020, methodology): “Narang et al. (2020) train a T5 model to generate explanations in a set-up analogous to our MT-R A setting.”