[Paper Review] WT5?! Training Text-to-Text Models to Explain their Predictions

Sharan Narang, Colin Raffel+4 more|arXiv (Cornell University)|Apr 30, 2020

Topic Modeling54 references103 citations

TL;DR

This paper trains text-to-text models to generate natural language explanations alongside predictions, achieving state-of-the-art explainability and enabling semi-supervised and cross-domain transfer without changing training loss.

ABSTRACT

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.

Motivation & Objective

• Motivate making neural predictions interpretable by enabling natural-language explanations.
• Leverage the text-to-text framework to train models to output explanations without altering training or decoding procedures.
• Demonstrate state-of-the-art performance on explainability benchmarks across multiple NLP tasks.
• Explore semi-supervised, cross-domain, and cross-task transfer of explanation abilities.

Proposed method

• Use the Text-to-Text Transfer Transformer (T5) as the base model and fine-tune it for classification with natural-language explanations.
• Prepends the input with the token explain to generate an accompanying explanation in the output when desired.
• Trains with a mixed objective: examples with explanations and examples without explanations to enable semi-supervised learning.
• Optionally supports extractive explanations by training the model to output explanation spans that can be aligned to input spans.
• Evaluates explanations using BLEU for abstractive ones and F1 on tokenized input spans for extractive explanations, complemented by human judgements via Mechanical Turk.

Experimental results

Research questions

• RQ1 Can a single text-to-text model generate coherent natural-language explanations for its predictions without modifying the loss objective?
• RQ2 How does including explanations affect predictive accuracy and explanation quality across datasets?
• RQ3 What is the effectiveness of semi-supervised learning where explanations are provided only on a subset of data?
• RQ4 To what extent can explanation abilities transfer across domains and tasks without explicit explanation annotations?

Key findings

• WT5-11B achieves high accuracy and competitive explanation quality across datasets, often matching or surpassing prior state-of-the-art on both predictions and explanations.
• The model attains human- or super-human levels for explanation quality on several datasets as measured by human judgments.
• Explanation generation remains effective under semi-supervised settings with limited annotated explanations, especially for abstractive explanations.
• Cross-domain transfer (e.g., MNLI, IMDb) yields plausible explanations with high human-rated quality and strong classification performance.
• Cross-task transfer (e-SNLI to CoS-E) is feasible with appropriate input formatting and beam-search decoding, producing reasonable explanations.