Source-linked AI summary
A survey on improving NLP models with human explanations
Mareike Hartmann, Daniel Sonntag
TL;DR
NLP research lacks sufficient comparisons of explanation types and integration methods, making method selection for specific use cases difficult. This paper surveys these approaches and the factors relevant to choosing them, highlighting explanation information content and human annotation effort as important considerations.
Problem
Methods for learning from human explanations are rarely compared across explanation types and integration mechanisms, making it difficult to choose an approach for a specific use case.
Method
The paper reviews highlight- and free-text-based methods for integrating human explanations into NLP model learning and discusses factors informing method selection.
Results
The survey identifies explanation information content as central to model benefit and describes evidence that different integration methods and explanation types have rarely been compared directly.
Takeaways & Limitations
Model-independent measures of explanation information content could guide annotation instructions or filter explanations that might harm model performance.
Takeaways & Limitations
Providing explanations adds annotation overhead; one study found highlight explanations took twice as long as labels for document classification.
Abstract
from arXiv · showhide
Training a model with access to human explanations can improve data efficiency and model performance on in- and out-of-domain data. Adding to these empirical findings, similarity with the process of human learning makes learning from explanations a promising way to establish a fruitful human-machine interaction. Several methods have been proposed for improving natural language processing (NLP) models with human explanations, that rely on different explanation types and mechanism for integrating these explanations into the learning process. These methods are rarely compared with each other, making it hard for practitioners to choose the best combination of explanation type and integration mechanism for a specific use-case. In this paper, we give an overview of different methods for learning from human explanations, and discuss different factors that can inform the decision of which method to choose for a specific use-case.
1 Introduction
Learning from human explanations is presented as a way to improve NLP models while supporting richer human–machine interaction. The paper surveys explanation types and integration methods to help choose approaches for specific use cases.
- Human explanations can help models learn more efficiently, focus on relevant features, and reduce reliance on spurious correlations.They may reduce annotated-data needs and support adaptation to new domains or personalization.
- Human explanations align with users’ preference to provide richer feedback than correct labels alone.
- The survey reviews NLP methods using different explanation types and mechanisms for integrating explanations into learning.It focuses on highlight and free-text explanations and compares how explanation information can be incorporated.
2 Learning from Explanations
The survey organizes methods for learning from human highlight and free-text explanations by how explanation information is integrated into NLP models, and identifies factors for choosing among them.
- Explanation types: Highlight explanations assign attribution scores to input elements, whereas free-text explanations are variable-length natural-language sequences that explain why an instance receives a label.Highlight attributions are often binary in practice, while free-text explanations are not constrained to input elements.
- Regularizing feature importance: Regularizing feature importance trains with classification and explanation losses, encouraging model attributions to match human highlight attributions.The explanation loss measures the distance between gold and model attribution scores.
- Semantic parsing to obtain noisy labels: Semantic parsing converts free-text explanations into executable labeling functions that assign noisy labels to unlabeled inputs for training-data augmentation.The approach depends on a semantic parser, although rule-based parsers can provide decent results according to cited work.
- Multi-task learning and explain-and-predict: Multi-task learning jointly trains target-task and explanation-producing models with shared parameters, while explain-and-predict pipelines generate explanations before predicting from them.The latter was introduced to improve interpretability rather than primarily to improve the target task.
- Comparative studies: Comparative evidence remains scarce: one study found multi-task learning needed more examples and slightly underperformed regularization, while another reported similar results for highlight and free-text explanations on E-SNLI.The similar E-SNLI results may reflect annotators being encouraged to include highlighted words in their free-text explanations.
- Information content and human factors: Useful explanations may be those that reveal relevant latent information beyond the input alone, but human explanations can be costly, incomplete, subjective, or low quality.Zaidan et al. (2007) found that adding highlight explanations doubled annotation time, while the survey notes no direct timing comparison with free-text explanations.
3 Take-Aways
The paper highlights a lack of comparative studies across explanation types and integration methods, while identifying explanation information content as a promising direction for improving learning approaches.
- Comparative studies are lacking across explanation types and integration methods.
- Initial studies suggest that an explanation’s information content plays a central role in its effect on model quality.
- Model-independent measures could guide annotators toward informative explanations or filter explanations that could harm model performance.