Source-linked AI summary
Towards Interpretable Natural Language Understanding with Explanations as Latent Variables
Wangchunshu Zhou, Jinyi Hu, Hanlin Zhang, Xiaodan Liang, Maosong Sun, Chenyan Xiong, Jian Tang
TL;DR
Existing explanation-based approaches often require many costly human annotations. ELV treats explanations as latent variables and jointly optimizes explanation generation and prediction, extending the framework to semi-supervised self-training. Experiments report effective prediction in both settings and meaningful explanations.
Problem
Existing approaches to generating explanations often require large sets of costly human-annotated explanations for training.
Source: §1 · p. 2Method
ELV treats explanations as latent variables and uses variational EM to jointly train an explanation generator and an explanation-augmented predictor, with ELV-EST extending the framework to semi-supervised learning.
Source: §1 · p. 2, §3.1 · p. 4, §3.4 · p. 5Results
Takeaways & Limitations
The framework supports interpretable natural-language understanding with limited annotated explanations in the evaluated supervised and semi-supervised settings.
Source: §5 · p. 9Generated explanations may be biased or ad hoc and not necessarily informative about the model's predictions, a concern the framework only partially mitigates.
Source: Broader Impact · p. 9
Abstract
from arXiv · showhide
Recently generating natural language explanations has shown very promising results in not only offering interpretable explanations but also providing additional information and supervision for prediction. However, existing approaches usually require a large set of human annotated explanations for training while collecting a large set of explanations is not only time consuming but also expensive. In this paper, we develop a general framework for interpretable natural language understanding that requires only a small set of human annotated explanations for training. Our framework treats natural language explanations as latent variables that model the underlying reasoning process of a neural model. We develop a variational EM framework for optimization where an explanation generation module and an explanation-augmented prediction module are alternatively optimized and mutually enhance each other. Moreover, we further propose an explanation-based self-training method under this framework for semi-supervised learning. It alternates between assigning pseudo-labels to unlabeled data and generating new explanations to iteratively improve each other. Experiments on two natural language understanding tasks demonstrate that our framework can not only make effective predictions in both supervised and semi-supervised settings, but also generate good natural language explanation.
1 Introduction
Natural-language explanations can make predictions more interpretable and provide additional supervision, but existing approaches often require many costly human annotations. ELV jointly trains explanation generation and explanation-augmented prediction with few annotated explanations, and extends this approach to semi-supervised learning.
1 Introduction
- Existing explanation-based approaches require many human-annotated explanations, which are expensive and sometimes impossible to obtain.The paper seeks effective prediction and good explainability with only a limited number of human explanations.Source: §1 · p. 2
- ELV treats explanations as latent variables and jointly trains an explanation generator and an explanation-augmented predictor using variational EM.The E-step trains the generator using annotated explanations or posterior inference; the M-step trains the predictor using generated explanations.Source: §1 · p. 2
- ELV-EST extends the framework to semi-supervised learning by using explanations as implicit logic rules to label unlabeled data.The paper evaluates its approach on relation extraction and sentiment analysis, reporting effectiveness for prediction and explainability in supervised and semi-supervised settings.Source: §1 · p. 2
2 Related Work
Prior work uses natural-language explanations as prediction features or generates them separately before prediction, but often requires abundant annotations or fails to model interactions between generation and prediction. ELV treats explanations as latent variables and jointly trains both modules; its semi-supervised approach uses distributed representations rather than parsed logical forms.
2 Related Work
- Some explanation-feature approaches require explanations at both training and test time, making annotation costly and prediction easier when test explanations are supplied.Earlier work also converted explanations into classifier features or incorporated them with BERT through ExpBERT [10].Source: §2 · p. 2
- Separate explanation-generation and prediction stages can improve interpretability and prediction [3] [4], but require many annotated explanations and do not model module interactions.The paper notes there is no guarantee generated explanations reflect or benefit the predictor; Camburu et al. [3] reported interpretability at a performance cost.Source: §2 · pp. 2–3
- ELV instead treats explanations as latent variables and jointly trains the generation and explanation-augmented prediction modules.Figure 1 describes an E-step for explanation generation and an M-step for predictor training.Source: §2 · p. 3, Figure 1 · p. 3
- Prior semi-supervised methods parse explanations into logical labeling functions [13] [7], whereas ELV's approach uses distributed representations and requires no semantic parsing.The paper notes that parser-compatible explanations may not be possible in many NLP applications.Source: §2 · p. 3
3 Methodology
The methodology treats explanations as latent variables and jointly optimizes an explanation generator and an explanation-augmented predictor. It extends this framework to semi-supervised learning by iteratively adding pseudo-labeled examples and generated explanations.
3.1 Natural Language Explanation as Latent Variables
- ELV treats explanations as latent variables and alternates variational E- and M-steps to jointly optimize the explanation generator and explanation-augmented predictor.The framework is initialized on the small set of examples with human explanations, then optimized using labeled data and those examples.Source: §3.1 · p. 4
3.2 E-step: Explanation Generation model
- The E-step uses UniLM to generate explanations from the input and label, with the predictor's feedback guiding explanations for labeled examples without annotations.For examples with annotated explanations, the E-step instead maximizes the likelihood of the ground-truth explanations.Source: §3.2 · p. 4
3.3 M-step: Explanation-Augmented Prediction model
- At test time, the predictor retrieves explanations from similar sentences because labels are unavailable and generating explanations without labels can yield irrelevant or misleading results.Similarity is measured using SentenceBERT embeddings; the predictor combines explanations with the input to make a label prediction.Source: §3.3 · pp. 4–5
3.4 Explanation-based Self-Training
- ELV-EST initializes on limited labeled data, then iteratively pseudo-labels unlabeled examples, generates explanations for them, and retrains both models.The growing explanation set supports retrieval during prediction.Source: §3.4 · pp. 5–6, §3.3 · p. 5
- New explanations can improve the generator and serve as implicit rules that help the predictor assign more accurate pseudo-labels in later iterations.Unlike conventional self-training, ELV-EST also produces explainable predictions and does not require task-specific semantic parsers or matching models.Source: §3.4 · p. 6
4 Experiments
Across four datasets, ELV and ELV-EST outperform their respective supervised and semi-supervised comparison methods, while human evaluation and analysis examine explanation quality and the effects of explanation quantity and quality. Experiments cover relation extraction and aspect-based sentiment classification with limited annotated explanations.
4.1 Datasets
- Experiments cover TACRED and SemEval relation extraction plus Restaurant and Laptop aspect-based sentiment classification, using explanations collected in [7].The supervised setting uses all labeled examples; the semi-supervised setting uses a small labeled fraction and treats the remaining original labels as unavailable.Source: §4.1 · p. 6, §4.2 · p. 6
4.2 Experimental Settings
- The prediction and generation backbones are BERT-base and UniLM-base, respectively, with early stopping by validation F1.Each EM iteration trains the prediction model for 3 epochs and the generator for 5; 10 explanations are retrieved for each task.Source: §4.2 · p. 6
4.3 Compared Methods
- Supervised comparisons include BERT-base, an M-step-only ELV variant, and recent task-specific methods; semi-supervised comparisons include Pseudo-Labeling, Self-Training, and Data Programming [13].All compared semi-supervised variants use BERT-base as backbone.Source: §4.3 · pp. 6–7
4.4 Experimental Results
- ELV significantly outperforms the BERT baseline on all four datasets and consistently outperforms the M-step-only variant in supervised settings.It also compares favorably with recent relation-extraction and aspect-based sentiment classification studies.Source: §4.4 · p. 7, Table 2 · p. 7, Table 3 · p. 7
- Around 7 F1 points on relation-extraction datasets and over 15 F1 points on aspect-based sentiment datasets separate ELV-EST from BERT-base plus self-training.ELV-EST also significantly outperforms other semi-supervised methods and data programming on all four datasets.Source: §4.4 · p. 7, Table 4 · p. 7, Table 5 · p. 7
- Human evaluation found ELV explanations substantially better than fine-tuned UniLM explanations and more relevant to model decisions.Five graduate students rated informativeness, correctness, and consistency; inter-rater agreement was 0.51 Kappa.Source: Table 6 · p. 7, §4.4 · p. 8
4.5 Analysis
- Randomly replacing 80% of explanation words significantly decreases performance, while ELV's advantage over M-step-only persists with as few as 15 annotated explanations.ELV continues improving with more explanations, while M-step-only starts to saturate at 45; across training-data proportions, ELV leads BERT most at 10% labeled data, and ELV-EST leads self-training especially with fewer unlabeled examples.Source: Table 7 · p. 7, §4.5 · pp. 8–9, Figure 2 · p. 8
5 Conclusion
ELV trains interpretable natural-language understanding models with limited human-annotated explanations by treating explanations as latent variables linking generation and classification. Results in supervised and semi-supervised settings show effective prediction and meaningful explanations; applying the framework to other tasks and pretrained models remains future work.
5 Conclusion
- ELV treats explanations as latent variables to link explanation generation and explanation-based classification while using limited human annotations.The authors report effective predictions and meaningful explanations in both supervised and semi-supervised settings.Source: §5 · p. 9
- The authors identify application to other natural-language understanding tasks and evaluation with other pretrained language models as future work.They name both stronger models and more computationally efficient models as candidates for future evaluation.Source: §5 · p. 9
Broader Impact
Generated explanations may carry biases and may not reflect the model’s actual decision process. The framework partially mitigates the latter risk by using retrieved explanations to train the explanation-augmented classifier.
Broader Impact
- Generated explanations may be biased, and the authors leave mitigation of that risk to future work.They also note that explanations may be ad hoc and not informative about how the model predicts, a drawback shared by current explanation-generation models.Source: Broader Impact · p. 9
- Retrieved generated explanations are used to train the explanation-augmented classifier, partially mitigating the risk that explanations do not reflect its predictions.Source: Broader Impact · p. 9
A Examples of Explanations
The appendix presents human-annotated and model-generated explanations from different datasets to illustrate the approach.
A Examples of Explanations
- The section presents human-annotated and model-generated explanations from different datasets to help readers understand the approach.Source: A Examples of Explanations · p. 13
A.1 TACRED
The appendix illustrates human-annotated and model-generated explanations across relation-extraction examples, showing how labels are paired with textual cues. Examples include both lexical patterns and subject–object position or distance conditions.
A.1.1 Human Annotated Explanations
- Human-annotated TACRED examples connect relation labels to lexical cues and subject–object position or distance conditions.Examples include “then known as” for alternate names, “’s mother” for parent relations, and “secretly married” for spouse relations.Source: §A.1.1 · p. 13
A.1.2 Machine Generated Explanations
- Model-generated TACRED examples use nearby phrases such as “family,” “senior research associate at,” and “based in” to justify relation labels.Source: §A.1.2 · p. 13
A.2.1 Human Annotated Explanations
- Human-annotated relation-extraction examples explain labels through phrases including “is the most common cause of,” “is a part of,” and “from past.”Source: §A.2.1 · pp. 13–14
A.2.2 Machine Generated Explanations
- Model-generated relation-extraction examples use cues such as “caused,” “was removed into,” and “attaches a” to support their labels.Source: §A.2.2 · p. 14
A.3.1 Human Annotated Explanations
- Human-annotated sentiment examples link positive or negative labels to nearby evaluative words such as “best,” “favorable,” and “annoying.”Source: §A.3.1 · p. 14
A.3.2 Machine Generated Explanations
These appendix examples show model-generated sentiment explanations and human-annotated and model-generated explanations for product-review terms. The explanations use nearby sentiment-bearing words or phrases as cues.
A.3.2 Machine Generated Explanations
- Model-generated sentiment examples use nearby words such as “wonderful,” “fine,” and “unbearable” to explain their labels.Source: §A.3.2 · p. 14
A.4.1 Human Annotated Explanations
- Human-annotated product-review examples associate labels with cues such as “annoying,” “very nicely,” and “awesome” near the relevant term.Source: §A.4.1 · p. 15
A.4.2 Machine Generated Explanations
- Model-generated product-review examples cite nearby phrases such as “disappointed,” “exceptionally easy,” and “nearly useless” to support their labels.Source: §A.4.2 · p. 15
B Human Evaluation Details
The human evaluation uses 100 Restaurant test examples predicted by ELV and has five graduate students score explanations generated by compared models. They assess informativeness, correctness, and consistency with the input sentence.
B Human Evaluation Details
- The evaluation samples 100 Restaurant test examples and uses ELV to predict their labels.Compared models generate explanations using the input sentences and predicted labels.Source: B Human Evaluation Details · p. 15
- Five graduate students with sufficient English proficiency score the generated explanations.Source: B Human Evaluation Details · p. 15
- The rubric rates informativeness, factual correctness, and consistency with the input sentence.Informativeness concerns usefulness for understanding predictions; correctness concerns factual accuracy; consistency concerns whether the explanation is true of the input.Source: B Human Evaluation Details · p. 15
C Addition Analysis Experiments
The additional analysis compares ELV trained with 60% of labeled data against BERT-base trained with varying labeled-data fractions. The authors report that ELV can match or exceed BERT-base trained with more labeled data.
C Addition Analysis Experiments
- ELV with 60% of labeled data can achieve or exceed BERT-base trained with 70%, 80%, 90%, or 100% of labeled data.The comparison also includes BERT-base trained with 60% of labeled data; the results are reported in Table 8.Source: C Addition Analysis Experiments · p. 16
- The authors conclude that ELV effectively leverages human-annotated explanations as additional information.Source: C Addition Analysis Experiments · p. 16