PaperlayerView exact revision on arXiv

Source-linked AI summary

Combat COVID-19 Infodemic Using Explainable Natural Language Processing Models

Jackie Ayoub, X. Jessie Yang, Feng Zhou

arXiv:2103.00747v1cs.CL

TL;DR

COVID-19 misinformation created a need for models that detect false claims while supporting public trust. The paper combines DistilBERT with SHAP explanations and evaluates whether explanations affect trust and sharing, finding greater trust and willingness to share with SHAP-based conditions than with text alone.

  • Problem

    COVID-19 misinformation made trustworthy information difficult to find, while limited labeled data and black-box NLP models constrained reliable detection and public trust.

  • Method

    The paper uses DistilBERT to detect COVID-19 misinformation, SHAP to explain predictions, and a three-condition human-subject experiment to evaluate trust and sharing.

  • Results

    Participants were significantly more likely to trust and share COVID-19 information with SHAP explanations, with or without source and evidence, than with text alone.

  • Takeaways & Limitations

    The results support using DistilBERT and SHAP to detect COVID-19 misinformation while improving public trust in model predictions.

  • Takeaways & Limitations

    Limited labeled data may restrict detection of new misinformation related to COVID-19.

Abstract

from arXiv · show

Misinformation of COVID-19 is prevalent on social media as the pandemic unfolds, and the associated risks are extremely high. Thus, it is critical to detect and combat such misinformation. Recently, deep learning models using natural language processing techniques, such as BERT (Bidirectional Encoder Representations from Transformers), have achieved great successes in detecting misinformation. In this paper, we proposed an explainable natural language processing model based on DistilBERT and SHAP (Shapley Additive exPlanations) to combat misinformation about COVID-19 due to their efficiency and effectiveness. First, we collected a dataset of 984 claims about COVID-19 with fact checking. By augmenting the data using back-translation, we doubled the sample size of the dataset and the DistilBERT model was able to obtain good performance (accuracy: 0.972; areas under the curve: 0.993) in detecting misinformation about COVID-19. Our model was also tested on a larger dataset for AAAI2021 - COVID-19 Fake News Detection Shared Task and obtained good performance (accuracy: 0.938; areas under the curve: 0.985). The performance on both datasets was better than traditional machine learning models. Second, in order to boost public trust in model prediction, we employed SHAP to improve model explainability, which was further evaluated using a between-subjects experiment with three conditions, i.e., text (T), text+SHAP explanation (TSE), and text+SHAP explanation+source and evidence (TSESE). The participants were significantly more likely to trust and share information related to COVID-19 in the TSE and TSESE conditions than in the T condition. Our results provided good implications in detecting misinformation about COVID-19 and improving public trust.

1. Introduction

COVID-19 misinformation created an urgent need for timely detection, trustworthy information, and explanations that could improve public trust. The paper proposed a DistilBERT-based prediction model with SHAP explanations and a human-subject evaluation of trust and sharing.

  • The COVID-19 infodemic made trustworthy sources and reliable advice difficult to find, motivating automated detection and debunking of false claims.
  • DistilBERT was adopted to detect COVID-19 misinformation because BERT is computationally intensive, while DistilBERT maintained almost similar performance with fewer resources.
  • Explainability was motivated by evidence that revealing model insights can improve acceptance and trust, especially in high-risk decision contexts such as COVID-19.
  • SHAP was used to explain DistilBERT predictions by calculating individual feature contributions, addressing concerns that machine-learning NLP models are black boxes.
  • The study evaluated text, text+SHAP explanation, and text+SHAP explanation+source and evidence conditions for trust and willingness to share information.

2. Related work

Prior misinformation-detection work spans traditional supervised models, deep and semi-supervised learning, and explanation methods. This paper addresses computational cost, limited domain labels, and the need for trustworthy predictions with DistilBERT, back-translation, and SHAP.

  • Supervised methods: Traditional misinformation detection has used supervised models, including decision trees, RNNs, CNNs, LSTMs, and generative adversarial learning.
  • Semi-supervised methods: Semi-supervised approaches reduce reliance on labeled data by combining small labeled sets with pre-trained or unlabeled information.
  • Deep learning methods: BERT-based approaches outperformed LSTM and gradient-boosted tree models, while additional news pretraining further improved results over BERT.
  • Research gaps: BERT is computationally expensive, with 110 million parameters in BERT base and 340 million in BERT large, limiting real-time use without accelerated hardware.
  • Research gaps: The paper proposes DistilBERT to address BERT’s computational burden, back-translation to double training data, and SHAP to explain predictions locally.
  • Explainability: Feature-based explanations provide global and local interpretability, with local explanations described as more helpful for user understanding and trust.

3. Methods

The study builds an explainable COVID-19 misinformation detector by collecting and augmenting fact-checked claims, fine-tuning BERT-derived models, and applying SHAP explanations. It evaluates whether these explanations improve trust in model predictions.

  • 3.2. Model Building: BERT was fine-tuned on labeled COVID-19 data, then DistilBERT learned to mimic BERT through knowledge distillation.The distillation used the teacher’s raw predictions and a cross-entropy-based distillation loss.
  • 3.4. Model Evaluation: A three-condition between-subjects experiment compared text-only predictions with predictions supplemented by SHAP explanations, with or without source and evidence.The evaluation was conducted through an AMT survey to assess trust and willingness to share.
  • 3.1. Data Preprocessing: 984 fact-checked COVID-19 claims were collected and doubled using back-translation, alongside a larger 8,560-claim dataset for additional testing.The collected set contained 575 true and 409 fake claims; the larger dataset came from the COVID-19 Fake News Detection Shared Task.
  • 3.3. Model Explanation: SHAP assigned local feature contributions to DistilBERT predictions, transforming log-odds explanations into predicted truth probabilities.The method represents each word’s contribution relative to a base value and aggregates contributions across feature combinations.

4. Results

The augmented DistilBERT approach performed strongly across the evaluated misinformation datasets and was more efficient and explainable than BERT. SHAP explanations were associated with greater trust and willingness to share than text-only predictions.

  • 4.1. Model Performance: BERT performed slightly better than DistilBERT, while DistilBERT was more efficient and more explainable through SHAP.The distilled model used logistic regression and lost little performance relative to BERT.
  • 4.1. Model Performance: The augmented DistilBERT method performed best among the selected traditional machine learning models across the evaluated performance measures.Performance was assessed using precision, recall, F1 score, accuracy, and AUC with 10-fold cross-validation.
  • 4.2. User Evaluation: Trust in model decisions was significantly higher with SHAP explanations than in the control condition for both TSE (p = .031) and TSESE (p = .005).The experiment found significant main effects of condition on trust, F(2, 241) = 5.628, p = .004.
  • 4.2. User Evaluation: Willingness to share information was significantly higher in both explanation conditions than in the control condition, with no significant difference between TSE and TSESE.The main effect of condition on willingness to share was F(2, 241) = 10.730, p = .000.

5. Discussions

The discussion finds that DistilBERT offers efficient misinformation detection, while SHAP explanations increase trust and willingness to share compared with text alone. Adding sources and evidence produced no significant further improvement over SHAP explanations, although qualitative responses suggest possible value.

  • 5.2. Trust and Willingness to Share: Participants trusted and were more willing to share information in the TSE and TSESE conditions than in the text-only control condition.The study evaluated text, text+SHAP explanation, and text+SHAP explanation+source and evidence conditions.
  • 5.2. Trust and Willingness to Share: There was no significant difference between TSE and TSESE in trust or willingness to share.The authors caution that only 10 randomly selected claims were tested and that the added source-and-evidence effect may require further study.
  • 5.2. Trust and Willingness to Share: For false claims, the three conditions significantly affected trust, F(2, 241) = 7.984, p = .000, and willingness to share, F(2, 241) = 11.918, p = .000.For true claims, the condition effect was significant for willingness to share but not trust.

6. Conclusion and Future Work

The conclusion presents the DistilBERT-and-SHAP model as useful for detecting COVID-19 misinformation and improving public trust. Future work should address limited and outdated data, domain awareness, participant representativeness, measurement of biases, and the difficulty of interpreting SHAP explanations.

  • 6. Conclusion and Future Work: The proposed model used DistilBERT and SHAP to support COVID-19 misinformation detection and public trust.The authors describe the method as having implications for both detecting misinformation and improving trust.
  • 6. Conclusion and Future Work: The model was built on a small dataset collected by April 2020, limiting its ability to detect new COVID-19 misinformation.The authors recommend adding data as the pandemic unfolds.
  • 6. Conclusion and Future Work: The AMT sample may not represent the broader population, and unmeasured political, ideological, and demographic factors could affect responses.The authors recommend additional survey questions and demographic measures.
  • 6. Conclusion and Future Work: Some participants found SHAP explanations confusing, motivating future exploration of more intuitive explanation formats.The difficulty involved making predictions from individual words despite an introductory training section.
Loading 2103.00747v1…