Deriving Machine Attention from Human Rationales

By Yujia Bao, Shiyu Chang, Mo Yu, Regina Barzilay. 2018

113 citations recorded by the source corpus.

Attention-based models are successful when trained on large amounts of data. In this paper, we demonstrate that even in the low-resource scenario, attention can be learned effectively. To this end, we start with discrete human-annotated rationales and map them into continuous attention. Our central hypothesis is that this mapping is general across domains, and thus can be transferred from resource-rich domains to low-resource ones. Our model jointly learns a domain-invariant representation and induces the desired mapping between rationales and attention. Our empirical results validate this hypothesis and show that our approach delivers significant gains over state-of-the-art baselines, yielding over 15% average error reduction on benchmark datasets.

Subfields: Artificial Intelligence, Model Architectures

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 Model Architectures

Citations in the corpus

When Can Models Learn From Explanations? A Formal Framework for Understanding the Roles of Explanation Data (2021, background): “, 2017; 2018), or through placing priors over how features are weighted or extracted (Zaidan et al., 2007; Small et al., 2011; Zhang et al., 2016; Ross et al., 2017; Bao et al., 2018; Selvaraju et al., 2019; Liang et al., 2020; Stammer et al., 2020).”
Attention is not Explanation (2019, background): “Tying attention weights to human-provided rationales is another potentially promising avenue (Bao et al., 2018).”
Interpretations are useful: penalizing explanations to align neural networks with prior knowledge (2019, background): “An alternative line of work proposes penalizing the gradients of a neural network to match human-provided binary annotations and shows the possibility to improve performance (Ross et al., 2017; Bao et al., 2018; Du et al., 2019) and adversarial robustness (Ross & Doshi-Velez, 2018).”
Evaluating Explanations: How Much Do Explanations from the Teacher Aid Students? (2020, methodology): “Several past approaches have used attention regularization to incorporate human rationales, with an aim to improve the overall performance of the system for classification tasks (Bao et al., 2018; Zhong et al., 2019) and machine translation (Yin et al.”
Few-shot Text Classification with Distributional Signatures (2019, background): “Recently, Bao et al. (2018) also successfully transferred task-specific attention through human rationales.”
Rational LAMOL: A Rationale-based Lifelong Learning Framework (2021, background): “This could be due to many incompatibilities between human rationales and machine attention scores, as mentioned in Bao et al. (2018), which made our algorithm choose sub-optimal layers/heads.”
Learning Credible Deep Neural Networks with Rationale Regularization (2019, background): “The most similar work to ours is using human rationales to improve neural predictions [26].”
Attention in Natural Language Processing (2019, background): “3) Transfer Learning: Furthermore, it is possible to perform transfer learning across different domains [1] or tasks [115].”
Do Human Rationales Improve Machine Explanations? (2019, background): “Bao et al. (2018) extend this work by learning a mapping from the human rationales to continuous attention.”
Few-Shot Semantic Parsing for New Predicates (2021, background): “, 2016), significant manual effort (Bao et al., 2018; Rabinovich et al., 2017), or it is designed only for applications other than semantic parsing (Liu et al.”
A Game Theoretic Approach to Class-wise Selective Rationalization (2019, methodology): “Following the same evaluation protocol [8], we consider a classiﬁcation setting by treating reviews with ratings ≤ 0.4 as negative and ≥ 0.6 as positive.”
Aligning Faithful Interpretations with their Social Attribution (2020, methodology): “We argue against widespread use of this term, as it refers to multiple concepts in NLP and ML (e.g., Zaidan et al., 2007; Bao et al., 2018; DeYoung et al., 2019; Das and Chernova, 2020), and importantly, ‘‘rationalization’’ attributes human intent to the highlight selection, which is not necessarily”
Teach Me to Explain: A Review of Datasets for Explainable Natural Language Processing (2021, background): “[11] Yujia Bao, Shiyu Chang, Mo Yu, and Regina Barzilay.”