Out of Order: How important is the sequential order of words in a sentence in Natural Language Understanding tasks?

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1145–1160 August 1–6, 2021. ©2021 Association for Computational Linguistics 1145 Out of Order: How Important Is The Sequential Order of Words in a Sentence in Natural Language Understanding Tasks? Thang M. Pham1 thangpham@auburn.edu Trung Bui2 bui@adobe.com Long Mai2 malong@adobe.com Anh Nguyen1 anh.ng8@gmail.com 1Auburn University 2Adobe Research Abstract Do state-of-the-art natural language under- standing models care about word order? Not always! We found 75% to 90% of the correct predictions of BERT-based classifiers, trained on many GLUE tasks, remain constant after in- put words are randomly shuffled. Although BERT embeddings are famously contextual, the contribution of each individual word to classification is almost unchanged even after its surrounding words are shuffled. BERT- based models exploit superficial cues (e.g. the sentiment of keywords in sentiment analysis; or the word-wise similarity between sequence- pair inputs in natural language inference) to make correct decisions when tokens are ran- domly shuffled. Encouraging models to cap- ture word order information improves the per- formance on most GLUE tasks and SQuAD 2.0. Our work suggests that many GLUE tasks are not challenging machines to understand the meaning of a sentence. 1 Introduction Machine learning (ML) models recently achieved excellent performance on state-of-the-art bench- marks for evaluating natural language understand- ing (NLU). In July 2019, RoBERTa (Liu et al., 2019) was the first to surpass a human baseline on GLUE (Wang et al., 2019). Since then, 13 more methods have also outperformed humans on the GLUE leaderboard. Notably, at least 8 out of the 14 solutions are based on BERT (Devlin et al., 2019)—a transformer architecture that learns repre- sentations via a bidirectional encoder. Given their superhuman GLUE-scores, how do BERT-based models solve NLU tasks? How do their NLU capa- bility differs from that of humans? We shed light into these important questions by examining model sensitivity to the order of words. Word order is one of the key characteristics of a Q1 Does marijuana cause cancer? Q2 How can smoking marijuana give you lung cancer? (a) Prediction: “duplicate” 0.96 Q1 Does marijuana cause cancer? Q2′ you smoking cancer How marijuana lung can give? (b) Prediction: “duplicate” 0.98 Q1 Does marijuana cause cancer? Q2′′ lung can give marijuana smoking How you cancer? (c) Prediction: “duplicate” 0.99 Q1 Does marijuana cause cancer? Q1′ Does cancer cause marijuana? (d) Prediction: “duplicate” 0.77 Figure 1: A RoBERTa-based model achieving a 91.12% accuracy on QQP, here, correctly labeled a pair of Quora questions “duplicate” (a). Interestingly, the predictions remain unchanged when all words in question Q2 is randomly shuffled (b–c). QQP models also often incorrectly label a real sentence and its shuf- fled version to be “duplicate” (d). We found evidence that GLUE models rely heavily on words to make deci- sions e.g. here, “marijuana” and “cancer” (more impor- tant words are highlighted by LIME). Also, there ex- ist self-attention matrices tasked explicitly with extract- ing word-correspondence between two input sentences regardless of the position of those words. Here, the top-3 pairs …

Content truncated to 500 words