Towards preserving word order importance through Forced Invalidation

Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2563–2570 May 2-6, 2023 ©2023 Association for Computational Linguistics Towards preserving word order importance through FORCED INVALIDATION Hadeel Al-Negheimish1, Pranava Madhyastha2,1 and Alessandra Russo1 1 Imperial College London 2 City, University of London {halnegheimish,a.russo}@imperial.ac.uk, pranava.madhyastha@city.ac.uk Abstract Large pre-trained language models such as BERT have been widely used as a frame- work for natural language understanding (NLU) tasks. However, recent findings have revealed that pre-trained language models are insensitive to word order. The performance on NLU tasks remains unchanged even after randomly per- muting the word of a sentence, where crucial syntactic information is destroyed. To help pre- serve the importance of word order, we propose a simple approach called FORCED INVALIDA- TION (FI): forcing the model to identify per- muted sequences as invalid samples. We per- form an extensive evaluation of our approach on various English NLU and QA based tasks over BERT-based and attention-based mod- els over word embeddings. Our experiments demonstrate that FI significantly improves the sensitivity of the models to word order.1 1 Introduction Ordering of words in a sentence is an important structural attribute for natural languages such as English, where subject-verb-object (SVO) structure is common and important to convey the meaning of the sentence. Understanding and comprehend- ing natural language without strict adherence to a systematic ordering of words would make it an extremely challenging task. Recent work has in- vestigated a surprising lack of sensitivity to word order information in state-of-the-art masked lan- guage models. Recent research has focused on the impact of word order perturbation during the evaluation of models trained on well-ordered data. The results show that masked language models exhibit a catas- trophic lack of sensitivity to word order permuta- tions or shuffles, even for complex tasks in which task-relevant syntactic properties are completely 1Our code and data for replication are available at https: //github.com/halnegheimish/ForcedInvalidation destroyed (Pham et al., 2020; Al-Negheimish et al., 2021; Sinha et al., 2021b; Gupta et al., 2021). These studies show that models are still predicting the gold label for examples even after sequences have been permuted, and they also do so with high confidence (Sinha et al., 2021b; Gupta et al., 2021). This anomalous behaviour can potentially result in undesirable shortcuts or can cause models to fail catastrophically in simple adversarial settings. Fur- thermore, Sinha et al. (2021a) study the effect of pre-training masked language models on shuffled data, and suggest that the model might simply be capturing higher-order word co-occurrence statis- tics, rather than uncovering sophisticated semantic and syntactic structures necessary for language un- derstanding. In this paper, we present a simple, yet general ap- proach called FORCED INVALIDATION (FI), where we force models to explicitly identify sequence per- mutations (§3). While our proposal is extensible to multiple types of models, here we present a con- trolled study over masked language model-based BERT models, and attention-based models over word-embeddings (§4). We present a large battery of experiments over a …

Content truncated to 500 words