Linguistic Knowledge in Data Augmentation for Natural Language Processing: An Example on Chinese Question Matching

Data augmentation (DA) is a common solution to data scarcity and imbalance problems, which is an area getting increasing attentions from the Natural Language Processing (NLP) community. While various DA techniques have been used in NLP research, little is known about the role of linguistic knowledge in DA for NLP; in particular, whether more linguistic knowledge leads to a better DA approach. To investigate that, we designed two adapted DA programs and applied them to LCQMC (a Large-scale Chinese Question Matching Corpus) for a binary Chinese question matching classification task. The two DA programs produce augmented texts by five simple text editing operations, largely irrespective of language generation rules, but one is enhanced with a n-gram language model to make it fused with extra linguistic knowledge. We then trained four neural network models and a pre-trained model on the LCQMC train sets of varying size as well as the corresponding augmented trained sets produced by the two DA programs. The test set performances of the five classification models show that adding probabilistic linguistic knowledge as constrains does not make the base DA program better, since there are no discernible performance differences between the models trained on the two types of augmented train sets. Instead, since the added linguistic knowledge decreases the diversity of the augmented texts, the trained models generalizability is hampered. Moreover, models trained on both types of the augmented trained sets were found to be outperformed by those directly trained on the associated un-augmented train sets, due to the inability of the underlying text editing operations to make paraphrastic augmented texts. We concluded that the validity and diversity of the augmented texts are two important factors for a DA approach or technique to be effective and proposed a possible paradigm shift for text augmentation.