Research On Key Technologies Of Relation Extraction In Low-resource Scenarios

Relation extraction is an essential task in the field of natural language processing,which aims to automatically capture the structured relations contained in the unstructured plain text.It is the cornerstone of various applications,including knowledge graph construction,question answering,and recommendation systems.With the rapid development of artificial intelligence and deep learning technology,relation extraction methods based on deep learning have been extensively studied.The neural network is trained through large-scale high-quality annotated data with rich relation categories,so that the model can obtain the ability to distinguish these relations.Although impressive performance has been achieved,these methods heavily rely on a large number of labeled data for training,making them difficult to adapt to novel relations that have never been seen in the training process.However,in real scenarios,due to the high cost of manpower and material resources for annotation and the scarcity of data in specific fields,it is very difficult to obtain a large amount of labeled data,which largely limits the practical application of existing deep learning-based relation extraction models.Therefore,it is of great and urgent significance to explore how to empower the model to generalize to novel relations with a handful of labeled instances.Few-shot relation extraction has attracted recent attention from domestic and foreign scholars.However,there are still some open problems that need to be addressed,including how to better learn discriminative and unbiased representations,solve hard few-shot relation extraction tasks,and mine deep knowledge in large-scale pre-trained language models.Hence,this paper conducts research on relation extraction task based on low-resource scenarios,and makes the following contributions:1.To address the insufficient supervision and spurious correlation issues in transfer-based few-shot relation extraction methods,this paper proposes a two-stage multi-branch network based on contrastive learning and prototypical networks,aiming to learn more discriminative and unbiased representations.Specifically,in the pre-training stage,a supervised contrastive pre-training strategy is employed to fully explore the rich relation categories in the global training data and obtain more discriminative representations using contrastive learning.In the meta-training stage,a prototypical network based on sentence branch and entity branch is designed,and the auxiliary entity branch encourages the sentence branch to learn unbiased representations.Experimental results demonstrate that the proposed model effectively alleviates insufficient supervision and spurious correlation in existing methods and improves the performance of few-shot relation extraction tasks.2.Existing few-shot relation extraction methods fail to effectively address hard few-shot tasks,as they overlook pivotal information that is crucial to distinguish confusing classes and train indiscriminately via randomly sampled tasks of varying difficulty.To fill this gap,this work injects semantic information of relation labels and presents a novel contrastive learning method based on hybrid prototypes and relations,which captures fine-grained prototype representations to better model hard tasks.Meanwhile,two training strategies are proposed,namely,hard negative prototype generation strategy and task-adaptive focal loss,to increase the difficulty of training tasks and force the model to focus on hard tasks during training,thereby improving the effectiveness and robustness of the model.The proposed method achieves state-of-the-art performance on three popular few-shot relation extraction datasets and significantly improves the performance on hard tasks.3.Towards generalized few-shot relation extraction task,this paper proposes a generative prompt tuning method to eliminate the rigid restrictions of current prompt based approaches.It innovatively converts relation extraction into a text infilling task and utilizes pre-trained generative models to jointly generate entity types and relation label sequences,thus fully leveraging the knowledge of pre-trained language models and achieving effective relation recognition.In the prediction stage,a simple yet effective entity-guided decoding and relation scoring mechanism is proposed to implicitly constrain the generated relation types by injecting entity information and efficiently align the generated sequences with the label set.We prove the effectiveness of our method on four relation extraction datasets based on three pre-trained models under both fully supervised and low-resource scenarios.