Research On Distantly Supervised Fine-grained Entity Typing

Fine-grained entity typing(FET)is an essential task in information extraction,and its goal is to classify entity mentions in sentences into a set of predefined fine-grained types.Understanding the type of entity mentions in sentences is helpful for various downstream applications,such as entity linking,question-answering systems,recommendation systems,etc.The current popular fine-grained entity typing models are mainly based on deep learning and require an amount of high-quality labeled data for training.However,manual labeling is expensive.Therefore,distant supervision has been adopted to automatically annotate a large number of unlabeled mentions in the training corpus.The basic idea is that an unlabeled entity mention will be linked to an existing entity in the knowledge base(KB),and then all types of the entity will be assigned to the entity mention.Despite its efficiency,since it does not consider the context information of the entity mention when it is annotated,distant supervision often suffers from the noisy labeling problem.The FET model trained with noisy labels can significantly degrade the performance.To address the noisy labeling problem on FET,most existing methods propose to divide the training data into “clean” set and “noisy” set and use different strategies to deal with them during the training process.Although the existing research work has made great progress,there are still two challenges:(1)The “clean” set used by current methods are mainly generated by manual annotation or heuristic rules.The former requires additional human annotation,and the latter produces a “clean” set that still suffers from the false positive problem(i.e.noisy labels).(2)The current method mainly adopts the method based on partial label loss to deal with the “noisy”set,that is,the model uses its own prediction result as the training optimization target,which will suffer from the confirmation bias problem.To address the challenges of current methods,this paper proposes two different solutions.Specifically,the research contents and main contributions of this paper are as follows:(1)This paper proposes a fine-grained entity typing method based on three-stage curriculum learning,which iteratively automatically filters out clean data from distantly supervised annotated datasets for training the FET model.This method solves the false positive problem on “clean” set without introducing additional human annotations.At the same time,in three different stages of curriculum learning,a hierarchical label smoothing method based on different strategies is proposed for the noisy labeling problem in the dataset,which smooths the labels of the training data to different degrees,and uses the smoothed labels to train the models and prevent models from fitting noisy data.(2)This paper proposes a fine-grained entity typing method based on semi-supervised learning.Specifically,in this paper,the “clean” set is considered labeled data,and the “noisy”set is considered unlabeled data.To solve the false positive problem on the “clean” set,this paper proposes a mixed label smoothing method to generate smoothed labels for the data in the“clean” set for model training,avoiding overfitting to noisy data.To address the confirmation bias problem on the “noisy” set,this paper does not consider the original labels of the “noisy”set but uses pseudo labeling techniques to generate reliable pseudo labels used to train the model.(3)This paper conducts experiments on three widely used distantly supervised fine-grained entity typing datasets.For the fine-grained entity typing method based on three-stage curriculum learning,the method is compared with the most relevant baseline methods at the time,and the experimental results show that using this method achieves better results than the baseline on all three datasets,and the case study shows that the method can filter the clean data.For the fine-grained entity typing method based on semi-supervised learning,the method is compared with the most relevant baselines at the time on three datasets.The experimental results demonstrate the effectiveness of the method and show that the method outperforms other semi-supervised learning methods on denoise.