Exploring Task Difficulty for Few-Shot Relation Extraction

Abstract

Few-shot relation extraction (FSRE) focuses on recognizing novel relations by learning with merely a handful of annotated instances. Meta-learning has been widely adopted for such a task, which trains on randomly generated few-shot tasks to learn generic data representations. Despite impressive results achieved, existing models still perform suboptimally when handling hard FSRE tasks, where the relations are fine-grained and similar to each other. We argue this is largely because existing models do not distinguish hard tasks from easy ones in the learning process. In this paper, we introduce a novel approach based on contrastive learning that learns better representations by exploiting relation label information. We further design a method that allows the model to adaptively learn how to focus on hard tasks. Experiments on two standard datasets demonstrate the effectiveness of our method.

A Simple yet Effective Relation Information Guided Approach for Few-Shot Relation Extraction

“…The comparable models contain two CNN-based models Proto-HATT (Gao et al, 2019a) and MLMAN (Ye and Ling, 2019), as well as nine BERT-based models BERT-PAIR (Gao et al, 2019b), REGRAB (Qu et al, 2020), TD-proto (Yang et al, 2020), CTEG (Wang et al, 2020), ConceptFERE (Yang et al, 2021), MTB (Baldini Soares et al, 2019), CP (Peng et al, 2020), MapRE (Dong et al, 2021), and HCRP (Han et al, 2021). Since our proposed approach is based on the Prototype Network with BERT, we also compare the Proto-BERT without relation information.…”

Learn from Relation Information: Towards Prototype Representation Rectification for Few-Shot Relation Extraction

“…MapRE (Dong et al, 2021) proposed a framework considering both label-agnostic and label-aware semantic mapping information for low resource relation extraction in both pre-training and fine-tuning. HCRP (Han et al, 2021) introduced three modules containing hybrid prototype learning, relationprototype contrastive learning, and task adaptive focal loss for the model improvement. However, these methods always introduced relation information implicitly, which may introduce more parameters and are limited in dealing with outlier samples.…”

“…MapRE (Dong et al, 2021) proposed a framework considering both label-agnostic and label-aware semantic mapping information in pre-training and fine-tuning. HCRP (Han et al, 2021) introduced three modules containing hybrid prototype learning, relation-prototype contrastive learning, and task adaptive focal loss for the model improvement. However, most existing methods mainly utilized instances given in each relation class to obtain the prototype representation (generally to average these instances).…”

A Simple yet Effective Relation Information Guided Approach for Few-Shot Relation Extraction

“…The comparable models contain two CNN-based models Proto-HATT (Gao et al, 2019a) and MLMAN (Ye and Ling, 2019), as well as nine BERT-based models BERT-PAIR (? ), REGRAB (Qu et al, 2020), TD-proto (Yang et al, 2020), CTEG (Wang et al, 2020), ConceptFERE (Yang et al, 2021), MTB (Baldini Soares et al, 2019), CP (Peng et al, 2020), MapRE (Dong et al, 2021), and HCRP (Han et al, 2021). Since our proposed approach is based on the Prototype Network with BERT, we also compare the Proto-BERT without relation information.…”

“…CTEG (Wang et al, 2020) proposed a model that learns to decouple high co-occurrence relations, where two types of external information are added. Another intuitive idea is to hope that the model can learn good prototypes or representations , that is, to reduce the distances of the intra-class while widening the ones among different classes (Han et al, 2021;Dong et al, 2021), where However, there are two limitations in how these works introduce relation information. The first is that most of them take implicit constraints, like contrastive learning or relation graphs, instead of the direct fusion, which can be weak facing the remote samples.…”