Unsupervised Domain Adaptation with Global and Local Graph Neural Networks in Limited Labeled Data Scenario: Application to Disaster Management

TL;DR

This work utilizes limited labeled data along with abundantly available unlabeled data, generated during a source disaster to propose a novel two-part graph neural network that outperforms state-of-the-art techniques in the area of disaster management.

Abstract

Identification and categorization of social media posts generated during disasters are crucial to reduce the sufferings of the affected people. However, lack of labeled data is a significant bottleneck in learning an effective categorization system for a disaster. This motivates us to study the problem as unsupervised domain adaptation (UDA) between a previous disaster with labeled data (source) and a current disaster (target). However, if the amount of labeled data available is limited, it restricts the learning capabilities of the model. To handle this challenge, we utilize limited labeled data along with abundantly available unlabeled data, generated during a source disaster to propose a novel two-part graph neural network. The first-part extracts domain-agnostic global information by constructing a token level graph across domains and the second-part preserves local instance-level semantics. In our experiments, we show that the proposed method outperforms state-of-the-art techniques by $2.74\%$ weighted F$_1$ score on average on two standard public dataset in the area of disaster management. We also report experimental results for granular actionable multi-label classification datasets in disaster domain for the first time, on which we outperform BERT by $3.00\%$ on average w.r.t weighted F$_1$. Additionally, we show that our approach can retain performance when very limited labeled data is available.

Figures (3)

• Figure 1: A block level illustration of our model. The components $H_G$ and $H_{G\mathbf{x}}$ represent the GCN on token graph (Section \ref{['subsec:global_token_graph']}) and GAT on instance graph (Section \ref{['subsec:local_instance_graph']}) respectively. Here $x_i$ is a token from instance $\mathbf{x}$. The BiLSTM layer (Section \ref{['subsec:classification_layer']}) uses concatenation of the two to learn an end-to-end model. For simplicity, we only show connections in one direction and the last forward hidden state $h_T$.
• Figure 2: Heatmap of normalized cosine similarity of a sample of token vocabulary from FIRE16 and SMERP17 dataset pair; (a) before and (b) after global token graph training. We observe that the training results in tokens forming clusters of similarity (indicated by clusters of warmer values on the right). The normalization scheme is discussed in Section \ref{['subsec:qual_results']}.
• Figure 3: Average absolute performance gain (weighted F$_1$) of GLEN over BERT as training data is reduced on NEQ-QFL (left) and FIRE16-SMERP17 (right) dataset pairs respectively.