Attentive Siamese LSTM for Low-Resource Text Abnormal Detection of Transportation Documents

Low-resource text anomaly detection in transportation documents poses a significant challenge due to the limited availability of labeled data for training. Transportation documents within freight networks are organized into 29 fields that often lack clear logical connections, complicating the identification of abnormal fields. This task necessitates the development of advanced algorithms capable of recognizing both similarities and differences within texts. This paper proposes an enhanced attentive Siamese Long Short-Term Memory (LSTM) network for anomaly detection in transportation documents. The approach begins with data augmentation using documents from the Ministry of Transport to expand the training dataset. A pre-detection method utilizing abstract document vectors is then introduced to enhance detection efficiency. Finally, an improved attentive Siamese LSTM network is proposed, specifically tailored for anomaly detection in transportation documents. Our extensive experiments demonstrate the effectiveness of the proposed framework in detecting anomalies.