WOS Hierarchical Text Classification

The WOS Hierarchical Text Classification are three dataset variants created from Web of Science (WOS) title and abstract data categorised into a hierarchical, multi-label class structure. The aim of the sampling and filtering methodology used was to create well-balanced class distributions (at chosen hierarchical levels). Furthermore, the WOS_JTF variant was also created with the aim to only contain publication data such that their class assignments results is classes instances that semantically more similar.

The three dataset variants have the following properties: 1. WOS_JT comprises 43,366 total samples (train=30356, dev=6505, test=6505) and only uses the journal-based classifications as labels. 2. WOS_CT comprises 65,200 total samples (train=45640, dev=9780, test=9780) and only uses citation-based classifications as labels. 3. WOS_JTF comprises 42,926 total samples (train=30048, dev=6439, test=6439) and uses a filtered set of papers based on journal and citation classification.

The dataset is available at:

https://huggingface.co/datasets/marcelsun/wos_hierarchical_multi_label_text_classification

Dataset details:
*.json: - concatenated title and abstract mapped to a list each associated class label.

depth2label.pt: dictionary where: - key = depth of classification hierarchy. - value = list of classes associated with depth.

path_list.pt: - list of tuples for every edge between classes in the hierarchical classification. This specifies the acyclic graph.

slot.pt: dictionary where: - key = label_id of parent class - value = label_ids of children classes

value2slot.pt: dictionary where: - key = label_id - value = label_id of parent class

value_dict.pt: dictionary where: - key = label_id - value = string representation of class.