NeuralClassifier is Tencent’s open-source toolkit for neural text classification, with a particular strength in hierarchical multi-label classification โ the harder, label-tree-structured problems common in production content and document systems. It packages a wide bench of proven text encoders behind one training and evaluation workflow. It is an open-source project on GitHub with around 1,900 stars, and it received updates in 2025.
Where most text classification demos stop at flat single-label problems, NeuralClassifier is designed for datasets where each document carries several labels arranged in a hierarchy, such as category trees for news, products or support tickets. The toolkit implements encoders including FastText, TextCNN, TextRNN, RCNN, VDCNN, DPCNN, DRNN, AttentiveConvNet and a Transformer encoder, and handles the hierarchical label structure during both training and prediction. Standard flat classification scenarios are supported too, making it a practical reference implementation for applied NLP teams.
Free. NeuralClassifier is open source and free to use for research and commercial text classification projects.
The toolkit ships with example datasets and evaluation utilities for multi-label problems, reporting metrics such as micro and macro F1 that reflect performance across both frequent and rare labels in the tree. Prediction outputs preserve the hierarchical label paths, which matters for downstream systems like taxonomy tagging and catalog classification where a parent category prediction is useful even when the leaf is uncertain. Its encoder zoo also makes controlled comparisons easy: the same dataset and label structure can be benchmarked across nine architectures with only configuration changes.
NLP engineers and researchers working on document categorization, content tagging, intent routing or any classification problem where labels form a tree rather than a flat list.