This repository provides a comprehensive collection of deep learning models for text classification, targeting NLP researchers and practitioners. It offers implementations of various classic and state-of-the-art architectures, enabling users to explore, benchmark, and apply them to their own datasets for tasks like sentiment analysis and multi-label classification.

How It Works

The project implements a wide array of text classification models, including fastText, TextCNN, RNNs, RCNNs, Hierarchical Attention Networks, Seq2Seq with attention, Transformers, Dynamic Memory Networks, and Entity Networks. It supports multi-label classification and offers ensemble methods like boosting. The models are designed to be independent of the dataset, with a focus on providing baseline implementations and exploring different architectural choices for language understanding.

Quick Start & Requirements

• Install: TensorFlow 1.8+ (compatible with 1.1-1.13), Python 2.7+ (Python 3.6+ with minor adjustments).
• Data: Pre-processed cached data (1.8GB zip) is available for quick setup, requiring ~8GB RAM during training. Alternatively, custom data can be pre-processed using provided Jupyter notebooks.
• Links: NLP API Demo, CLUE benchmark

Highlighted Details

• Implements 13 distinct text classification models, including BERT and Transformer.
• Provides performance benchmarks on a multi-label prediction task, showing scores and training times.
• Includes implementations for sequence-to-sequence tasks and memory networks capable of transitive inference.
• Offers a boosting ensemble method to improve model performance by stacking identical models.

Maintenance & Community

• Contact: brightmart@hotmail.com
• QQ Group: 836811304 for ML/DL/NLP discussions.

Licensing & Compatibility

• The repository does not explicitly state a license. TensorFlow 1.8 is a key dependency.

Limitations & Caveats

• Primarily developed for Python 2.7, with Python 3 compatibility requiring minor code adjustments.
• Some models are described as "simple" and may not achieve top-tier performance without further tuning.
• BERT implementation notes suggest potential memory constraints on standard GPUs for longer sequences.