README.md

AdaSeq: An All-in-One Library for Developing State-of-the-Art Sequence Understanding Models

English | 简体中文

Introduction

AdaSeq (Alibaba Damo Academy Sequence Understanding Toolkit) is an easy-to-use all-in-one library, built on ModelScope, that allows researchers and developers to train custom models for sequence understanding tasks, including part-of-speech tagging (POS Tagging), chunking, named entity recognition (NER), entity typing, relation extraction (RE), etc.

🌟 Features:

• Plentiful Models:

  AdaSeq provide plenty of cutting-edge models, training methods and useful toolkits for sequence understanding tasks.

• State-of-the-Art:

  Our aim to develop the best implementation, which can beat many off-the-shelf frameworks on performance.

• Easy-to-Use:

  One line of command is all you need to obtain the best model.

• Extensible:

  It's easy to register a module, or build a customized sequence understanding model by assembling the predefined modules.

⚠️Notice: This project is under quick development. This means some interfaces could be changed in the future.

📢 What's New

• 2022-03: [SemEval 2023] Our U-RaNER Won 1st place in 9 tracks at SemEval 2023 Task2: Multilingual Complex Named Entity Recognition! Source code will be released soon.
• 2022-12: [EMNLP 2022] Retrieval-augmented Multimodal Entity Understanding Model (MoRe)
• 2022-11: [EMNLP 2022] Ultra-Fine Entity Typing Model (NPCRF)
• 2022-11: [EMNLP 2022] Unsupervised Boundary-Aware Language Model (BABERT)

⚡ Quick Experience

You can try out our models via online demos built on ModelScope: [English NER] [Chinese NER] [CWS]

More tasks, more languages, more domains: All modelcards we released can be found in this page Modelcards.

🛠️ Model Zoo

Supported models:

• Transformer-based CRF
• Partial CRF
• Retrieval Augmented NER
• Biaffine NER
• Global-Pointer
• Multi-label Entity Typing
• ...

💾 Dataset Zoo

We collected many datasets for sequence understanding tasks. All can be found in this page Datasets.

📦 Installation

AdaSeq project is based on Python version >= 3.7 and PyTorch version >= 1.8.

• installation via pip：

pip install adaseq

• installation from source：

git clone https://github.com/modelscope/adaseq.git
cd adaseq
pip install -r requirements.txt -f https://modelscope.oss-cn-beijing.aliyuncs.com/releases/repo.html

Verify the Installation

To verify whether AdaSeq is installed properly, we provide a demo config for training a model (the demo config will be automatically downloaded).

adaseq train -c demo.yaml

You will see the training logs on your terminal. Once the training is done, the results on test set will be printed: test: {"precision": xxx, "recall": xxx, "f1": xxx}. A folder experiments/toy_msra/ will be generated to save all experimental results and model checkpoints.

📖 Tutorials

• Quick Start
• Basics
  • Learning about Configs
  • Customizing Dataset
  • [TODO] Common Architectures
  • [TODO] Useful Hooks
  • Hyperparameter Optimization
  • Training with Multiple GPUs
• Best Practice
  • Training a Model with Custom Dataset
  • Reproducing Results in Published Papers
  • [TODO] Uploading Saved Model to ModelScope
  • [TODO] Customizing your Model
  • [TODO] Serving with AdaLA

📝 Contributing

All contributions are welcome to improve AdaSeq. Please refer to CONTRIBUTING.md for the contributing guideline.

📄 License

This project is licensed under the Apache License (Version 2.0).