SOTA

Performance: Highest F1 Score/ UAS(LAS)

Speed: Fastest

This repo's target

Faster and Accurate Syntactic Parsing both on Constituency and Dependency.

Naive Motivation

Conversion: Dependency Tree could be converted from constituency by utilizing head rule.
Span: where the head rule is used.

Implementation

Self-Attentive Con Parser, start from scratch with:

  • Phrase Structure Tree Load/Conversion/Expr, Oracle Design
  • Pretrain Model Loaded: ELMO
  • Multi-Head Self-Attentive Encoder
  • Max-Span Tree Inference
  • Training

Biaffine Dep Parser, start from scratch with:

  • Dependency Relation Tree Load/Conversion/Expr, Oracle Design
  • Pretrain Word Embedding Loaded
  • Bi-LSTM Encoder
  • MST Inference with Attention Module
  • Training

FAParser's new features:

  • Enhanced Pretrain Module of Word Representation
  • Enhanced MST Inference with Multi-Head Attention Module
  • Conditional modeling on interaction between information flows of phrase structure and dependency relation

Similar to the design of fairseq, we organize our FAParser as:

FAParser
β”‚   README.md
β”‚   train.py
β”‚   inference.py
β”‚   preprocess.py
β”‚
└───evaluation: for validation or testing
β”‚   β”‚   F1
β”‚   β”‚   Accuracy
β”‚   β”‚       β”‚ UAS
β”‚   β”‚       β”” LAS
β”‚   β””  ...
β”‚
└───data: 
β”‚   β”‚   tree loaded or structure utils
β”‚   β”‚   
β”‚   β”” ...
β”‚   
└───criterion: 
β”‚   β”‚   cross entropy. etc
β”‚   β”‚   
β”‚   β”” ...
β”‚
└───models: 
β”‚   β”‚   three parser
β”‚   β”‚   
β”‚   β”” ...
β”‚
└───modules: 
β”‚   β”‚   series of module used in models
β”‚   β”‚   
β”‚   β”” ...
β”‚
└───optim: for optimizer
β”‚   β”‚   lr_shedule
β”‚   β”‚   adam...
β”‚   β”” ...
β”‚
└───tasks: for loss computing
β”‚   β”‚   Constituency Parser
β”‚   β”‚   Dependenecy Parser
β”‚   β”‚   FAParser
β”‚   β”” ...
β”‚
└───utlis:
β”‚   β”‚   command/preprocess/meters...
β”‚   β”‚   
β”‚   β”” ...

criterion,modules,optim and data could be initialized with the fairseq's sub-modules. we just need build some task-specific sub-packages.

Faparser

A Fast(er) and Accurate Syntactic Parsing by Exacter Searching.

Faparser Info

⭐ Stars 15
πŸ”— Source Code github.com
πŸ•’ Last Update a year ago
πŸ•’ Created 3 years ago
🐞 Open Issues 0
βž— Star-Issue Ratio Infinity
😎 Author baoy-nlp