FAParser

FAParser copied to clipboard

SOTA

Performance: Highest F1 Score/ UAS(LAS)

• Constituency Parsing with a Self-Attentive Encoder
• Deep Biaffine Attention for Neural Dependency Parsing

Speed: Fastest

• Straight to the Tree: Constituency Parsing with Neural Syntactic Distance

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

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