PyTorch Implementation of "Bridging Pre-trained Language Models and Hand-crafted Features for Unsupervised POS Tagging" (Findings of ACL 2022)

Last update: Apr 29, 2022

Related tags

Text Data & NLP FeatureCRFAE

Overview

Feature_CRF_AE

Feature_CRF_AE provides a implementation of Bridging Pre-trained Language Models and Hand-crafted Features for Unsupervised POS Tagging:

@inproceedings{zhou-etal-2022-Bridging,
  title     = {Bridging Pre-trained Language Models and Hand-crafted Features for Unsupervised POS Tagging},
  author    = {Zhou, houquan and Li, yang and Li, Zhenghua and Zhang Min},
  booktitle = {Findings of ACL},
  year      = {2022},
  url       = {?},
  pages     = {?--?}
}

Please concact Jacob_Zhou \at outlook.com if you have any questions.

Contents
Installation
Performance
Usage

Installation

Feature_CRF_AE can be installing from source:

$ git clone https://github.com/Jacob-Zhou/FeatureCRFAE && cd FeatureCRFAE
$ bash scripts/setup.sh

The following requirements will be installed in scripts/setup.sh:

python: 3.7
allennlp: 1.2.2
pytorch: 1.6.0
transformers: 3.5.1
h5py: 3.1.0
matplotlib: 3.3.1
nltk: 3.5
numpy: 1.19.1
overrides: 3.1.0
scikit_learn: 1.0.2
seaborn: 0.11.0
tqdm: 4.49.0

For WSJ data, we use the ELMo representations of elmo_2x4096_512_2048cnn_2xhighway_5.5B from AllenNLP. For UD data, we use the ELMo representations released by HIT-SCIR.

The corresponding data and ELMo models can be download as follows:

# 1) UD data and ELMo models:
$ bash scripts/prepare_data.sh
# 2) UD data, ELMo models as well as WSJ data 
#    [please replace ~/treebank3/parsed/mrg/wsj/ with your path to LDC99T42]
$ bash scripts/prepare_data.sh ~/treebank3/parsed/mrg/wsj/

Performance

WSJ-All

Seed	M-1	1-1	VM
0	84.29	70.03	78.43
1	82.34	64.42	77.27
2	84.68	62.78	77.83
3	82.55	65.00	77.35
4	82.20	66.69	77.33
Avg.	83.21	65.78	77.64
Std.	1.18	2.75	0.49

WSJ-Test

Seed	M-1	1-1	VM
0	81.99	64.84	76.86
1	82.52	61.46	76.13
2	82.33	61.15	75.13
3	78.11	58.80	72.94
4	82.05	61.68	76.21
Avg.	81.40	61.59	75.45
Std.	1.85	2.15	1.54

Usage

We give some examples on scripts/examples.sh. Before run the code you should activate the virtual environment by:

$ . scripts/set_environment.sh

Training

To train a model from scratch, it is preferred to use the command-line option, which is more flexible and customizable. Here are some training examples:

$ python -u -m tagger.cmds.crf_ae train \
    --conf configs/crf_ae.ini \
    --encoder elmo \
    --plm elmo_models/allennlp/elmo_2x4096_512_2048cnn_2xhighway_5.5B \
    --train data/wsj/total.conll \
    --evaluate data/wsj/total.conll \
    --path save/crf_ae_wsj

$ python -u -m tagger.cmds.crf_ae train \
    --conf configs/crf_ae.ini \
    --ud-mode \
    --ud-feature \
    --ignore-capitalized \
    --language-specific-strip \
    --feat-min-freq 14 \
    --language de \
    --encoder elmo \
    --plm elmo_models/de \
    --train data/ud/de/total.conll \
    --evaluate data/ud/de/total.conll \
    --path save/crf_ae_de

For more instructions on training, please type python -m tagger.cmds.[crf_ae|feature_hmm] train -h.

Alternatively, We provides some equivalent command entry points registered in setup.py: crf-ae and feature-hmm.

$ crf-ae train \
    --conf configs/crf_ae.ini \
    --encoder elmo \
    --plm elmo_models/allennlp/elmo_2x4096_512_2048cnn_2xhighway_5.5B \
    --train data/wsj/total.conll \
    --evaluate data/wsj/total.conll \
    --path save/crf_ae

Evaluation

$ python -u -m tagger.cmds.crf_ae evaluate \
    --conf configs/crf_ae.ini \
    --encoder elmo \
    --plm elmo_models/allennlp/elmo_2x4096_512_2048cnn_2xhighway_5.5B \
    --data data/wsj/total.conll \
    --path save/crf_ae

Predict

$ python -u -m tagger.cmds.crf_ae predict \
    --conf configs/crf_ae.ini \
    --encoder elmo \
    --plm elmo_models/allennlp/elmo_2x4096_512_2048cnn_2xhighway_5.5B \
    --data data/wsj/total.conll \
    --path save/crf_ae \
    --pred save/crf_ae/pred.conll

PyTorch Implementation of "Bridging Pre-trained Language Models and Hand-crafted Features for Unsupervised POS Tagging" (Findings of ACL 2022)

Related tags

Overview

Feature_CRF_AE

Contents

Installation

Performance

WSJ-All

WSJ-Test

Usage

Training

Evaluation

Predict

Owner

Jacob Zhou

This python module is an easy-to-use port of the text normalization used in the paper "Not low-resource anymore: Aligner ensembling, batch filtering, and new datasets for Bengali-English machine translation". It is intended to be used for normalizing / cleaning Bengali and English text.

scikit-learn wrappers for Python fastText.

Neural network sequence labeling model

An end to end ASR Transformer model training repo

LOT: A Benchmark for Evaluating Chinese Long Text Understanding and Generation

Contains analysis of trends from Fitbit Dataset (source: Kaggle) to see how the trends can be applied to Bellabeat customers and Bellabeat products

A Facebook Messenger Chatbot using NLP

Natural language computational chemistry command line interface.

A2T: Towards Improving Adversarial Training of NLP Models (EMNLP 2021 Findings)

🍊 PAUSE (Positive and Annealed Unlabeled Sentence Embedding), accepted by EMNLP'2021 🌴

[ICCV 2021] Instance-level Image Retrieval using Reranking Transformers

Official codebase for Can Wikipedia Help Offline Reinforcement Learning?

Implementation of Multistream Transformers in Pytorch

Yomichad - a Japanese pop-up dictionary that can display readings and English definitions of Japanese words

APEACH: Attacking Pejorative Expressions with Analysis on Crowd-generated Hate Speech Evaluation Datasets

GrammarTagger — A Neural Multilingual Grammar Profiler for Language Learning

Code Implementation of "Learning Span-Level Interactions for Aspect Sentiment Triplet Extraction".

Blackstone is a spaCy model and library for processing long-form, unstructured legal text

A multi-lingual approach to AllenNLP CoReference Resolution along with a wrapper for spaCy.

Tokenizer - Module python d'analyse syntaxique et de grammaire, tokenization