upunaprosk / grammar-checker

Essay Grammar Checker trained on REALEC Corpus using SpaCy

✒️ Essay Grammar Checker

Essay Grammar Checker trained on Russian Error-Annotated Learner English Corpus using SpaCy.

Training

The checker consists of 6 pipelines each trained on specific error types. Error Categories used for pipeline mapping:

    "spelling":{"Spelling", "Capitalisation"},  
    "punctuation": {"Punctuation"},
    "articles": {"Articles"},  
    "vocabulary": {"lex_item_choice", "lex_part_choice",
                   'Category_confusion','Formational_affixes'},
    "grammar_major": {'Tense_choice','Prepositions','Agreement_errors', 'Redundant_comp'},
    "grammar_minor": {'Word_order','Noun_number', 'Numerals','Verb_pattern', 'Determiners'}

The project.yml defines the data assets required by the project, as well as the available commands and workflows. For details, see the spaCy projects documentation. nlp.rehearse method can be also used to update trained models.

Commands

The following commands are defined by the project. They can be executed using spacy project run [name]. Commands are only re-run if their inputs have changed.

Command	Description
preprocess	Convert the data to spaCy format required
generate_configs	Configs class weight update
train_pipelines	Launch training
evaluate_pipelines	Evaluate models
assemble_pipelines	Assemble model
package	Package the resulting model

Workflows

The following workflows are defined by the project. They can be executed using spacy project run [name] and will run the specified commands in order. Commands are only re-run if their inputs have changed.

Workflow	Steps
all	preprocess → generate_configs → train_pipelines → evaluate_pipelines → assemble_pipelines → package

Assets

The following assets are defined by the project. They can be fetched by running spacy project assets in the project directory. The data used for training can be extracted from the corpus using the following code.

File	Source
assets/realec/data_realec.tar.bz2	REALEC

Performance

Metric	Scores
f1-scores	punctuation:0.779, spelling:0.939, capitalisation:0.902, articles:0.852, lex_part_choice: 0.235, lex_item_choice: 0.685, Category_confusion: 0.705, Formational_affixes: 0.742, Verb_pattern:0.629, Noun_number:0.920, Word_order:0.527, Numerals:0.736, Determiners:0.044, Agreement_errors:0.835, Prepositions:0.710, Redundant_comp:0.495, Tense_choice:0.825

Usage

Install

!pip install https://huggingface.co/iproskurina/en_grammar_checker/resolve/main/en_grammar_checker-any-py3-none-any.whl

# Using spacy.load().
import spacy
nlp = spacy.load("en_grammar_checker")

# Importing as module.
import en_grammar_checker
nlp = en_grammar_checker.load()

Streamlit

streamlit run streamlit_app.py

SpanCategorizer-based rendering