MenuNER

This work is done on the paper 'MenuNER: Domain-Adapted BERT Based NER Approach for a Domain with Limited Dataset and Its Application to Food Menu Domain' ( https://www.mdpi.com/2076-3417/11/13/6007 ). We started implementing from scratch all the methods present in the paper and then performing some additional experiments. All the experiments and the method details can be found in the report file.

Abstract

Named entity recognition (NER) is a sub-task of information extraction that seeks to locate and classify named entities in text, into pre-defined categories such as the names of persons, organizations, locations, etc. In this paper, we focus on Menu entity extraction from online user reviews for the restaurant. The approach involves first adapting NER task on a new domain where large datasets are rarely available, embedding and fine-tuning the network model BiLSTM with CRF, very popular in NER tasks, using extended feature vectors.

Our Architecture

MenuNER architecture used in this work. Each word of each sentence is embedded with POS, BERT and Char embedding layer (green). Then the BiLSTM (orange), along with the MultiHeadAttention (cyan), the Linear layer(lilac) and the CRF module (turquoise) work to predict the label of the word between 'O' and 'MENU'.

Setup

First, clone the GitHub and create a new virtual environment in the project directory and activate it.

git clone https://github.com/eleonorapoeta/MenuNER/
python3 -m venv .env
source .env/bin/activate

Then, install all requirements

pip install -r requirements.txt

Reviews Retrieval and manipulation

After the setting the virtual environment and installation of the requirements, to perform this part of the project you need to search for the Yelp Dataset available at https://www.yelp.com/dataset/ .

cd ./MenuNER/bert_training
reviews.py --reviews_filename --business_filename 

BERT domain

To train the BERT domain on the reviews extracted, you can also add a previous BERT checkpoint from which you can resume the training

bert_domain.py --filtered_reviews_filename  --checkpoint_path 

Training

You can train the BiLSTM_CRF model with the possibility of choosing, for each module, whether it contributes or not. Train and Validation files are provided.

cd ./MenuNER/ner
train.py 
--data_dir 
--seed 
--max_length 
--train 
--val
--lr
--epochs
--bert_checkpoint
--pos
--pos_embedding_dim
--char
--char_embedding_dim
--attention
--eval_and_save_steps

Test

You can test your trained model on the test file already provided.

cd ./MenuNER/ner
test.py 
--data_dir 
--pos2ix 
--model_path 
--max_length 
--test
--bert_checkpoint
--pos
--pos_embedding_dim
--char
--char_embedding_dim
--attention

Usage

To easily test all the experiments aforementioned we also provide the following Google Colab notebook. All of our experiments are run on Google Colab framework.