AR-BERT: Aspect-relation enhanced Aspect-level Sentiment Classification with Multi-modal Explanations

Code for our TheWebConf 2022 paper:
Title: AR-BERT: Aspect-relation enhanced Aspect-level Sentiment Classification with Multi-modal Explanations
Authors: Sk Mainul Islam and Sourangshu Bhattacharya

Software Requirements:

• Python 3.6.8
• Tensorflow-gpu 1.15.0 (See suitable tensorflow 1.x GPU version based on your system CUDA: https://www.tensorflow.org/install/source#gpu)
• Spacy 2.3.2
• Snappy 1.1.7 (http://snap.stanford.edu/snappy/index.html)

Hardware Requirements:

• Quadro P5000 Single core GPU (16278MiB)
• CUDA version: 10.0

Data:

Go to folder /data/

Usage:

Please follow the follwing steps sequentially (as mentioned in teh paper also)

1. Aspect to entity mapping:

Go to folder: /src/aspect_to_entity_mapping/

1.1. To prepare raw data for wikification:

python data_preprocessing_for_wikification_input.py

1.2 To extract entities (wikification):

python wikification_of_input_data.py wikifier_input_data_directory dataset_name(laptop/rest/twitter) mode(train/test) wikifier_output_directory

2. Two level aspect entity embediing generation:

Go to folder: /src/two_level_aspect_entity_embedding_generation/

2.1 Clustering of DBPedia Page-link Knowledge Graph:

Go to folder: ./kg_clustering/

2.1.1 For preparing numeric edge list (src, dest) and dictionary entity label --> id and vice-versa from KG:

python kg_pre_processing_create_edge_tuple_list.py path of the DBPedia Page-link English KG

2.1.2 For creating SNAP undirected graph and retrieving maximum connected componenet required for next step clustering:

python dbpedia_pagelink_edge_tuple_list_to_snap_graph.py

2.1.3 For preparing numeric edge list (src, dest) and dictionary entity label --> id and vice-versa from maximum connected componenet of KG:

python connected_kg_pre_processing_create_edge_tuple_list.py

2.1.4 Clustering of KG using Louvain algorithm:

Go to folder: ./kg_clustering/louvain_clustering/ See readme there to perform clustering step-by-step.

2.2. Create node id to cluster id and cluster id to list of node ids dictionaries after clustering:

python create_lovain_dictionary_cluster_id_to_node_list.py

2.3 Prepare cluster weighted graph and check connectivity:

python build_community_graph_and_check_connectivity.py
python clusterd_knowledge_graph_statistics.py (to get the graph statistics)

2.4 Prepare aspect entity subgraph

create_entity_graph_for_graphsage.py

2.5 For cluster graph embedding generation:

Go to folder: ./weighted_graphsage/

python -m weighted_graphsage.unsupervised_train

2.5 For aspect entity sub-graph embedding generation:

Go to folder: ./graphsage/

python -m graphsage.unsupervised_train --train_prefix path of the entity subgraph --epochs 100 --max_degree maximum node degree of the entity sungraph (varies wrt different dataset)

3. ALSC:

Go to folder /src/alsc/

3.1. AR-TNET:

cd ./ar-tnet/
python main_no_split_lstm_after_cpt_first_and_last_hidden_reshape_concat_GraphSAGE.py \
-ds_name 14semeval_rest \
-connection_type AS \

Change the parameter values based on the dataset name [14semeval_laptop/14semeval_rest/twitter] and connection type [AS/LF]

3.2. AR-BERT:

cd ./ar-bert/bert_tensorflow/
python run_classifier_alsc.py \
--data_dir path of the raw data file (.xml) \
--task_name SemEval2014AtscProcessor \
--vocab_file path of the bert-base-uncased vocab file\
--bert_config_file path of the bert-base-uncased config file\
--output_dir path of the output directory

3.3 AR-BERT-S:

cd ./ar-bert-s/
python run_bert.py

3.4 Incorrect disambiguation detection:

cd ./incorrect_disambiguation_detection/

python sampling_entity_embedding_from_graph.py \
unique_entity_list_file_path \
train_raw_data_file_path \
test_raw_data_file_path \
dataset_name \
10

python prepare_bert_embedding_for_nodes_i_j_k.py \
raw_sentence_file_path raw_aspect_file_path \
raw_character_start_and_end_index_of-aspect_in_sentence_file_path \
bert-base-uncased_vocab_file_path \
bert_token_embedding_file_path \
entity_sampling_file_path \
entity_index_to_list_of_sent_index_file_path \
output_dir_path \
dataset_name 

python semantic_probe.py \
--train_file_path path of the training data \
--output_dir path of the output directory \
--ds_name laptop/rest/twitter \

3.5 AR-BERT-idd:

cd ./ar-bert-idd/
python bert_gs_e_probe_main.py \
--data_dir path of raw data directory(.xml files)
--cluster_embedding_fp path of cluster graph embedding \
--bert_pretrained_embedding_fp node i, j, k embedding file path (train data) \
--bert_pretrained_embedding_fp_test node i, j, k embedding file path (test data) \
--dataset_name laptop \
--bert_config_file bert_config.json file path \
--task_name semeval2014-atsc \
--vocab_file vocab.txt file path \
--output_dir ./laptop/ \
--init_checkpoint bert_model.ckpt file path \
--init_checkpoint_graphsage GraphSAGE_best_model.ckpt subgraph embedding model initial checkpoint path \
--init_checkpoint_probe laptop_semantic_probe_best_model.ckpt initial sematic probe model path \
--do_lower_case --do_train --do_predict --model graphsage_mean --train_prefix path of the entity subgraph  --num_train_epochs 10

4. Multi-modal explanation generation:

Go to folder /src/multi_modal_explanation_generation/

python multi_modal_explanation_extraction.py \
--data_dir path of the raw data directory \
--dataset_name laptop/rest/twitter \
--output_dir path of the output directory \
--vocab_file path of bert-base-uncased vocab.txt file \
--do_lower_case 

Citation

If you use this work, please cite our paper using the following Bibtex tag:

@article{DBLP:journals/corr/abs-2108-11656,
author    = {Sk Mainul Islam and
             Sourangshu Bhattacharya},
title     = {AR-BERT: Aspect-relation enhanced Aspect-level Sentiment Classification with Multi-modal Explanations},
journal   = {CoRR},
volume    = {abs/2108.11656},
year      = {2022},
url       = {https://arxiv.org/abs/2108.11656},
eprinttype = {arXiv},
eprint    = {2108.11656},
timestamp = {Fri, 27 Aug 2021 15:02:29 +0200},
biburl    = {https://dblp.org/rec/journals/corr/abs-2108-11656.bib},
bibsource = {dblp computer science bibliography, https://dblp.org}

}