liangxiao / ABSA-BERT-pair

Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence (NAACL 2019)

ABSA as a Sentence Pair Classification Task

Codes and corpora for paper "Utilizing BERT for Aspect-Based Sentiment Analysis via Constructing Auxiliary Sentence" (NAACL 2019)

Requirement

• pytorch: 1.0.0
• python: 3.7.1
• tensorflow: 1.13.1 (only needed for converting BERT-tensorflow-model to pytorch-model)
• numpy: 1.15.4
• nltk
• sklearn

Step 1: prepare datasets

SentiHood

Since the link given in the dataset released paper has failed, we use the dataset mirror listed in NLP-progress and fix some mistakes (there are duplicate aspect data in several sentences). See directory: data/sentihood/.

Run following commands to prepare datasets for tasks:

cd generate/
bash make.sh sentihood

SemEval 2014

Train Data is available in SemEval-2014 ABSA Restaurant Reviews - Train Data and Gold Test Data is available in SemEval-2014 ABSA Test Data - Gold Annotations. See directory: data/semeval2014/.

Run following commands to prepare datasets for tasks:

cd generate/
bash make.sh semeval

Step 2: prepare BERT-pytorch-model

Download BERT-Base (Google's pre-trained models) and then convert a tensorflow checkpoint to a pytorch model.

For example:

python convert_tf_checkpoint_to_pytorch.py \
--tf_checkpoint_path uncased_L-12_H-768_A-12/bert_model.ckpt \
--bert_config_file uncased_L-12_H-768_A-12/bert_config.json \
--pytorch_dump_path uncased_L-12_H-768_A-12/pytorch_model.bin

Step 3: train

For example, BERT-pair-NLI_M task on SentiHood dataset:

CUDA_VISIBLE_DEVICES=0,1,2,3 python run_classifier_TABSA.py \
--task_name sentihood_NLI_M \
--data_dir data/sentihood/bert-pair/ \
--vocab_file uncased_L-12_H-768_A-12/vocab.txt \
--bert_config_file uncased_L-12_H-768_A-12/bert_config.json \
--init_checkpoint uncased_L-12_H-768_A-12/pytorch_model.bin \
--eval_test \
--do_lower_case \
--max_seq_length 512 \
--train_batch_size 24 \
--learning_rate 2e-5 \
--num_train_epochs 6.0 \
--output_dir results/sentihood/NLI_M \
--seed 42

Note:

• For SentiHood, --task_name must be chosen in sentihood_NLI_M, sentihood_QA_M, sentihood_NLI_B, sentihood_QA_B and sentihood_single. And for sentihood_single task, 8 different tasks (use datasets generated in step 1, see directory data/sentihood/bert-single) should be trained separately and then evaluated together.
• For SemEval-2014, --task_name must be chosen in semeval_NLI_M, semeval_QA_M, semeval_NLI_B, semeval_QA_B and semeval_single. And for semeval_single task, 5 different tasks (use datasets generated in step 1, see directory : data/semeval2014/bert-single) should be trained separately and then evaluated together.

Step 4: evaluation

Evaluate the results on test set (calculate Acc, F1, etc.).

For example, BERT-pair-NLI_M task on SentiHood dataset:

python evaluation.py --task_name sentihood_NLI_M --pred_data_dir results/sentihood/NLI_M/test_ep_4.txt

Note:

• As mentioned in step 3, for sentihood_single task, 8 different tasks should be trained separately and then evaluated together. --pred_data_dir should be a directory that contains 8 files named as follows: loc1_general.txt, loc1_price.txt, loc1_safety.txt, loc1_transit.txt, loc2_general.txt, loc2_price.txt, loc2_safety.txt and loc2_transit.txt
• As mentioned in step 3, for semeval_single task, 5 different tasks should be trained separately and then evaluated together. --pred_data_dir should be a directory that contains 5 files named as follows: price.txt, anecdotes.txt, food.txt, ambience.txt and service.txt
• For the rest 8 tasks, --pred_data_dir should be a file just like that in the example.