[EMNLP2021] Aspect-Controllable Opinion Summarization
This PyTorch code was used in the experiments of the research paper
Reinald Kim Amplayo, Stefanos Angelidis, and Mirella Lapata. Aspect-Controllable Opinion Summarization. EMNLP, 2021.
The code is cleaned post-acceptance and may run into some errors. Although I did some quick check and saw the code ran fine, please create an issue if you encounter errors and I will try to fix them as soon as possible.
We used two different datasets from two different papers: Space (Angelidis et al., 2021) and Oposum (Angelidis and Lapata, 2018). We extended the latter to create Oposum+. For convenience we provide the train/dev/test datasets here which are preprocessed accordingly and saved in separate json or jsonl files. The training corpora are saved with filename train.jsonl, while the dev/test sets are named dev.json/test.json accordingly.
A single line (in jsonl files) or a single example (in json files) is formatted as follows:
{
"..."
"reviews": [
{
"sentences": [
"this is the first sentence.",
"this is the second one.",
"..."
],
"..."
},
"..."
],
"summary": [
"this is the first summary.",
"this is the second summary.",
"this is the last summary."
],
"..."
}In the example above, reviews is a list of reviews represented as a dictionary with sentences as keys, and summary is a list of reference summaries. There can be other information included in the files but are not used in the code (e.g., review and product identifiers, review rating, etc.). When using the datasets, please also cite the corresponding papers (see below).
AceSum has three steps: induce aspect controllers, create synthetic dataset, and fine-tune T5-small. Below, we show a step-by-step proceedure on how to replicate the results in the paper.
To do this, we need to train an aspect controller induction model.
We first create a training dataset with weak supervision from seed words (in the seeds directory). This can be done by running:
python src/create_mil_data.py [dataset]
where [dataset] can be space or oposum/[domain], where [domain] can be any of the six domains of the Oposum dataset. This creates a train.mil.jsonl file in the data/[dataset] directory. To skip this step, download the corresponding files in the link above.
We then train the aspect controller induction model by running:
python src/train_mil.py -mode=train -dataset=[dataset] -num_aspects=[num_aspects] -model_name=mil
where [num_aspects] is 6 for Space and 3 for Oposum.
Given the aspect controller model mil.model, we (1) create the synthetic training dataset and also (2) preprocess the dev and test splits. These can all be done using the src/create_sum_data.py script/
Run the following to create the synthetic training data:
python src/create_sum_data.py -mode=train -dataset=[dataset] -num_aspects=[num_aspects] -load_model=[mil_model_dir]
where [mil_model_dir] is the location of mil.model.
Run the following to create the dev/test general and aspect-specific summarization datasets:
python src/create_sum_data.py -mode=eval-[type] -dataset=[dataset] -num_aspects=[num_aspects] -load_model=[mil_model_dir]
where [type] is either general or aspect.
Additionally for Oposum, we need to combine the domain-specific datasets into one. Use the combine_oposum.py script to do this:
python src/combine_oposum.py [dataset].
where [dataset] can be one of the following: train.sum.jsonl, dev.sum.general.jsonl, dev.sum.aspect.jsonl, test.sum.general.jsonl, test.sum.aspect.jsonl.
Again, all of these can be skipped by downloading the files in the link above.
This is done by simply running src/train_sum.py:
python src/train_sum.py -mode=train -dataset=[space or oposum] -num_aspects=[6 or 18] -model_name=sum
Note that the number of aspects of oposum is now 3*6=18 since we have combined the datasets in the previous step. This returns a sum.model file after training.
Skip this step by downloading the files in the link above.
Generating the summaries can be done by running:
python src/train_sum.py -mode=eval-[type] -dataset=[space or oposum] -num_aspects=[6 or 18] -load_model=[sum_model_dir]
where [type] is either general or aspect and [sum_model_dir] is the location of sum.model.
This will create an output file acesum.[general or aspect].out in the corresponding directory.
If you want to skip everything and just want AceSum, this repo also includes an output directory containing AceSum general and aspect-specific summaries. The ith line in the file is the summary of the ith test example.
To cite the paper/code/data splits, please use this BibTeX:
@inproceedings{amplayo-etal-2021-aspect,
title = "Aspect-Controllable Opinion Summarization",
author = "Amplayo, Reinald Kim and
Angelidis, Stefanos and
Lapata, Mirella",
booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
month = nov,
year = "2021",
address = "Online and Punta Cana, Dominican Republic",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.emnlp-main.528",
doi = "10.18653/v1/2021.emnlp-main.528",
pages = "6578--6593",
}
@inproceedings{angelidis-lapata-2018-summarizing,
title = "Summarizing Opinions: Aspect Extraction Meets Sentiment Prediction and They Are Both Weakly Supervised",
author = "Angelidis, Stefanos and
Lapata, Mirella",
booktitle = "Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing",
month = oct # "-" # nov,
year = "2018",
address = "Brussels, Belgium",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/D18-1403",
doi = "10.18653/v1/D18-1403",
pages = "3675--3686",
}
@article{angelidis-etal-2021-extractive,
title = "Extractive Opinion Summarization in Quantized Transformer Spaces",
author = "Angelidis, Stefanos and
Amplayo, Reinald Kim and
Suhara, Yoshihiko and
Wang, Xiaolan and
Lapata, Mirella",
journal = "Transactions of the Association for Computational Linguistics",
volume = "9",
year = "2021",
address = "Cambridge, MA",
publisher = "MIT Press",
url = "https://aclanthology.org/2021.tacl-1.17",
doi = "10.1162/tacl_a_00366",
pages = "277--293",
}
If there are any questions, please send me an email: reinald.kim at ed dot ac dot uk