Unsupervised Opinion Summarization with Content Planning

Authors: Reinald Kim Amplayo, Stefanos Angelidis, Mirella Lapata12489-12497

AAAI 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results on three domains show that our approach outperforms competitive models in generating informative, coherent, and fluent summaries that capture opinion consensus.
Researcher Affiliation Academia Reinald Kim Amplayo, Stefanos Angelidis, Mirella Lapata Institute for Language, Cognition and Computation School of Informatics, University of Edinburgh reinald.kim@ed.ac.uk, s.angelidis@ed.ac.uk, mlap@inf.ed.ac.uk
Pseudocode No The paper describes the model architecture and training process but does not include any explicit pseudocode or algorithm blocks.
Open Source Code Yes Our code can be downloaded from https://github.com/ rktamplayo/Plan Sum.
Open Datasets Yes We performed experiments on three opinion summarization benchmarks. These include the Rotten Tomatoes dataset1 (RT; Wang and Ling 2016) which contains a large set of reviews for various movies written by critics... Our second dataset is Yelp2 (Chu and Liu 2019)... Finally, the Amazon dataset3 (Braˇzinskas, Lapata, and Titov 2019)...
Dataset Splits Yes After applying these constraints we obtained 100k (Yelp), 25k (RT), and 90k (Amazon) review-summary pairs. Statistics of these datasets are reported in Table 1.
Hardware Specification Yes Our model was trained on a single Ge Force GTX 1080Ti GPU and is implemented using Py Torch.
Software Dependencies No Our model was trained on a single Ge Force GTX 1080Ti GPU and is implemented using Py Torch. We used the subword tokenizer of BERT (Devlin et al. 2019), which has a 30k token vocabulary trained using Word Piece (Wu et al. 2016).
Experiment Setup Yes Across models, we set all hidden dimensions to 256, the dropout rate to 0.1, and batch size to 16. We used the Adam optimizer (Kingma and Ba 2015) with a learning rate of 3e 4, l2 constraint of 3, and warmup of 8,000 steps.