Human-Level Interpretable Learning for Aspect-Based Sentiment Analysis

Authors: Rohan K Yadav, Lei Jiao, Ole-Christoffer Granmo, Morten Goodwin14203-14212

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

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate interpretability and accuracy, we conducted experiments on two widely used ABSA datasets from Sem Eval 2014: Restaurant 14 and Laptop 14. The experiments show how each relevant feature takes part in conjunctive clauses that contain the context information for the corresponding aspect word, demonstrating human-level interpretability. At the same time, the obtained accuracy is on par with existing neural network models, reaching 78.02% on Restaurant 14 and 73.51% on Laptop 14.
Researcher Affiliation Academia Rohan K Yadav, Lei Jiao, Ole-Christoffer Granmo, Morten Goodwin Centre for Artificial Intelligence Research, University of Agder, 4879, Grimstad, Norway rohan.k.yadav@uia.no, lei.jiao@uia.no, ole.granmo@uia.no, morten.goodwin@uia.no
Pseudocode Yes Algorithm 1 Training Process of TM based ABSA; Algorithm 2 Testing Process of TM based ABSA
Open Source Code Yes The code and the datasets are available online1. 1https://github.com/rohanky/tm_absa
Open Datasets Yes The datasets are obtained from Sem Eval-2014 Task 4. The task has two domain-specific datasets, namely, Restaurant 14 (res14) and Laptop 14 (lap14). These datasets are provided with training and testing data. The statistics of the two datasets is shown in Table 3. The code and the datasets are available online1. 1https://github.com/rohanky/tm_absa
Dataset Splits No The paper provides train and test split statistics in Table 3, but does not explicitly mention or provide details for a separate validation split. It states 'best reproducible results by running the ABSA TM for 100 epochs' but does not specify how validation was handled.
Hardware Specification No The paper does not provide any specific details about the hardware used to run the experiments, such as GPU or CPU models.
Software Dependencies No The paper does not provide specific version numbers for any software dependencies or libraries used in the experiments.
Experiment Setup Yes As we have used the integer weighted TM (Abeyrathna, Granmo, and Goodwin 2020), the parameters available are the number of clauses, the threshold T, and the specificity s, which are configured as 700, 90 100, and 15 respectively for both datasets. For pre-processing of text, we substitute the short form to its full form, such as isn t to is not . Additionally, we stem the words to reduce the vocabulary size created due to spelling mistakes and variants of words2. The remaining pre-processing procedure has already been explained before. We train the TM model on both the datasets for 100 epochs each.