It's a Jungle Out There: Filtering High Quality Reviews From Amazon Using Attention-Based Modeling
Abstract
Today, a large amount of misinformation proliferates on the internet, including online review websites. Because of the prevalence of fake or unknowledgeable reviewers, there is a need for new models to order and filter reviews. Hence, this thesis explores the potential for new machine learning methods trained over carefully curated “expert” reviews to automatically uncover high-quality reviews from large review collections. Concretely, we leverage machine learning and deep learning models to capture the semantic, grammatical, and argumentative structure of a high-quality review such that we are able to filter out fake or unknowledgeable reviews. We provide evidence showing that attention-based modeling is able to capture this semantic and argumentative structure such that we can produce high performance in delineating high-quality reviews from low-quality reviews. Specifically, we leverage different training methods coupled with multiple machine learning and deep learning models to identify the highest performing training method and model. We find that cross-domain training coupled with attention-based modeling on sentence-based high-quality review filtering produces the highest performance, outperforming other models and training methods by just under five percent. Additionally, we find that this model shows strong evidence of generalizing our task of high-quality review filtering outside of the initial domains on which the model was trained. Thus, we are able to show proof-of-concept that automated high-quality review filtering can now be captured with advanced modeling techniques.
Subject
BERTattention
modeling
machine
learning
deep
transformer
review
reviews
crowd
sentiment
expert
quality
Citation
Innis, Jonathan Edward (2020). It's a Jungle Out There: Filtering High Quality Reviews From Amazon Using Attention-Based Modeling. Undergraduate Research Scholars Program. Available electronically from https : / /hdl .handle .net /1969 .1 /188417.