LADy: a system for Latent Aspect DetectionÂ

Submitter and Co-author information

Christine Wong, Faculty of Science

Standing

Undergraduate

Type of Proposal

Oral Research Presentation

Challenges Theme

Open Challenge

Faculty Sponsor

Hossein Fani

Proposal

Analyzing publicly expressed reviews in social media platforms enhances customer satisfaction and retention, and identifies business opportunities. An aspect is what a review targets to convey a sentiment whose detection is crucial in review analysis. Thus, aspect detection has long gained much attention among researchers in Natural Language Processing and Social Network Analysis. Existing methods analyze labeled datasets with explicitly-mentioned aspects in reviews and forego the latent (implicit) occurrences of aspects. Indeed, online reviews are short and informal, hence relying on social background knowledge and context rather than mentioning the aspects explicitly. In this proposal, we have contributed LADy, an open-source library for online unsolicited review analysis that focuses on finding latent (implicit) aspects in customers’ reviews. We employ latent Dirichlet allocation and biterm topic model to learn the latent aspects’ distributions for generating the reviews. We assume that a customer undergoes a two-stage hypothetical generative process to write a review about an aspect of a product or service: 1) deciding on an aspect amongst the set of aspects available, and 2) selecting the words from the set of words in a language that is more interrelated to the aspect. Leveraging an object-oriented structure, LADy readily accommodates the addition of new topic modeling methods and training datasets. Our early experimental results on benchmark datasets show that our proposed models improve analysis accuracy when the aspects are latent with no surface form in reviews. LADy also allows business owners to optimize sustainable business production and growth based on Sustainable Development Goals.

Grand Challenges

Viable, Healthy and Safe Communities

Share

COinS
 

LADy: a system for Latent Aspect DetectionÂ

Analyzing publicly expressed reviews in social media platforms enhances customer satisfaction and retention, and identifies business opportunities. An aspect is what a review targets to convey a sentiment whose detection is crucial in review analysis. Thus, aspect detection has long gained much attention among researchers in Natural Language Processing and Social Network Analysis. Existing methods analyze labeled datasets with explicitly-mentioned aspects in reviews and forego the latent (implicit) occurrences of aspects. Indeed, online reviews are short and informal, hence relying on social background knowledge and context rather than mentioning the aspects explicitly. In this proposal, we have contributed LADy, an open-source library for online unsolicited review analysis that focuses on finding latent (implicit) aspects in customers’ reviews. We employ latent Dirichlet allocation and biterm topic model to learn the latent aspects’ distributions for generating the reviews. We assume that a customer undergoes a two-stage hypothetical generative process to write a review about an aspect of a product or service: 1) deciding on an aspect amongst the set of aspects available, and 2) selecting the words from the set of words in a language that is more interrelated to the aspect. Leveraging an object-oriented structure, LADy readily accommodates the addition of new topic modeling methods and training datasets. Our early experimental results on benchmark datasets show that our proposed models improve analysis accuracy when the aspects are latent with no surface form in reviews. LADy also allows business owners to optimize sustainable business production and growth based on Sustainable Development Goals.