AUTOMATIC AND ADAPTIVE EXTRACTION OF ACTION KNOWLEDGE FROM PRODUCT REVIEWS

Abstract

Recommending products based on user experience and feedback was proved to be an e↵ective marketing strategy. Product reviews are a promising source of knowl- edge in the product recommendation process. Much of the research done was mining textual data from reviews to extract sentiments, satisfaction level, and ratings. Lit- tle work was invested for extracting action knowledge and other semantics that can help with the recommendation process. Annotating action knowledge from customer reviews was done manually using the contribution of human annotators to read, an- notate, and extract the tags and labels stated or deduced from content. How can we automate the extraction of action knowledge entities from unstructured product reviews, whether stated or predicted from context, in order to replace manual an- notation tools? Moreover, How to automatically integrate product reviews in the recommendation process through representative knowledge graphs constructed from the predicted entities? How can human intervention help in the adaptation of ML models and to what incremental level of updating can the framework reach from the human-in-the-loop mechanism? This work proposes a framework to automate ac- tion knowledge extraction from customer reviews through machine learning models that will be incrementally improved through ‘Human-in-the-loop‘ technique. This framework semantically annotates the actions expressed in product reviews, captures other related entities, links these entities to form a knowledge graph that serves the development of action-aware recommendation apps. The system is adapted and incrementally updated throughout continually retraining the models with approved correct data. To validate the solution proposed, an experimental evaluation protocol is applied to train the models with updated sets of approved annotations. The ex- periment revealed an improvement in the performance of the predictive models. In addition, the datasets collected by human annotators and used for model retraining were improved in terms of reducing the gap between classes of the models. The contribution of our work is to introduce a full consolidated automatic framework, joining multiple components, to construct an end-to-end prototype that generates the input of action-aware recommendation app. This automatic framework is pro- posed to automate the action knowledge extraction from product reviews and the construction of knowledge graphs used in recommendation systems, in an adaptive manner.