Some details
The Challenge
Analysing millions of public comments on social media is important for marketing campaigns to help companies to understand better what exactly clients like or dislike in the products. Currently, this work is done or by hired experts who have some elementary domain and language understanding or is outsourced to third parties. The first option is getting expensive on the scale since it
requires more and more people that also have to be trained. The second option is risky since such gathered data has R&,D value and can',t be shared very easily. Using natural language processing (NLP) algorithms leveraged by the latest development in AI, especially deep learning and state-of-the-art results in natural language understanding and generation (NLU, NLG) seems like a great fit for the given mission.The Solution
Typically, the classification of a piece of text into different emotional or sentiment groups is called sentiment analysis in the literature and it is well solved with the latest AI developments. Although, the classical sentiment analysis doesn’t fit here since it just tells the average tonality of sentences instead of a detailed explanation of people’s preferences about entities and words that appear in the text. For instance, how would you classify such a sentence ",Thanks, the coffee was good, but I had to wait for too long for it",? Positive, Negative, maybe Neutral class? In fact, you would rather say that the coffee thing is positive, but service is negative. This slight change of concept arises into a totally new task class called aspect-based sentiment analysis (ALSA).
The solution we have built was a completely novel NLP system for fine-grained aspect-based sentiment analysis, topic modelling, and text categorization for different languages and data sources including automated machine learning pipeline. The core of it was the beforementioned BERT deep learning model that was trained to detect different entities, their sentiment polarity, and corresponding trigger words.
The Result
The final value of the project was in cutting costs and optimising time and process for social media analysis with optimising routine work for more than 30% of the staff with a fully automated AI pipeline.
We have observed not just the replacement of the manual work, but also an overwhelming speedup of the work: a dataset that could have been processed for weeks by the human team is processed by the algorithms in several minutes.
