)
The Social Media and CRM Marketing Intelligence division of the main Italian media broadcaster had expressed the need to have an automated software system capable of understanding the 'sentiment' (the polarity) expressed by the public through 'interactions' in Italian (for example tweets, posts, etc.) on the main social channels, regardless of the specific channel or program to which they referred.
The main objectives of the project were:
Move from a quantitative analysis to a qualitative analysis of social interactions attributable to television and radio programs
Increase audience satisfaction
Build a scalable solution that can easily integrate with new flows of interactions coming from social channels
)
)
Language and sentiment are highly time-dependent phenomena, so all resources and linguistic models developed and released must be subject to continuous monitoring and re-training, in order to ensure the maintenance of the system's predictive abilities.
To achieve this objective and allow the necessary adjustments to the analytical models to be made, a “continuous learning” functionality has been developed, which allows users, whenever they find the need, to analyze and modify the answers provided by the SMSA model in order to obtain a better model.
Target Reply developed the SMSA system through the composition of PaaS services provided by the Microsoft Azure cloud, creating both a general-purpose Data Lake and a Sentiment Analysis solution tailored to the customer's needs. For the SMSA task, Target Reply Rome decided to use Natural Language Processing (NLP), Machine Learning (ML) and Sentiment Analysis (SA) engines developed by the AI research center of the University of Tor Vergata. Together with the Tor Vergata team of researchers and experts, Target Reply Roma has built a specially trained, optimized and specialized analytical solution for the domain and specific customer needs. The SA engine consists of a battery of classifiers specialized in identifying the 4 polarity classes required by the customer + 1 class of non-polarity/objectivity and assigning each of them a measure of confidence (also produced by an additional specially trained ML model). The optimization and fine-tuning of the SMSA solution were carried out starting from a set of (so-called) Golden Tweets, whose polarity was established by a pool of evaluators selected by the customer, enriched by a additional set of Silver Tweets, whose polarity has instead been determined through the application of Distant Supervision and Active Learning techniques.