Sentiment Analysis Microservice

Our Sentiment Analysis Microservice enables enterprises to monitor any type of written communication: social media posts and messages, emails, incoming customer service requests, surveys, or any other type of “unstructured text”. Unlock key benefits from this data to better understand your customers' behavior.

Key Features

Dataset Adaptability

This highly adaptable solution can be “trained” using many different datasets varied in source, type, length, and language. As a fully customizable solution, there is no information you won't have access to.

Dynamic Re-training

Because new free-text data is classified based on sentiment, it is possible to re-train the underlying classification model so that its accuracy improves over time – even when a new communication channel is added to the analysis. This all happens simultaneously with the regular sentiment classification process with no downtime or scheduled model re-training.

Simple & Quick Implementation and Deployment

Our Sentiment Analysis Microservice operates as a stand-alone module that can be easily integrated using a REST API, message broker, or other commonly used interface. It can be deployed in any hosting model (on-premises/off-premises/cloud). The components used are light-weight, open-source, and globally supported. For the development and testing purposes, the module is available as a Docker container that can be launched in seconds. Depending on the dataset used for model training, the entire implementation of the sentiment microservice can be finalized in just a few days.

The Science Behind the Solution

WordNetLemmatizer and custom tokenizer are used for data pre-processing (text tokenization and lemmatization). For the actual sentiment analysis, there are two different approaches, chosen automatically depending on the accuracy results:

Hybrid ensemble classification model
Sentiment microservice uses three independent classifiers combined. Plus, a voting process determines the final sentiment detected. Techniques used in this approach include Bayesian classifier, Support Vector Machine and linear regression. The voting process is handled by the EnsembleVoteClassifier.
Neural net model
In this model, free-text data is first vectorized using Word2Vec model and then passed through a trained LSTM (long short-term memory) artificial neural network with an output of the probability of belonging to the given sentiment class.