13.3 In summary

Take care when reporting the results of lexicon-based approaches such as NLTK’s Vader and TextBlob. For instance, a non-technical stakeholder who sees a polarity or compound score of -0.1 and -0.6 will naturally ask questions such as, “why is the second review more negative than the first one? What happened? How are these scores assigned?” Rather than diving deep into a technical explanation about the inner workings of Vader or TextBlob, we suggest grouping the sentiment analysis results into three broad categories (positive, neutral, negative) and explaining the ‘so what’ for your stakeholders. Focusing on the implications of your findings for the business will ensure the audience takes away your key points.

It is possible to train your own sentiment analysis model using machine learning methods like Naive Bayes. However, this is a delicate process that requires time to refine. Like any machine learning method, the model will need to be regularly updated with new data as the business evolves.