Social media have been established as platforms from which one may extract public opinion about events, products, and services. Twitter is a principal generator for a huge amount of unstructured information, and corporations are beginning to comprehend the power of Twitter and understand the views of their customers by analyzing the related posted data. Examples of potential applications of such analysis include improving the accuracy of sales forecasting, supporting significant marketing decisions, improving existing products etc.

The focus of this Thesis is to investigate appropriate neural network architectures and refine their parameters to maximize the effectiveness of sentiment identification in sentences posted online on the Twitter social network. For this, we have developed and tested various neural network models and used them to extract sentiment from a set of 8,600 tweets that have been posted online between June 01st and Sept 30th, 2021. This dataset contains tweets -identified and collected using the Apple hashtag: #apple- posted by consumers expressing opinions or experiences regarding the company’s products or services. These tweets have been processed into a form that is appropriate for text analysis and have been manually labelled to either contain “positive” or “negative” sentiment. The labeled dataset was used as the input to the neural network for the training process. 

Having selected an appropriate, simple, NN architecture, we refined the parameters of the training process and of the model itself to maximize the effectiveness of the network and correctly predict positive or negative sentiment.  Our method includes systematic training of model variations under different training conditions.  The training and model structure effectiveness was analyzed statistically to arrive at an optimized model.

The proposed experimentation and analysis method may be applied to fine tune the training of networks used in similar applications, such as movie reviews, product assessments etc.