Background
As long as there has been a written language, there has also been tones communicated through that language that make the reader feel what the writer meant, rather than just reading the words without meaning.
Neural Networks can be used to interpret the sentiment of a message through the types of words that are used in the communication. Neural Networks are a machine learning algorithm that uses a process called backproppogation to train layers of algorithms and fine tune the connections that exist. Each layer consists of functions called “neurons” that allow the data to be processed at each level.
By using these neural networks, and fine tuning the connections between words and sentiment, we can make a good guess at what each inputted message may mean, rather than say.
Training
In the training for this model, a prebuilt dataset from IMDB that contained reviews for movies was used, with some of the data specifically meant for training containing predefined “positive” or “negative” values. These values are important because the model that is being trained has to know whether it is getting the guesses that it is making right or wrong. The training data is used to create different “weights” for values, which create the different pathways through the layers of the model.
As the training progresses over multiple iterations, the model becomes more competent at its job, where we can see the accuracy rise for every iteration. This trend will eventually turn into a relatively straight line as the model reaches its informational capacity, and training reaches a critical time/benefit point. This model has been edited and fine tuned to best serve the needs of the task that has been set before. To do this, layers have been added, density altered, and connections added as well. All of these efforts together created a model with ~93% accuracy on this task.
Results
The result of this process is a trained model that can be used in various applications. The model that is created will be able to take input as a string of text, which it will then vectorize and process through the model, and return the corresponding value for the model.
(Higher # = More positive sentiment)�(Lower # = More negative sentiment)
This specific model could be used by companies that collect reviews to not only allow users to rate by star, but also by the sentiment that is conveyed in their review. The applications for this type of technology are widespread, and a large portion are concerned directly with consumer interaction.
Sentiment Analysis Through Neural Networks
Team member: Zeke Hudson | Faculty advisor: Changqing Luo
2022