Week #5: So Much To Choose From

1. Recommendation engine

Bearing in mind that one of the main objectives of the digital sales assistant is product recommendation, the team worked on a module that not only offers excellent performance on this task, but is also easily integrated with the rest of the system.

At its core, the Recommendation Engine resorts to ElasticSearch, one of the most popular document-based databases, commonly used in the implementation of search engines, due to its maturity, performance and ease of use.

When provided with the slots tagged by the GPT-3 model (e.g. clothing category, color, size), the Recommendation Engine is able to query the ElasticSearch database and retrieve a list of recommended items to be presented to the user, based on the preferences and needs shown during the conversation.

2. Sentiment analysis

Last week, the team worked on the training of an emotion analysis model, which did not meet the standards set by the team for this project, which ultimately led to the decision of training a sentiment analysis model instead.

Thankfully, although the efforts to find a dataset on the domain of retail ultimately ended in nothing, the team found a mature IMDB movie review dataset which at least allowed the team to study the impact of neural network architectures, layer dimension and the application of NLP techniques (e.g. Word2Vec, text pre-processing) on the performance of the neural networks.

By the end of this model's training, a neural network with a bi-directional GRU layer at its core was obtained with an accuracy value of around 83%, which left the team more satisfied and confident that it would not drag the solution down.

3. Scientific article

Having had the intermediate delivery last week, the team received some feedback from the teachers, which allowed the already written chapters to be improved upon.

Furthermore, this week the team moved on with the Proposal chapter of the article, where an overview of the solution's architecture and usage is presented, as well as the benefits and limitations identified by the team when compared to other existing solutions.

4. Week retrospective

All in all, the team found this week to be rather productive, with the goals set for this week being accomplished and the results on the sentiment analysis model meeting the team's expectations.

Next week, the team expects to work on the chat bot's back-end and dialogue manager, which will integrate the models trained until now as well as the recommendation engine, resulting in the system architecture presented in week #3.