[TIL] Keep Working on RNN Model Training

1. Validation of the RNN Model with a Normalized Dataset for Binary Classification

You tested the RNN model using the Naver Movie Review Dataset (NMRS) for sentiment analysis with binary labels. The accuracy obtained was 0.86, indicating that the model performed well in determining positive and negative sentiments. Therefore, there doesn't seem to be any issue with the RNN model itself. However, since the model was not trained specifically on Korean language, you concluded that a larger dataset, specifically for multiclass classification (num_classes = 5), is required for the CodeBlue project.

2. Issue with Sentence Understanding When the Sentence Structure Differs

You encountered an issue where the model couldn't properly understand sentences that included symptoms but had different grammatical structures. To address this, you applied preprocessing by removing stopwords from the sentences. This helped to some extent in resolving the issue, although the accuracy slightly decreased. The model now correctly identified the severity level even with slight grammatical changes.

3. Cumulative Training on the Same Dataset and its Impact on Accuracy

You conducted an experiment to determine if the accuracy improves when performing cumulative training on the same dataset. However, the results showed that the accuracy decreased from 0.5799 to 0.4000 when training on the same dataset repeatedly. Therefore, it can be concluded that cumulative training does not necessarily lead to increased accuracy.

4. Training with a Narrowed Dataset of Approximately 130 Normalized Examples

You trained the model with a narrowed dataset of around 130 normalized examples, which had a more specific topic range. The test accuracy was significantly high. However, the model's ability to predict severity levels accurately was not ideal due to the limited amount of data available.

In conclusion, you determined that more normalized datasets are needed to improve the model's performance.