One of the main fields of AI is Natural Language Processing and its applications in the real world. Here on Amalgam.ai we are building different models to solve some of the problems around the NLP world, and by consequence, trying to make the world a better place.
Can machine learning identify the appropriate reading level of a passage of text, and help inspire learning? Reading is an essential skill for academic success. When students have access to engaging passages offering the right level of challenge, they naturally develop reading skills. If a passage is too easy, the student can get bored, if is too hard the student can quit, so, we need the optimal level of challenge to keep them interested in the reading.
In this competition we were asked to build machine learning algorithms to rate the complexity of reading passages for grade 3-12 classroom use, thus, helping administrators, teachers, and students in their day to day in the classroom.
The results of this competition will be scored on the root mean squared error. RMSE is defined as:
For this competition the organization provided excerpts from several time periods and a wide range of reading ease scores.
The data has these columns:
- id: unique ID for excerpt;
- url_legal: URL of source;
- license: license of source material;
- excerpt: text to predict reading ease of;
- target: reading ease;
- standard_error: measure of spread of scores among multiple raters for each excerpt.
A curious case in this dataset is that the rating of complexity was made by different people. So, in some cases the standard deviation of rating was very big, showing that even for humans it is hard to reach a consensus about the complexity of a passage.
As shown in the description of the competition we are dealing with a regression problem. To tackle this, our approach is basically divided in two steps:
1. Further pre-train a Transformer model on competition dataset;
2. Fine tune a Transformer to the regression task of the competition.
Our solution can be better visualized on the diagram below:
Part 1 – Further Pre-Train Transformer
In this competition, we decided to use a RoBERTa Large and a RoBERTa Base, due to a better performance on the task in hand..
To do the pre-train we use the HuggingFace transformers library in association with our open-source libraries Aurum and Stripping. Together we were able to track every experiment that we made. We pre-train our model for 5, 8 and 10 epochs. The best results for RoBERTa Large were achieved with 8 epochs and for RoBERTa Base the best results were achieved with 5 epochs.
Part 2 – Fine tuning for the task:
Once our transformer models were ready, the second part consisted in the fine tuning for the specific task of rate the complexity of a text passage. To do this we trained a neural network with an attention head and a linear layer predicting the values of target.
To this part, we also used Aurum and Stripping to experiment control and PyTorch Lightning to create the models and for distributed training.
Part 2.1 – The Good Ol’ Classic ML:
One of the approaches, not for fine tuning but for prediction, we trained SVM’s regressors using as input the output of the transformers models. We did that using ScikitLearn library and 5 fold cross validation.
Ensemble and Predictions:
To generate the final submission for the competition we made a weighted ensemble using 4 different models with different settings of pre-trained RoBERTa models and fine tuned regression models. We also used SVM’s as regressors either.