Homework 1A
is about data preprocessing that is used inHomework 1B
to train NLP models on the Italian Language.
- This report walks through two tasks: task 0 (EmotivITA) and subtask A of task 1 (HODI).
- It provides a brief explanation of the two tasks, their input and output formats, how the data are formated, the prompts and the motivation for those specific prompt choices.
- It also explains how to run the scripts succesfully.
- Read the requirements files here for what to expect from this part: EmotivITA and HODI.
- This part walks through the key details of the implementation of multiple models on the EmotivITA dataset for sentence classification task for the Italian language.
- This work aims to establish
baseline models
and then implement more robust models from theRNN family
that can outperform them. - I experimented with 3
statistics-based
baselines, 2Logistics Regressions
(one with embedding layer), and some combinations ofBiLSTM
andBiGRU
models. - I was able to achieve varying but good results as elaborated in the Results section of its report.
- I also experimented with
triple head
-- likeSiamese networks
-- where, instead of having a single output layer, it has a triple. This is the same as having 3 different networks. Note that 3 different networks require more compute but more stable and faster to converge. Thus, I implemeneted the later for network training stability. - For further details on this, see the report here