Computing a baseline

[AndreiBarsan]

According to the CIL project description, we must compare our solution with a baseline.

You compare your novel algorithm to at least two baseline algorithms. For the baselines, you can use the implementations you developed as part of the programming assignments.

So, I suppose we can just use something like logistic regression over the averaged/concatenated embeddings of the (fixed length) sentences?

I think in the end we could have a table in the paper with the following rows:

• Baseline: LogReg + Avg/Concat embeddings
• CNN without preprocessing
• CNN with preprocessing
• Bigger (256/512 fileters @ 3,4,5,7 filters) CNN?
• Shallow LSTM (1 layer, gradient clipping for stability, etc.)
• Deep LSTM (2/3 layer, large-ish hidden representation, etc.)

What do you think?

[AndreiBarsan]

I'll compute the simple baseline.

[AndreiBarsan]

Computed simple baseline with averaging in PR #23. Will also work on tf-idf classification.

[AndreiBarsan]

PR #25 also contains a tf-idf baseline.

[bernhard2202]

Now we have 2 baselines computed SGD (averaging, concatenation) and tf-idf - therefore closed.