Below are a few questions that people have asked (or might ask).
How can I get started?
If you want to try something more interesting, there are a number of existing methods for language-to-code generation that are more sophisticated:
- Syntactic Neural Model (Yin and Neubig 2017): Paper Code
- Abstract Syntax Networks (Rabinovich et al. 2017): Paper
- Neural Coarse-to-fine Parsing (Dong and Lapata 2018): Paper Code
How are Outputs Evaluated?
When evaluating the output of a code generation program, the most important thing is whether the output accurately reflects the semantics of the true program. One way to do this evaluation would be to create unit tests or fuzz tests that check the correctness of the output of each program. However, given that CoNaLa covers a large variety of programs that span software libraries and environments, creating these tests was not feasible given our time and resource constraints.
Thus, the official CoNaLa evaluation script uses BLEU score, a measure of similarity based on n-gram overlap. This has been used in previous work on code generation, and is a proxy measure that tells how close the generated programs match a gold-standard program.
What version of Python is CoNaLa?
Python 3. All code found online was converted to Python3 using the
2to3 utility, specifically this script.