For the course Machine Learning for Programming at Cambridge, I proposed a neural re-ranking model to improve results of a robust baseline for semantic code search (SCS), a retrieval task where a code snippet must be found given a particular natural language query. Specifically, I proposed an architecture where ElasticSearch rankings, which have been shown to be a strong baseline for SCS (Sachdev et al., 2018), are improved using a neural system that is trained to re-rank them. Results were promising, though further work is needed to determine if this holds using more established benchmarks (e.g., which use user queries instead of comments). Nonetheless, it was a nice opportunity to explore graph neural networks (Li et al., 2016; Fernandes et al., 2019) to obtain code-snippet-level representations of code, in a setting slightly different to natural language.
Happy to share the report and code upon request.