Estimating Comma Placement in Natural Language

We study the feasibility of identifying comma locations using both n-gram models and stochastic contextfree grammars (SCFGs). Specifically, our algorithms take an input sentence without commas and returns the positions where commas should be inserted, along with probability or confidence estimates. This can be generalized to correcting comma placement with minor modifications. However, we focus on this simpler comma insertion problem. Two widely used tools for processing natural language are n-gram models and SCFG parsers. N-grams provide a linear Markov model, while SCFG parsers build a rich hierarchical structure. In English, commas are typically used to separate phrases in a sentence. Hence, it would seem logical that SCFGs should have better performance than the n-gram model, which examines sentences in n-gram length chunks. However, n-grams can easily be trained on very large data sets, and thus can provide a rich source of statistical information. Since these language models are very di↵erent, this paper evaluates both.