If you are a hacker who is interested in linguistics, you might want to check out the work we are doing at my startup Ozora Research (http://ozoraresearch.com h

If you are a hacker who is interested in linguistics, you might want to check out the work we are doing at my startup Ozora Research (http://ozoraresearch.com http://ozoraresearch.com). Our goal is to build high-quality, sophisticated NLP solutions for "cleantext" : text that has been professionally written and copy-edited, so that it doesn't have obvious errors of spelling, grammar, punctuation, etc. What makes our work different is that, unlike almost all other research in NLP, we don't depend on human-annotated data (like the Penn TreeBank) for training and evaluation. Instead, we build our models into lossless data compressors, and evaluate by invoking the compressor on a large text corpus. In addition to this change in evaluation methodology, we are also deeply interested in the empirical study of text - that is, in the questions of traditional syntax research. For example, an important part of English syntax relates to verb argument structure: verbs differ in the types of arguments they can accept. Some verbs can accept infinitive complements, that-complements, or indirect objects; others cannot. Our system contains a module that handles argument structure and forwards this information to the parser and compressor. I wrote a book presenting a rationale for why scientists should be deeply interested in large-scale, lossless data compression. Roughly, empirical data compression research is a variation of the scientific method: if theory A achieves better compression than theory B, then A is closer to the truth. Scientists can use this principle to search systematically through theory-space and find high-quality theories. http://arxiv.org/abs/1104.5466 http://arxiv.org/abs/1104.5466 Interestingly, compression has a direct link to generation. By sending random bits into the decompressor, you obtain sample sentences that illustrate the model's "idea" of what English text looks like. By studying the discrepancy between the sample sentences and real text, you can discover ways to improve the model. This is a very productive technique that we use routinely. Chomsky declared that his goal was to build a machine that generates grammatical English sentences, but none of his followers actually tried to build such a device. So in some sense our research is actually closer to Chomsky's stated goal than what Chomsky's followers are doing. I would love to hear from people who might want to collaborate on linguistics work, or use the data compression idea to explore other fields (eg. computer vision, bioinformatics, astronomy).