Error when parsing the American Constitution with spacy about self-attentive-parser HOT 3 CLOSED

Thanks for reporting!

The issue here is due to Section 8, which is a single (very long) sentence. The parser was not trained on sentences of this length, and there is a hard-coded limit of 300 words per sentence.

I'll work on updating the code to throw a more understandable error in this situation.

Unfortunately I don't see a straightforward way to relax the length restriction. The training sentences are much shorter (most are less than 40 words long), and the way word positions are represented in the parser has theoretical issues with generalizing to longer lengths. Moreover, the memory used by decoding algorithm scales as the square of the sentence length, which is why you are seeing OOM (out-of-memory) errors on GPU.

I'll think about what to do in these situations, but for now my recommendation would be to modify the text in some way to omit super-long sentences. For example, you can take turns deleting some of the "To ...;" lines and parsing the rest, and then combine everything to build a parse tree.

from self-attentive-parser.

Thanks!

from self-attentive-parser.

I'm closing this issue since I've added more understandable error messages when a sentence that is too long is given to the parser. Unfortunately there's no easy way to relax the length restrictions themselves, especially now that the recommended parser models use BERT (which has length limits of its own, for many of the same reasons).

from self-attentive-parser.