On failure, the parser tends to produce seemingly correct output when only quickly looking on its first few pages (as demonstrated many times before). We should check the output sanity. 1) structurally, using the CoNLL-U sanity checker and 2) by comparing the number of tokens and sentences which entered the pipeline against the final number which left the pipeline. That way we can at least guarantee we have parsed everything into something that looks reasonable.
On failure, the parser tends to produce seemingly correct output when only quickly looking on its first few pages (as demonstrated many times before). We should check the output sanity. 1) structurally, using the CoNLL-U sanity checker and 2) by comparing the number of tokens and sentences which entered the pipeline against the final number which left the pipeline. That way we can at least guarantee we have parsed everything into something that looks reasonable.