On 06/08/13 13:34, Christopher Kotfila wrote:
No, not the maxent model, What I meant was the evaluation of percision and recall are based on exact matches of start and end of sentence, rather than correctly identifying only start or end.
hmmm I can't say with full confidence without looking at the source but I would expect no... basically, the evaluation step involves comparing what your model says against gold data. The gold data have to already be in a 1-sentence-per-line format so I'm guessing it first reads the gold sentences to produce a gold set and then just asks whether each prediction is contained in that. If it finds the exact same sentence in the gold set then it is correct otherwise wrong. On second thoughts your initial rationale may actually be valid here...if the model recognises 2 sentences as one then you actually get 2 errors cos none of the 2 sentences will ever be recognised in any line. However, if it predicts 1 sentence for 2 then there is only 1 error...
Jim
