question on accuracy

[languagemachines]

question on accuracy

[kosloot]

hopefully @antalvdb can shine a light on this?

[antalvdb]

This paper from 2007 has the basic performance estimations for POS tagging, morphological analysis, and dependency parsing. The latter is computed by a predecessor of the CoNLL dependency parsing evaluator. The paper does not specify scores for the lemmatizer (but see this paper), the shallow parser / XP chunker (current score on test data: 91.3 precision, 92.5 recall, 91.9 F-score) or the named entity recognizer (current score on test data: overall F-score 82.1, persons 81.7, locations 90.9, organizations 75.1). The latter scores have not been published yet.

[jwijffels]

Thank you for links to the papers and the scores. That already provides some information.

I'm trying to compare frog to the model accuracies reported with udpipe models (https://ufal.mff.cuni.cz/udpipe/users-manual#universal_dependencies_20_models_performance) which were either built on the UD_Dutch or UD_Dutch-LassySmall corpus (see http://universaldependencies.org/treebanks/nl-comparison.html or details on these 2 corpora). As the numbers reported in the papers are highly likely driven by the corpus used I wonder if there are accuracy related metrics (precision/recall/f/uas/las) scores available for a model which was also trained on these corpora from universaldependencies? Or is this wishfull thinking that someone would have done this?

[antalvdb]

Unfortunately we do not have the time to do these types of comparative evaluations. We always welcome anyone willing to put in time to do these types of exercises, and are happy to assist where possible.

Frog's parser is described in more detail here. The memory-based parser emulates the Alpino parser. Its inference (constraint satisfaction inference) is fast, but produces parses that are less accurate than those of Alpino. We trade accuracy for (predictable, relatively high) speed.

[jwijffels]

Thank you for the input. I understand completely that you don't have time for this. It's not a small task.
I'm basically asking because recently I wrote an R wrapper around UDPipe (https://github.com/bnosac/udpipe) and I'm now investigating how good UDPipe is in comparison to other similar parsers, e.g. the Alpino parser or frog for Dutch, opennlp or the python pattern nlp package.

I was recently making a comparison between UDPipe & Spacy (https://github.com/jwijffels/udpipe-spacy-comparison) but I would like to have e.g. Frog added as well as the python pattern library, Alpino and OpenNLP. Do you know if such research has already been done so that maybe I can take a short-cut in this analysis.

[proycon]

As far as I know, it has not been done and would be a very welcome study..

[jwijffels]

Last week, I got into contact with Gertjan van Noord & Gosse Bouma as they had a paper on evaluating Alpino versus Parsey/Parseysaurus on the UD_Lassy-Small treebank http://aclweb.org/anthology/W17-0403 . I received the output of the Alpino results in CONLLU format which allowed a comparison to UDPipe. There were some nitty-gritty details on the evaluation but it already gave an indication of accuracy.

All it takes for making a comparison is providing the annotation result of some text for which we know the annotation in conllu output after which the evaluation script used by the CONLL17 shared task available at https://github.com/ufal/conll2017/blob/master/evaluation_script/conll17_ud_eval.py can be used. But the tricky part is to get the annotation result in conllu format :)

[kosloot]

as this is not a REAL issue, I close this