Hi, I had some trouble running TER recently as well, mostly due to the sparse documentation.
In order to run TER, there are two main file formats that you can use (there’s a third format as well, but I’m not familiar with it). Since you are already running BLEU and METEOR, the simplest would be to use the same SGML files with TER that you produced for those metrics. This, however, won’t work if your text contains actual XML tags within the sentences. In this case, simply add to the end of each line in your plain text ref and hyp files the sentence number in parenthesis, like this: word1 word2 word3 … wordn (sentID1) word1 word2 word3 … wordn (sentID2) There’s more information that can be put in the parenthesis (like document ID, etc.), but simply using the sentence ID would be enough. If several lines have the same sentence ID, they will be treated as multiple reference translations of the same sentence. Cheers, Ventzi ––––––– Dr. Ventsislav Zhechev EuroMatrix+, CNGL Dublin City University Dublin, Ireland http://ventsislavzhechev.eu 22.05.2010, в 17:04, [email protected] написал(а): > Hi, > > I have planned my research to depend on three metrics "BLEU","TER" and > METEOR" for output evaluation, rather than just depending on BLEU. As manual > (found on moses web site) specifies, I downloaded tercom-0.7.25.tgz from > www.cs.umd.edu/~snover/tercom/tercom-0.7.25.tgz. > The following command has been used for evaluation as TER Readme file > suggests, > % java -jar tercom.7.25.jar -r <ref file in plain text format> -h <hypothesis > file in plain text format> > > but , > there were many warnings "invalid lines" in reference file and finally, > ***ERROR*** No reference for segment <a phrase> was shown on the output. > > Could you please help me? to overcome this issue?.. I am thanking for your > attention and help... > > S.Sripirakas. _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
