Hi Giulio, Happy New Year!
I have fixed the bugs in t3.c. It turned out to be two issues: 1) size_t is unsigned, whereas a lot of code depended on various types being signed. 2) In t3.c's "viterbi" function, I (or someone else) had introduced a bug somewhere between version 1.8.6 and 2.0.0.pre01, in how the "b" array was created. This led to all kinds of memory problems, hence the issue you saw with long sentences. I have pushed the commits to master. Please let me know if you find anything strange in the commits. Best wishes, Ulrik On 05/01/16 18:28, Giulio Paci wrote: > -----BEGIN PGP SIGNED MESSAGE----- > Hash: SHA512 > > Hi to all, > how are you? I hope you started well the new year. :-) > > My clean-up work is still going on, but I am starting to see the light. > I have implemented a few functions that helps to make the taggers re-entrant > (although some work is still needed in order to achieve this goal). > > In the meanwhile I setup a few experiments, so that I can check that > everything is working properly. I tested et and t3 tagging capabilities. > They are both producing output in the correct format. However t3 tends to > produce less stable results. > In particular I noticed that t3 tends to produce a lot of errors (~100%, very > often the assigned tag was not present in the dataset for each given word) at > the beginning of > some sentences, while it performs well at the end of the sentences (usually > the last 10 tokens seems to provide reasonable results). Shortening the > sentences improve the > situation. > > Did you experienced this issue in the past? Do you have any idea why this is > happening (I suspect it is due to probability calculation errors)? > > Bests, > Giulio. > -----BEGIN PGP SIGNATURE----- > Version: GnuPG v1 > > iQIcBAEBCgAGBQJWi/0qAAoJEM7JlC23rbhvsm8QAK0LmQx0Yz5+r4V4kETqq10O > XDf16mdPVBphNPeRSjuCe9/qmOXC43q8rf8BLRbVGt1u5BVH7gXL1aOxgic6oarO > QnHYih1HwpFlMI+WPDff03ZBPINlmA1++wOZmXw6nd/FU10mDAb7cRZ4ExVLuBia > QE9Yfjo22XtNXWK1ieKEMXM/FiwXGgrJkipBlahIOA/SSaRZVbzzyaVbOetBJKZ6 > 88IAlfqENdEp+K7QnnJnI1j56JfZVHctzjWkTIgfGToKc2XYUfIB5PJYOoJip0xL > U6jG8nDbmKMRxx88kOUZR6+3vuhHy2cRaHQuaSu3P0gftVxfbf5IHuBaq62nEsbd > tapDCCEZx6BkrELjDraiioqD6kRW7HGEgFXooqTksPsNrLzOYEIv37pWg4jNOs59 > 6HLL1R40gNn7FIp1bodZtiyqZ5P4i8XC7t/4KCqOcs7FGOMG2udknQfR32YsrKQa > KSC3eBRx5uzx7bmy+Jkr9AGCwkKKB0igsBwOVlDCaUtFtFMmSthNomfr63oz/JAh > 6XwGEBo1BUy8aaKwzzGM+mfdhVzzNTDm+DF5ciqJ64reOiyTny8oF6KqrN/Njt2A > bkB1NY8Sx1bw5l/n2TQuBG8rlMk0p8MeCDF+JFeKEatK+/1gYu2V516J1hW323UD > q572VcBAUp9klNR4JlKn > =WpPc > -----END PGP SIGNATURE----- > > ------------------------------------------------------------------------------ > _______________________________________________ > acopost-devel mailing list > acopost-devel@lists.sourceforge.net > https://lists.sourceforge.net/lists/listinfo/acopost-devel ------------------------------------------------------------------------------ _______________________________________________ acopost-devel mailing list acopost-devel@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/acopost-devel
