This is an old revision of the document!
Table of Contents
Addicter
The introductory page on the Addicter project is here.
This page lies in the external name space and is intended for collaboration with people outside of ÚFAL.
Word Alignment -- Progress and Results
Latest best results
Alternative model comparison
| Precision/Recall/F-score: | ||||
|---|---|---|---|---|
| Lex | Order | Punct | Miss | |
| ter* | 0.106/0.387/0.167 | 0.025/0.191/0.044 | 0.132/0.936/0.232 | 0.026/0.170/0.046 |
| meteor | 0.092/0.251/0.135 | 0.047/0.229/0.078 | 0.248/0.665/0.361 | 0.020/0.382/0.038 |
| hmm | 0.162/0.426/0.234 | 0.069/0.309/0.112 | 0.281/0.793/0.415 | 0.025/0.400/0.047 |
| lcs | 0.168/0.462/0.247 | 0.000/0.000/0.000 | 0.293/0.848/0.435 | 0.026/0.374/0.049 |
| gizainter | 0.170/0.483/0.252 | 0.049/0.137/0.072 | 0.284/0.878/0.429 | 0.029/0.409/0.054 |
| gizadiag* | 0.183/0.512/0.270 | 0.044/0.250/0.075 | 0.285/0.784/0.417 | 0.038/0.224/0.065 |
| berkeley* | 0.200/0.540/0.291 | 0.050/0.330/0.087 | 0.292/0.844/0.434 | 0.039/0.267/0.068 |
Explicit wrong lex choice detection
- “Dan's method”
- align input+czeng to reference+czeng and input+czeng to hypotheses+czeng
- extract hypothesis-to-reference alignments from there
| Precision/Recall/F-score: | ||||
|---|---|---|---|---|
| Lex | Order | Punct | Miss | |
| czengdiag* | 0.187/0.514/0.275 | 0.069/0.455/0.120 | 0.230/0.883/0.365 | 0.035/0.234/0.061 |
| czenginter | 0.197/0.543/0.290 | 0.108/0.475/0.176 | 0.233/0.926/0.372 | 0.032/0.402/0.060 |
Alignment combinations
via weighed HMM
| Precision/Recall/F-score: | ||||
|---|---|---|---|---|
| Lex | Order | Punct | Miss | |
| ter+hmm | 0.116/0.402/0.180 | 0.030/0.184/0.051 | 0.145/0.912/0.251 | 0.026/0.181/0.046 |
| meteor+hmm | 0.162/0.426/0.234 | 0.068/0.309/0.112 | 0.286/0.794/0.421 | 0.025/0.400/0.047 |
| gizadiag+hmm | 0.186/0.515/0.273 | 0.040/0.215/0.067 | 0.297/0.836/0.438 | 0.039/0.238/0.067 |
| gizainter+hmm | 0.194/0.505/0.281 | 0.062/0.282/0.101 | 0.299/0.806/0.436 | 0.033/0.382/0.061 |
| berkeley+hmm | 0.203/0.548/0.297 | 0.049/0.320/0.085 | 0.290/0.816/0.428 | 0.041/0.277/0.071 |
| czengdiag+hmm | 0.190/0.517/0.278 | 0.073/0.457/0.126 | 0.291/0.841/0.432 | 0.039/0.238/0.067 |
| czenginter+hmm | 0.214/0.545/0.307 | 0.093/0.525/0.158 | 0.304/0.818/0.443 | 0.038/0.363/0.068 |
| berk+czengint+hmm | 0.219/0.568/0.316 | 0.070/0.432/0.120 | 0.298/0.817/0.436 | 0.048/0.290/0.082 |
| berk+czengint+gizaint+hmm | 0.220/0.569/0.317 | 0.068/0.420/0.118 | 0.298/0.812/0.436 | 0.048/0.290/0.083 |
| berk+czengint+meteor+hmm | 0.220/0.569/0.317 | 0.070/0.440/0.121 | 0.295/0.810/0.433 | 0.048/0.290/0.083 |
| berk+czengint+meteor+gizaint+hmm | 0.221/0.571/0.318 | 0.068/0.424/0.118 | 0.298/0.808/0.436 | 0.049/0.292/0.084 |
TODOs
- test alignment with synonym detection (cz_wn required) = separating
lex::anddisam:: - order evaluation
- a lot of background research
- currently finds misplaced items, but their shift distances are off
- to fix – for every misplaced token
- if it (and only it) were to be moved in the original permutation, what would be the best place?
- evaluate with nr. of intersections
- try domain adaptation for word alignment, EMNLP 2011 paper
- comb and comment the code
- add help files
- integrate with the rest of Addicter
- approach applicable to learner's corpora
- see Anne Lüdelig, TLT9
- adapt Addicter to Sara's program
- alternative to reference-based evaluation: “Inconsistencies in Penn parsing”, M. Dickinson