forked from wmt-conference/wmt16-website
-
Notifications
You must be signed in to change notification settings - Fork 0
/
quality-estimation-task.html
369 lines (274 loc) · 28.7 KB
/
quality-estimation-task.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
<HTML>
<HEAD>
<title>Translation Task - ACL 2016 First Conference on Machine Translation</title>
<style> h3 { margin-top: 2em; } </style>
</HEAD>
<body>
<center>
<script src="title.js"></script>
<p><h2>Shared Task: Quality Estimation</h2></p>
<script src="menu.js"></script>
</center>
<p>This shared task will build on its previous four editions to further examine automatic methods for estimating the quality of machine translation output at run-time, without relying on reference translations. We include <b>word-level</b> (and a variant at phrase-level), <b>sentence-level</b> and <b>document-level</b> estimation. The sentence, phrase and word-level tasks will explore a large dataset produced from post-editions by professional translators (as opposed to crowdsourced translations as in the previous year). For the first time, the data will be domain-specific (IT domain). The document-level task will use, for the first time, entire documents, which have been human annotated for quality indirectly in two ways: through reading comprehension tests and through a two-stage post-editing exercise. Our tasks have the following <b>goals</b>:
<ul>
<li>To advance work on sentence and word-level quality estimation by providing domain-specific, larger and professionally annotated datasets.</li>
<li>To study the utility of detailed information logged during post-editing (time, keystrokes, actual edits) for different levels of prediction.</li>
<li>To analyse the effectiveness of different types of quality labels provided by humans for longer texts in document-level prediction. </li>
<li>To investigate quality estimation at a new level of granularity: phrases. </li>
<!--
<li>To explore differences between sentence-level and document-level
prediction.</li>
<li>To analyse the effect of training data sizes and quality for sentence and word-level prediction, particularly the use of annotations obtained from crowdsourced post-editing. </li>
<li>To explore word-level quality prediction at different levels of granularity. </li>
<li> To investigate the effectiveness of different quality labels. </li>
<li>To push current work on sentence-level quality estimation towards robust models that can work across MT systems;</li>
<li> To study the effects of training and test datasets with mixed domains, language pairs and MT systems. </li>
<li>To test work on sentence-level quality estimation for the task of selecting the best translation amongst multiple systems;</li>
<li>To evaluate the applicability of quality estimation for post-editing tasks;</li>
<li>To provide a first common ground for development and comparison of quality estimation systems at word-level.</li>
-->
</ul>
This year's shared task provides new training and test datasets for all tasks, and allows participants to explore any additional data and resources deemed relevant. A in-house MT system was used to produce translations for the sentence and word-level tasks, and multiple MT systems were used to produce translations for the document-level task. Therefore, MT system-dependent information will be made available where possible.
<p><br><hr>
<!-- BEGIN SENTENCE-LEVEL-->
<h3><font color="blue">Task 1: Sentence-level QE</font></h3>
<!--<p><b><font color="purple">Results <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/results/task1.pdf">here<a/></font></b>, <b>gold-standard labels</b> <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/gold/Task1_gold.tar.gz">here<a/>
-->
<p> This task consists in scoring (and ranking) sentences according to post-editing effort. Multiple labels will be made available, including the percentage of edits need to be fixed (HTER), post-editing time, and keystrokes. Prediction according to each label will be evaluated independently, and any of these can outputs (or their combination) can be used to produce a ranking of translations. The data consists of 15,000 segments on the IT domain translated by an in-house phrase-based SMT system and post-edited by professional translators. The <a =href="https://github.com/ghpaetzold/PET">PET</a> tool was used to collect these various types of information during post-editing. HTER labels are computed using <a href="http://www.umiacs.umd.edu/~snover/terp/">TER</a> (default settings: tokenised, case insensitive, exact matching only, but with scores capped to 1).
</p>
<p>As <i><font color="green">training</font></i> and <i><font color="green">development</font></i> data, we provide English-German datasets with <b>12,000</b> and <b>1,000</b> source sentences, their machine translations, their post-editions (translations) and HTER as post-editing effort scores (other scores, such as post-editing time can be provided on request).
<b><font color="red">New</font></b>: <a href="http://hdl.handle.net/11372/LRT-1631">Download development and training data</a>. The data is publicly available but since it has been provided by our industry partners it is subject to specific terms and conditions. However, these have no practical implications on the use of this data for research purposes. Download the baseline features for <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task1_en-de_training.baseline17.features">training</a> and <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task1_en-de_dev.baseline17.features">development</a> sets.</li>
<p>As <i><font color="green">test data</font></i>, we provide a new set of <b>2,000</b> English-German translations produced by the same SMT system used for the training data. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task1_en-de_test.tar.gz">Download test data</a> and <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task1_en-de_test.baseline17.features"> the baseline features</a>.
</li>
<p><p>
The usual <a href="http://www.quest.dcs.shef.ac.uk/quest_files/features_blackbox_baseline_17">17 features</a> used in WMT12-15 is considered for the <b>baseline system</b>. This system uses SVM regression with an RBF kernel, as well as grid search algorithm for the optimisation of relevant parameters. <a href="http://www.quest.dcs.shef.ac.uk/">QuEst++</a> is used to build prediction models.
<!--
and this <a href="http://www.quest.dcs.shef.ac.uk/wmt13_files/evaluateWMTQP2013-Task1_1.pl">script</a> is used to evaluation the models. For significance tests, we use the bootstrap resampling method with <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/bootstrap-hypothesis-difference-significance.pl">this code</a>.
<br>
<p>As in previous years, two variants of the results can be submitted:
<ul>
<li><b>Scoring</b>: An absolute quality score for each sentence translation according to the type of prediction, to be interpreted as an error metric: lower scores mean better translations.</li>
<li><b>Ranking</b>: A ranking of sentence translations for all source sentences from best to worst. For this variant, it does not matter how the ranking is produced (from HTER predictions, likert predictions, post-editing time, etc.). The reference ranking will be defined based on the true HTER scores.</li>
</ul>
-->
<p><i><font color="green">Evaluation</font></i> is performed against the true label and/or ranking using as metrics:
<ul>
<li><b>Scoring</b>: Pearson's correlation (primary), Mean Average Error (MAE) and Root Mean Squared Error (RMSE).</li>
<li><b>Ranking</b>: Spearman's rank correlation (primary) and DeltaAvg. </li>
</ul>
<p><br><hr>
<!-- BEGIN WORD-LEVEL-->
<h3><font color="blue">Task 2: Word and phrase-level QE</font></h3>
<!--
<p><b><font color="purple">Results <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/results/task2.pdf">here<a/></font></b>, <b>gold-standard labels</b> <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/gold/Task2_gold.tar.gz">here<a/>
-->
<p>The goal of this task is to study the prediction of word and phrase-level errors in MT output. For practical reasons, we frame the problem as the binary task of distinguishing between 'OK' and 'BAD' tokens. The data for this task is the same as provided in Task 1, with English-German machine translations.</p>
<p>For the word-level variant, as in previous years, all segments are automatically annotated for errors with binary word-level labels by using the alignments provided by the <a href="http://www.cs.umd.edu/~snover/tercom/">TER</a> tool (settings: tokenised, case insensitive, exact matching only, disabling shifts by using the `-d 0` option) between machine translations and their post-edited versions. Shifts (word order errors) were not annotated as such (but rather as deletions + insertions) to avoid introducing noise in the annotation.</p>
As <i><font color="green">training</font></i> and <i><font color="green">development</font></i> data, we provide the tokenised translation outputs with tokens annotated with 'OK' or 'BAD' labels.
<a href="http://hdl.handle.net/11372/LRT-1631">Download development and training data</a>. <font color="red"><b>Please download baseline features for <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2_en-de_training.features">training<a/> and <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2_en-de_dev.features">development</a> sets from here (updated 4/4/16)</b></font>. The data is publicly available but since it has been provided by our industry partners it is subject to specific terms and conditions. However, these have no practical implications on the use of this data for research purposes.</li>
<p>As <i><font color="green">test data</font></i>, we provide tokens from additional 2,000 English-German sentences, produced in the same way. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2_en-de_test.tar.gz">Download test data (and baseline features)</a>.
</li>
<p>Submissions are <i><font color="green">evaluated</font></i> in terms of classification performance via the multiplication of F1-scores for the 'OK' and 'BAD' classes against the original labels.
F1-score for the 'BAD' class, which has been used as a primary metric in previous years, is biased towards 'pessimistic' labellings. In other words, it favours systems which tend to label more words as 'BAD'.
In contrast, the multiplication of F1-OK and F1-BAD has two components which penalise different labellings and balance each other. 'Unfair' labellings (ones where either F1-OK or F1-BAD are close to zero) will have a score close to zero, and the overall score is never greater than any of its components.
We will also report the F1-BAD score.
<a href="https://gist.github.com/varvara-l/028e4439fb992d089935">Evaluation script.</a>
<!--We also provide an <a href=" https://gist.github.com/chrishokamp/e2b5bcc07d1006026c48">alternative evaluation script </a>that takes as input labels in the exact same format as the labels distributed for training and dev sets, i.e.: one line per sentence, one tag per word, whitespace separated, with tags in the set {'OK', 'BAD'}.
For significance tests, we used the approximate randomisation method with <a href="http://www.nlpado.de/~sebastian/software/sigf.shtml">this code</a>. -->
<p>As <b>baseline system</b> for this task we use the baseline features provided above to train a binary classifier using a standard logistic regression algorithm (available for example in the scikit-learn toolkit).</p>
<p>As an extension of the word-level task, we introduce a new task: <font color="blue">phrase-level prediction</font></b>. For this task, given a "phrase" (segmentation as given by the SMT decoder), participants are asked to label it as 'OK' or 'BAD'. Errors made by MT engines are interdependent and one incorrectly chosen word can cause more errors, especially in its local context. Phrases as produced by SMT decoders can be seen as a representation of this local context and in this task we ask participants to consider them as atomic units, using phrase-specific information to improve upon the results of the word-level task.</p>
<p>The data to be used is exactly the same as for task 1 and the word-level task. The labelling of this data was adapted from word-level labelling by assigning the 'BAD' tag to any phrase that contains at least one 'BAD' word.</p>
<p>As <i><font color="green">training</font></i> and <i><font color="green">development</font></i> data, we provide the tokenised translation outputs with phrase segmentation for both source and machine-translated sentences. We also provide target-source phrase alignments and phrase-level labels in separate files.
<a href="http://hdl.handle.net/11372/LRT-1631">Download development and training data</a>. <font color="red"><b>Please download baseline features for <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2p_en-de_training.features">training<a/> and <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2p_en-de_dev.features">development</a> sets from here (updated 4/4/16)</b></font>. The data is publicly available but since it has been provided by our industry partners it is subject to specific terms and conditions. However, these have no practical implications on the use of this data for research purposes.</li>
</li>
<p>As <i><font color="green">test data</font></i>, we provide tokens from additional 2,000 English-German sentences, produced in the same way. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task2p_en-de_test.tar.gz">Download test data (and baseline features)</a>.
</li>
<p>The phrase-level submissions are <i><font color="green">evaluated</font></i> are evaluated at the <b>word level</b> using the multiplication of F1-OK and F1-BAD as a primary metric.
<!--
We use word-level scores here because the main goal of the phrase-level QE task is to check whether subsentence-level QE can be improved by using phrase-level features. Therefore, the word-level and phrase-level systems need to be directly comparable
-->
<p>As <b>baseline system</b> for this task we use the baseline features provided above to train a CRF model with a CRF++ tool. </p>
<!-- BEGIN DOCUMENT-LEVEL-->
<p><br><hr>
<h3><font color="blue">Task 3: Document-level QE</font></h3>
<!--
<p><b><font color="purple">Results <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/results/task3.pdf">here<a/></font></b>, <b>gold-standard labels</b> <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/gold/Task3_gold.tar.gz">here<a/>
-->
<p>This task consists in predicting the quality of units larger than sentences. Different from WMT14, this year we consider entire documents, instead of paragraphs.
<!--We consider as application a scenario where the reader needs to process the translation of an entire text, as opposed to individual sentences, and has no knowledge of the source language. -->
The data was extracted from the WMT 2008-2013 English-Spanish translation shared task datasets. The machine translation for each source document was randomly picked from the set of all systems that participated in the task.
<p>The quality labels were computed based on human annotation. It is an adaptation of HTER obtained from a two-stage post-editing approach similar to the one described in <a href="https://aclweb.org/anthology/W/W15/W15-4916.pdf">(Scarton et al., 2014)<a>. <!--and reading comprehension tests. -->
These labels attempt to capture the quality of different documents translated by various MT systems
<!--without penalised by the fact that similar systems can present similar results for different documents,
When automatic metrics are considered as quality labels (such as METEOR, used as quality label on last year document-level task).
The two-stage post-editing method (Scarton et al., 2014) aims at -->
by isolating quality issues that can only be fixed when the entire document is available, from other types of errors, which can be fixed based on the sentence only.
<!-- and give different weights for phenomena that can only be solved with document context. -->
In the first stage, sentences are post-edited in isolation and in random order. In the second stage, these post-edited sentences are reorganised into their original document order and further edited (by the same post-editor), now given the document as context. The difference in percentage of edits between the first and second stages is then used to weight final quality HTER score.
<!--the difference from the first post-editing and the machine translation (in both cases, TER metrics is used to compute the differences). -->
The goal is to penalise documents that needed more editing in the second stage.
Post-editing was done by professional translators.
<p>For the <i><font color="green">training</font></i> of prediction models, we provide a new dataset consisting of source documents and their machine translations (English-Spanish), all in the <it>news domain</it>, extracted from the test set of WMT 2008-2013 and MT systems that participated in the translation shared tasks:
<ul>
<li>146 English→Spanish documents. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task3_en-es_training.tar.gz">Download training data</a>. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task3_en-es_training.baseline17.features">Download 17 baseline feature set</a>.</li>
</ul>
<p>As <i><font color="green">test data</font></i>, we provide a new set of translations for English→Spanish documents produced in the same way as for the training data. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task3_en-es_test.tar.gz">Download test data</a>. <a href="http://www.quest.dcs.shef.ac.uk/wmt16_files_qe/task3_en-es_test.baseline17.features">Download 17 baseline feature set</a>.
<!--
<ul>
<li>415 English source documents → 415 German translation suggestions. <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/task3_en-de_test.tar.gz">Download data</a>. <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/task3_en-de_test.baseline17.features">Download 17 baseline feature set</a>.</li>
<li>415 German source documents → 415 English translation suggestions. <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/task3_de-en_test.tar.gz">Download data</a>. <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/task3_de-en_test.baseline17.features">Download 17 baseline feature set</a>.</li>
</ul>
<p>Two variants of the results can be submitted:
<ul>
<li><b>Scoring</b>: An absolute quality score for each document translation according to the type of prediction, to be interpreted as an error metric: higher scores mean better translations.</li>
<li><b>Ranking</b>: A ranking of document translations for all source documents from best to worst. For this variant, it does not matter how the ranking is produced (from METEOR predictions, likert predictions, post-editing time, etc.). The reference ranking will be defined based on the true METEOR scores.</li>
</ul>
-->
<p> <i><font color="green">Evaluation</font></i> is performed against the true quality label and/or ranking using the following metrics:
<ul>
<li><b>Scoring</b>: Pearson's correlation, Mean Average Error (MAE) and Root Mean Squared Error (RMSE).</li>
<li><b>Ranking</b>: Spearman's rank correlation and DeltaAvg. </li>
</ul>
<p><p>
QuEst++ <a href="http://www.quest.dcs.shef.ac.uk/quest_files/features_blackbox_baseline_17">17 baseline features</a> for document-level will be used as the baseline system. As with sentence-level, the <b>baseline system</b> is trained using SVM regression with an RBF kernel, as well as grid search algorithm for the optimisation of relevant parameters. <!--We will use the same <a href="http://www.quest.dcs.shef.ac.uk/wmt13_files/evaluateWMTQP2013-Task1_1.pl">evaluation script</a> as for sentence-level. For significance tests, we use the bootstrap resampling method with <a href="http://www.quest.dcs.shef.ac.uk/wmt15_files/bootstrap-hypothesis-difference-significance.pl">this code</a>.-->
<br>
<p><br><hr>
<!-- EXTRA STUFF -->
<h3>Additional resources</h3>
<p>These are the resources we have used to extract the baseline features in Task 1, which can also be useful for Task 2. If you require other resources/info from the MT system, let us know:
<p>
<b>English</b>
<lu>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/lm.tok.en.tar.gz">language model</a></li>
<!--<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/nc.pos.1.en.lm">language model of POS tags</a></li>-->
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/ngram-count.tok.en.out.clean.tar.gz">n-gram counts</a></li>
</lu>
<p><b>German</b>
<lu>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/lm.tok.de.tar.gz">language model</a></li>
<!--<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/corpus.de.pos.lm">language model of POS tags</a></li>-->
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/ngram-count.tok.de.out.clean.tar.gz">n-gram counts</a></li>
</lu>
<p><b>Giza tables</b>
<lu>
<li>English-German (and v.v.) <a href="http://www.quest.dcs.shef.ac.uk/quest_files_16/EN-DE.lex.tar.gz">lexical translation table</a></li>
</lu>
<p>Tasks 3 uses multiple MT systems on WMT data, so the usual <a href="http://www.statmt.org/wmt16/translation-task.html">news translation task</a> data resources can be used:
<p>We also suggest the following <b><font color="green">interesting resources</font></b> that can be used as additional data for training (notice the difference in language pairs and/or text domains and/or MT systems):
<ul>
<li><a href="http://www.statmt.org/wmt15/quality-estimation-task.html">WMT15</a>, <a href="http://www.statmt.org/wmt14/quality-estimation-task.html">WMT14</a>, <a href="http://www.statmt.org/wmt13/quality-estimation-task.html">WMT13</a>, <a href="http://www.statmt.org/wmt12/quality-estimation-task.html">WMT12</a> Quality Estimation shared-task datasets.</li></li>
<li><a href="http://www-clips.imag.fr/geod/User/marion.potet/index.php?page=download">LIG corpus</a> of 10,881 French-English SMT translations and their human post-editions (HTER scores can be easily derived). <a href="http://www.lrec-conf.org/proceedings/lrec2012/pdf/506_Paper.pdf">Description</a>.</li></li>
<li><a href="http://anrtrace.limsi.fr/trace_postedit.tar.bz2">LISMI's TRACE corpora</a> of approximately 7,000 French-English and 7,000 English-French translations by different MT systems, for various text domains, and their post-editions by professionals translators. <a href="http://www.mtsummit2013.info/files/proceedings/main/mt-summit-2013-wisniewski-et-al.pdf">Description</a>.</li>
<li><a href="http://bridge.cbs.dk/platform/?q=CRITT_TPR-db">CRITT Translation Process Research Database<a> with user activity data of translators behavior collected in several translation studies with Translog-II and with the CASMACAT workbench.</li>
</ul>
<!--
<p>These are the resources we have used to extract the baseline features in Tasks 1 and 3:
<p>
<b>English</b>
<lu>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/europarl-nc.en">source training corpus</a></li>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lm.europarl-nc.en">language model</a></li>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/nc.pos.1.en.lm">language model of POS tags</a></li>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/ngram-counts.europarl-nc.en.proc">n-gram counts</a></li>
<li>English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/truecase-model.en">truecase model</a></li>
</lu>
<p><b>Spanish</b>
<lu>
<li>Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/training.es">source training corpus</a></li>
<li>Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lm.europarl-interpolated-nc.es">language model</a></li>
<li>Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/pos_lm.es">language model of POS tags</a></li>
<li>Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/ngram-counts.europarl-nc.es">n-gram counts</a></li>
<li>Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/truecase-model.es">truecase model</a></li>
</lu>
<p><b>German</b>
<lu>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/source_corpus.de">source training corpus</a></li>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/news.3gram.de.lm">language model</a></li>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/corpus.de.pos.lm">language model of POS tags</a></li>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/news.3gram.de.counts.proc">n-gram counts</a></li>
<li>German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/truecase-model.de">truecase model</a></li>
</lu>
<p><b>Giza tables</b>
<lu>
<li>English-Spanish <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lex.e2s">Lexical translation table src-tgt</a></li>
<li>English-German <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lex.de-en">Lexical translation table src-tgt</a></li>
<li>Spanish-English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lex.s2e">Lexical translation table src-tgt</a></li>
<li>German-English <a href="http://www.quest.dcs.shef.ac.uk/quest_files/lex.de-en">Lexical translation table src-tgt</a></li>
</lu>
-->
<p><br><hr>
<!-- SUBMISSION INFO -->
<h3>Submission Format</h3>
<h4><font color="red">Tasks 1 and 3: Sentence- and document-level</font></h4>
<p> The output of your system <b>a given subtask</b> should produce scores for the translations at the <i>segment-level</i> of the relevant task (sentence or document) formatted in the following way: </p>
<pre><METHOD NAME> <SEGMENT NUMBER> <SEGMENT SCORE> <SEGMENT RANK><br><br></pre>
Where:
<ul>
<li><code>METHOD NAME</code> is the name of your
quality estimation method.</li>
<li><code>SEGMENT NUMBER</code> is the line number
of the plain text translation file you are scoring/ranking.</li>
<li><code>SEGMENT SCORE</code> is the predicted (HTER/METEOR) score for the
particular segment - assign all 0's to it if you are only submitting
ranking results. </li>
<li><code>SEGMENT RANK</code> is the ranking of
the particular segment - assign all 0's to it if you are only submitting
absolute scores. </li>
</ul>
Each field should be delimited by a single tab character.
<h4><font color="red">Task 2: Word-level QE</font></h4>
<p> The output of your system should produce scores for the translations at the <i>word-level</i>
formatted in the following way: </p>
<pre><METHOD NAME> <SEGMENT NUMBER> <WORD INDEX> <WORD> <BINARY SCORE> <br><br></pre>
Where:
<ul>
<li><code>METHOD NAME</code> is the name of your quality estimation method.</li>
<li><code>SEGMENT NUMBER</code> is the line number of the plain text translation file you are scoring (starting at 0).</li>
<li><code>WORD INDEX</code> is the index of the word in the tokenised sentence, as given in the training/test sets (starting at 0).</li>
<li><code>WORD</code> actual word.</li>
<li><code>BINARY SCORE</code> is either 'OK' for no issue or 'BAD' for any issue.</li>
</ul>
Each field should be delimited by a single tab character.
<h3>Submission Requirements</h3>
Each participating team can submit at most 2 systems for each of the language pairs of each subtask. These should be sent
via email to Lucia Specia <a href="mailto:[email protected]" target="_blank">[email protected]</a>. Please use the following pattern to name your files:
<p>
<code>INSTITUTION-NAME</code>_<code>TASK-NAME</code>_<code>METHOD-NAME</code>, where:
<p> <code>INSTITUTION-NAME</code> is an acronym/short name for your institution, e.g. SHEF
<p><code>TASK-NAME</code> is one of the following: 1, 2, 3.
<p><code>METHOD-NAME</code> is an identifier for your method in case you have multiple methods for the same task, e.g. 2_J48, 2_SVM
<p> For instance, a submission from team SHEF for task 2 using method "SVM" could be named SHEF_2_SVM.
<p>You are invited to submit a short paper (4 to 6 pages) to WMT
describing your QE method(s). You are not required to
submit a paper if you do not want to. In that case, we ask you
to give an appropriate reference describing your method(s) that we can cite
in the WMT overview paper.</p>
<h3>Important dates</h3>
<table>
<tr><td>Release of training data </td><td>January 30, 2016</td></tr>
<tr><td>Release of test data </td><td>April 10 2016</td></tr>
<tr><td>QE metrics results submission deadline </td><td>April 30 2016</td></tr>
<tr><td>Paper submission deadline</td><td>May 8, 2016</td></tr>
<tr><td>Notification of acceptance</td><td>June 5, 2016</td></tr>
<tr><td>Camera-ready deadline</td><td>June 22, 2016</td></tr>
</table>
<h3>Organisers</h3>
<br>
Varvara Logacheva (University of Sheffield)
<br>
Carolina Scarton (University of Sheffield)
<br>
Lucia Specia (University of Sheffield)
<br>
<h3>Contact</h3>
<p> For questions or comments, email Lucia
Specia <a href="mailto:[email protected]" target="_blank">[email protected]</a>.
</p>
<p align="right">
Supported by the European Commission under the
<a href="http://expert-itn.eu/"><img align=right src="figures/expert.png" border=0 width=100 height=40></a>
<a href="http://www.qt21.eu/"><img align=right src="figures/qt21.png" border=0 width=100 height=40></a>
<br>projects (grant numbers 317471 and 645452) <p>
</body>
</HTML>