File size: 12,879 Bytes
2ea1065 |
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 |
# ToTTo Dataset
ToTTo is an open-domain English table-to-text dataset with over 120,000 training examples that proposes a controlled generation task: given a Wikipedia table and a set of highlighted table cells, produce a one-sentence description.
During the dataset creation process, tables from English Wikipedia are matched with (noisy) descriptions. Each table cell mentioned in the description is highlighted and the descriptions are iteratively cleaned and corrected to faithfully reflect the content of the highlighted cells.
We hope this dataset can serve as a useful research benchmark for high-precision conditional text generation.
You can find more details, analyses, and baseline results in [our paper](https://arxiv.org/abs/2004.14373). You can cite it as follows:
```
@inproceedings{parikh2020totto,
title={{ToTTo}: A Controlled Table-To-Text Generation Dataset},
author={Parikh, Ankur P and Wang, Xuezhi and Gehrmann, Sebastian and Faruqui, Manaal and Dhingra, Bhuwan and Yang, Diyi and Das, Dipanjan},
booktitle={Proceedings of EMNLP},
year={2020}
}
```
## Getting Started
**Download the ToTTo data**
The ToTTo dataset is released under the [Creative Commons Share-Alike 3.0](https://creativecommons.org/licenses/by-sa/3.0/) license.
To download the data from the command line:
```
wget https://storage.googleapis.com/totto-public/totto_data.zip
unzip totto_data.zip
```
(or alternatively copy the above url into your browser address bar.)
Inside the `totto_data` directory you should see three files: `totto_train_data.jsonl`, `totto_dev_data.jsonl`, and `unlabeled_totto_test_data.jsonl` for the training, development, and unlabeled test sets respectively.
**Download the evaluation scripts**
You can find evaluation scripts and some exploratory processing scripts at [this repository](https://github.com/google-research/language/tree/master/language/totto). It also includes a separate README file with instruction on how to run the evaluation.
## Dataset Description
The ToTTo dataset consists of three `.jsonl` files, where each line is a JSON dictionary with the following format:
```json
{
"table_page_title": "'Weird Al' Yankovic",
"table_webpage_url": "https://en.wikipedia.org/wiki/%22Weird_Al%22_Yankovic",
"table_section_title": "Television",
"table_section_text": "",
"table": "[Described below]",
"highlighted_cells": [[22, 2], [22, 3], [22, 0], [22, 1], [23, 3], [23, 1], [23, 0]],
"example_id": 12345678912345678912,
"sentence_annotations": [{"original_sentence": "In 2016, Al appeared in 2 episodes of BoJack Horseman as Mr. Peanutbutter's brother, Captain Peanutbutter, and was hired to voice the lead role in the 2016 Disney XD series Milo Murphy's Law.",
"sentence_after_deletion": "In 2016, Al appeared in 2 episodes of BoJack Horseman as Captain Peanutbutter, and was hired to the lead role in the 2016 series Milo Murphy's Law.",
"sentence_after_ambiguity": "In 2016, Al appeared in 2 episodes of BoJack Horseman as Captain Peanutbutter, and was hired for the lead role in the 2016 series Milo Murphy's 'Law.",
"final_sentence": "In 2016, Al appeared in 2 episodes of BoJack Horseman as Captain Peanutbutter and was hired for the lead role in the 2016 series Milo Murphy's Law."}],
}
```
The `table` field is a `List[List[Dict]]`. The outer lists represents rows and the inner lists columns. Each `Dict` has the fields `column_span: int`, `is_header: bool`, `row_span: int`, and `value: str`. The first two rows for the example above look as follows:
```json
[
[
{ "column_span": 1,
"is_header": true,
"row_span": 1,
"value": "Year"},
{ "column_span": 1,
"is_header": true,
"row_span": 1,
"value": "Title"},
{ "column_span": 1,
"is_header": true,
"row_span": 1,
"value": "Role"},
{ "column_span": 1,
"is_header": true,
"row_span": 1,
"value": "Notes"}
],
[
{ "column_span": 1,
"is_header": false,
"row_span": 1,
"value": "1997"},
{ "column_span": 1,
"is_header": false,
"row_span": 1,
"value": "Eek! The Cat"},
{ "column_span": 1,
"is_header": false,
"row_span": 1,
"value": "Himself"},
{ "column_span": 1,
"is_header": false,
"row_span": 1,
"value": "Episode: 'The FugEektive'"}
], ...
]
```
-The table metadata consists of the `table_page_title`, `table_section_title`, and `table_section_text` strings to help give the model more context about the table.
-The `highlighted_cells` field is a `List[[row_index, column_index]]` where each `[row_index, column_index]` pair indicates that `table[row_index][column_index]` is highlighted.
-The `example_id` is simply a unique id for this example.
-The `sentence_annotations` field consists of the `original sentence` and the sequence of revised sentences performed in order to produce the `final_sentence`. See our [paper](https://arxiv.org/abs/2004.14373) for more details.
To help understand the dataset, you can find a sample of the train and dev sets in the `sample/` folder of our [supplementary repository](https://github.com/google-research/language/tree/master/language/totto). It additionally provides the `create_table_to_text_html.py` script that visualizes examples, the output of which you can also find in the `sample/` folder.
### Official Task
The official task described in our [paper](https://arxiv.org/abs/2004.14373) is given the `table`, `highlighted cells`, and table metadata (`table_page_title`, `table_section_title`, and `table_section_text`) as input, to generate the `final_sentence`.
### Dev and Test Set
The dev and test set have between two and three references for each example, which are added to the list at the `sentence_annotations` key. The test set annotations are *private* and thus not included in the data.
If you want us to evaluate your model on the development or the private test set, please submit your files [here](https://forms.gle/AcF9TRqWrPhPzztt7). You can find more submission information [below](https://github.com/google-research-datasets/ToTTo#leaderboard-submission). By emailing us or by submitting prediction files, you consent to being contacted by Google about your submission, this dataset or any related competitions.
We provide two splits within the dev and test sets - one uses previously seen combinations of table headers and one uses unseen combinations. The sets are marked using the `overlap_subset: bool` flag that is added to the JSON representation. By filtering the evaluation to examples with the flag set to `true`, you will be able to test the generalization ability of your model.
****
# Leaderboard
We are maintaining a leaderboard with official results on our test set.
The leaderboard indicates whether or not a model was trained on any auxiliary Wikipedia data. This is because our tables and (unrevised) test targets are from Wikipedia and thus we would like to study the effect of using additional Wikipedia data to train models.
We ask you to **not** incorporate any part of the ToTTo development set into the training data, and only use it for validation/hyperparameter tuning as development sets are typically used.
In addition to [BLEU](https://www.aclweb.org/anthology/P02-1040/) and [PARENT](https://arxiv.org/abs/1906.01081), we also report a learnt metric [BLEURT](https://arxiv.org/abs/2004.04696). The checkpoint used was BLEURT-base-128 which can be found [here](https://github.com/google-research/bleurt). To handle multiple references, we take the average of the scores as suggested by [Sellam et al. 2020](https://arxiv.org/abs/2010.04297).
<table>
<tr>
<th></th>
<th></th>
<th></th>
<th colspan="3">Overall</th>
<th colspan="3">Overlap Subset</th>
<th colspan="3">Non-Overlap Subset</th>
</tr>
<tr>
<th>Model</th>
<th>Link</th>
<th>Uses Wiki</th>
<th>BLEU</th>
<th>PARENT</th>
<th>BLEURT</th>
<th>BLEU</th>
<th>PARENT</th>
<th>BLEURT</th>
<th>BLEU</th>
<th>PARENT</th>
<th>BLEURT</th>
</tr>
<tr>
<td> LATTICE </td>
<td> <a href="https://arxiv.org/abs/2205.03972">[Wang et al. 2022]</a> </td>
<td>yes</td>
<td>48.4</td>
<td>58.1</td>
<td>0.222</td>
<td>56.1</td>
<td>62.4</td>
<td>0.345</td>
<td>40.4</td>
<td>53.9</td>
<td>.099</td>
</tr>
<tr>
<td> SKY </td>
<td> in preparation </td>
<td>yes</td>
<td><b>49.9</b></td>
<td><b>59.8</b></td>
<td>0.212</td>
<td><b>57.8</b></td>
<td>64.0</td>
<td>0.334</td>
<td><b>42.0</b></td>
<td><b>55.7</b></td>
<td>0.091</td>
</tr>
<tr>
<td> CoNT </td>
<td> <a href="https://arxiv.org/abs/2205.14690v2">[An et al., 2022]</a></td>
<td>yes</td>
<td>49.1</td>
<td>58.9</td>
<td>0.238</td>
<td>56.7</td>
<td>63.2</td>
<td>0.355</td>
<td>41.3</td>
<td>54.6</td>
<td>0.121</td>
</tr>
<tr>
<td> Supervised+NLPO </td>
<td> <a href="https://arxiv.org/abs/2210.01241">[Ramamurthy et al. 2022]</a></td>
<td>yes</td>
<td>47.4</td>
<td>59.6</td>
<td>0.192</td>
<td>55.0</td>
<td><b>64.3</b></td>
<td>0.315</td>
<td>39.2</td>
<td>55.0</td>
<td>0.068</td>
</tr>
<tr>
<td> Anonymous 3 </td>
<td> in preparation </td>
<td>yes</td>
<td>49.3</td>
<td>58.8</td>
<td>0.235</td>
<td>57.1</td>
<td>63.4</td>
<td>0.358</td>
<td>41.5</td>
<td>54.1</td>
<td>0.112</td>
</tr>
<tr>
<td> ProEdit </td>
<td> Paper in preparation </td>
<td>yes</td>
<td> 48.6 </td>
<td>59.18</td>
<td>0.202</td>
<td>55.9</td>
<td>63.3</td>
<td>0.325</td>
<td>41.3</td>
<td>55.1</td>
<td>0.078</td>
</tr>
<tr>
<td> Anonymous 2</td>
<td> Paper in preparation </td>
<td>yes</td>
<td>49.4 </td>
<td>59.0</td>
<td><b>0.253</b></td>
<td>57.0</td>
<td>62.9</td>
<td><b>0.370</b></td>
<td>41.7</td>
<td>55.1</td>
<td><b>0.136</b></td>
</tr>
<tr>
<td> PlanGen (University of Cambridge, Apple) </td>
<td> <a href="https://arxiv.org/abs/2108.13740"> [Su et al. 2021] </a> </td>
<td>yes</td>
<td>49.2 </td>
<td>58.7</td>
<td>0.249</td>
<td>56.9</td>
<td>62.8</td>
<td>0.371</td>
<td>41.5</td>
<td>54.6</td>
<td>0.126</td>
</tr>
<tr>
<td> T5-based (<i>Google</i>) </td>
<td> <a href="https://arxiv.org/abs/2005.10433">[Kale, 2020]</a></td>
<td>yes</td>
<td>49.5</td>
<td>58.4</td>
<td>0.230</td>
<td>57.5</td>
<td>62.6</td>
<td>0.351</td>
<td>41.4</td>
<td>54.2</td>
<td>0.1079</td>
</tr>
<tr>
<td>BERT-to-BERT (Wiki+Books)</td>
<td><a href="https://arxiv.org/abs/1907.12461">[Rothe et al., 2019]</a></td>
<td>yes</td>
<td>44.0</td>
<td>52.6</td>
<td>0.121</td>
<td>52.7</td>
<td>58.4</td>
<td>0.259</td>
<td>35.1</td>
<td>46.8</td>
<td>-0.017</td>
</tr>
<tr>
<td>BERT-to-BERT (Books)</td>
<td><a href="https://arxiv.org/abs/1907.12461">[Rothe et al., 2019]</a></td>
<td>no</td>
<td>43.9</td>
<td>52.6</td>
<td>0.104</td>
<td>52.7</td>
<td>58.4</td>
<td><b>0.255</b></td>
<td>34.8</td>
<td>46.7</td>
<td>-0.046</td>
</tr>
<tr>
<td>Pointer Generator</td>
<td><a href="https://www.aclweb.org/anthology/P17-1099/">[See et al., 2017]</a></td>
<td>no</td>
<td>41.6</td>
<td>51.6</td>
<td>0.076</td>
<td>50.6</td>
<td>58.0</td>
<td>0.244</td>
<td>32.2</td>
<td>45.2</td>
<td>-0.0922</td>
</tr>
<tr>
<td>Content Planner</td>
<td><a href="https://www.aaai.org/ojs/index.php/AAAI/article/view/4668">[Puduppully et al., 2019]</a></td>
<td>no</td>
<td>19.2</td>
<td>29.2</td>
<td>-0.576</td>
<td>24.5</td>
<td>32.5</td>
<td>-0.491</td>
<td>13.9</td>
<td>25.8</td>
<td>-0.662</td>
</tr>
</table>
## Leaderboard Submission
If you want to submit dev and test outputs, please format your predictions as a single `.txt` file with line-separated predictions. The predictions should be in the same order as the examples in the `test.jsonl` file.
You can upload your prediction files [here](https://forms.gle/AcF9TRqWrPhPzztt7) and email us at totto@google.com to tell us you have submitted. By emailing us or by submitting prediction files, you consent to being contacted by Google about your submission, this dataset or any related competitions.
|