Commit
•
12a6fed
1
Parent(s):
d3736f4
Update README.md
Browse files
README.md
CHANGED
|
@@ -79,6 +79,219 @@ SLIMER performs comparably to these state-of-the-art models on OOD input domains
|
|
| 79 |
We extend the standard zero-shot evaluations (CrossNER and MIT) with BUSTER, which is characterized by financial entities that are rather far from the more traditional tags observed by all models during training.
|
| 80 |
An inverse trend can be observed, with SLIMER emerging as the most effective in dealing with these unseen labels, thanks to its lighter instruction tuning methodology and the use of definition and guidelines.
|
| 81 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 82 |
<table>
|
| 83 |
<thead>
|
| 84 |
<tr>
|
|
|
|
| 79 |
We extend the standard zero-shot evaluations (CrossNER and MIT) with BUSTER, which is characterized by financial entities that are rather far from the more traditional tags observed by all models during training.
|
| 80 |
An inverse trend can be observed, with SLIMER emerging as the most effective in dealing with these unseen labels, thanks to its lighter instruction tuning methodology and the use of definition and guidelines.
|
| 81 |
|
| 82 |
+
<!DOCTYPE html>
|
| 83 |
+
<html>
|
| 84 |
+
<head>
|
| 85 |
+
<style>
|
| 86 |
+
table {
|
| 87 |
+
width: 100%;
|
| 88 |
+
border-collapse: collapse;
|
| 89 |
+
font-size: 12px;
|
| 90 |
+
}
|
| 91 |
+
th, td {
|
| 92 |
+
border: 1px solid black;
|
| 93 |
+
padding: 4px;
|
| 94 |
+
text-align: center;
|
| 95 |
+
}
|
| 96 |
+
th {
|
| 97 |
+
background-color: #f2f2f2;
|
| 98 |
+
}
|
| 99 |
+
.col-model { width: 10%; }
|
| 100 |
+
.col-backbone { width: 15%; }
|
| 101 |
+
.col-params { width: 10%; }
|
| 102 |
+
.col-mit, .col-crossner, .col-buster, .col-avg { width: 7%; }
|
| 103 |
+
</style>
|
| 104 |
+
</head>
|
| 105 |
+
<body>
|
| 106 |
+
|
| 107 |
+
<table>
|
| 108 |
+
<thead>
|
| 109 |
+
<tr>
|
| 110 |
+
<th class="col-model">Model</th>
|
| 111 |
+
<th class="col-backbone">Backbone</th>
|
| 112 |
+
<th class="col-params">#Params</th>
|
| 113 |
+
<th class="col-mit" colspan="2">MIT</th>
|
| 114 |
+
<th class="col-crossner" colspan="5">CrossNER</th>
|
| 115 |
+
<th class="col-buster">BUSTER</th>
|
| 116 |
+
<th class="col-avg">AVG</th>
|
| 117 |
+
</tr>
|
| 118 |
+
<tr>
|
| 119 |
+
<th></th>
|
| 120 |
+
<th></th>
|
| 121 |
+
<th></th>
|
| 122 |
+
<th class="col-mit">Movie</th>
|
| 123 |
+
<th class="col-mit">Restaurant</th>
|
| 124 |
+
<th class="col-crossner">AI</th>
|
| 125 |
+
<th class="col-crossner">Literature</th>
|
| 126 |
+
<th class="col-crossner">Music</th>
|
| 127 |
+
<th class="col-crossner">Politics</th>
|
| 128 |
+
<th class="col-crossner">Science</th>
|
| 129 |
+
<th class="col-buster"></th>
|
| 130 |
+
<th class="col-avg"></th>
|
| 131 |
+
</tr>
|
| 132 |
+
</thead>
|
| 133 |
+
<tbody>
|
| 134 |
+
<tr>
|
| 135 |
+
<td class="col-model">ChatGPT</td>
|
| 136 |
+
<td class="col-backbone">gpt-3.5-turbo</td>
|
| 137 |
+
<td class="col-params">-</td>
|
| 138 |
+
<td class="col-mit">5.3</td>
|
| 139 |
+
<td class="col-mit">32.8</td>
|
| 140 |
+
<td class="col-crossner">52.4</td>
|
| 141 |
+
<td class="col-crossner">39.8</td>
|
| 142 |
+
<td class="col-crossner">66.6</td>
|
| 143 |
+
<td class="col-crossner">68.5</td>
|
| 144 |
+
<td class="col-crossner">67.0</td>
|
| 145 |
+
<td class="col-buster">-</td>
|
| 146 |
+
<td class="col-avg">-</td>
|
| 147 |
+
</tr>
|
| 148 |
+
<tr>
|
| 149 |
+
<td class="col-model">InstructUIE</td>
|
| 150 |
+
<td class="col-backbone">Flan-T5-xxl</td>
|
| 151 |
+
<td class="col-params">11B</td>
|
| 152 |
+
<td class="col-mit">63.0</td>
|
| 153 |
+
<td class="col-mit">21.0</td>
|
| 154 |
+
<td class="col-crossner">49.0</td>
|
| 155 |
+
<td class="col-crossner">47.2</td>
|
| 156 |
+
<td class="col-crossner">53.2</td>
|
| 157 |
+
<td class="col-crossner">48.2</td>
|
| 158 |
+
<td class="col-crossner">49.3</td>
|
| 159 |
+
<td class="col-buster">-</td>
|
| 160 |
+
<td class="col-avg">-</td>
|
| 161 |
+
</tr>
|
| 162 |
+
<tr>
|
| 163 |
+
<td class="col-model">UniNER-type</td>
|
| 164 |
+
<td class="col-backbone">LLaMA-1</td>
|
| 165 |
+
<td class="col-params">7B</td>
|
| 166 |
+
<td class="col-mit">42.4</td>
|
| 167 |
+
<td class="col-mit">31.7</td>
|
| 168 |
+
<td class="col-crossner">53.5</td>
|
| 169 |
+
<td class="col-crossner">59.4</td>
|
| 170 |
+
<td class="col-crossner">65.0</td>
|
| 171 |
+
<td class="col-crossner">60.8</td>
|
| 172 |
+
<td class="col-crossner">61.1</td>
|
| 173 |
+
<td class="col-buster">34.8</td>
|
| 174 |
+
<td class="col-avg">51.1</td>
|
| 175 |
+
</tr>
|
| 176 |
+
<tr>
|
| 177 |
+
<td class="col-model">UniNER-def</td>
|
| 178 |
+
<td class="col-backbone">LLaMA-1</td>
|
| 179 |
+
<td class="col-params">7B</td>
|
| 180 |
+
<td class="col-mit">27.1</td>
|
| 181 |
+
<td class="col-mit">27.9</td>
|
| 182 |
+
<td class="col-crossner">44.5</td>
|
| 183 |
+
<td class="col-crossner">49.2</td>
|
| 184 |
+
<td class="col-crossner">55.8</td>
|
| 185 |
+
<td class="col-crossner">57.5</td>
|
| 186 |
+
<td class="col-crossner">52.9</td>
|
| 187 |
+
<td class="col-buster">33.6</td>
|
| 188 |
+
<td class="col-avg">43.6</td>
|
| 189 |
+
</tr>
|
| 190 |
+
<tr>
|
| 191 |
+
<td class="col-model">UniNER-type+sup.</td>
|
| 192 |
+
<td class="col-backbone">LLaMA-1</td>
|
| 193 |
+
<td class="col-params">7B</td>
|
| 194 |
+
<td class="col-mit">61.2</td>
|
| 195 |
+
<td class="col-mit">35.2</td>
|
| 196 |
+
<td class="col-crossner">62.9</td>
|
| 197 |
+
<td class="col-crossner">64.9</td>
|
| 198 |
+
<td class="col-crossner">70.6</td>
|
| 199 |
+
<td class="col-crossner">66.9</td>
|
| 200 |
+
<td class="col-crossner">70.8</td>
|
| 201 |
+
<td class="col-buster">37.8</td>
|
| 202 |
+
<td class="col-avg">58.8</td>
|
| 203 |
+
</tr>
|
| 204 |
+
<tr>
|
| 205 |
+
<td class="col-model">GoLLIE</td>
|
| 206 |
+
<td class="col-backbone">Code-LLaMA</td>
|
| 207 |
+
<td class="col-params">7B</td>
|
| 208 |
+
<td class="col-mit">63.0</td>
|
| 209 |
+
<td class="col-mit">43.4</td>
|
| 210 |
+
<td class="col-crossner">59.1</td>
|
| 211 |
+
<td class="col-crossner">62.7</td>
|
| 212 |
+
<td class="col-crossner">67.8</td>
|
| 213 |
+
<td class="col-crossner">57.2</td>
|
| 214 |
+
<td class="col-crossner">55.5</td>
|
| 215 |
+
<td class="col-buster">27.7</td>
|
| 216 |
+
<td class="col-avg">54.6</td>
|
| 217 |
+
</tr>
|
| 218 |
+
<tr>
|
| 219 |
+
<td class="col-model">GLiNER-L</td>
|
| 220 |
+
<td class="col-backbone">DeBERTa-v3</td>
|
| 221 |
+
<td class="col-params">0.3B</td>
|
| 222 |
+
<td class="col-mit">57.2</td>
|
| 223 |
+
<td class="col-mit">42.9</td>
|
| 224 |
+
<td class="col-crossner">57.2</td>
|
| 225 |
+
<td class="col-crossner">64.4</td>
|
| 226 |
+
<td class="col-crossner">69.6</td>
|
| 227 |
+
<td class="col-crossner">72.6</td>
|
| 228 |
+
<td class="col-crossner">62.6</td>
|
| 229 |
+
<td class="col-buster">26.6</td>
|
| 230 |
+
<td class="col-avg">56.6</td>
|
| 231 |
+
</tr>
|
| 232 |
+
<tr>
|
| 233 |
+
<td class="col-model">GNER-T5</td>
|
| 234 |
+
<td class="col-backbone">Flan-T5-xxl</td>
|
| 235 |
+
<td class="col-params">11B</td>
|
| 236 |
+
<td class="col-mit">62.5</td>
|
| 237 |
+
<td class="col-mit">51.0</td>
|
| 238 |
+
<td class="col-crossner">68.2</td>
|
| 239 |
+
<td class="col-crossner">68.7</td>
|
| 240 |
+
<td class="col-crossner">81.2</td>
|
| 241 |
+
<td class="col-crossner">75.1</td>
|
| 242 |
+
<td class="col-crossner">76.7</td>
|
| 243 |
+
<td class="col-buster" style="color: red;">27.9</td>
|
| 244 |
+
<td class="col-avg">63.9</td>
|
| 245 |
+
</tr>
|
| 246 |
+
<tr>
|
| 247 |
+
<td class="col-model">GNER-LLaMA</td>
|
| 248 |
+
<td class="col-backbone">LLaMA-1</td>
|
| 249 |
+
<td class="col-params">7B</td>
|
| 250 |
+
<td class="col-mit">68.6</td>
|
| 251 |
+
<td class="col-mit">47.5</td>
|
| 252 |
+
<td class="col-crossner">63.1</td>
|
| 253 |
+
<td class="col-crossner">68.2</td>
|
| 254 |
+
<td class="col-crossner">75.7</td>
|
| 255 |
+
<td class="col-crossner">69.4</td>
|
| 256 |
+
<td class="col-crossner">69.9</td>
|
| 257 |
+
<td class="col-buster" style="color: red;">23.6</td>
|
| 258 |
+
<td class="col-avg">60.8</td>
|
| 259 |
+
</tr>
|
| 260 |
+
<tr>
|
| 261 |
+
<td class="col-model">SLIMER w/o D&G</td>
|
| 262 |
+
<td class="col-backbone">LLaMA-2-chat</td>
|
| 263 |
+
<td class="col-params">7B</td>
|
| 264 |
+
<td class="col-mit">46.4</td>
|
| 265 |
+
<td class="col-mit">36.3</td>
|
| 266 |
+
<td class="col-crossner">49.6</td>
|
| 267 |
+
<td class="col-crossner">58.4</td>
|
| 268 |
+
<td class="col-crossner">56.8</td>
|
| 269 |
+
<td class="col-crossner">57.9</td>
|
| 270 |
+
<td class="col-crossner">53.8</td>
|
| 271 |
+
<td class="col-buster">40.4</td>
|
| 272 |
+
<td class="col-avg">49.9</td>
|
| 273 |
+
</tr>
|
| 274 |
+
<tr>
|
| 275 |
+
<td class="col-model"><b>SLIMER</b></td>
|
| 276 |
+
<td class="col-backbone"><b>LLaMA-2-chat</b></td>
|
| 277 |
+
<td class="col-params"><b>7B</b></td>
|
| 278 |
+
<td class="col-mit"><b>50.9</b></td>
|
| 279 |
+
<td class="col-mit"><b>38.2</b></td>
|
| 280 |
+
<td class="col-crossner"><b>50.1</b></td>
|
| 281 |
+
<td class="col-crossner"><b>58.7</b></td>
|
| 282 |
+
<td class="col-crossner"><b>60.0</b></td>
|
| 283 |
+
<td class="col-crossner"><b>63.9</b></td>
|
| 284 |
+
<td class="col-crossner"><b>56.3</b></td>
|
| 285 |
+
<td class="col-buster"><b>45.3</b></td>
|
| 286 |
+
<td class="col-avg"><b>52.9</b></td>
|
| 287 |
+
</tr>
|
| 288 |
+
</tbody>
|
| 289 |
+
</table>
|
| 290 |
+
|
| 291 |
+
</body>
|
| 292 |
+
</html>
|
| 293 |
+
|
| 294 |
+
|
| 295 |
<table>
|
| 296 |
<thead>
|
| 297 |
<tr>
|