File size: 12,173 Bytes
26b46c0 fc743ec 26b46c0 d069798 fc743ec e0ae669 5dc088a b5c1274 5b28de0 30be815 0fcd9a7 e61d685 30be815 6a9c26c 30be815 0fcd9a7 30be815 0fcd9a7 30be815 497b810 e0ae669 7068ade e0ae669 2d7e780 4946b33 12a6fed e42769f 12a6fed f4fa6dc 51cc6c6 0979f85 51cc6c6 |
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 |
---
license: llama2
datasets:
- Universal-NER/Pile-NER-type
language:
- en
pipeline_tag: text-generation
---
# SLIMER: Show Less Instruct More Entity Recognition
SLIMER is an LLM specifically instructed for zero-shot NER.
Instructed on a reduced number of samples, it is designed to tackle never-seen-before Named Entity tags by leveraging a prompt enriched with a DEFINITION and GUIDELINES for the NE to be extracted.
<!DOCTYPE html>
<html>
<head>
<style>
.container {
border: 2px solid #000;
padding: 5px;
width: 300px;
margin: 0 auto;
font-family: Arial, sans-serif;
font-size: 8px;
}
.header {
background-color: black;
color: white;
padding: 5px;
text-align: center;
font-weight: bold;
font-size: 14px;
}
.content {
padding: 5px;
}
.definition, .guidelines {
padding: 5px;
}
.definition {
background-color: #ccffcc;
}
.guidelines {
background-color: #ffffcc;
}
</style>
</head>
<body>
<div class="container">
<div class="header">Instruction Tuning Prompt</div>
<div class="content">
<p><b>[INST]</b></p>
<p>You are given a text chunk (delimited by triple quotes) and an instruction.<br>
Read the text and answer to the instruction in the end.</p>
<p>"""<br>
{input text}<br>
"""</p>
<p><b>Instruction:</b> Extract the Named Entities of type <b>DATE</b> from the text chunk you have read.</p>
<p>You are given a <b>DEFINITION</b> and some <b>GUIDELINES</b>.</p>
<div class="definition">
<p><b>DEFINITION:</b> <b>DATE</b> refers to specific points in time, including days, months, years, and relative time expressions like 'Week 2'.</p>
</div>
<div class="guidelines">
<p><b>GUIDELINES:</b> Avoid labeling non-specific time references like 'recently' or 'soon'. Exercise caution with ambiguous terms like 'May' (month or verb) and 'Wednesday Adams' (person's name which includes a day of the week).</p>
</div>
<p>Return a JSON list of instances of this Named Entity type. Return an empty list if no instances are present.</p>
<p><b>[/INST]</b></p>
</div>
</div>
</body>
</html>
Currently existing approaches fine-tune on an extensive number of entity classes (around 13K) and assess zero-shot NER capabilities on Out-Of-Distribution input domains.
SLIMER performs comparably to these state-of-the-art models on OOD input domains, while being trained only a reduced number of samples and a set of NE tags that overlap in lesser degree with test sets.
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.
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.
<!DOCTYPE html>
<html>
<head>
<style>
table {
width: 100%;
border-collapse: collapse;
font-size: 12px;
}
th, td {
border: 1px none;
padding: 4px;
text-align: center;
}
th {
background-color: #f2f2f2;
}
.col-model { width: 10%; }
.col-backbone { width: 15%; }
.col-params { width: 10%; }
.col-mit, .col-crossner, .col-buster, .col-avg { width: 7%; }
</style>
</head>
<body>
<table>
<thead>
<tr>
<th class="col-model">Model</th>
<th class="col-backbone">Backbone</th>
<th class="col-params">#Params</th>
<th class="col-mit" colspan="2">MIT</th>
<th class="col-crossner" colspan="5">CrossNER</th>
<th class="col-buster">BUSTER</th>
<th class="col-avg">AVG</th>
</tr>
<tr>
<th></th>
<th></th>
<th></th>
<th class="col-mit">Movie</th>
<th class="col-mit">Restaurant</th>
<th class="col-crossner">AI</th>
<th class="col-crossner">Literature</th>
<th class="col-crossner">Music</th>
<th class="col-crossner">Politics</th>
<th class="col-crossner">Science</th>
<th class="col-buster"></th>
<th class="col-avg"></th>
</tr>
</thead>
<tbody>
<tr>
<td class="col-model">ChatGPT</td>
<td class="col-backbone">gpt-3.5-turbo</td>
<td class="col-params">-</td>
<td class="col-mit">5.3</td>
<td class="col-mit">32.8</td>
<td class="col-crossner">52.4</td>
<td class="col-crossner">39.8</td>
<td class="col-crossner">66.6</td>
<td class="col-crossner">68.5</td>
<td class="col-crossner">67.0</td>
<td class="col-buster">-</td>
<td class="col-avg">-</td>
</tr>
<tr>
<td class="col-model">InstructUIE</td>
<td class="col-backbone">Flan-T5-xxl</td>
<td class="col-params">11B</td>
<td class="col-mit">63.0</td>
<td class="col-mit">21.0</td>
<td class="col-crossner">49.0</td>
<td class="col-crossner">47.2</td>
<td class="col-crossner">53.2</td>
<td class="col-crossner">48.2</td>
<td class="col-crossner">49.3</td>
<td class="col-buster">-</td>
<td class="col-avg">-</td>
</tr>
<tr>
<td class="col-model">UniNER-type</td>
<td class="col-backbone">LLaMA-1</td>
<td class="col-params">7B</td>
<td class="col-mit">42.4</td>
<td class="col-mit">31.7</td>
<td class="col-crossner">53.5</td>
<td class="col-crossner">59.4</td>
<td class="col-crossner">65.0</td>
<td class="col-crossner">60.8</td>
<td class="col-crossner">61.1</td>
<td class="col-buster">34.8</td>
<td class="col-avg">51.1</td>
</tr>
<tr>
<td class="col-model">UniNER-def</td>
<td class="col-backbone">LLaMA-1</td>
<td class="col-params">7B</td>
<td class="col-mit">27.1</td>
<td class="col-mit">27.9</td>
<td class="col-crossner">44.5</td>
<td class="col-crossner">49.2</td>
<td class="col-crossner">55.8</td>
<td class="col-crossner">57.5</td>
<td class="col-crossner">52.9</td>
<td class="col-buster">33.6</td>
<td class="col-avg">43.6</td>
</tr>
<tr>
<td class="col-model">UniNER-type+sup.</td>
<td class="col-backbone">LLaMA-1</td>
<td class="col-params">7B</td>
<td class="col-mit">61.2</td>
<td class="col-mit">35.2</td>
<td class="col-crossner">62.9</td>
<td class="col-crossner">64.9</td>
<td class="col-crossner">70.6</td>
<td class="col-crossner">66.9</td>
<td class="col-crossner">70.8</td>
<td class="col-buster">37.8</td>
<td class="col-avg">58.8</td>
</tr>
<tr>
<td class="col-model">GoLLIE</td>
<td class="col-backbone">Code-LLaMA</td>
<td class="col-params">7B</td>
<td class="col-mit">63.0</td>
<td class="col-mit">43.4</td>
<td class="col-crossner">59.1</td>
<td class="col-crossner">62.7</td>
<td class="col-crossner">67.8</td>
<td class="col-crossner">57.2</td>
<td class="col-crossner">55.5</td>
<td class="col-buster">27.7</td>
<td class="col-avg">54.6</td>
</tr>
<tr>
<td class="col-model">GLiNER-L</td>
<td class="col-backbone">DeBERTa-v3</td>
<td class="col-params">0.3B</td>
<td class="col-mit">57.2</td>
<td class="col-mit">42.9</td>
<td class="col-crossner">57.2</td>
<td class="col-crossner">64.4</td>
<td class="col-crossner">69.6</td>
<td class="col-crossner">72.6</td>
<td class="col-crossner">62.6</td>
<td class="col-buster">26.6</td>
<td class="col-avg">56.6</td>
</tr>
<tr>
<td class="col-model">GNER-T5</td>
<td class="col-backbone">Flan-T5-xxl</td>
<td class="col-params">11B</td>
<td class="col-mit">62.5</td>
<td class="col-mit">51.0</td>
<td class="col-crossner">68.2</td>
<td class="col-crossner">68.7</td>
<td class="col-crossner">81.2</td>
<td class="col-crossner">75.1</td>
<td class="col-crossner">76.7</td>
<td class="col-buster" style="color: red;">27.9</td>
<td class="col-avg">63.9</td>
</tr>
<tr>
<td class="col-model">GNER-LLaMA</td>
<td class="col-backbone">LLaMA-1</td>
<td class="col-params">7B</td>
<td class="col-mit">68.6</td>
<td class="col-mit">47.5</td>
<td class="col-crossner">63.1</td>
<td class="col-crossner">68.2</td>
<td class="col-crossner">75.7</td>
<td class="col-crossner">69.4</td>
<td class="col-crossner">69.9</td>
<td class="col-buster" style="color: red;">23.6</td>
<td class="col-avg">60.8</td>
</tr>
<tr>
<td class="col-model">SLIMER w/o D&G</td>
<td class="col-backbone">LLaMA-2-chat</td>
<td class="col-params">7B</td>
<td class="col-mit">46.4</td>
<td class="col-mit">36.3</td>
<td class="col-crossner">49.6</td>
<td class="col-crossner">58.4</td>
<td class="col-crossner">56.8</td>
<td class="col-crossner">57.9</td>
<td class="col-crossner">53.8</td>
<td class="col-buster">40.4</td>
<td class="col-avg">49.9</td>
</tr>
<tr>
<td class="col-model"><b>SLIMER</b></td>
<td class="col-backbone"><b>LLaMA-2-chat</b></td>
<td class="col-params"><b>7B</b></td>
<td class="col-mit"><b>50.9</b></td>
<td class="col-mit"><b>38.2</b></td>
<td class="col-crossner"><b>50.1</b></td>
<td class="col-crossner"><b>58.7</b></td>
<td class="col-crossner"><b>60.0</b></td>
<td class="col-crossner"><b>63.9</b></td>
<td class="col-crossner"><b>56.3</b></td>
<td class="col-buster"><b>45.3</b></td>
<td class="col-avg"><b>52.9</b></td>
</tr>
</tbody>
</table>
</body>
</html>
```python
from vllm import LLM, SamplingParams
vllm_model = LLM(model="expertai/SLIMER")
sampling_params = SamplingParams(temperature=0, max_tokens=128, stop=['</s>'])
prompts = [prompter.generate_prompt(instruction, input) for instruction, input in instruction_input_pairs]
responses = vllm_model.generate(prompts, sampling_params)
```
## Citation
If you find SLIMER useful in your research or work, please cite the following paper:
``` latex
@misc{zamai2024lessinstructmoreenriching,
title={Show Less, Instruct More: Enriching Prompts with Definitions and Guidelines for Zero-Shot NER},
author={Andrew Zamai and Andrea Zugarini and Leonardo Rigutini and Marco Ernandes and Marco Maggini},
year={2024},
eprint={2407.01272},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2407.01272},
}
``` |