File size: 16,694 Bytes
7a056d3 a4e88ec 7a056d3 9bab57b 7a056d3 9bab57b 7a056d3 9bab57b 7a056d3 1e12620 7a056d3 b106941 7a056d3 ff5c91a 7a056d3 93a422a 7a056d3 1e12620 a4e88ec 1e12620 a4e88ec 6354a79 7a056d3 42d788f 7a056d3 42d788f 7a056d3 4bb81d8 6a9e84a af3e2cf 4673016 7a056d3 6a9e84a 7a056d3 5237ac5 6a9e84a 7a056d3 5237ac5 6a9e84a 5237ac5 6a9e84a 7a056d3 6a9e84a 7a056d3 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 |
---
license: llama3.1
datasets:
- Universal-NER/Pile-NER-type
language:
- en
pipeline_tag: text-generation
tags:
- zero-shot NER
- NER
base_model:
- meta-llama/Llama-3.1-8B-Instruct
---
# SLIMER-PARALLEL-LLaMA3
SLIMER is an LLM specifically instructed for zero-shot NER on English language.
This LLaMA-3 based SLIMER scores __+17 %__ over paper's original SLIMER LLaMA-2,
while allowing up to 16 NEs to be extracted __in parallel__ per prompt.
GitHub repository: https://github.com/andrewzamai/SLIMER/tree/v2.0
SLIMER for Italian language can be found at: https://huggingface.co/expertai/LLaMAntino-3-SLIMER-IT
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 NEs to be extracted.
<!DOCTYPE html>
<html>
<head>
<title>Instruction Tuning Prompt</title>
<style>
.container {
border: none;
padding: 5px;
width: 300px;
margin: 0 auto;
font-family: Arial, sans-serif;
font-size: 8px;
border-radius: 10px; /* Rounded borders for container */
overflow: hidden; /* Ensure child elements respect container's rounded borders */
background-color: #f0f0f0
}
.header {
background-color: black;
color: white;
padding: 5px;
text-align: center;
font-weight: bold;
font-size: 14px;
border-top-left-radius: 10px; /* Rounded top-left corner */
border-top-right-radius: 10px; /* Rounded top-right corner */
}
.content {
padding: 5px;
}
.definition, {
padding: 5px;
border-radius: 10px; /* Rounded borders for definition and guidelines */
}
.definition {
background-color: #ffc773;
}
.footer {
background-color: black;
color: white;
padding: 10px;
font-weight: bold;
border-bottom-left-radius: 10px;
border-bottom-right-radius: 10px;
}
</style>
</head>
<body>
<div class="container">
<div class="header">SLIMER-3-PARALLEL prompt</div>
<div class="content">
<p><b><|start_header_id|>user<|end_header_id|></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 entities of type <b>[NEs_list]</b> from the text chunk you have read. Be aware that not all of these entities are necessarily present. Do not extract entities that do not exist in the text, return an empty list for that tag. Ensure each entity is assigned to only one appropriate class.</p>
<p>To help you, here are dedicated Definition and Guidelines for each entity tag.</p>
<div class="definition">
<p>{</p>
<p>{<b>"NE_type_1":</b> {"Definition": "", "Guidelines": ""}</p>
<p><b> ... </b>
<p>{<b>"NE_type_N":</b> {"Definition": "", "Guidelines": ""}</p>
<p>}</p>
</div>
<p>Return only a JSON object. The JSON should strictly follow this format: {"NE_type_1": [], ..., "NE_type_N":[]}. DO NOT output anything else, just the JSON itself.</p>
<p><b><|eot_id|><|start_header_id|>assistant<|end_header_id|></b></p>
</div>
<div class="footer"></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">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" style="color: red;">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" style="color: red;">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"><b>SLIMER</b></td>
<td class="col-backbone"><b>LLaMA-3.1-Instruct</b></td>
<td class="col-params"><b>8B</b></td>
<td class="col-mit"><b>58.4</b></td>
<td class="col-mit"><b>45.3</b></td>
<td class="col-crossner"><b>58.0</b></td>
<td class="col-crossner"><b>65.0</b></td>
<td class="col-crossner"><b>77.0</b></td>
<td class="col-crossner"><b>71.2</b></td>
<td class="col-crossner"><b>67.3</b></td>
<td class="col-buster"><b>39.32</b></td>
<td class="col-avg"><b>60.2</b></td>
</tr>
</tbody>
</table>
</body>
</html>
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<title>JSON Template</title>
<style>
body {
font-family: Arial, sans-serif;
line-height: 1.6;
padding: 20px;
}
.description {
font-weight: bold;
color: #333;
margin-bottom: 10px;
}
.template {
background-color: #f0f0f0;
padding: 10px;
border-radius: 5px;
margin-bottom: 20px;
}
.highlight-orange {
color: orange;
font-weight: bold;
}
</style>
</head>
<body>
<div class="description">JSON SLIMER prompt</div>
<div class="template">
<pre>{
"description": "SLIMER PARALLEL 3 prompt",
"prompt_input": "<|start_header_id|>system<|end_header_id|>\n\nYou are a helpful NER assistant designed to output JSON.<|eot_id|>\n<|start_header_id|>user<|end_header_id|>\n\nYou are given a text chunk (delimited by triple quotes) and an instruction.\nRead the text and answer to the instruction in the end.\n\"\"\"\n{<span class="highlight-orange">input</span>}\n\"\"\"\nInstruction: Extract the entities of type {ne_tags} from the text chunk you have read. Be aware that not all of these entities are necessarily present. Do not extract entities that do not exist in the text, return an empty list for that tag. Ensure each entity is assigned to only one appropriate class.\nTo help you, here are dedicated Definition and Guidelines for each entity tag.\n{Def_and_Guidelines}\nReturn only a JSON object. The JSON should strictly follow this format:\n{expected_json_format}.\nDO NOT output anything else, just the JSON itself."
}</pre>
</div>
</body>
</html>
```python
from vllm import LLM, SamplingParams
vllm_model = LLM(model="expertai/SLIMER-PARALLEL-LLaMA3")
tokenizer = vllm_model.get_tokenizer()
# suggested temperature 0, max_tokens hyperparam
cutoff_len = 4096
sampling_params = SamplingParams(temperature=0, max_tokens=1000, stop=tokenizer.eos_token)
# given list of NE types and dictionary of Def and Guidelines for each --> returns instruction
slimer_prompter = SLIMER_PARALLEL_instruction_prompter("SLIMER_PARALLEL_instruction_template", './src/SFT_finetuning/templates')
# create a dictionary of dictionaries, each NE_type as key should have a {Definition: str, Guidelines: str} value
ne_types_list = ['ORGANIZATION', 'UNIVERSITY', 'LOCATION', 'PERSON', 'CONFERENCE']
def_guidelines_per_NE_dict = {'ORGANIZATION': {'Definition': "'organization' refers to structured groups, institutions, companies, or associations.", 'Guidelines': "Avoid labeling generic terms like 'team' or 'group'. Exercise caution with ambiguous entities like 'Apple' (company vs. fruit) and 'Manchester United' (sports team vs. fan club)."}, 'UNIVERSITY': {'Definition': 'UNIVERSITY represents educational institutions that offer higher education and academic research programs.', 'Guidelines': "Avoid labeling general concepts such as 'education' or 'academia' as UNIVERSITY. Exercise caution with ambiguous terms like 'Cambridge' (can refer to different institutions) and 'Harvard' (can refer to a person)."}, 'LOCATION': {'Definition': 'LOCATION refers to specific geographic entities such as venues, facilities, and institutions that represent physical places with distinct addresses or functions.', 'Guidelines': "Exercise caution with ambiguous terms, e.g., 'Amazon' (company, river, and region) and 'Cambridge' (U.S. city, UK city, and university). Consider the context and specificity to accurately classify locations."}, 'PERSON': {'Definition': 'PERSON refers to individuals, including public figures, celebrities, and notable personalities.', 'Guidelines': 'If a person is working on research (including professor, Ph.D. student, researcher in companies, and etc) avoid labeling it as PERSON entity.'}, 'CONFERENCE': {'Definition': 'CONFERENCE refers to specific events or gatherings where experts, researchers, and professionals convene to present and discuss their work in a particular field or discipline.', 'Guidelines': "Exercise caution when labeling entities that could refer to institutions, organizations, or associations rather than specific events. Take care with ambiguous terms like 'International Journal of Computer Vision', which may refer to a publication rather than a conference."}}
instruction = slimer_prompter.generate_prompt(
ne_tags=", ".join(ne_types_list),
def_and_guidelines=json.dumps(def_guidelines_per_NE_dict, indent=2),
expected_json_format=json.dumps({k: [] for k in def_guidelines_per_NE_dict.keys()}, indent=2)
)
input_text = 'Typical generative model approaches include naive Bayes classifier s , Gaussian mixture model s , variational autoencoders and others .'
# this promper formats the input text to analize with SLIMER instruction
input_instruction_prompter = Prompter('LLaMA3-chat-NOheaders', template_path='./src/SFT_finetuning/templates')
system_message = "You are a helpful NER assistant designed to output JSON."
conversation = [
{"role": "system", "content": system_message},
{"role": "user", "content": input_instruction_prompter.generate_prompt(input=input_text, instruction=instruction)}, # the input_text + instruction
]
prompt = tokenizer.apply_chat_template(conversation, tokenize=False, truncation=True, max_length=cutoff_len, add_generation_prompt=True)
responses = vllm_model.generate(prompt, 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},
}
``` |