File size: 13,639 Bytes
8868390
5391ab8
8868390
 
 
 
 
 
 
 
5391ab8
 
8868390
 
5cf3fa1
8868390
b4d7717
81f342b
639cb91
c7e5ddd
 
8868390
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
2331e72
8868390
 
 
 
 
 
 
 
 
2331e72
8868390
2331e72
8868390
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5391ab8
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
---
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-3: Show Less Instruct More Entity Recognition LLaMA3

This LLaMA-3 based model scores +12.5 % wrt paper's original SLIMER LLaMA-2.

⚠️ A more powerful model, which scores +17% and allows parallel NEs extraction, can be found at:
https://huggingface.co/expertai/SLIMER-PARALLEL-LLaMA3

GitHub repository: https://github.com/andrewzamai/SLIMER

SLIMER is an LLM specifically instructed for zero-shot NER on English language.

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 NE 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, .guidelines {
            padding: 5px;
            border-radius: 10px; /* Rounded borders for definition and guidelines */
        }
        .definition {
            background-color: #ffc773;
        }
        .guidelines {
            background-color: #73d7ff;
        }
        .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">Instruction Tuning 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 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 (for example [\"text_span_1\", \"text_span_2\"]. Return an empty list [] if no instances are present. Return only the JSON list, no further motivations or introduction to the answer.</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>56.4</b></td>
            <td class="col-mit"><b>44.8</b></td>
            <td class="col-crossner"><b>55.6</b></td>
            <td class="col-crossner"><b>63.3</b></td>
            <td class="col-crossner"><b>68.8</b></td>
            <td class="col-crossner"><b>69.2</b></td>
            <td class="col-crossner"><b>67.3</b></td>
            <td class="col-buster"><b>45.9</b></td>
            <td class="col-avg"><b>59.0</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 prompt",
  "prompt_input": "<|start_header_id|>system<|end_header_id|>\n\nYou are an expert in Named Entity Recognition designed to output JSON only.<|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 Named Entities of type {<span class="highlight-orange">NE_name</span>} from the text chunk you have read. You are given a DEFINITION and some GUIDELINES.\nDEFINITION: {<span class="highlight-orange">definition</span>}\nGUIDELINES: {<span class="highlight-orange">guidelines</span>}\nReturn a JSON list of instances of this Named Entity type (for example [\"text_span_1\", \"text_span_2\"]. Return an empty list [] if no instances are present. Return only the JSON list, no further motivations or introduction to the answer.<|eot_id|>\n<|start_header_id|>assistant<|end_header_id|>\n\n"
}</pre>
  </div>
</body>
</html>


```python
from vllm import LLM, SamplingParams

vllm_model = LLM(model="expertai/SLIMER-LLaMA3")

sampling_params = SamplingParams(temperature=0, max_tokens=128)

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}, 
}
```