|
--- |
|
license: apache-2.0 |
|
library_name: paddlenlp |
|
language: |
|
- en |
|
- zh |
|
--- |
|
|
|
[](https://github.com/PaddlePaddle/PaddleNLP) |
|
|
|
# PaddlePaddle/uie-m-large |
|
|
|
Information extraction suffers from its varying targets, heterogeneous structures, and demand-specific schemas. The unified text-to-structure generation framework, namely UIE, can universally model different IE tasks, adaptively generate targeted structures, and collaboratively learn general IE abilities from different knowledge sources. Specifically, UIE uniformly encodes different extraction structures via a structured extraction language, adaptively generates target extractions via a schema-based prompt mechanism - structural schema instructor, and captures the common IE abilities via a large-scale pre-trained text-to-structure model. Experiments show that UIE achieved the state-of-the-art performance on 4 IE tasks, 13 datasets, and on all supervised, low-resource, and few-shot settings for a wide range of entity, relation, event and sentiment extraction tasks and their unification. These results verified the effectiveness, universality, and transferability of UIE. |
|
|
|
UIE Paper: https://arxiv.org/abs/2203.12277 |
|
|
|
PaddleNLP released UIE model series for Information Extraction of texts and multi-modal documents which use the ERNIE 3.0 models as the pre-trained language models and were finetuned on a large amount of information extraction data. |
|
|
|
 |
|
|
|
## Available Models |
|
|
|
| Model Name | Usage Scenarios | Supporting Tasks | |
|
| :----------------------------------------------------------: | :--------------------------------------------------------- | :--------------------------------------------------- | |
|
| `uie-base`<br />`uie-medium`<br />`uie-mini`<br />`uie-micro`<br />`uie-nano` | For **plain text** The **extractive** model of the scene supports **Chinese** | Supports entity, relation, event, opinion extraction | |
|
| `uie-base-en` | An **extractive** model for **plain text** scenarios, supports **English** | Supports entity, relation, event, opinion extraction | |
|
| `uie-m-base`<br />`uie-m-large` | An **extractive** model for **plain text** scenarios, supporting **Chinese and English** | Supports entity, relation, event, opinion extraction | |
|
| <b>`uie-x-base`</b> | An **extractive** model for **plain text** and **document** scenarios, supports **Chinese and English** | Supports entity, relation, event, opinion extraction on both plain text and documents/pictures/tables | |
|
|
|
|
|
## Performance on Text Dataset |
|
|
|
We conducted experiments on the in-house test sets of the three different domains of Internet, medical care, and finance: |
|
|
|
<table> |
|
<tr><th row_span='2'><th colspan='2'>finance<th colspan='2'>healthcare<th colspan='2'>internet |
|
<tr><td><th>0-shot<th>5-shot<th>0-shot<th>5-shot<th>0-shot<th>5-shot |
|
<tr><td>uie-base (12L768H)<td>46.43<td>70.92<td><b>71.83</b><td>85.72<td>78.33<td>81.86 |
|
<tr><td>uie-medium (6L768H)<td>41.11<td>64.53<td>65.40<td>75.72<td>78.32<td>79.68 |
|
<tr><td>uie-mini (6L384H)<td>37.04<td>64.65<td>60.50<td>78.36<td>72.09<td>76.38 |
|
<tr><td>uie-micro (4L384H)<td>37.53<td>62.11<td>57.04<td>75.92<td>66.00<td>70.22 |
|
<tr><td>uie-nano (4L312H)<td>38.94<td>66.83<td>48.29<td>76.74<td>62.86<td>72.35 |
|
<tr><td>uie-m-large (24L1024H)<td><b>49.35</b><td><b>74.55</b><td>70.50<td><b>92.66</b ><td>78.49<td><b>83.02</b> |
|
<tr><td>uie-m-base (12L768H)<td>38.46<td>74.31<td>63.37<td>87.32<td>76.27<td>80.13 |
|
<tr><td>🧾🎓<b>uie-x-base (12L768H)</b><td>48.84<td>73.87<td>65.60<td>88.81<td><b>79.36</b> <td>81.65 |
|
</table> |
|
|
|
0-shot means that no training data is directly used for prediction through paddlenlp.Taskflow, and 5-shot means that each category contains 5 pieces of labeled data for model fine-tuning. Experiments show that UIE can further improve the performance with a small amount of data (few-shot). |
|
|
|
## Performance on Multimodal Datasets** |
|
|
|
We experimented on the zero-shot performance of UIE-X on the in-house multi-modal test sets in three different domains of general, financial, and medical: |
|
|
|
<table> |
|
<tr><th ><th>General <th>Financial<th colspan='2'>Medical |
|
<tr><td>🧾🎓<b>uie-x-base (12L768H)</b><td>65.03<td>73.51<td>84.24 |
|
</table> |
|
|
|
The general test set contains complex samples from different fields and is the most difficult task. |
|
|
|
> Detailed Info: https://github.com/PaddlePaddle/PaddleNLP/blob/develop/applications/information_extraction/README_en.md |
