|
--- |
|
library_name: paddlenlp |
|
license: apache-2.0 |
|
language: |
|
- zh |
|
tags: |
|
- zero-shot-classification |
|
--- |
|
|
|
[](https://github.com/PaddlePaddle/PaddleNLP) |
|
|
|
# PaddlePaddle/utc-large |
|
|
|
Text classification technology is widely used in various industries such as dialogue intention recognition, bill archiving, and event detection. |
|
However, there are many challenges in industrial-level text classification practices, including diverse tasks, limited data availability and label transfer difficulty. |
|
To address these issues, UTC models text classification as a matching task between labels and text, based on the idea of Unified Semantic Matching (USM). |
|
Thus, it can handle multiple classification tasks with a single model, reducing development and machine costs and achieving good zero/few-shot transfer performance. |
|
Specifically, UTC won the 1st place on both [ZeroCLUE](https://www.cluebenchmarks.com/zeroclue.html) and [FewCLUE](https://www.cluebenchmarks.com/fewclue.html) benchmarks. |
|
|
|
|
|
USM Paper: https://arxiv.org/abs/2301.03282 |
|
|
|
PaddleNLP released UTC model for various text classification tasks which use ERNIE models as the pre-trained language models and were finetuned on a large amount of text classification data. |
|
|
|
|
|
![UTC-diagram]() |
|
|
|
![UTC-benchmarks]() |
|
|
|
## Available Models |
|
|
|
| Model Name | Usage Scenarios | Supporting Tasks | |
|
| :--------------: | :------------------------- | :---------------------------- | |
|
| `utc-large` | A **text classification** model supports **Chinese** | Supports intention recognition, semantic matching, natural language inference, semantic analysis, etc. | |
|
|
|
|
|
## Performance on Text Dataset |
|
|
|
We conducted experiments on the in-house test sets of |
|
|
|
|
|
|
|
**Detailed Info:** https://github.com/PaddlePaddle/PaddleNLP/tree/develop/applications/zero_shot_text_classification |
|
|
