广告文案生成数据集

## 数据集描述
AdvertiseGen以商品网页的标签与文案的信息对应关系为基础构造，是典型的开放式生成任务，在模型基于key-value输入生成开放式文案时，与输入信息的事实一致性需要得到重点关注。

## 数据预览

任务描述：给定商品信息的关键词和属性列表kv-list，生成适合该商品的广告文案adv；
数据规模：训练集114k，验证集1k，测试集3k；
数据来源：清华大学CoAI小组；
数据样例：
```
{
"content": "上衣 牛仔布 白色 简约 刺绣 外套 破洞",
"summary": "简约而不简单的牛仔外套，白色的衣身十分百搭。衣身多处有做旧破洞设计，打破单调乏味，增加一丝造型看点。衣身后背处有趣味刺绣装饰，丰富层次感，彰显别样时尚。"
}
```

## 基线系统

本数据集提供的基线系统，基于百度提出的ERNIE-UNIMO统一模态预训练框架。在本次比赛的三个文本生成任务中，我们基于本基线使用的模型是UNIMO-text,是基于ERNIE-UNIMO框架在文本数据上预训练得到模型。

## Dataset Description
AdvertiseGen is constructed based on the correspondence between product webpage tags and advertising copy information, which is a typical open-ended generation task. When the model generates open-ended copy based on key-value inputs, the factual consistency with the input information requires focused attention.

## Data Preview

Task Description: Given the keyword and attribute list (kv-list) of product information, generate appropriate advertising copy adv for the product;
Data Scale: 114k training samples, 1k validation samples, 3k test samples;
Data Source: CoAI Group of Tsinghua University;
Data Example:

{
"content": "Top, denim, white, minimalist, embroidery, jacket, ripped",
"summary": "A minimalist denim jacket that is far from ordinary. The white body is extremely versatile. Multiple worn and ripped designs are added to the jacket, breaking the monotony and adding a touch of visual appeal. Interesting embroidery decorations are placed on the back of the jacket, enriching the layering and showcasing a unique fashion sense."
}

## Baseline System

The baseline system provided for this dataset is based on the ERNIE-UNIMO unified-modal pre-training framework proposed by Baidu. Among the three text generation tasks in this competition, the model we used based on this baseline is UNIMO-text, which is a model pre-trained on text data using the ERNIE-UNIMO framework.

搜集汇总

背景与挑战

背景概述

广告文案生成数据集是一个用于开放式生成任务的数据集，基于商品网页的标签与文案的信息对应关系构造，包含大量训练、验证和测试数据。任务要求根据商品关键词和属性列表生成广告文案，基线系统基于ERNIE-UNIMO框架。