All Projects → harveyaot → Dianjing

harveyaot / Dianjing

Licence: mit
点睛 - 头条号文章标题生成工具 (Dianjing, AI to write Title for Articles)

Projects that are alternatives of or similar to Dianjing

Mldl
Machine Learning and Deep Learning Resources
Stars: ✭ 211 (-1.4%)
Mutual labels:  jupyter-notebook
Python Business Analytics
Python solutions to solve practical business problems.
Stars: ✭ 209 (-2.34%)
Mutual labels:  jupyter-notebook
Tensorflow Without A Phd
A crash course in six episodes for software developers who want to become machine learning practitioners.
Stars: ✭ 2,488 (+1062.62%)
Mutual labels:  jupyter-notebook
Deepconvlstm
Deep learning framework for wearable activity recognition based on convolutional and LSTM recurretn layers
Stars: ✭ 211 (-1.4%)
Mutual labels:  jupyter-notebook
Statistical Learning Method Solutions Manual
《统计学习方法》(第一版)习题解答,在线阅读地址:https://datawhalechina.github.io/statistical-learning-method-solutions-manual
Stars: ✭ 211 (-1.4%)
Mutual labels:  jupyter-notebook
Tfwss
Weakly Supervised Segmentation with Tensorflow. Implements instance segmentation as described in Simple Does It: Weakly Supervised Instance and Semantic Segmentation, by Khoreva et al. (CVPR 2017).
Stars: ✭ 212 (-0.93%)
Mutual labels:  jupyter-notebook
Coloring T Sne
Exploration of methods for coloring t-SNE.
Stars: ✭ 211 (-1.4%)
Mutual labels:  jupyter-notebook
Bitcoin prediction
This is the code for "Bitcoin Prediction" by Siraj Raval on Youtube
Stars: ✭ 214 (+0%)
Mutual labels:  jupyter-notebook
Godot oculus quest toolkit
An easy to use VR toolkit for Oculus Quest development using the Godot game engine
Stars: ✭ 207 (-3.27%)
Mutual labels:  jupyter-notebook
Coursera Stanford
Stanford
Stars: ✭ 212 (-0.93%)
Mutual labels:  jupyter-notebook
Best Of Jupyter
🏆 A ranked list of awesome Jupyter Notebook, Hub and Lab projects (extensions, kernels, tools). Updated weekly.
Stars: ✭ 200 (-6.54%)
Mutual labels:  jupyter-notebook
Python Examples
Python技术 公众号文章代码实例
Stars: ✭ 212 (-0.93%)
Mutual labels:  jupyter-notebook
Skift
scikit-learn wrappers for Python fastText.
Stars: ✭ 213 (-0.47%)
Mutual labels:  jupyter-notebook
Deeplearning
Some practices about deep learning
Stars: ✭ 210 (-1.87%)
Mutual labels:  jupyter-notebook
Machine Learning Interview Enlightener
This repo is meant to serve as a guide for Machine Learning/AI technical interviews.
Stars: ✭ 207 (-3.27%)
Mutual labels:  jupyter-notebook
Mdrepo
Repositório para armazenamento de código e notebooks de postagens do blog e cursos.
Stars: ✭ 210 (-1.87%)
Mutual labels:  jupyter-notebook
Squad
Building QA system for Stanford Question Answering Dataset
Stars: ✭ 213 (-0.47%)
Mutual labels:  jupyter-notebook
Pythonnumericaldemos
Well-documented Python demonstrations for spatial data analytics, geostatistical and machine learning to support my courses.
Stars: ✭ 213 (-0.47%)
Mutual labels:  jupyter-notebook
Kekoxtutorial
전 세계의 멋진 케라스 문서 및 튜토리얼을 한글화하여 케라스x코리아를 널리널리 이롭게합니다.
Stars: ✭ 213 (-0.47%)
Mutual labels:  jupyter-notebook
Neural decoding
A python package that includes many methods for decoding neural activity
Stars: ✭ 212 (-0.93%)
Mutual labels:  jupyter-notebook

DianJing

点睛 - 头条号文章标题生成工具

  1. 功能: 自动为头条的文章生成一个题目候选列表(Automatically Generate Article Title in TouTiao Style)
  2. 展现形式: 初期是linux 的客户端,后期开发一个前端页面,或者一个chrome 插件的形式存在。
  3. 主要技术: 使用encoder-decoder的技术对 头条的 摘要 和 文章对(abstract-title pair)标题进行训练
  4. 数据来源: 主要使用 头条的 数据接口,抓万级别的训练样本。

数据使用和爬取 Data Usage and Crawl

  1. 可用的训练数据,./data/basic_data_80k_v2.pkl.tgz 包含了约61K的(abstract, title)数据。通过tar -xzvf basic_data_80k_v2.pkl.tgz 解压之后,可以使用 ./scripts 下的 data_utils.py 来check 数据数量和展示sample的样本。
  2. 另有一份large data set 包含 700k左右的training 样本, 将稍候公开 (ETA Mar. 2018)。
  3. 使用 ./scripts/crawl.py 来爬取头条数据,但是需要指定头条feed 流中的 as 和 cp 两个参数,这两个参数,最好每三天更新一次,获取方法如下 从chrom 浏览器的 network 中可以看到最新feed 流地址的这两个参数

实验日志 Experiment Log

  1. 2017/05/27 使用大约30K的训练样本,摘要-标题对,对每个汉字做100 维 embeding 使用CNN做encoder,GRU unit 的RNN 做decoer. 一天500个epoch 之后训练效果如下 :
    • 分析:
      • 基本可以分析出描述中的关键语义
      • 但是语言可读性较差
    • 改进方向
      • 训练样本可能不足
      • 基于中文分词做,不是汉字粒度
      • LSTM 在生成长文本上的能力并不好,可以考虑基于大量语料库的language model
  2. 2017/06/01
    • 提升:
      • 使用search api 爬去了8000(dict/keywords.select)关键字的600K 文章
      • 使用jieba 进行分词,进行
      • 在30K 训练样本上的语言可读性提高
    • 问题:
      • 在600K 数据上OOV 问题严重,模型难以收敛
    • 解决:
      • 增大vocabulary size
      • 先训练rnn 的language model,能说好话
      • 然后是基于condition的条件下说话
  3. 2017/06/20
    • 这预测标题写的,以假乱真,不去查一下,还真的被他骗了
    • 分析:
      • 增大vacobulary size 到 212K
      • pre-train rnn 的decoder,然后标题可读性大大提升
      • 700k 的training samples, 对语义的理解准确性提升
    • 提升:
      • 调低learning rate 和 batch_size 增加探索能力
      • 考虑如何解决 OOV 的问题
      • 主要是push 当前的 search,abstract-embedding,AI,三个版本上线
Note that the project description data, including the texts, logos, images, and/or trademarks, for each open source project belongs to its rightful owner. If you wish to add or remove any projects, please contact us at [email protected].