https://github.com/LLluoling/PENS-Personalized-News-Headline-Generation

Code for PENS: A Dataset and Generic Framework for Personalized News Headline Generation
https://github.com/LLluoling/PENS-Personalized-News-Headline-Generation

Last synced: 3 months ago
JSON representation

Code for PENS: A Dataset and Generic Framework for Personalized News Headline Generation

• Host: GitHub
• URL: https://github.com/LLluoling/PENS-Personalized-News-Headline-Generation
• Owner: LLluoling
• Created: 2021-07-26T16:39:18.000Z (over 3 years ago)
• Default Branch: main
• Last Pushed: 2022-11-10T09:09:57.000Z (almost 2 years ago)
• Last Synced: 2024-06-24T05:36:28.524Z (5 months ago)
• Language: Python
• Size: 129 KB
• Stars: 32
• Watchers: 3
• Forks: 10
• Open Issues: 5
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

• StarryDivineSky - PENS-Personalized-News-Headline-Generation

README

        
# PENS - ACL2021

## {PENS}: A Dataset and Generic Framework for Personalized News Headline Generation

This is a Pytorch implementation of [PENS](https://www.microsoft.com/en-us/research/uploads/prod/2021/06/ACL2021_PENS_Camera_Ready_1862_Paper.pdf). 

## I. Guidance

### 0. Enviroment

- Install pytorch version >= '1.4.0'

- Install the pensmodule package under ''PENS-Personalized-News-Headline-Generation'' using code ``` pip install -e . ```

### 1. Data Prepare

- Download the PENS dataset [here](https://msnews.github.io/pens.html) and put the dataset under data/.

- (optional) Download glove.840B.300d.txt under data/ if you choose to use pretrained glove word embeddings.

### 2. Running Code

- ```cd pensmodule ```

- Follow the order: Preprocess --> UserEncoder --> Generator and run the pipeline**.ipynb notebook to preprocess, train the user encoder and the train generator, individually.

More infor please refer to the homepage of the [introduction of PENS dataset](https://msnews.github.io/pens.html).

## II. Training Tips

Here we take NRMS as user encoder, the followings are some experiment detailes that are not illustrated in the paper.

### 0. TIPS

- In this paper, we used mento carlo search for RL training, which is very slow in training and sometimes hard to converge. Thus we provide ac training in this provided code.

- If you pretrain the generator for a couple of epoches, you should set a very small learning rate during RL training.

- **Large improvements can be made compared with the baselines that we provided, the importance always lies in the design of reward functions.**

### 1. Training Reward

![image info](./docs/reward.png)

### 2. Test performance on different training steps

![image info](./docs/rouge1.png)

![image info](./docs/rouge2.png)

![image info](./docs/rougel.png)

### 3. Cases

| epoch | generated headline | 

| :-----| ----: |

| Case 1 |  | 

| 1000 | top stockton news arrests 2 impaired drivers | 

| 5000 | top stockton news arrests 2 impaired drivers who had unrestrained children in their cars | 

| Case 2 |  | 

| 1000 | trump says tens of thousands of people couldn t get in 2020 rally | 

| 5000 | trump says tens of thousands of people outside his 2020 campaign rally at orlando | 

**Noted:**

- With the training process goes, the generated sentences are more fluent and contains more rich information.

-  **Rouge scores is not the best evaluation scores, but a compromising choice. Of course the best evaluation is to check out the real clicks of users to see if they are more interested. Thus sometimes a more fluent and human-like generated sentence gets lower rouge scores.**