# MITweet

**Repository Path**: dufefintech/MITweet

## Basic Information

- **Project Name**: MITweet
- **Description**: No description available
- **Primary Language**: Python
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2024-05-25
- **Last Updated**: 2024-05-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# MITweet

Ideology Takes Multiple Looks: A High-Quality Dataset for Multifaceted Ideology Detection (EMNLP 2023)

## Multifaceted Ideology Schema

The multifaceted ideology schema contains **five domains** that reflect different aspects of society. Under the five domains, there are **twelve facets** with ideological attributes of left- and right-leaning.

<p align="center"><img src="pics/schema_table.png" alt="Multifaceted Ideology Schema" title="Multifaceted Ideology Schema" width="700" /></p>

<p align="center">Multifaceted Ideology Schema</p>

<p align="center"><img src="pics/schema.png" alt="Multifaceted Ideology Schema" title="Illustration of Multifaceted Ideology Schema" width="500" /></p>

<p align="center">Illustration of Multifaceted Ideology Schema</p>

## The MITweet Dataset

Based on the schema, we construct a new high-quality dataset, **MITweet**, for a new **multifaceted ideology detection (MID)** task. MITweet contains **12,594** English Twitter posts, each manually annotated with a Relevance label, and an Ideology label if the Relevance label is “Related”, along each facet. Meanwhile, MITweet covers 14 highly controversial topics in recent years (e.g., abortion, covid-19 and Russo-Ukrainian war).

### Label Distribution

<img src="pics/label distribution.png" alt="Label Distribution of MITweet" title="Label Distribution of MITweet" width="700" />

## Baselines

we develop baselines for the new MID task based on three widely-used PLMs (BERT, RoBERTa, BERTweet)  under both **in-topic** and **cross-topic** settings. We split the multifaceted ideology detection procedure into two sub-tasks in a pipeline manner:

1. **Relevance Recognition**
2. **Ideology Detection**

### In-topic Setting

<img src="pics/results_in-topic.png" alt="results_in-topic" title="results_in-topic" width="450" />

### Cross-topic Setting

<img src="pics/results_cross-topic.png" alt="results_cross-topic" title="results_cross-topic" style="zoom:70%;" />

## Reproduce

We provide the dataset and code for reproducing.

In the directory  `data` ,  `MITweet.csv`  is the complete dataset.

Each  `.csv`  data file contains the following columns: 

- `topic`
- `tweet`

- `tokenized tweet`  : tokenized tweets using the tweet segmentation tool in nltk
- `R1` ~ `R5` : relevance labels for the 5 domains.  `1`  means "Related",  `0`  means "Unelated"
- `R1-1-1` ~ `R512-5-3` : relevance labels for the 12 facets.  `1`  means "Related",  `0`  means "Unrelated"
- `I1` ~ `I12` : ideology labels for the 12 facets.  `0` , `1` , `2`  mean left-leaning, center, right-leaning, respectively.  `-1`  means "Unrelated", so no ideology label

### How to Run

- **Indicator Detection**

  ```
  python log_odds_ratio.py
  ```

- **Relevance Recognition**

  ```
  python train_relevance.py \
  	--train_data_path your_path \
  	--val_data_path your_path \
  	--test_data_path your_path
  ```

- **Ideology Detection**

  ```
  python train_ideology.py \
  	--train_data_path your_path \
  	--val_data_path your_path \
  	--test_data_path your_path \
  	--indicator_file_path your_path
  ```