Awesome

Directed Sentiment Analysis in News Text

This repository provides a dataset and code for extracting sentiment relationships between political entities in news text.

Task

Given a sentence s that contains two entities p and q, the problem aims to detect the sentiment relation from p to q among five classes: neutral, p holds a positive or negative opinion towards q, and the reverse direction.

For example, for a given sentence like Donald Trump emphatically blamed China for the coronavirus pandemic, a model is required to understand Trump is the source of the negative sentiment toward China by predicting a type among the five sentiment classes. The answer is "p holds a negative opinion toward q."

Data

We construct a dataset of 16,288 sentences by collecting news articles and crowdsourcing the annotation task. The class distribution is as follows.

We split the dataset into 13144, 1461, and 1623 instances for train, validation, and test set through a stratified split. We also provide resampled versions of training sets for experimental purposes.

Model

We present a new approach of utilizing pretrained BERT-like transformer models (e.g., RoBERTa). We transform the task into multiple sub-tasks to answer yes/no questions on whether a target sentiment is embedded in the text. The basic idea is we inquire about an intelligent machine that can answer yes/no questions on whether a target sentiment exists and then combine the answers corresponding to each sentiment class for making a final guess.

We present the overall framework in the following figure. Technically, taking auxiliary input in BERT-like transformers enables implementing the intelligent machine by making a different prediction with the same sentence input, according to the question fed as additional input.

Example Usage

python dse2qa.py \
  --input_type T \
  --resample up

Required parameters:

• input_type: augmented input's type ("T(emplate)" or "P(seudo)")
• resample: resampling method ("none", "up", "down")

Reference

You can get more information on the dataset and method in our upcoming paper in ACL'21. You can use the resources for your purposes by citing the paper.

@inproceedings{park2021directed,
  title={Who Blames or Endorses Whom? Entity-to-Entity Directed Sentiment Extraction in News Text},
  author={Park, Kunwoo and Pan, Zhufeng and Joo, Jungseock},
  booktitle={Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 13th International Joint Conference on Natural Language Processing},
  year={2021}
}