News sentiment development on the example of “Migration”

Analyzing the sentiment development of news articles with the topic "migration" over time. This project was done in the course of the Text Analytics lecture at Heidelberg University.

Code structure

• pipeline.py: All main components of the project are included in the pipeline. The pipeline is controlled by the config.ini.

Pipeline components

• article_selection: Selects articles which are relevant for the analysis by a keyword search

• sentiment_analysis: Contains the different approaches to analyze the sentiment

  • bert.py: Evaluation of sentiment by the BERT model
  • sentiment_dictionary.py: Evaluation of sentiment through the SentiWS dictionary
  • negation_handling.py: Improvement of the dictionary approach by handling the negation of words
  • word2vec_sentiment.py: Word2Vec Model to get synonyms of search words for qualitative analysis
  • inference.py: functions for applying and evaluating sentiment analysis methods on large batches of data

• training: Training code to fine-tune the BERT Model (the result of the training is published here)

• visualization:

  • dash_plot.py: Dash application to show sentiment timelines
  • wordcloud.py: Generates word clouds of results from word2vec model

Other components

• annotation: Tool to annotate articles to generate training and test data
• scraping: Generating the articles
• pipeline_test.py: code tests

Setup Instructions:

Setup requirements: Linux, Python 3.8

1. Clone this repository

2. Create a new virtual environment and activate it:

    virtualenv env
    source env/bin/activate
   

3. Install the dependencies from the frozen-requirements.txt and then install the german language-package for spacy:

   pip install -r frozen-requirements.txt
   python -m spacy download de
   

Instructions to run the code

Obtain the news article data dataset

Either ask us for the scraped articles or use scraping/collect_articles.py to build the dataset yourself (can take a few days). For detailed instructions, refer to the docstring of collect_articles.py.

The expected article source files (*-sources.txt files) can be obtained from https://wortschatz.uni-leipzig.de/en/download/German). They are located inside of the .tar.gz files listed there. For this project we used the following archives:

• deu_news_2007_100k.tar.gz
• deu_news_2008_100k.tar.gz
• deu_news_2009_100k.tar.gz
• deu_news_2010_100k.tar.gz
• deu_news_2011_100k.tar.gz
• deu_news_2012_100k.tar.gz
• deu_news_2013_100k.tar.gz
• deu_news_2014_100k.tar.gz
• deu_news_2015_100k.tar.gz
• deu_newscrawl_2017_100k.tar.gz
• deu_newscrawl_2018_100k.tar.gz
• deu_newscrawl-public_2019_100k.tar.gz

Run the pipeline

The pipeline is controlled by the config.ini file. Configure it as you wish. Then run

pipeline.py config.ini

Note on data availability

We cannot upload our article data publicly due to copyright reasons. If you are interested in our dataset version and/or in the intermediate results, please email us so we can help you. Our finetuned BERT model can be found at https://huggingface.co/mdraw/german-news-sentiment-bert

Project report

We share a redacted version of our final project report here. Please refer to this document for more details on the background, methods and results of the associated project for which the code was written.

Team members

• Simon Lüdke (simon.luedke at gmx.de)
• Josephine Grau (josephine.grau at web.de)
• Martin Drawitsch (martin.drawitsch at gmail.com)