Willinki / Just-News

Small project on word embeddings, using Compass Aligned Distributional Embeddings.

Just-News

The aim of the project is to analyze the language used by famous american newspapers using distributional methods.

Our research focuses on the use of subjective language, opposed to a more neutral* and distant writing style.

In other words, we are trying to determine which newspapers tend to rely on emotions and morals, rather than merely reporting facts.

Results presented in data_semantics.pdf

Data

We conduct our research on the following newspapers:

• New York Times
• CNN
• ABC News
• Breitbart
• Slate
• The Federalist
• NewsMax

Our dataset consists of around 40000 articles, plus ~20000 Wikipedia pages, that we use as control corpus, under the hypothesis that it uses neutral language.

Phases

Here are presented the phases of our work.

• Data retrieval. Consists in getting the data from various sources (mainly web-scraping). The code related to this phase is not included in the repo. A link shall be added once some clean up process has been executed.
• Data cleaning and preprocessing. Consist of creating a dataset suitable for analysis. (CADE_testing.ipynb).
• CADE training. We make use of CADE in order to obtain aligned word-embeddings, one for each newspaper. (CADE_testing.ipynb).
• Score induction. In order to obtain a subjectivity score for the words in our vocabulary, we apply a score induction procedure. This procedure allow us to propagate the score coming from a an annotated lexicon to all the words in the vocabulary.
• Score propagation. How do we determine the subjectivity score of an entire article, starting from the score of the words in it? This process will require some manual benchmarking.
• Final results and visualization.
• (Extra) Following the indications contained in this article, we could push on the the aspect of interpretability of results.

Modifications to SWEAT

• enrich and lexicon_refinement inside utils.py need to be updated to be used in more than 2 models.
• Also, we must try to relax some of the constraints in the lexicon refinement function. For example, we might set a lower value of zipf_cutoff. Another option is to set topn = n inside the building of the comparative framework, thus maintaining the n nearest across the various representations, instead of just one.

Note

* We are aware of the fact that there is no such thing as a neutral way of reporting news, and that every journal and journalist tend to have its own views. Our research, in fact, focuses on the use of language only.