• Origin:
  • Crowdflower Data for Everyone dataset: https://www.crowdflower.com/data-for-everyone/
  • https://www.kaggle.com/crowdflower/political-social-media-posts/data
• Meta-data:
  • _unit_id: a unique id for the message
  • _golden: always FALSE; (presumably whether the message was in Crowdflower's gold standard)
  • _unit_state: always "finalized"
  • _trusted_judgments: the number of trusted human judgments that were entered for this message; an integer between 1 and 3
  • _last_judgment_at: when the final judgment was collected
  • audience: one of national or constituency
  • audience:confidence: a measure of confidence in the audience judgment; a float between 0.5 and 1
  • bias: one of neutral or partisan
  • bias:confidence: a measure of confidence in the bias judgment; a float between 0.5 and 1
  • message: the aim of the message. one of:
    • attack: the message attacks another politician
    • constituency: the message discusses the politician's constituency
    • information: an informational message about news in government or the wider U.S.
    • media: a message about interaction with the media
    • mobilization: a message intended to mobilize supporters
    • other: a catch-all category for messages that don't fit into the other
    • personal: a personal message, usually expressing sympathy, support or condolences, or other personal opinions
    • policy: a message about political policy
    • support: a message of political support
  • message:confidence: a measure of confidence in the message judgment; a float between 0.5 and 1
  • orig__golden: always empty; presumably whether some portion of the message was in the gold standard
  • audience_gold: always empty; presumably whether the audience response was in the gold standard
  • bias_gold: always empty; presumably whether the bias response was in the gold standard
  • bioid: a unique id for the politician
  • embed: HTML code to embed this message
  • id: unique id for the message WITHIN whichever social media site it was pulled from
  • label: a string of the form "From: firstname lastname (position from state)"
  • message_gold: always blank; presumably whether the message response was in the gold standard
  • source: where the message was posted; one of "facebook" or "twitter"
  • text: the text of the message

Seeking to predict relationships between certain words and certain message types as determined by qualified human assessors. To do so, we will apply both a Naive Bayes (NB) Classifier and a Support Vector Machine (SVM).

• Hypothesis #1: Our SVM classifier will be a better predictor than our NB classifier.
• Hypothesis #2: Certain words will be highly correlated with certain messages. For instance, we anticipate the following text/message pairs to be highly correlated:
  • Obama > attack
  • veterans > support
  • proud > support
  • bill > policy
  • ICYMI > media