21.7 Other tasks: Personality

What are the “Big Five” dimensions of human personality?
  1. Extroversion vs Introversion

  2. Emotional stability vs Neuroticism

  3. Agreeableness vs Disagreeableness

  4. Conscientiousness vs Unconscientiousness

  5. Openness to experience

What are the corpus for personality?

The essay corpus of Pennebaker and King (1999) has around 2500 essays of students “writing whatever comes to mind”. The EAR corpus require volunteers to wear a recorder and randomly recorded short snippets of conversation throughout the day. The Facebook corpus has over 300 million of words from 75K volunteers.

What is the interpersonal stance?

It is the affective stance taken towards another person in a specific interaction. This means given an interaction, label the participant as friendly, supportive, distant, etc. For example, in a romantic interaction, we can measure how flirtatious, friendly, awkward, or assertive a person is.

21.8 Affect Recognition

How can we detect affect?

Affect detection is similar to sentiment detection. We can use standard supervised text classification techniques, using all the words or bigrams as features. For example, we could detect introversion vs extroversion by simply identify and use the common words or bigrams as features. The word clouds for both groups is shown below.

21.9 Lexicon-based methods for Entity-Centric Affect

How can we get the affect score for a particular entity?

We can use the Entity-Centric affect method which combines the affect lexicons with contextual embeddings to assign an affect score to an entity. The algorithm works as follows:

  1. For each word in the training corpus, use ELMo or BERT to extract contextual embedding for each instance of the word. We would then average over the embeddings to obtain a single embedding vector. User the NRC VAD lexicon to get the sentiment, agency, and power scores for the word

  2. Train three regression models on all words to predict the sentiment, agency, and power scores from the averaged word embeddings

During inference for an entity mention, we would get the contextual embeddings of the entity mention and feed the embeddings to the three trained regression models to get the three different scores for the entity. This means we would have a three-tuple score for each entity mention.

21.10 Connotation Frames

What is a connotation frame?

A connotation frame is a lexicon that incorporates a richer kind of grammatical structure by combining affective lexicons with the frame semantic lexicons. The basic idea is that a predicate like a verb can signal connotations about the verb’s arguments. For example, with the sentence “Country A violated the soverignty of Country B”. By using the verb “violate”, the author is expressing sympathies towards Country B, portraying Country B as a victim. This is illustrated in the figure below.



Data Scientist

Leave a Reply