Feature Extraction from Social Media Posts for Psychometric Typing of Participants

摘要

Sentiment analysis is an important tool for assessing the dynamic emotional terrain of social media interactions and behaviors [1]. Underlying the shallow emotional phenomenology are deeper and more stable strata, such as culture and psychology. This work addresses the latter, by applying text mining methods to the assessment of individual psychometrics. A methodology is described for reducing bulk, unstructured text to low-dimensional numeric feature vectors, from which components of the Myers-Briggs Typology Indicator (MBTI) [2J of the text's author can be reliably inferred. MBTI is a psychometric schema that emerged from the personality theories of Freud and Jung in the early 20th Century, refined and codified by K. C. Briggs and her daughter, I. Briggs-Myers in the 1940's and 50's. This schema positions people along four (nominally independent) axes between pairs of polar motivations/ preferences: Extroversion vs. Introversion (E-I); Intuition vs. Sensing (N-S); Feeling vs. Thinking (F-T); and, Judging vs Perceiving (J-P). Under this schema, each person falls into one of 16 psychometric groups, each designated by a four-character string (e.g., INTJ) [3]. Empirical results are shown for text generated during the social media interaction of over 8,600 PersonalityCafe users [4], all of whom are of known MBTI type. Blind tests to validate the features were conducted for a population (balanced by MBTI type), with exemplars based upon text samples having several thousand words each. The feature extraction method presented supports partial (1-letter) MBTI psychometric typing: E-I 95%; J-P 76.25%; F-T 91.25%, N-S 90%. Other results are reported.