AAU Student Projects - visit Aalborg University's student projects portal
A master's thesis from Aalborg University
Book cover


Exploring emotion detection in textual data, using knowledge based methods and web scraping, and communication of the results through affective visualizations created by generative algorithm

Authors

;

Term

4. term

Education

Publication year

2019

Submitted on

Abstract

This exploratory study investigates natural language processing approaches and proposes an alternative way to visualize and communicate multivariate data, aiming to augment recommender systems with an emotional category. The work proceeds in two phases: In the NLP phase, both traditional (valence, subjectivity) and experimental (arousal, entropy, color) features are extracted from the IMDB reviews dataset; t-SNE is used for dimensionality reduction to produce a spatial organization of the texts, which is examined through basic statistics, value distributions, visualizations, and exploratory clustering. In the visualization phase, the affective capabilities of computationally generated visuals are explored to communicate selected data qualities, mapping them to preattentive elements in generative visualizations—one per film, for 20 films in total. The communicativeness of the visuals and the effectiveness of the generative algorithm are evaluated via a quantitative online survey with 79 participants. While further work is needed, results from feature extraction suggest that the experimental features influence the organization of textual data, and the survey indicates that affective visualizations can convey data characteristics. Together, the two phases point to the potential of emotional signatures to enrich emotion-aware recommender systems.

Dette eksplorative arbejde undersøger metoder til naturlig sprogbehandling og foreslår en alternativ måde at visualisere og formidle multivariate data på med henblik på at styrke anbefalingssystemer med en følelseskategori. Designet er todelt: I NLP-fasen udtrækkes både traditionelle (valens, subjektivitet) og eksperimentelle (arousal, entropi, farve) træk fra IMDB-anmeldelser; t-SNE anvendes til dimensionsreduktion for at skabe en rumlig organisering af teksterne, som inspiceres via grundlæggende statistik, fordelinger, visualiseringer og udforskende klyngeanalyse. I visualiseringsfasen undersøges affektive, computergenererede billeder som middel til at kommunikere udvalgte datakvaliteter; disse kortlægges til præattentive elementer i generative visualiseringer, én pr. film, i alt 20. Visualiseringernes kommunikativitet og algoritmens effektivitet vurderes gennem en kvantitativ online-undersøgelse med 79 deltagere. Foreløbige resultater indikerer, at de eksperimentelle træk påvirker den rumlige organisering af tekstdata, og at affektive visualiseringer kan formidle datakarakteristika; yderligere arbejde er dog nødvendigt. Studiet peger på et potentiale for at udnytte følelsessignaturer til at berige emotionelt bevidste anbefalingssystemer.

[This apstract has been generated with the help of AI directly from the project full text]