Improving Feature Representation Based on a Neural Network for Author Profiling in Social Media Texts

Area of Science:

• Natural Language Processing
• Computational Linguistics
• Social Media Analysis

Background:

• Preprocessing social media data presents unique challenges due to informal language.
• Existing methods may not adequately capture the nuances of social media text.
• Author profiling tasks require robust feature representation for accurate results.

Purpose of the Study:

• To introduce a novel lexical resource for social media data preprocessing.
• To demonstrate the enhancement of neural network-based feature representations using this resource.
• To evaluate the resource's effectiveness in author profiling tasks.

Main Methods:

• Development of a lexical resource comprising dictionaries of slang, contractions, abbreviations, and emoticons.
• Creation of language-specific dictionaries for English, Spanish, Dutch, and Italian.
• Integration of the lexical resource into the preprocessing pipeline for neural network models.
• Experimental evaluation on the PAN 2015 and PAN 2016 author profiling corpora.

Main Results:

• The lexical resource significantly enhanced neural network-based feature representations.
• Improved performance was observed in author profiling tasks when using the developed resource for preprocessing.
• Experiments on PAN 2015 and PAN 2016 corpora confirmed the resource's utility.

Conclusions:

• The developed lexical resource is effective for preprocessing social media data.
• The resource improves the performance of neural network models in author profiling.
• This freely available resource offers a valuable tool for researchers in social media analysis.