Demographic Factors Improve Classification Performance

Demographic Factors Improve Classification Performance

1. Introduction

   • Socio Lingusits
     • Language Use
       • Demographic Factors
         • Age
         • Gender
       • Results
         • Yond people and womenare are more creative
         • Men and older people are less creative
   • Natural language Processing
     • Initial Assumption
       • Language is Uniform
     • Test Data
       • Social Media
         • Model
           • Bad performance
             • Without demographic factors
           • Good performance
             • With demographic factors
             • Multiple studies
         • NON UNIFORM
       • News
         • UNKNOWN
           • ?
     • Text classification tasks
       • Sentiment analysis
         • Determining the popularity of a document
       • Topic identification
         • Assigning a high level concept to a document that captures its content
       • Author attribute clasification
         • Inferring demographic factors from linguistic features

2. Research Focus

   • Encoding demographic factors?
     • Embeddings conditioned on respective demographic variable
       • Distributed representations of words in a vector space, capturing syntactic and semantic regularities among words
   • Effect of demographic factors on performance?
     • F1 performance of classifiers
       • Text classification tasks
       • With demographic info
         • CONSISTENTLY BETTER
           • Small improvements
             • On average
             • In 8/30 cases statistically significant
               • p<0.005 bootstrap sampling tests
             • ?
       • Without demographic info
       • Five different languages
         • Danish
         • Franch
         • German
         • English
           • US
           • British
       • ?
     • Using Logistic regression models
       • ?