Deriving Sentiments from Opinions or Product Choices

[#1]

See how to:

• Quantify positive or negative sentiment in unstructured text data
• Understand basics of Lexical Sentiment Analysis
  • Scores sentiment from individual words in each doc when no external measure of sentiment is available
  • Uses a Sentiment dictionary (aka "lexicon") that specifies scores (e.g., wonderful = +90, disappointed = -30)
  • Scores individual sentiment phrase to calculate document sentiment
  • Create three dictionary lists - Sentiment term: a term bearing sentiment such as ‘wonderful’; Intensifier term: sentiment multiplier such as ‘very’; Negator term: sentiment polarity reversal such as not
    • Customize dictionaries for your situation, domain, and add scores
    • Import and export dictionaries
  • Understand scoring
    • Individual phrases are scored: Sentiment term score  x  intensifier weighting  x  negation (-1)
    • Documents are scored: Scaled: mean sentiment phrase score and Min/Max: Sum of max and min sentiment
  • Understand rules
    • A term can belong to only one class: sentiment, intensifier, or negator
    • Parse Documents option uses natural language processing (NLP) to better assign negators and intensifiers
    • Stemming does not affect sentiment scores (and for good reason)
    • If a previously declared stop word is added to the dictionary, it is temporarily removed as a stop word in every analysis within the Text Explorer window
  • Explore results
    • Sentiment Distribution provides overall view of sentiment across all docs
    • Sentiment Term List shows most frequent sentiment terms are most frequent and you can click to browse all documents with selected term(s)
    • Document Table presents document-level scoring (hover for explanations)
    • Text pane shows terms in context (hover for explanations)
    • Features section finds terms that frequently co-occur with sentiment terms and rescores documents with respect to selected features
  • Export sentiment scores for use in other analyses                                                       

• Understand basics of Derived Sentiment Analysis
  • Uses external sentiment measure (e.g. satisfaction rating) to drive sentiment analysis
  • Uses Text Explorer's Term Selection to build a model of sentiment measure predicted by document-term matrix (DTM)
  • Uses data to derive term, document sentiment scores
  • Identifies sentiment terms we may otherwise miss
  • Model can be applied to future text, even when external sentiment measure is not available
• Run the Derived Sentiment Analysis using Term Selection
  • Use nominal, ordinal, or continuous response variables
    • Nominal with >2 levels requires choosing target class (i.e. reduces to binary outcome of target vs not-target)
    • Ordinal must be numeric data type, is modeled as continuous response with normal distribution
  • Select additional up-front options: Estimation and validation methods, DTM weighting, Max terms in DTM)
  • Access Generalized Regression controls after running initial model and build multiple models to find the best
• Explore results
  • Top-level summary is similar to Sentiment Analysis
  • ‘Contribution’ is generic word for response scale (e.g., log odds)
  • Document's contribution values are sum of term coefficients (regression model coefficient)
  • Term LogWorth is function of p-value (higher LogWorth corresponds to lower p-value)

Resources

• JMP Discovery Summit 2017 presentation
• JMP Discovery Summit 2018 presentation on Modeling Consumer Feedback With Text Mining Features in JMP® Pro 
• Using Generalized Regression (Mastering JMP video)
• Using Text Explorer to Extend Analysis (Mastering JMP video)

Sentiment Summary