1. Seshadri Tirunillai University of Houston
Mining Marketing Meaning from Online Chatter:
Strategic Brand Analysis of Big Data Using Latent Dirichlet Allocation
Seshadri Tirunillai
University of Houston
Gerard J. Tellis
University of Southern California
Presented by:
Porter Jenkins
Smeal College of Business

2. Agenda Introduction Motivation Goals Contributions Methodology
Background: LDA
Text Preprocessing
Dimension Extract using LDA
Dimension Labelling
Validation
Empirical Results

3. Introduction: Motivation
Research shows that product quality an important factor:
Brand Loyalty
Market Share
Value
etc…
Surveys are typical measure of quality
Limited samples
Infrequent

4. Introduction: Motivation
Alternative approach?
User-generated content (UGC)
Cheap (often free)
Frequent, even contemporaneous
Passionate, honest responses

5. Introduction: Motivation
How can we specify a model-based approach to extract meaning from thousands (millions) of text documents online?
Even if we have a model for text, how can evaluate quality of products?
From a practical perspective, there are many ways to define quality (multidimensional)
If we have dimensions, we would want to understand:
Valence (sentiment)
Labels of Dimensions
Validity
Importance
Dynamics
Heterogeneity

6. Introduction:
Goals
Extract latent dimensions of quality from online text data
Evaluate the latent dimensions on the criteria previously mentioned
Utilize latent dimensions for strategic analysis
Brand positioning, segmentation etc….

7. Introduction: Contributions
Unified framework to:
Simultaneously extract the latent dimensions of quality and the valence (sentiment) of those dimensions
Analyze the dynamics of quality at frequent intervals
Show importance of latent dimensions using time- varying intensity in conversations
Identify optimal number of latent dimensions
Estimate heterogeneity among customers along latent dimensions
Make generalizable inferences over multiple markets

8. Methodology

9. Methodology: Background of LDA
How can we extract meaning from text? —> Non-trivial
Challenges:
A single word can have different meaning in different contexts
Words can have different meanings over different markets
i.e., ‘Small’ television vs. ‘Small’ computer memory
We can’t simply use universal mapping of positive/negative sentiment
People use different vocabulary
Results in a very sparse corpus matrix
Average corpus matrix: 201 x 2571 (documents x words)
Most elements will be 0
Can’t use dimension reduction techniques

10. Methodology: Background of LDA
Enter Latent Dirichlet Allocation (Blei, Ng, Jordan 2003)
Model/Algorithm from the machine learning literature
Unsupervised learning algorithm:
Don’t have to know dimensions in advance
Doesn’t require elaborate mapping dictionaries
Extracts meaning of words within context of use
Extends well to big data, sparse matrices
Measures valence and dimensions jointly

11. Methodology: Example of LDA1
1: Adapted from Carl Rasmussen

12. Methodology: Example of LDA1
1: Adapted from Carl Rasmussen

13. Methodology: Example of LDA1
1: Adapted from Carl Rasmussen

14. Methodology: LDA as a graphical model1
1: Adapted from Carl Rasmussen

15. Methodology: LDA as a Graphical Model
Key contribution: Addition of valence to model (ν)
Per word topic assignment (z) is dependent on the document topics AND the valence of the word
Valence of word (ν) is dependent on valence of topic (π) in document
φ is the multinomial distribution of the topics (dimensions) with the associated valence over the vocabulary of the words in the reviews

16. Methodology: Steps in Paper
Text Preprocessing
Dimension extraction using LDA
Dimension Labeling
Cross-validation from other data sources

17. Methodology: Text Preprocessing
Treat each review as separate document
Eliminate non-english characters; i.e., URL’s, punctuation etc…
Replace pronouns with corresponding noun; a.k.a Anaphoric resolution
Separate strings into individual sentences. Tokenize sentences
Replace negative words with ‘not’
Stem words; i.e., ‘swimming’ —> ‘swim’
Porter’s algorithm
Remove stop words: ‘the’, ‘and’, ‘is’, etc….
Remove all words that appear in less than 2% of documents

18. Methodology: Dimension Extraction (LDA)
Joint density of observed words, dimensions (topics), and valence given hyper parameters
Word-level
parameters
Document-level
parameters
Latent dimension, valence
parameters

19. Methodology: Dimension Extraction (LDA)
Likelihood of getting w, the words in a given document (3)
We can sum across all valence levels and topics to get (4)

20. Methodology: Dimension Extraction (LDA)
Posterior Distribution
Analytically intractable
We can use standard MCMC techniques (Gibbs Sampler) sampling to approximate this

21. Methodology: Dimension Extraction (LDA)
A few items to train in the algorithm
Seed words to identify valence
We specify a small dictionary of seed words that are unambiguous
i.e., ‘good’, ‘great’, ‘bad’, ‘horrible’, etc…
Iteratively estimate probability of valence of newly encountered words by their co-occurrence with seed words
Select optimal number of dimensions (topics)
For each market
Begin with two dimensions, increase the number of dimensions until log likelihood reaches maximum

22. Methodology: Dimension Labeling
Goal: Choose a topic label so that dimensions reflect topic of discussion
We want an unsupervised approach
Assign score for a word given a dimension (topic)
Mutual Information (MI)
The amount of information gained by the dimension as a result of the presence of the word
Genetics?
Biology?
Neurology?
Technology?

23. Methodology: Dimension Labeling
We can use entropy as a metric to identify MI
Entropy measures the probability that dimension k generates a randomly chosen document (6)
η represents event that the document discussed the kth dimension
Entropy conditional on words observed in document (7)
Finally, calculate how much a given word, w*, reduced uncertainty in the entropy of a dimension, k. Equation (8) is positive if w* reduces uncertainty
Select word(s) with highest MI as label(s)

24. Methodology: Validation
How can we know if our automated algorithm is any good?
Compare to ratings of quality from humans
Have humans read each review
Construct set of dimensions and associated valence by presence of words
Compute kappa statistic to measure agreement
Found ‘strong’ to ‘quite strong’ agreement
Compare to Consumer reports
Jaccard Coefficient to measure degree of overlapping dimensions
Fairly high for all markets
Conclusion: MI works well
Alternatively, you could have humans read a few random articles in each dimension for deeper meaning

25. Results

26. Results: Dimension of Quality (Mobile Phones)
What does the LDA dimensions extraction give us?

27. Results: Dimension of Quality (All Markets)
Some dimensions of quality are common across industries
i.e., performance, customer service, visual appeal
Others are unique
Toys: Safety
Footwear: Comfort
Data Storage: Portability

28. Results Heterogeneity of Dimensions
Don’t use estimated parameters to assess heterogeneity
Herfindahl Index:
Concentration of reviews that mention a given dimension within a brand
Relative to all other dimensions extracted for that brand
H can be interpreted as the average concentration of dimensions (within a brand)

29. Results Heterogeneity of Dimensions
Herfindahl Index is an inverse measure of heterogeneity:
High H for a brand, means low heterogeneity of dimensions, Low H means high heterogeneity
Key insight: In vertical markets, consumers agree more on definition of quality

30. Results Heterogeneity of Dimensions Over Time
Herfindahl Index does not take into account the time- varying nature of heterogeneity in dimensions
Calculate the percentage of instability in the Herfindhal index over time
Let ρ be the correlation between percentage share of consumers citing the dimensions within a given brand and all other dimensions between two periods
Let σ be the standard deviation of shares of dimensions, and n is the total number of dimensions at time t

31. Results Heterogeneity of Dimensions Over Time
Key point:
Vertically integrated markets are more stable
Indicators of quality do not change
Horizontal markets are less stable

32. Results
Brand Mapping
Calculate distance between two brands on a given dimension
Dissimilarity between vector of words underlying that dimension
Non-euclidean
Use Hellinger Distance
Measure distance for all combinations of brands within a given market to construct similarity matrix
Use multidimensional scaling to map to 2-D space
Use top two most important dimensions (frequency of occurrence)

33. Results Static Brand Mapping

34. Results Within-brand Segmentation
Segment size represented by size of circles
Brand perception represented by location on map

35. Results Dynamic Brand Mapping

36. Results: Dynamics of Dimensions
iPhone
Results:
Dynamics of Dimensions
Blackberry Storm
How is the perception of ‘Ease of Use’ changing over time?
Can be used in conjunction with exogenous shocks:
Blackberry storm
iPhone
Volume spikes at these points as well (panel B)

37. Conclusions
User-generated Content is a rich form of data. It can be used to extract important latent dimensions of product quality
Dimensions differ across brands in a given market. They also differ across markets
The valence associated with the dimensions varies across markets
The dimensions are corroborated by human raters and Consumer Reports
These dimensions can be used to construct brand perceptual maps. Maps can be static, segmented, or dynamic
In vertically differentiated markets (e.g., mobile phones, computers), dimensions are more objective and heterogeneity is low, and stability is high over time. The opposite is generally true of horizontally integrated markets