Interview, Podcast
a16z Podcast | It's Not What You Say, It's How You Say It -- When Language Meets Big Data
- Kieran Snyder is CEO and co-founder of Textio, a company applying machine learning and natural language processing (NLP) to optimize text-based content for specific outcomes, primarily in recruitment.
- Textio analyzes over 10,000 job listings from diverse companies to measure application rates and candidate demographics based on linguistic patterns.
- Kickstarter Case Study: Before analyzing job listings, Textio achieved >90% predictive accuracy on a project's fundraising success by minute zero.
- Successful projects favored longer text, mixed typography, and multiple headings over clean, minimalist design.
- Success factors correlated to celebrity endorsements and social media strategy, though text attributes were more significant than expected.
- Real Estate Application: Vocabulary shifts based on target demographics; for example, "off-street parking" positively impacts low-income home sales but negatively impacts high-income home sales.
- Lexical Volume: Textio has identified over 25,000 unique phrases that influence application volume and candidate demographics in job postings.
- Dynamic Language Trends: Phrases evolve in effectiveness over time; "big data" shifted from a positive signal to neutral by June 2015 as usage saturated the market.
- Negative Triggers: The word "synergy" acts as a "gateway term" that statistically correlates with other clichés (e.g., "value add") and actively deters candidates across all demographics.
- Current High-Performing Terms: The phrase "at scale" is currently a top positive signal, having migrated from tech to other industries; conversely, "workforce analytics" has been replaced by "people analytics".
- Authorship Quality: Job listings authored originally perform better than those stitched together from multiple sources, as the latter lack a coherent point of view.
- Gender Bias in Performance Reviews: Analysis of hundreds of tech performance reviews revealed that "abrasive" was used in 17/100+ women's reviews and 0/100+ men's reviews.
- "Aggressive" was used with men as a constructive exhortation to increase the trait, but as a negative judgment for women.
- Resume Style Differences: Analysis of 1,100 tech resumes showed systematic gender-based presentation styles:
- Men's resumes were shorter, focused on quantifiable details and deliverables, and rarely used bullets for narrative.
- Women's resumes utilized prose narratives, included executive summaries, and contained personal interest statements twice as long as men's.
- Subtle Linguistic Nuances: The phrase "fast-paced" statistically deters more women from applying compared to the neutral "rapidly moving."
- Product Mechanism: Textio provides real-time feedback during document composition, annotating text with suggestions and scoring based on a training set of documents and their outcomes.
- Infrastructure Shift: The democratization of this technology is attributed to cloud computing (AWS, Google Cloud, Azure), which lowered barriers for startups to process large text corpora.
- Future Predictions: Snyder suggests the technology could evolve to generate optimized content (e.g., resumes) based on user inputs, though not to fabricate qualifications.
- Diversification: Users are applying the tool beyond job listings to pitch decks, syllabi, marketing copy, loan applications, and toy product descriptions.
- Market Adaptation: As users optimize language using the tool, successful patterns degrade; the system relies on continuous data ingestion to track shifting "signal vs. noise" trends.