A fundamental equity PM at a large-cap growth fund wants to systematically compare risk factor disclosures across 200 10-K filings within a single sector. She deploys an LLM-based pipeline to extract and normalize language around supply chain concentration, regulatory exposure, and competitive moat language. Which architectural choice BEST ensures that cross-company comparisons are semantically consistent rather than superficially lexical?
Select an answer to reveal the explanation.
Short Explanation and Infographic
Think of it like translating 200 different foreign-language menus into one standardized format before comparing dishes — the embedding approach puts every company's risk language on the same semantic coordinate grid, so 'concentration risk' described in five different ways by five different CFOs all lands in the same neighborhood of meaning.
Full explanation below image
Full Explanation
Option B is correct because semantic consistency in cross-company NLP analysis requires a shared representational space. A domain-fine-tuned financial language model (such as one pre-trained on SEC filings, earnings calls, or Bloomberg corpora) learns the nuanced vocabulary of capital markets — terms like 'sole-source supplier,' 'single customer concentration,' or 'supply chain dependency' will map close together even though they differ lexically. Cosine similarity in this embedding space measures conceptual proximity, not word overlap.
Option A (keyword frequency) is a bag-of-words approach that treats language as counts of tokens rather than meaning. Two companies disclosing the same risk with different vocabulary will score very differently, producing false signal. This approach also suffers from boilerplate inflation — attorneys often use dense legalese that raises word counts without increasing actual risk.
Option C (sentiment model from movie reviews) is a domain mismatch error. General-purpose sentiment models are calibrated for consumer language. Financial disclosures use hedging language, passive voice, and legalistic understatement that will be misclassified at high rates. The SEC's EDGAR filings are specifically known to use tonal consistency that fools out-of-domain models.
Option D (manual tagging then LLM summarization) inverts the workflow — it uses expensive human labor for the classification step and LLM only for output generation. This defeats the purpose of scalable AI-driven comparative analysis across 200 filings and introduces inter-rater variability in the manual tagging phase.
Best practice follows the CFIA framework for AI alpha pipelines: embed with a financially domain-adapted model, structure output into normalized schema (risk theme, severity qualifier, geographic scope), and apply clustering or retrieval-augmented comparison to surface cross-sectional patterns invisible to human analysts reviewing filings sequentially.