PaperlayerView exact revision on arXiv

Source-linked AI summary

A Training-Free, Alignment-Free Approach to Corporate Intelligence: Application to SEC Filings

Jean-François Delpech

arXiv:2609.11620v1cs.CL

TL;DR

Financial-disclosure analysis faces context, hallucination, cost, and embedding-alignment constraints. The paper introduces deterministic sparse seed vectors whose context accumulation creates a shared, linearly composable space for SEC filings. The framework supports traceable corporate-intelligence analyses without domain-specific training or LLM inference, while its outputs still require human interpretation.

  • Problem

    SEC-filings analysis is limited by context-window constraints, hallucination risk, high processing cost, and arbitrary rotations between independently trained embedding spaces.

  • Method

    The framework hashes words into deterministic sparse seed vectors and accumulates sentence contexts in a fixed shared basis, eliminating training and alignment.

  • Results

    The shared construction supports issuer vocabulary extraction, thematic sentence retrieval, filing and issuer comparison, and tracking language changes across filings on ordinary hardware.

  • Takeaways & Limitations

    The method provides a compact, searchable, and traceable upstream representation of disclosure that can support human readers or smaller models more efficiently and reliably.

  • Takeaways & Limitations

    The method reports vocabulary and associations but does not answer an analyst’s question on its own, and interpreting coded filing terms can require reading definitions.

Abstract

from arXiv · show

High-dimensional dense text embeddings and large language models face real obstacles in financial-disclosure analysis: context-window limits, hallucination risk, high computational cost, and the arbitrary rotation of vector spaces across independently trained models. We present a training-free, alignment-free framework for corporate intelligence built on deterministic sparse seed vectors. Hashing word strings into a fixed high-dimensional basis places all documents and all temporal epochs in a common coordinate system by construction, removing any need for training or alignment. Accumulating these seed vectors across sentence contexts yields corpus-specific semantic signatures that compose linearly, supporting sub-second document comparison, issuer fingerprinting, tracking of how an issuer's vocabulary shifts between filings, and thematic sentence extraction, all on ordinary CPU hardware. Demonstrating the approach on a multi-year corpus of SEC filings (10-K, 10-Q, 8-K), we show how material corporate events, among them Boeing's 737 MAX crisis, Intel's supply-chain disruptions, and Bunge's acquisition of Viterra, emerge as distinct, interpretable semantic profiles, each traceable to the exact source sentences that produced it, with no domain-specific training and no LLM inference.

1. Introduction

SEC-filings analysis is constrained by scale, hallucination risk, cost, and the alignment problems of learned embeddings. The paper addresses these limitations with deterministic sparse seed vectors whose accumulated contexts form a shared, linearly composable semantic space.

  • Motivation: Large SEC filings can exceed practical context windows, while external processing, hallucination, and the cost of analyzing tens of thousands of filings create additional obstacles.These constraints make direct LLM-based analysis difficult in compliance-sensitive and large-scale settings.
  • Limitations of existing representations: Learned embeddings require complex alignment across independently trained models and retraining for new documents, while their similarity scores are not directly traceable to source sentences.Alignment becomes less reliable across long time spans or divergent corpora.
  • Fixed seed-vector basis: The method hashes each word into a fixed high-dimensional sparse basis, using quasi-orthogonal vectors to represent large vocabularies without training or alignment.Random projection preserves distances approximately, allowing a modest dimension to support many terms with limited interference.
  • Context accumulation: Sentence vectors are weighted sums of word seed vectors, and semantic vectors accumulate the sentence contexts in which words occur.This makes semantic distinctiveness depend on distributional neighborhoods rather than the written form alone.
  • Compositional semantic space: Because words, sentences, documents, and corpora share one additive basis, their representations can be compared, combined, and updated directly across corpora and time.The construction supports linear composition and enables document differences to be read through words whose vectors diverge most.

2. A word meaning depends on its context

The method represents words through corpus-specific usage, so even unambiguous terms acquire different nearest neighbors across industries and reporting contexts. These context-dependent profiles remain traceable to the filings and sentences that produced them, while absent terms yield null results rather than forced similarities.

  • 2.1. Neighbors of ‘digital’: Digital likewise reflects company-specific contexts, spanning tax policy at Alnylam, drone hardware at AeroVironment, and chip engineering at Mercury Systems and AMD.The method captures these associations from the words appearing near digital in each issuer’s filings.
  • 2.2. Neighbors of cloud: Cloud separates accounting-standard disclosure from operational infrastructure: non-semiconductor filings associate it with FASB and capitalization, whereas AMD associates it with hyperscalers and workloads.The comparison reflects different business and regulatory contexts rather than the familiar weather-versus-computing distinction.
  • 2.2. Neighbors of cloud: When a term is absent, the linear method returns an empty neighbor set instead of artificially assigning weak similarities.Bunge has no cloud signal and AMD has no significant africa co-occurrence signal in the covered filings.
  • 2. A word meaning depends on its context: The same word can acquire different semantic neighbors across industries because its vector encodes corpus usage rather than a fixed dictionary sense.This applies even to monosemous terms such as africa, whose representation shifts with surrounding vocabulary and reporting habits.
  • 2.3. Neighbors of africa: Africa produces industry-specific profiles, from Alnylam’s administrative geographies and AeroVironment’s defense-market vocabulary to Mercury’s hardware monitoring and Bunge’s client-management context.These differences arise despite africa denoting the same continent in every filing.

3. Semantic profiles and fingerprints

The paper constructs issuer fingerprints by contrasting an issuer’s filings with a reference SEC corpus, producing ranked distinctive terms that expose corporate-specific information and link directly back to source disclosures.

  • 3. Semantic profiles and fingerprints: Fingerprints reveal product portfolios, technology architectures, acquisitions, partners, facilities, executive names, and legal entities that generic summaries may obscure.
  • 3. Semantic profiles and fingerprints: AeroVironment’s profile captures BlueHalo and related autonomy and unmanned-systems vocabulary, showing how a major acquisition appears directly in the issuer profile.The fingerprint updates almost instantaneously when a new filing enters the corpus.
  • 3. Semantic profiles and fingerprints: Alnylam’s compact fingerprint centers on its RNAi franchise, while AMD’s reflects product families, computing architectures, and acquired or partnered capabilities.
  • 3. Semantic profiles and fingerprints: Mercury Systems’ fingerprint combines facilities, an acquisition, financing language, and executive names, demonstrating that useful profiles extend beyond consumer-facing product names.
  • 3. Semantic profiles and fingerprints: Because each word remains localized to filings, sections, and sentences, analysts can retrieve verbatim SEC passages and verify the context behind any distinctive term.The fingerprint therefore functions as a navigable index into the underlying disclosure database.

4. Time Evolution of SEC Filings

The method builds dynamic issuer profiles by comparing cumulative semantic spaces across filing dates, clustering newly distinctive words, and retrieving their exact source sentences. Applications show interpretable themes across acquisitions, COVID disclosures, and Boeing’s legal and operational events.

  • Time-evolving profiles: Comparing cumulative semantic spaces across filing dates identifies words that a new filing introduced or reinforced, enabling dynamic issuer profiles.Selected words are clustered by pairwise semantic similarity, then linked to exact matching sentences in decreasing thematic order.
  • Method: Word-level clustering produces coherent, self-labelling themes with fully attributed supporting passages rather than unlabeled sentence groups.The approach retains comparability across filings and corpora without alignment while preserving traceability from cluster to term to source sentence.
  • Mercury Systems: Mercury’s acquisition disclosure separates Pentek’s identity and product vocabulary from the company’s own corporate and operational identity.The extraction covers target identity and capabilities, strategic rationale, and transaction mechanics; the term “saddle” becomes distinctive because of the company’s address context.
  • Intel: Intel’s COVID fingerprint captures pandemic-related supply, containment, employee, personnel, and economic-recovery risks absent from earlier filings.The extracted vocabulary includes travel bans, quarantines, shelter-in-place, social distancing, and shutdowns.
  • Boeing: Boeing’s filing clusters distinguish 737 MAX conduct, prosecution mechanics, and the $2.51 billion financial settlement, with each theme tied to source sentences.The clusters also demonstrate that extraction remains consistent across structurally different filings and does not depend on a fixed document format.
  • Traceability: The method’s fully attributed clusters can support analysts directly or constrain a language model to quote retrieved passages rather than summarize from memory.Every cluster resolves to specific sentences in the source document.

5. Semantic vector querying

Semantic vector querying retrieves issuer-specific neighborhoods for words and documents, then returns the exact passages associated with those neighborhoods. The examples show product, corporate-history, and business-model fingerprints that reflect how each issuer uses its vocabulary.

  • Term querying: Querying a term in an issuer’s semantic space returns neighboring vocabulary that captures the company-specific register surrounding that term.For Alnylam, “onpattro” retrieves product, revenue, launch, and commercial terms rather than merely identifying the drug.
  • Term querying: Alnylam’s Onpattro neighborhood leads a boolean query back to exact sentences describing the product’s commercial performance and launch history.The path from term to neighbors to verbatim passages is directly auditable.
  • Issuer fingerprints: Mercury and Kratos place “drones” near different vocabularies, reflecting sensor and signal-processing payloads in one issuer’s filings and complete airframes in the other’s.Retrieving the corresponding sentences provides an attributed basis for comparing the two business models.
  • Traceability: Word-level clustering and sentence-level provenance together provide source material for analysts or grounded input for language models instructed to use only retrieved passages.The method combines interpretable vocabulary with explicit passage attribution.

6. Retrieval and corpus-relativity

The retrieval experiments compare vocabulary-anchored semantic neighborhoods with transformer similarity across sentence pools, filing epochs, and corpora. Results show sharper, more attributable profiles at lower computational cost, while making corpus and epoch choice explicit analytical parameters.

  • Retrieval comparison: The method forms sharper retrieval profiles than pooled BGE embeddings, isolating vocabulary-close sentences with an interpretable elbow rather than diffuse topical relatedness.In the Alnylam example, the method’s top neighboring similarities are 0.75 and 0.60 after the query, while BGE returns a broader 0.72-to-0.65 neighborhood.
  • Computational cost: Embedding and searching 168 Alnylam sentences took 1.76 s with BGE on a GPU versus 0.010 s with the method on one CPU core.The reported ratio is hardware-dependent, but the passage attributes the difference to semantic-space lookups and vector additions versus transformer inference.
  • Corpus-relativity: A word’s semantic vector is corpus-local because it accumulates usage contexts, although the shared seed basis keeps spaces comparable across corpora.Cross-issuer searches are therefore deliberate projections rather than the default search operation.
  • Epoch comparison: For the Boeing query, the top neighbor falls from around 0.83 under 2021 vectors to 0.52 under 2026 vectors as the later vocabulary becomes larger and more varied.The lower similarity reflects reduced centrality of the specific 2021 query in the 2026 corpus.
  • Corpus-relativity: The construction supports historical sweeps and nested corpus spaces, allowing analysts to follow neighborhood changes while storing only spaces broad enough to cover intended search scopes.The paper describes the latest cumulative issuer space as incorporating earlier contributions.

7. Conclusion

The method provides a fast, deterministic, auditable upstream foundation for corporate intelligence by representing disclosures in one shared space and tracing outputs to source text.

  • A training-free, alignment-free representation maps words to deterministic seeds and accumulates sentence contexts for corporate-disclosure analysis.
  • The method supports semantic-neighbor retrieval but does not support fuzzy matching and remains complementary to locally computed transformers or LLMs.
  • The shared basis makes words, sentences, filings, and corpora directly comparable and allows heterogeneous sources to be integrated without alignment.
  • The database can incorporate earnings calls, presentations, press releases, patents, technical information, and news beyond SEC filings.
  • Every score traces to specific source words and sentences, while absent terms produce null results instead of fabricated content.
  • Its role is to reduce large disclosures to a compact, searchable, traceable database rather than answer analysts’ questions independently.

A.1.1. Word weights

Word weights are derived from corpus frequency rather than TF-IDF, emphasizing rare informative terms while discounting boilerplate.

  • The construction avoids TF-IDF because fixed word directions do not compete for representational space or require document-frequency discounting.
  • Weights use normalized corpus frequency, giving common words low or zero weight and rare informative words high weight.
  • Boilerplate terms receive tiered negative weights, from effective exclusion of legal filler to lighter discounts for generic filing language.

A.1.2. From word to seed

Each word receives a reproducible 64-bit FNV-1a seed that identifies it and deterministically generates its sparse vector.

  • The hash is initialized from a fixed offset basis, updated byte by byte with the FNV prime modulo arithmetic, and returned as the word seed.
  • FNV-1a hashing maps each word to a 64-bit seed that serves as both its database identifier and the source of its sparse seed vector.

A.1.3. From seed to seed vector

The appendix builds the representation from deterministic sparse vectors, accumulates sentence contexts into word semantics, and derives filing-level retrieval and complexity measures.

  • A.1.3. From seed to seed vector: Sparse seed vectors are generated mechanically from word seeds using xorshift-derived coordinates, alternating signs, and fixed magnitude to obtain quasi-orthogonal representations.
  • A.1.4. Semantic accumulation: Semantic vectors accumulate weighted sentence vectors for each distinct word occurrence, updating incrementally without over-weighting repetitions within one sentence.
  • A.1.5. Sentence vectors: Sentence vectors sum unit-normalized word semantics with frequency-based weights, then normalize the result so sentences are comparable by inner product.
  • A.2.1. Issuer profiles: Issuer fingerprints rank distinctive vocabulary using issuer frequency, intrinsic word weight, reference-corpus frequency, and issuer exclusivity while excluding boilerplate.
  • A.2.2. Key-word extraction from a filing: Filing reinforcement is measured by the unnormalized inner product between pre-filing and post-filing word vectors, emphasizing accumulated magnitude rather than direction.
  • A.2.3–A.2.4. Named entities and filing complexity: Collocation scoring joins sufficiently recurrent adjacent token pairs into multiword tokens, while effective rank summarizes fingerprint spread but is not established as a complexity measure.
Loading 2609.11620v1…