From Structured Representation to Technology Space

Date: 2026-07-08
Stage: Representation

Objective

Determine how structured technical features can be compared across different patent publications after they have been decomposed into Problem, Solution, Component, and Effect aspects.

Background

The previous experiment showed that a patent technical feature can be decomposed into several technological aspects, including Problem, Solution, Component, and Effect.

This decomposition improves the representation of technical meaning.

Instead of treating a feature as a single undifferentiated sentence, the same feature can now be examined from several different perspectives.

However, decomposition alone does not solve the comparison problem.

For example, two Problem descriptions may be:

Prevent propagation of thermal runaway between adjacent cells.

and

Reduce heat transfer from an abnormal cell to neighboring cells.

A reader can recognize that these descriptions are related.

A computer system, however, still requires a quantitative method to determine:

how similar they are,
whether they should belong to the same technical group,
which one is closer to another problem,
and how they are positioned relative to thousands of other features.

PSCE therefore provides a structured semantic representation, but it does not yet provide a measurable technology space.

From Description to Comparable Representation

A conventional keyword-based method can determine whether two descriptions contain the same words.

However, technical similarity does not always depend on identical terminology.

For example:

thermal propagation
heat transfer between cells
spreading of thermal runaway
inter-cell thermal influence

may describe related technical problems even though their wording differs.

Conversely, two descriptions may share the same terminology while referring to different mechanisms or contexts.

Keyword matching is therefore insufficient for comparing technical meaning at the feature level.

What is required is a representation that preserves semantic similarity while allowing numerical comparison.

This is the role of embedding.

Embedding as a Technology-Space Representation

An embedding converts a textual description into a numerical vector.

The vector itself is not intended to be directly interpreted by a reader.

Its purpose is to represent the semantic position of the description relative to other descriptions.

Conceptually:

Technical description
        ↓
    Embedding
        ↓
Numerical vector
        ↓
Similarity / distance

If two technical descriptions are semantically similar, their vectors are expected to be located relatively close to each other.

If their meanings differ significantly, the vectors should be farther apart.

The exact coordinates are less important than the relative relationships among vectors.

This makes it possible to transform a collection of patent features into a technology space.

Why PSCE Should Be Embedded Separately

A technical feature contains several different types of meaning.

Two features may describe:

the same Problem but different Solutions,
similar Solutions applied to different Problems,
the same Component used for different purposes,
or similar Effects achieved through different mechanisms.

If the entire feature is embedded as a single text block, these distinctions may be mixed together.

PSCE decomposition makes it possible to create separate semantic spaces.

For example:

Feature A
 ├─ Problem vector
 ├─ Solution vector
 ├─ Component vector
 └─ Effect vector

This allows comparison to occur at the aspect level rather than only at the whole-feature level.

A system can therefore ask different questions:

Which patents address similar technical problems?

Which patents use similar technical solutions?

Which inventions produce similar technical effects?

These are different analytical questions and may produce different relationships.

Technology Space

The purpose of embedding is therefore not merely to convert text into numbers.

The broader objective is to create a space in which technical concepts can be positioned relative to one another.

For example, in a Problem Space:

P1 ───── P2
 \
  \
   P3
             P4

P1 and P2 may represent closely related technical problems.

P3 may be partially related.

P4 may represent a substantially different problem.

Once such a space exists, additional analysis becomes possible.

These include:

similarity search,
nearest-neighbor retrieval,
clustering,
community detection,
identification of recurring technical patterns,
comparison of technical approaches,
and visualization of technology structures.

Embedding therefore acts as the bridge between semantic representation and computational comparison.

Interpretation

PSCE decomposition and embedding solve different problems.

PSCE answers:

What aspect of the technical feature is being described?

Embedding answers:

How does this aspect relate to similar or different aspects across other patents?

The relationship can therefore be expressed as:

Patent
  ↓
Technical Feature
  ↓
PSCE decomposition
  ↓
Structured semantic representation
  ↓
Embedding
  ↓
Technology space
  ↓
Similarity / clustering / pattern discovery

Without PSCE decomposition, the technology space may mix different types of technical meaning.

Without embedding, PSCE remains a structured textual representation but cannot easily support large-scale comparison.

The two stages therefore serve complementary purposes.

Upcoming Issues

Several questions remain.

First, it is necessary to determine whether each PSCE aspect should be embedded independently or whether some aspects should be combined.

Second, the level of detail in each aspect description may influence the resulting vector.

Third, similarity thresholds must be evaluated to determine when two representations should be considered meaningfully related.

Finally, embedding quality depends on the consistency of the textual representation produced by the LLM. If similar technical concepts are described at inconsistent abstraction levels or with unstable terminology, the resulting technology space may also become unstable.

This introduces the next problem:

LLM representation drift.

Next Step

Evaluate how consistently the LLM represents similar technical concepts across different patents, and determine whether terminology normalization is necessary before constructing the technology space.

Research Records