A plain-language guide

How do we find possible matches?

A four-step look at how a bilingual episode moves from a promising lead to evidence a person can review.

Illustrative example only
Step one · Find episodes

Start with meaning, not a verdict.

A Chinese episode about 战争与国内政治危机 (“war and a crisis at home”) is compared with descriptions of indexed English episodes.

Target episode · Chinese

战争与国内政治危机

How war can be used to manage a political crisis at home—and how that choice can deepen division.

A lead is not a match.
Topic similarity tells us where to look next. It does not establish copying.
Possible source episodes · English
A
War and the crisis at homeStronger fit for the example
B
The politics of military spendingPossible fit
C
A history of railway engineeringWeaker fit
What else can create a lead?

Names, places, and distinctive examples in available target transcripts can help narrow the search. A human may also bring a manual lead from reading, listening, or public reporting.

Step two · Compare passages

Compare nearby windows of speech.

When transcripts are public or have been transcribed, short excerpts from overlapping roughly 30-second windows can be compared across languages.

Illustrative similarity matrix · click a cell
Made-up embedding similarity scores, not copying probabilities
Source 1Source 2Source 3
Target A
Target B
Target C

These are window excerpts: short stand-ins for larger passages. Finer sentence display is a later review aid.

Selected illustrative pair

Choose a cell to see the two excerpts it connects.

Step three · Select matches

Keep links that hold together.

The strongest-looking links are considered alongside competing candidates and the order in which passages appear. This is a selection step, not a single-score cutoff.

Illustrative links after selection

The map is a visual summary of this made-up example, not a result from a real pair.

Why B–2 stays

It clears several checks together.

Its illustrative score is above the working floor, it is the best available match in both directions for this example, and it fits the selected order.

Similarity is one signal.
Competing candidates, order, and gaps can change which links are retained. Reordered material can be harder to capture cleanly.
A little more context

The selection process first considers high-quality candidates, then chooses a coherent monotonic sequence, filters weak pairs, and fills compatible residual links. Not every high score survives every stage.

Step four · Review

Let a person inspect the evidence.

Accepted windows can open into finer sentence or group rows when those are available, while the broader window remains as context and fallback.

Illustrative excerpts
Target · Chinese
他把战争看作解决国内政治危机的一种办法。但这场战争反而加深了国内的分裂。
Source · English
“He saw war as a way to resolve the political crisis at home. Instead, the war deepened divisions within the country.”

Short excerpts stand in for a larger window. Any timestamps used in real work must be marked exact, estimated, or absent—never invented.

Four human questions
CorrespondenceDo the passages express the same idea?
Source relationshipIs this the likely source?
ExtentHow much of the episode is accounted for?
AttributionDoes the target credit the source?
A possible match is not a conclusion.
Candidate matching helps a reviewer find passages. It does not, by itself, establish copying.
What happens next?

A reviewer can record separate judgments for correspondence, source relationship, extent, and attribution, with the surrounding context available for a fair reading.

Step 1 of 4