Video answer review

Lecture answer comparison

Two questions. Six ways to assemble the lecture answer. Listen to each version, then compare how the source clips fit together.

Question

Loading question…

Loading comparison data…

Original baseline

01

Select relevant transcript lines and stitch their lecture clips.

Actual duration
…
Joins
…
Distinct lectures
…
Biggest skip
…

New methods

02

Judge neighboring passages together and keep their surrounding context.

Actual duration
…
Joins
…
Distinct lectures
…
Biggest skip
…

Use the player controls for audio and English captions. Playing one video pauses the other. Captions may contain recognition/timing errors.

Largest time jump

The biggest stretch of source time skipped between consecutive clips from the same lecture. Smaller bars mean less time skipped at that join.

Original medium…
Whole blocks…
0 seconds skippedLoading scale…

Both bars share a scale, set by the larger jump in the selected pair. This measures structural continuity, not an expert coherence score; switches between lectures are not time gaps on this scale.

Whole blocks reduce large time jumps in these two examples. Sentence reranking does not consistently improve continuity. Perceived coherence and answer completeness still need review.

All six methods for this question
Prepared clip metrics. Highlighted rows are the selected pair. Time values are in seconds.
MethodActual duration (s)JoinsDistinct lecturesLecture switchesBiggest skip (s)

The original baseline was locally reproduced from unchanged outputs, not professor-supplied finished videos. These comparisons use two diagnostic cases and different durations; original variants have no 60-second cap.

Use the source time maps to inspect each join. Annotation worked examples are constructed examples for the review process.

Supporting reading

VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videos

KCVR: Knowledge-Centric Video Reconstruction for Structured Pedagogical Summarization via Dynamic Graph Planning