Real-Time Comment Analysis

2001 Roadless Rule Revision

The docket in numbers: how the pile of submissions was collected, deduplicated, and cleaned — volume, length, duplication, and the organized campaigns running. Sentiment lives on the analysis home; topics on Topics in Depth.

2026-08-202026-09-09 · 118,562 comments to date
118,562
Comments collectediEvery public submission pulled from regulations.gov for this docket, before any processing — the raw count.
83,183
Unique after dedupiDistinct comment texts after exact duplicates are collapsed — identical copies (the same letter submitted many times) folded into one.
35,379
Duplicate submissionsiSubmissions that were exact copies of another comment: collected minus unique-after-dedup.
53,110
Truly uniqueiDistinct after a SECOND pass that also collapses template letters — comments sharing an identical body under personalized openings and signatures. Fewer than 'unique after dedup', which only catches exact copies. Runs on comments of at least 100 characters.
17
Topics trackediThe number of topic categories the classifier can assign to a comment (a comment can raise several).
117,304
Classified for sentimentiComments read by the model and labeled for their position — supports, opposes, or neutral. Identical duplicates inherit their group's label.

Comment length distribution

How long the comments run — a proxy for substance versus one-line form submissions.

9,118
18,236
27,355
36,473
2505007501k1k1k1k2k2k2k2k3k3k3k3k4k4k4k4k5k5k+
Comment length (characters, upper bound of bin)Bar height = number of comments

About the comments

Duplicate and template submissions are collapsed so each distinct comment counts once. Deduplication and uniqueness scoring run on the comments of at least 100 characters — shorter ones carry too little text to compare meaningfully — which is why this funnel’s total is below the collected count.

55% duplicate or template
Collectedall submissions
118,562
Unique textafter exact dedup
83,183−35,379 copies
Truly uniqueafter middle-content pass
53,110−30,073 near-copies

Organized campaigns

Beyond exact duplicates, comments are grouped into campaigns — sets built from the same template, whether copied word-for-word or lightly reworded. Each is counted once in the unique totals above; this is who is running them and how large they are.

77 distinct campaigns

Exact form letters — 19,086 comments share a verbatim template block

16,908“keep the roadless rule in place”
1,887“keep the roadless rule intact exactly”
40“keep the roadless rule in place”
23“drinking water for millions of americans”
16“please keep the roadless rule”
16“keep the roadless rule in place”

Reworded templates — 6,569 comments match a paraphrased version (caught by meaning, not exact words)

5,459strongly oppose rescinding the Roadless Rule and urge the Forest Service to keep these protections in place.<br/><br/>Our national forests are public…
167Dear Secretary Rollins,<br/><br/>I strongly oppose the Proposed Rule rescinding the 2001 Roadless Rule.<br/><br/>Rolling back the Roadless Rule will o…
123As a person who appreciates the many values national forests bring to my state, my wellbeing, our wildlife and more, support Alternative 1: The &ldquo…
96have lived in Western North Carolina since 2011 and DEEPLY oppose rescinding the 2001 Roadless Area Conservation Rule. Please select the No Action alt…
90Dear Secretary Rollins, <br/>I strongly oppose the USDA’s proposal to eliminate or weaken the Roadless Rule. This reckless action would devastat…
76live in Tennessee and oppose rescinding the 2001 Roadless Area Conservation Rule. Please select the No Action alternative and retain the rule. <br/><b…

Sharing & methodology

These results are meant to be shared.

Please cite us. This analysis is the intellectual property of Roadless.org and Nicholas Holshouser. You are welcome to reproduce it — every reproduction must include this citation:
Roadless.org and Nicholas Holshouser, “2001 Roadless Rule Revision — Public Comment Analysis,” https://roadless.org.

How it works. Sentiment and topic classification is performed by a large language model (an open-weight Qwen 2.5 model) that reads each unique comment and labels its position — supports, opposes, or neutral — and the topics it raises. Lexical analysis, which finds duplicate and template submissions and scores each comment for originality, uses TF-IDF and n-gram phrase matching, sentence-transformer embeddings, and syntactic and stylometric features (built with spaCy, scikit-learn, and sentence-transformers).