alea-institute/fcc-submissions

FCC ECFS Submissions

Text of 2,566,814 filings submitted to the U.S. Federal Communications Commission's Electronic Comment Filing System (ECFS): comments, reply comments, ex parte notices, petitions, orders and other docket filings. These are public docket records.

Version 2: text repair

This is version 2. Version 1 is preserved unchanged on the v1 branch: load_dataset("alea-institute/fcc-submissions", revision="v1").

The first version of this dataset took its text from the ECFS API's text_data field (2016 export). That field has four defects:

1. Line breaks removed. Every newline was deleted, so headings and address blocks run into the body (COMMISSIONWashington, D.C. 20554In the Matter of) and many words are glued (andthe).
2. Non-ASCII characters replaced by ?. For example, 47 C.F.R. ? 73.1650 and U.S.C. ?? 303(s).
3. Straight double quotes removed. For example, the digital divide where the filing says the "digital divide".
4. Long filings truncated at about 29,000–32,700 characters.

Version 2 rebuilds the text of 1,905,188 rows (74.2%) from an earlier (2013) capture of the ECFS text layer for the same filings. That capture keeps line breaks, double quotes, the full document length, and many of the original characters.

How rows were rebuilt

A row is rebuilt only if a strict, deterministic alignment proves that the published text can be derived from the archived text using nothing but the known defects:

• delete line breaks;
• delete ";
• replace a non-ASCII character with ?;
• cut the text at the truncation point.

Any other difference, even a single character, leaves the row unchanged (recovered = false). So outside the defect positions, every character of a rebuilt row is identical to the published text.

Rules applied to rebuilt rows:

• Line endings are normalised to \n.
• A published ? is replaced only by a character on a fixed allow-list: § ¶ ’ ‘ “ ” – — • … © ® ° ± ½ ¼ ¾, no-break space, common accented letters, × ÷ µ ¹ ² ³.
• A run of consecutive ? is restored only if every character in the run is on the allow-list. Otherwise the whole run stays ?. This keeps OCR noise (·, £, ¥, «, ») and half-decoded byte sequences (â??, ï¼?) out of the text.
• Truncated rows carry the full archived text, with the same character policy applied to the part beyond the old truncation point.
• No character is inferred or guessed. A ? that was already ? in the archived text stays ?. This is common for curly quotes and apostrophes (Commission?s, (?Order?)).

Coverage

	rows
total	2,566,814
rebuilt (recovered = true)	1,905,188 (74.2%)
not in the 2013 capture (mostly filings received after early 2013)	660,105
in the capture but failed strict alignment	160
in the capture but kept unchanged because the published text has control-byte placeholders absent from the archive	1,361

Totals over the rebuilt rows:

• 151.4M line breaks restored;
• 8.66M double quotes restored;
• 7.08M characters restored (mostly no-break space, §, ¶, °, curly quotes, ®, ©, dashes);
• 4.41M characters withheld and left as ?;
• 67,472 previously truncated rows now complete, adding about 6.2B characters.

Columns

column	meaning
submission_id, submission_type, date_received	unchanged from version 1
text	rebuilt text if recovered, otherwise the version 1 text
recovered	true if the text was rebuilt from the 2013 capture
text_source	ecfs_text_2013:<YYYY/MM/DD/id>.txt for a rebuilt row (its file in the 2013 capture), or hf_v1_unchanged
source_encoding	encoding of the archived bytes (every archived file is valid UTF-8)
n_newlines_restored, n_quotes_restored, n_chars_restored	counts of each repair in this row
n_chars_withheld	archived characters not restored (left as ?)
hard_wrapped	the rebuilt text is hard-wrapped at a fixed column, so line breaks can fall inside words (see below)
hf_truncated	the version 1 text was a truncated prefix; text is now the full document
hf_text_len	length of the version 1 text
n_qmarks	number of ? in text
qmark_residual	number of ? in text that do not look like sentence-final question marks

qmark_residual counts every ? except one that directly follows a letter, digit, ., ) or ] and is followed by whitespace, a straight quote or the end of the text. It is a count only; no text was changed from it. Use it to filter or down-weight rows with many lossy ?.

Known limitations

• Unrecovered rows (25.8%, recovered = false) keep all four defects. Filter on recovered if you need line structure.
• Scanned paper filings (pre-2000 especially) are OCR text with heavy noise in both versions.
• Some rows hold encoded attachments (base64 or RTF captured as text), in both versions.
• Line breaks follow the source page layout. That includes superscript ordinals on their own line (445 12\nth\n Street), footnote numbers on their own line, and docket numbers split at a line end (05-\n65).
• Hard-wrapped rows. 9,244 rebuilt rows (0.5%), hard_wrapped = true, were captured with lines cut at a fixed column, so a break can fall inside a word (who wo\nuld like). Most are a few 2002 form-letter campaigns.
  • The flag means: at least 10 line breaks fall between two letters at the document's most common line width, which covers at least half of its lines.
  • A check of 30 distinct flagged texts found:
    • 24 with words split by the wrap;
    • 4 encoded attachments (base64 or RTF blocks) captured as text;
    • 2 wrapped only between words.
  • The text of these rows is not edited. Drop or rejoin them if line breaks inside words matter for your use.
• Some restored double quotes sit inside OCR noise.
• Rows with ? from the source are not fixed. Many curly quotes and apostrophes were already ? in every available copy of the text.

Quality checks

• Strict re-verification. Every rebuilt row reproduces the version 1 text under the defect operations only (110 of 110 sampled rows re-verified independently).
• Hand review of 50 rebuilt rows: no character outside the defect positions differs. Of 233 restored characters, 231 are correct in context; the other 2 (°, ±) sit inside OCR noise.
• Blind review of 60 further rows by an independent reviewer: 0/60 fidelity violations, 0/60 mojibake, 0/60 wrong replacement characters, 0/60 corruption in clean text.

Personal data

The rebuild adds no fields that carry personal data. The text of public docket comments can include names and contact details that commenters submitted to a public docket, as in version 1.