Skip to main content

Config.Metadata.Title

Path: build/Config/Metadata/Title.hs | Language: Haskell | Lines: ~189

Configuration lists for filtering and cleaning HTML page titles

Overview​

This module contains configuration data for the title-cleaning pipeline. When gwern.net scrapes metadata from external URLs (for annotations, link previews, etc.), the raw <title> elements often contain junk: navigation boilerplate, error messages, site branding suffixes, or malformed Unicode. This module defines the filter lists that identify and remove such noise.

The data here feeds into the broader title-cleaning system, which includes both Haskell processing (via Utils.deleteMixedMany and similar) and Python machine-learning cleanup (/static/build/title-cleaner.py). The configuration is intentionally kept in a pure-data Haskell module so it can be easily extended without touching processing logic.

Design philosophy: be aggressive about false positives. A slightly over-cleaned title (losing a suffix like "- YouTube") is preferable to displaying "404 Not Found" or "CAPTCHA" as an annotation title.

Public API​

separators :: String​

Characters used as title/site-name delimiters: "—·|"

Used to split titles at brand suffixes (e.g., "Article Title | Site Name").

Called by: Title processing in LinkMetadata Calls: None (pure data)

badStringPatterns :: [String]​

Substrings that indicate an unusable title when present anywhere:

["Redirecting to ", "404 ", "Page Unavailable", "Page not found",
 "Page Not Found", "CAPTCHA", "s shortform feed", "Item not found"]

Titles containing any of these are rejected entirely.

Called by: Title validation logic Calls: None

badStrings :: [String]​

Exact-match blocklist (~290 entries). Titles matching any string exactly are rejected.

Categories include:

  • Error pages: "404 Not Found", "403 Forbidden", "500 Internal Server Error"
  • Generic/placeholder: "Home", "Search", "Untitled", "Welcome"
  • Paywalls/challenges: "CAPTCHA", "Are you a robot?", "Access Denied"
  • Site boilerplate: Long navigation dumps from Quanta, Slate, Archive.org
  • Domain names used as titles: "nytimes.com", "wsj.com", "x.com"
  • Personal sites with uninformative titles: "niplav", "ribbonfarm", "Dario Amodei"

Called by: Title validation (exact match) Calls: None

stringReplace :: [(String, String)]​

Character encoding fixes (6 entries). Repairs common mojibake:

[("  ", " "),           -- double space → single
 (" � ", " - "),        -- replacement char → dash
 ("°", "°"),           -- UTF-8/Latin-1 collision
 (" ", " "),           -- ditto (A-circumflex + NBSP)
 (" â\200\224 ", "—"),  -- em-dash encoding
 ("\128\200\231", "’")  -- curly quote encoding
]

Called by: Title normalization Calls: None

stringDelete :: [String]​

Suffix/prefix patterns to strip (~180 entries). Uses leading/trailing space to determine deletion direction via Utils.deleteMixedMany:

  • Trailing space → prefix deletion: "The Technium: ", "GitHub - "
  • Leading space → suffix deletion: " - YouTube", " - Wikipedia"

Categories include:

  • Site branding: " - LessWrong", " - Marginal REVOLUTION", " - Astral Codex Ten"
  • Navigation junk: Long Archive.org/Stripe/DEV Community boilerplate
  • Publisher suffixes: " - The New York Times", " - IEEE Spectrum"
  • Blog prefixes: "Abandoned Footnotes: ", "The Splintered Mind: "

Called by: Utils.deleteMixedMany Calls: None

Internal Architecture​

This module is pure configuration—no functions, just data declarations. The structure is:

separators      : String      (3 chars)
badStringPatterns: [String]   (8 patterns)
badStrings      : [String]    (~290 exact matches)
stringReplace   : [(String,String)] (6 pairs)
stringDelete    : [String]    (~180 patterns)

All lists are tested for uniqueness per the inline comments. The badStrings list is not unit-tested because entries are often used as few-shot examples in the Python ML cleaner.

Key Patterns​

Encoding-aware cleanup​

The stringReplace list handles common UTF-8/Latin-1 collision artifacts. These appear when pages serve UTF-8 content but declare Latin-1, or vice versa. The patterns like "°" → "°" fix degree symbols that were double-encoded.

Directional deletion via whitespace​

The stringDelete entries cleverly encode deletion direction:

  • " - YouTube" (leading space) → suffix, delete from end
  • "GitHub - " (trailing space) → prefix, delete from start

This is processed by Utils.deleteMixedMany which inspects whitespace position.

Several entries are massive multi-line strings containing entire navigation menus scraped as "titles". For example, the Quanta Magazine entry is ~400 characters of "Homepage\nSaved articles\nLogin\nSearch...". These represent pathological scraping failures.

Configuration​

This is the configuration. To add new filters:

  1. Exact bad titles: Add to badStrings
  2. Partial patterns (any position): Add to badStringPatterns
  3. Site branding to strip: Add to stringDelete with appropriate leading/trailing space
  4. Encoding fixes: Add to stringReplace

After changes, rebuild and optionally re-run the annotation scraper on affected URLs.

Integration Points​

Upstream consumers​

  • LinkMetadata: Main consumer, applies filters during annotation creation
  • /static/build/title-cleaner.py: Python ML model uses badStrings as few-shot examples

Utils dependency​

Relies on Utils.deleteMixedMany for the directional deletion logic in stringDelete.

See Also​