taiwan-cultural-center

name	taiwan_cultural_center
description	Platform rules, 4-tier date extraction, report detection, and troubleshooting for the Taiwan Cultural Center scraper
applyTo	scraper/sources/taiwan_cultural_center.py

Taiwan Cultural Center (TCC) — Platform Reference

Platform Profile

Field	Value
Base URL	https://www.tcc.go.jp/event/
Rendering	Static HTML
Auth required	No
Source name	taiwan_cultural_center

Field Mappings

Event Field	TCC Source
source_id	MD5 hash of URL path (first 16 chars)
raw_title	<h1> or page title
start_date	4-tier cascade (see below)
end_date	Same tier result; None if single-day
location_address	Fixed: 東京都港区虎ノ門1-1-12
raw_description	Full body text with date prefix prepended

Date Extraction — 4-Tier Cascade

Apply in strict order. Fall through to the next tier only when the current tier returns no date.

Tier	Method	Pattern
1	Body labels	日時: / 日 時: / 会期: / 開催日: / 期間: / 開催期間:
2	Prose date	MM月DD日(曜) in body; year inferred from post_date (−180 to +365 day window)
3	Title slash	M/DD(曜) in title; year inferred from post_date
4	Publish fallback	日付：YYYY-MM-DD at page bottom — use ONLY when tiers 1–3 all fail

Month-Only Dates (Tier 1 extension)

• Pattern 期間：2026年5月～10月 contains month-only dates without day numbers.
• _parse_date() handles YYYY年M月 → first day of that month (day=1).
• _extract_event_dates_from_body() advances month-only end_raw to the last day of the end month via calendar.monthrange(y, m)[1].
• Commit reference: 039f532

Month-Only Dates (Tier 1 extension)

• Pattern 期間：2026年5月～10月 contains month-only dates without day numbers.
• _parse_date() handles YYYY年M月 → first day of that month (day=1).
• _extract_event_dates_from_body() advances month-only end_raw to the last day of the end month via calendar.monthrange(y, m)[1].
• Commit reference: 039f532

Critical Parsing Rules

• _parse_date() must strip parenthetical day-of-week markers （月） / (土・祝) before strptime
• End dates with only a day number (〜5日): inject year + month from start_date
• Never overwrite raw_title or raw_description with translated or processed content

raw_description Convention

Prepend 開催日時: YYYY年MM月DD日\n\n when event date ≠ post_date.

Report Detection

Auto-add "report" to category when title contains:

レポート|レポ|報告|記録|アーカイブ|recap

Dedup

• Key: MD5 hash of URL path (first 16 chars)

Chrome MCP (Local Dev Only)

When selectors break or a new page structure appears:

1. Open Chrome DevTools, navigate to a TCC event page.
2. Use mcp_browser_snapshot to capture the accessibility tree.
3. Look for: .list-text.detail body container, date label pattern, 日付： publish date element.
4. Update taiwan_cultural_center.py with corrected selectors.
5. Never use Chrome MCP in CI — production scraping always uses Playwright.

連続上映企画 (Series Screenings)

For annual film screening series (e.g. 台湾映画上映会), the event page lists all individual screenings in a long description (10+ films, 13,000+ chars). Do NOT rely on GPT annotation to generate sub-events from this description — GPT-4o-mini reliably extracts only the first 2 entries even with 20,000-char input limit.

Recommended approach:

1. In the scraper, parse each screening entry (title, date, time, venue) from the page body.
2. Emit each screening as a separate Event with parent_event_id pointing to the series parent.
3. Set source_id = f"{parent_source_id}_sub{n}" for each screening.

2026 映画上映会 reference:

• 16 sub-events manually inserted 2026-04-29 (commit 603d7c4 history); see taiwan_cultural_center/history.md.
• Pattern: 『{title}』\n{M月D日(曜) HH:MM開演}／{venue}.

連続上映企画 (Series Screenings)

For annual film screening series (e.g. 台湾映画上映会), the event page lists all individual screenings in a long description (10+ films, 13,000+ chars). Do NOT rely on GPT annotation to generate sub-events from this description — GPT-4o-mini reliably extracts only the first 2 entries even with a 20,000-char input limit.

Recommended approach:

1. In the scraper, parse each screening entry (title, date, time, venue) from the page body.
2. Emit each screening as a separate Event with parent_event_id pointing to the series parent.
3. Set source_id = f"{parent_source_id}_sub{n}" for each screening.

2026 映画上映会 reference:

• 16 sub-events manually inserted 2026-04-29; see taiwan_cultural_center/history.md.
• Screening entry pattern: 『{title}』\n{M月D日(曜) HH:MM開演}／{venue}.
• Always extract name_zh from 原題： line in the description. Do NOT guess or translate from Japanese.
• source_url for sub-events = parent's source_url — never use a guessed URL.
• After inserting sub-events, update parent start_date/end_date to MIN(sub.start_date) / MAX(sub.end_date).

Troubleshooting

Symptom	Cause	Fix
Report articles have start_date = publish date	Tiers 1–3 all failed; fell through to Tier 4	Inspect body text; check that Tier 1 label regex matches current markup
start_date off by one year	Year inference window too narrow	Verify −180 to +365 day window is applied around post_date
end_date null when 〜5日 present	End date parser not injecting year + month	Patch parser to carry over year + month from start_date
Parse error on (土・祝)	strptime receiving day-of-week suffix	Ensure _parse_date() strips all （...） / (...) groups first
start_date = publish date for 期間：2026年5月～10月	_parse_date() didn’t handle month-only YYYY年M月 format	Add YYYY年M月 → day=1 case; advance end-date to last day via calendar.monthrange
sub-events only 2 of 16 for film series	GPT truncates sub-event output for long descriptions	Parse screenings in scraper layer; emit each as Event(parent_event_id=…)
start_date = page publish date when period is 2026年5月～10月	_parse_date() didn't handle month-only format	Ensure YYYY年M月 → day=1 case exists in _parse_date(); end_date → last day of end month via calendar.monthrange
sub-events only 2/16 for film series	GPT truncates sub-event output for long descriptions	Parse screenings in scraper layer; emit each as Event with parent_event_id