Remove CONLLU_PARSING_ISSUES.md

This documentation has been moved to the ud-hf-parquet-tools library
where it belongs, as CONLLU_PARSING.md:

https://github.com/bot-zen/ud-hf-parquet-tools/blob/main/CONLLU_PARSING.md

The library documentation is more comprehensive and includes:
- All 7 parsing challenges with examples
- Affected treebank statistics
- Implementation details and code locations
- Testing and validation procedures

Co-Authored-By: Claude Sonnet 4.5 <noreply@anthropic.com>

tools/CONLLU_PARSING_ISSUES.md

@@ -1,226 +0,0 @@
-# CoNLL-U Parsing Issues and Workarounds
-
-This document describes the various idiosyncrasies and bugs we encounter when parsing Universal Dependencies CoNLL-U files, and how we work around them.
-
-## 1. Double Equals Sign in FEATS/MISC Fields
-
-**Issue**: The `conllu` library incorrectly parses fields with double equals signs.
-
-**Example**:
-```
-Original: Gloss==POSS.1SG.NOM|RX==[PRO]|TokenType=Clit
-Parsed by conllu: {'Gloss': None, 'RX': None, 'TokenType': 'Clit'}
-Expected: {'Gloss': '=POSS.1SG.NOM', 'RX': '=[PRO]', 'TokenType': 'Clit'}
-```
-
-**Affected Treebanks**:
-- `bej_autogramm` (Beja)
-- Any treebank using values that start with `=`
-
-**Workaround**: Parse FEATS, XPOS, DEPS, and MISC fields directly from raw TSV lines instead of using conllu's parsed values.
-
-**Implementation**:
-- `extract_raw_fields_from_sentence()` in `04_generate_parquet.py`
-- `_extract_raw_fields()` in `templates/universal_dependencies.tmpl`
-
-## 2. Duplicate Metadata Keys
-
-**Issue**: The `conllu` library stores metadata as a Python dictionary, which cannot have duplicate keys. When a sentence has multiple metadata entries with the same key, only the last value is kept.
-
-**Example**:
-```conllu
-# media = Photo 1280x720, 83.5 KB
-# media = <a_href="https://...">...</a>
-# sent_id = BelarusDocs-257
-```
-
-After parsing: `{'media': '<a_href="https://...">...</a>', 'sent_id': 'BelarusDocs-257'}`
-(First media entry is lost!)
-
-**Affected Treebanks**:
-- `be_hse` (Belarusian-HSE): 1,216 sentences with duplicate `media` keys
-- `sa_ufal` (Sanskrit-UFAL): 18 sentences with duplicate keys
-- `br_keb` (Breton-KEB): 15 sentences
-- `pt_gsd` (Portuguese-GSD): 8 sentences with duplicate `generator`/`udpipe_model`
-- `tr_gb` (Turkish-GB): 3 sentences with duplicate `en`
-- Total: **1,323 sentences** across **14 treebanks**
-
-**Workaround**: Parse comment lines directly from raw file text before conllu processes them.
-
-**Implementation**:
-- `extract_raw_comments_from_sentence()` in `04_generate_parquet.py`
-- `_extract_raw_comments()` in `templates/universal_dependencies.tmpl`
-
-## 3. Metadata Keys Without Values
-
-**Issue**: The `conllu` library stores metadata keys without values (like `# newpar`) as `{'newpar': None}`. When reconstructing, we need to output just `# newpar`, not `# newpar = None`.
-
-**Example**:
-```conllu
-# newpar
-# sent_id = 1
-```
-
-Parsed as: `{'newpar': None, 'sent_id': '1'}`
-
-**Affected Treebanks**: Many treebanks use `# newpar`, `# newdoc` without values
-
-**Workaround**: Check if value is `None` and output just the key.
-
-**Implementation**: In comment reconstruction, check `if value is None: output key only`
-
-## 4. Empty Metadata Values
-
-**Issue**: The `conllu` library completely ignores metadata entries with empty values.
-
-**Example**:
-```conllu
-# text_en =
-# sent_id = 1
-```
-
-Parsed as: `{'sent_id': '1'}` (text_en is completely missing!)
-
-**Affected Treebanks**: 36 files across multiple treebanks have empty metadata values
-
-**Workaround**: Parse raw comment lines and preserve `"key ="` format when value is empty.
-
-**Implementation**: Store `"key ="` in comments list when value is empty string.
-
-## 5. Double-Hash Comments
-
-**Issue**: Comments with double hashes (like `# # newpar`) are ignored by the `conllu` library.
-
-**Example**:
-```conllu
-# # newpar
-# sent_id = 11
-```
-
-Parsed as: `{'sent_id': '11'}` (double-hash comment is lost)
-
-**Affected Treebanks**:
-- `ajp_madar` (South Levantine Arabic-MADAR): 1 occurrence
-- `sa_ufal` (Sanskrit-UFAL): 1 occurrence
-
-**Workaround**: These are treated as non-parseable comments by conllu. We parse raw lines and store as `"# newpar"` (preserving the second hash as part of content).
-
-**Status**: ⚠️ These will appear as `# # newpar` vs `# # newpar` differences (extra space removed)
-
-## 6. File-Level Comments
-
-**Issue**: Comments before the first sentence (like encoding declarations) are not associated with any sentence and are lost.
-
-**Example**:
-```conllu
-# -*- coding : UTF-8 -*-
-# sent_id = 1
-```
-
-The encoding line is not part of any sentence's metadata.
-
-**Affected Treebanks**: Various treebanks with encoding declarations at file start
-
-**Workaround**: None. These are file-level metadata, not sentence-level.
-
-**Status**: ⚠️ Accepted loss - encoding declarations are informational only and don't affect linguistic data.
-
-## 7. Empty Nodes Before Token 1
-
-**Issue**: Empty nodes (enhanced dependencies) with decimal IDs less than 1 (like `0.1`, `0.2`) must be inserted before the first token, not after.
-
-**Example**:
-```conllu
-# sent_id = CESS-CAT-A-19981201-124-s7B
-# text = No crec que la nostra vida corri riscos...
-0.1	_	_	PRON	p	_	_	_	2:nsubj	ArgTem=arg0:agt|Entity=(...)
-1	No	no	ADV	rn	Polarity=Neg	2	advmod	2:advmod	_
-```
-
-The empty node `0.1` comes BEFORE token 1, not after token 0 (which doesn't exist).
-
-**Affected Treebanks**:
-- `ca_ancora` (Catalan-AnCora): 445 sentences with empty nodes at position 0.x
-- Any treebank using empty nodes for pro-drop subjects or other zero elements
-
-**Workaround**: Special handling in reconstruction code to insert empty nodes with ID < 1 before the token loop starts.
-
-**Implementation**:
-- Template: Lines 128-144 in `templates/universal_dependencies.tmpl`
-- Validation: Lines 107-122 in `05_validate_parquet.py`
-
-**Status**: ✅ Fixed - empty nodes with any ID (including < 1) are now correctly reconstructed
-
-## Summary of Limitations
-
-### ✅ Fully Fixed
-1. Double equals parsing bug
-2. Duplicate metadata keys
-3. Metadata keys without values
-4. Empty metadata values
-5. Empty nodes with ID < 1 (0.x positions)
-
-### ⚠️ Known Acceptable Limitations
-1. File-level comments (encoding declarations) - not sentence-level data
-2. Double-hash comments - rare edge case (2 treebanks)
-
-### 📊 Impact
-- **100% fidelity** for linguistic data (tokens, annotations, dependencies)
-- **~99.8% fidelity** for metadata (minor formatting differences only)
-- **0 data loss** for linguistic annotations
-
-## Implementation Notes
-
-### Raw Parsing Strategy
-
-We use a hybrid approach:
-1. **Comments/Metadata**: Parse raw lines before conllu to preserve duplicates and empty values
-2. **Token fields (FEATS/MISC/DEPS/XPOS)**: Parse raw TSV fields to bypass double-equals bug
-3. **Token structure (form/lemma/upos/head/deprel)**: Use conllu's parsed values (work correctly)
-
-### Performance Impact
-
-Raw parsing adds minimal overhead:
-- One additional file read per CoNLL-U file
-- String splitting operations only (very fast)
-- No significant impact on generation time
-
-### Code Locations
-
-**Generation script**: `tools/04_generate_parquet.py`
-- `extract_raw_fields_from_sentence()` - lines 248-283
-- `extract_raw_comments_from_sentence()` - lines 286-332
-- Usage in extraction - lines 359-470
-
-**Dataset loader template**: `tools/templates/universal_dependencies.tmpl`
-- `_extract_raw_fields()` - lines 464-490
-- `_extract_raw_comments()` - lines 492-528
-- Usage in `_generate_examples()` - lines 541+
-
-**Validation script**: `tools/05_validate_parquet.py`
-- Uses `example_to_conllu()` which handles marker-based reconstruction
-
-## Testing
-
-**Validation script**: `tools/05_validate_parquet.py`
-
-```bash
-# Validate all local treebanks
-uv run tools/05_validate_parquet.py --local
-
-# Validate specific treebanks with detailed diffs
-uv run tools/05_validate_parquet.py --treebanks be_hse,bej_autogramm --local -vv
-
-# Test 3 diverse treebanks
-uv run tools/05_validate_parquet.py --test --local
-```
-
-**Known passing treebanks**:
-- `bej_autogramm`: 763/763 sentences (double equals fix verified)
-- `fr_gsd`, `en_ewt`, `it_isdt`: All test splits pass (47,131 sentences)
-
-## References
-
-- Universal Dependencies format specification: https://universaldependencies.org/format.html
-- Python conllu library: https://github.com/EmilStenstrom/conllu
-- Universal Dependencies v2.17 release: https://lindat.mff.cuni.cz/repository/xmlui/handle/11234/1-5150