Restructure technical report to ACL format, add Phase 0 gold eval methodology

- Rewrite TECHNICAL_REPORT_v1.1.md to ACL Long Paper format: Abstract,
Introduction, Related Work, Dataset, AL Framework, Experiments, Results,
Analysis, Conclusion, Limitations, Ethics Statement, References, Appendices
- Add Phase 0 gold evaluation set (500 random sentences, 250 dev + 250 test)
to PLAN_v1.1 and technical report, following AL evaluation best practices
- Add BKTreebank (Nguyen 2018) to Related Work with treebank comparison table
- Add comprehensive Limitations section (8 items) and Ethics Statement
- Add 7 new citations: VLSP 2013, VnCoreNLP, Zhang 2022 survey, Bouma 2017,
Farquhar 2021, Luth 2023, Marcus 1993, BKTreebank
- Add AL Cycle 1 report, gold annotations, references, and research materials
- Update annotation tooling and regenerate dataset with fix pipeline

AL_CYCLE1_REPORT.md

@@ -0,0 +1,357 @@
+# AL Cycle 1 Report — Word Segmentation
+
+## Overview
+
+| Item | Value |
+|------|-------|
+| Cycle | 1 (first gold annotation round) |
+| Source | Top 100 highest-uncertainty sentences from AL Cycle 0 |
+| Annotation tool | Label Studio (project 4, labels.ankiren.com) |
+| Annotator | 1 person |
+| Tasks imported | 100 |
+| Tasks annotated | 95 (5 not annotated) |
+| Tasks exported | 92 (3 skipped due to annotation gaps) |
+| Output | `gold_ws_cycle1.txt` (BIO format), `ls_export_cycle1.json` |
+
+## Sentence Selection
+
+Sentences were selected by CRF token marginal uncertainty scoring from AL Cycle 0:
+- CRF model: `tree-1/models/word_segmentation/udd_ws_v1_1-20260212_065135/`
+- Scored 20,000 sentences (dev + test splits)
+- Top 100 by composite score: mean uncertainty × (1 + boundary_weight) + 0.01 × n_long_tokens
+- These represent the **hardest** sentences for the CRF, not a random sample
+
+## Domain Distribution
+
+| Domain | Count | % |
+|--------|------:|---|
+| Fiction | 28 | 30.4% |
+| Wikipedia | 28 | 30.4% |
+| Non-fiction | 24 | 26.1% |
+| News | 8 | 8.7% |
+| Legal | 4 | 4.3% |
+
+Fiction and Wikipedia are overrepresented because they contain more difficult segmentation cases (literary names, Hán-Việt compounds, foreign names).
+
+## Skipped Sentences
+
+3 sentences skipped due to annotation gaps (syllables left untagged in Label Studio):
+
+| sent_id | Issue |
+|---------|-------|
+| `uvn-3488` | Foreign product names (Derma peel One) left untagged |
+| `uvb-n-10657` | Single syllable "là" left untagged |
+| `uvb-f-10312` | Punctuation "»" left untagged |
+
+## CRF Silver vs Gold Evaluation
+
+### Overall Metrics (92 sentences, 1,506 syllables)
+
+| Metric | Score |
+|--------|-------|
+| Syllable Accuracy | **0.8665** (1,305/1,506) |
+| Word Precision | 0.7934 |
+| Word Recall | 0.7685 |
+| Word F1 | **0.7808** |
+
+**Context**: The CRF baseline on silver data achieves Word F1 = 0.9834. The much lower F1 = 0.7808 on these 92 gold sentences confirms that AL successfully selected the most error-prone sentences.
+
+### Boundary Change Statistics
+
+| Change | Count | Meaning |
+|--------|------:|---------|
+| B→I | 83 | Silver has B (word start), gold has I (continuation) → silver over-splits |
+| I→B | 118 | Silver has I (continuation), gold has B (word start) → silver over-merges |
+| **Total changes** | **201** | 13.3% of all syllables have wrong boundary tags |
+
+### Per-Domain Breakdown
+
+| Domain | N | Syllables | Syl Acc | Precision | Recall | F1 |
+|--------|---|-----------|---------|-----------|--------|-----|
+| News | 8 | 213 | 0.9061 | 0.8485 | 0.8296 | **0.8390** |
+| Non-fiction | 24 | 414 | 0.8889 | 0.8235 | 0.8235 | 0.8235 |
+| Fiction | 28 | 412 | 0.8786 | 0.8264 | 0.8041 | 0.8151 |
+| Legal | 4 | 81 | 0.8395 | 0.7347 | 0.6923 | 0.7129 |
+| Wikipedia | 28 | 386 | 0.8317 | 0.7289 | 0.6868 | **0.7072** |
+
+**Wikipedia is the hardest domain** (F1 = 0.707) due to Hán-Việt classical text, foreign names, and specialized terminology. Legal is also low (F1 = 0.713) but with only 4 sentences.
+
+## Error Analysis
+
+75 out of 92 sentences (81.5%) have at least one segmentation difference between silver and gold.
+
+### Error Type Summary
+
+| Error Type | Count | Description |
+|------------|------:|-------------|
+| Over-merge | 134 | CRF merged syllables across word boundaries |
+| Over-split | 119 | CRF split syllables that belong to one word |
+| **Total** | **253** | |
+
+Over-merge is slightly more common (53%) than over-split (47%).
+
+### Error Category Taxonomy
+
+#### 1. Cross-boundary proper name merge (most frequent)
+
+The CRF frequently merges the last syllable of a proper name with the next word:
+
+| Silver (wrong) | Gold (correct) | Domain |
+|----------------|----------------|--------|
+| `Nguyễn_Bính_Mưa_dầm` | `Nguyễn_Bính` \| `Mưa` \| `dầm` | fiction |
+| `Chu_Du_tính_tình` | `Chu_Du` \| `tính_tình` | non-fiction |
+| `Tăng_Hoàng_Vinh_giữ` | `Tăng_Hoàng_Vinh` \| `giữ` | news |
+| `Khổng_Dĩnh_Đạt_thời` | `Khổng_Dĩnh_Đạt` \| `thời` | wikipedia |
+| `Hà_Tiên_trấn` | `Hà_Tiên` \| `trấn` | wikipedia |
+| `Quỳnh_nghiến` | `Quỳnh` \| `nghiến_răng` | fiction |
+| `Khải_nheo` | `Khải` \| `nheo_nheo` | fiction |
+
+**Root cause**: CRF sees proper name continuation (I-W) and extends it to the next syllable, especially when the next word starts with a common syllable.
+
+#### 2. Hán-Việt classical text boundary errors (Wikipedia-specific)
+
+Many Wikipedia sentences contain classical Chinese quotations where the CRF fails completely:
+
+| Silver (wrong) | Gold (correct) | sent_id |
+|----------------|----------------|---------|
+| `thiên_địa_lưu` | `thiên_địa` \| `lưu` | uvw-11634 |
+| `tất_hữu_tụng` | `tất` \| `hữu` \| `tụng` | uvw-379 |
+| `chi_dĩ_tụng` | `chi` \| `dĩ` \| `tụng` | uvw-379 |
+| `ẩm_thực_chi` | `ẩm_thực` \| `chi` | uvw-379 |
+| `Đinh_Lê_nhị_gia` | `Đinh_Lê` \| `nhị_gia` | uvw-4040 |
+
+**Root cause**: Classical Hán-Việt text uses mostly single-syllable words, but the CRF trained on modern Vietnamese aggressively merges adjacent syllables.
+
+#### 3. Foreign word/abbreviation merges
+
+| Silver (wrong) | Gold (correct) | sent_id |
+|----------------|----------------|---------|
+| `I_LOVE` | `I` \| `LOVE` | uvb-f-16922 |
+| `BST_Love` | `BST` \| `Love` | uvn-16886 |
+| `DNA_replication` | `DNA` \| `replication` | uvw-16915 |
+| `Frankael-Zermelo_Tiên` | `Frankael-Zermelo` \| `Tiên_đề` | uvw-14316 |
+
+**Root cause**: CRF treats consecutive Latin-script tokens as one word.
+
+#### 4. Compound word boundary shifts
+
+CRF merges across true word boundaries, creating invalid compounds:
+
+| Silver (wrong) | Gold (correct) | Type |
+|----------------|----------------|------|
+| `lợi_nhuận_tích` + `lũy` | `lợi_nhuận` + `tích_lũy` | boundary shift |
+| `niên_đại_mẫu` + `vật` | `niên_đại` + `mẫu_vật` | boundary shift |
+| `phóng_vệ_tinh` | `phóng` + `vệ_tinh` | over-merge |
+| `quy_phạm_pháp_luật` | `quy_phạm` + `pháp_luật` | over-merge (legal) |
+| `trình_bày_văn_bản` | `trình_bày` + `văn_bản` | over-merge (legal) |
+
+#### 5. Common compound under-merges (over-splits)
+
+CRF fails to merge well-known compound words:
+
+| Silver (wrong) | Gold (correct) |
+|----------------|----------------|
+| `Ủy` \| `ban` | `Ủy_ban` (committee) |
+| `lính` \| `thú` | `lính_thú` (soldier) |
+| `mu` \| `rùa` | `mu_rùa` (turtle shell) |
+| `trêu` \| `ghẹo` | `trêu_ghẹo` (tease) |
+| `sương` \| `mai` | `sương_mai` (morning dew) |
+| `mái` \| `nhà` | `mái_nhà` (roof) |
+| `đầm` \| `đuôi` \| `cá` | `đầm_đuôi_cá` (fishtail dress) |
+| `cha` \| `truyền_con` \| `nối` | `cha_truyền_con_nối` (hereditary) |
+| `siêu` \| `máy_tính` | `siêu_máy_tính` (supercomputer) |
+| `lượng` \| `tử_lai` | `lượng_tử` (quantum) + `lai` |
+
+#### 6. Proper name under-merges
+
+| Silver (wrong) | Gold (correct) |
+|----------------|----------------|
+| `Trương_Hán` \| `Siêu` | `Trương_Hán_Siêu` |
+| `Nữ_minh` \| `tinh` | `Nữ_minh_tinh` |
+| `Nicolas` \| `Fatio_de` \| `Duillier` | `Nicolas_Fatio_de_Duillier` |
+| `Nigeria_Ibrahim` \| `Baré_Maïnassara` | `Nigeria_Ibrahim_Baré_Maïnassara` |
+| `World_Music` \| `Awards` | `World_Music_Awards` |
+
+## All Sentence-Level Differences
+
+### Fiction (20/28 sentences with differences)
+
+**uvb-f-10494**: `dạo_khúc` → `dạo` | `khúc`
+
+**uvb-f-1078**: `Mặt_Kỳ` + `Nam_xụ` → `Mặt` | `Kỳ_Nam` | `xụ`
+
+**uvb-f-11528**: `thần_linh_giữ` → `thần_linh` | `giữ`
+
+**uvb-f-116**: `lính` | `thú` → `lính_thú`
+
+**uvb-f-12571**: `nỡ_tẩy` + `chay` → `nỡ` | `tẩy_chay`
+
+**uvb-f-14791**: `như` + `điên_sợ` → `như_điên` | `sợ`
+
+**uvb-f-16030**: `Quỳnh_nghiến` + `răng` → `Quỳnh` | `nghiến_răng`
+
+**uvb-f-16202**: `Nữ_minh` + `tinh_khịt` + `mũi` → `Nữ_minh_tinh` | `khịt_mũi`
+
+**uvb-f-16922**: `Ahern_PS` + `I_LOVE` + `YOU_Chương` → `Ahern` | `PS` | `I` | `LOVE` | `YOU` | `Chương`
+
+**uvb-f-17901**: `gia_nhân_túc_trực` + `chăn` | `lợn` → `gia_nhân` | `túc_trực` | `chăn_lợn`
+
+**uvb-f-18098**: `Khải_nheo` + `nheo` → `Khải` | `nheo_nheo`
+
+**uvb-f-2195**: `người` | `thương` + `Bờ` | `xe` | `nước_tạc` + `chi_bao` → `người_thương` | `Bờ_xe_nước` | `tạc` | `chi` | `bao`
+
+**uvb-f-3070**: `Xin` | `chào_sĩ_quan` → `Xin_chào` | `sĩ_quan`
+
+**uvb-f-6015**: `Liên_xua` | `tay` + `nói` | `gở` → `Liên` | `xua_tay` | `nói_gở`
+
+**uvb-f-6292**: `Trương_Hán` | `Siêu_bưng` + `khay` | `trà` + `lên_án` → `Trương_Hán_Siêu` | `bưng` | `khay_trà` | `lên` | `án`
+
+**uvb-f-6893**: `Nguyễn_Bính_Mưa_dầm` → `Nguyễn_Bính` | `Mưa` | `dầm`
+
+**uvb-f-7016**: `Nguyễn_Bính_Công` + `diều_mướn` → `Nguyễn_Bính` | `Công` | `diều` | `mướn`
+
+**uvb-f-7435**: `Mặt_San` + `y_dơn` | `dớt` → `Mặt` | `San` | `y` | `dơn_dớt`
+
+**uvb-f-8734**: `có_học` + `bơi` | `chó` → `có` | `học` | `bơi_chó`
+
+**uvb-f-9174**: `tạo_nghiệp_lành` → `tạo` | `nghiệp_lành`
+
+### Non-fiction (18/24 sentences with differences)
+
+**uvb-n-10024**: `Sống_biến` + `Lạnh_biến` → `Sống` | `biến` | `Lạnh` | `biến`
+
+**uvb-n-11587**: `phải_biết` → `phải` | `biết`
+
+**uvb-n-11976**: `Lệ` | `thủy_trình` → `Lệ_thủy` | `trình`
+
+**uvb-n-12321**: `sương` | `mai_phớt` + `mái` | `nhà` → `sương_mai` | `phớt` | `mái_nhà`
+
+**uvb-n-13233**: `lạnh_đặc` → `lạnh` | `đặc`
+
+**uvb-n-13288**: `I._Lương` + `Huệ_vương` → `I.` | `Lương_Huệ_vương`
+
+**uvb-n-16553**: `đương_ngủ` | `ngày` → `đương` | `ngủ_ngày`
+
+**uvb-n-18226**: `Ngữ` | `đoạn` + `Khái_niệm_ngữ` | `đoạn` → `Ngữ_đoạn` | `Khái_niệm` | `ngữ_đoạn`
+
+**uvb-n-18736**: `Lục` | `Giả_đáp` → `Lục_Giả` | `đáp`
+
+**uvb-n-19309**: `Lạc` | `Hồng_định` + `trêu` | `ghẹo` → `Lạc_Hồng` | `định` | `trêu_ghẹo`
+
+**uvb-n-2201**: `quyết_phục_thù` → `quyết` | `phục_thù`
+
+**uvb-n-3769**: `Chu_Du_tính_tình` → `Chu_Du` | `tính_tình`
+
+**uvb-n-3847**: `răng` | `nanh_sắc` | `nhọn` + `triết_gia_lùi` → `răng_nanh` | `sắc_nhọn` | `triết_gia` | `lùi`
+
+**uvb-n-3874**: `thành` | `thực_cầu` | `mong` + `cao_minh_phủ` | `chính` → `thành_thực` | `cầu_mong` | `cao_minh` | `phủ_chính`
+
+**uvb-n-3968**: `đồng` | `tự` + `làm_âm` → `đồng_tự` | `làm` | `âm`
+
+**uvb-n-511**: `Ủy` | `ban` + `Hành` | `chánh` → `Ủy_ban` | `Hành_chánh`
+
+**uvb-n-6010**: `chơi_trò` → `chơi` | `trò`
+
+**uvb-n-7402**: `mu` | `rùa` → `mu_rùa`
+
+**uvb-n-7669**: `Cung` | `Khôn_Thái` + `Cần_Thánh_chỗ` → `Cung_Khôn_Thái` | `Cần_Thánh` | `chỗ`
+
+### Wikipedia (22/28 sentences with differences)
+
+**uvw-10198**: `Hữu_thưởng` + `vu_đại_quốc` → `Hữu` | `thưởng` | `vu` | `đại_quốc`
+
+**uvw-11634**: `Thệ_tâm` + `thiên_địa_lưu` + `xỉ_giang` + `sơn_thổ` + `thiệt_hồng` → `Thệ` | `tâm` | `thiên_địa` | `lưu` | `xỉ` | `giang_sơn` | `thổ` | `thiệt` | `hồng`
+
+**uvw-13194**: `Quân_xướng` + `thần_họa` → `Quân` | `xướng` | `thần` | `họa`
+
+**uvw-13430**: `Đường` | `thẳng` → `Đường_thẳng`
+
+**uvw-14316**: `Hệ_tiên` | `đề` (×2) + `đề_số` | `học` + `Frankael-Zermelo_Tiên` | `đề_chọn` + `Tiên` | `đề` + `tiên` | `đề` → `Hệ_tiên_đề` (×2) | `số_học` | `Frankael-Zermelo` | `Tiên_đề` (×2) | `chọn` | `tiên_đề`
+
+**uvw-1451**: `Quan_Âm_Diệu_Thiện` → `Quan_Âm` | `Diệu_Thiện`
+
+**uvw-15121**: `Nhà_Habsburg` + `bắt_đầu_tích` | `lũy` + `cha` | `truyền_con` | `nối` → `Nhà` | `Habsburg` | `bắt_đầu` | `tích_lũy` | `cha_truyền_con_nối`
+
+**uvw-15306**: `Tống_tiêu_diệt` → `Tống` | `tiêu_diệt`
+
+**uvw-15392**: `Hải_quốc_văn_kiến` | `lục_》` → `Hải_quốc_văn_kiến_lục` | `》`
+
+**uvw-15619**: `Long` | `Hồ_dinh_đổi` + `Hà_Tiên_trấn` → `Long_Hồ` | `dinh` | `đổi` | `Hà_Tiên` | `trấn`
+
+**uvw-16915**: `DNA_replication` → `DNA` | `replication`
+
+**uvw-17218**: `Điểm_nóng` | `chảy` → `Điểm` | `nóng_chảy`
+
+**uvw-17235**: `Khổng_Dĩnh` | `Đạt_thời` + `Hoa_Hạ_vi` → `Khổng_Dĩnh_Đạt` | `thời` | `Hoa_Hạ` | `vi`
+
+**uvw-18245**: `ngăn_ắc` + `quy_ổn_định` → `ngăn` | `ắc_quy` | `ổn_định`
+
+**uvw-19010**: `nhà_toán_học` + `Nicolas` | `Fatio_de` | `Duillier` → `nhà` | `toán_học` | `Nicolas_Fatio_de_Duillier`
+
+**uvw-2178**: `Đồng_nguyên` + `chất_mềm` + `đồng_tươi` → `Đồng` | `nguyên_chất` | `mềm` | `đồng` | `tươi`
+
+**uvw-3214**: `Thiên_địa_hòa` | `xướng` + `chi_tượng` + `giao` | `hòa` → `Thiên_địa` | `hòa_xướng` | `chi` | `tượng` | `giao_hòa`
+
+**uvw-379**: `Phục` | `Hy_ghi` + `ẩm_thực_chi` + `tất_hữu_tụng` + `cố_thụ` + `chi_dĩ_tụng` → `Phục_Hy` | `ghi` | `ẩm_thực` | `chi` (×2) | `tất` | `hữu` | `tụng` (×2) | `cố` | `thụ` | `dĩ`
+
+**uvw-4040**: `Đinh_Lê_nhị_gia` + `võng_đạo` + `an_quyết` + `đại_phất` | `thế` + `toán` | `số` → `Đinh_Lê` | `nhị_gia` | `võng` | `đạo` | `an` | `quyết` | `thế_đại` | `phất` | `toán_số`
+
+**uvw-415**: `Chỉ` | `dã` → `Chỉ_dã`
+
+**uvw-4428**: `Nhu_vu` + `trinh_cát` → `Nhu` | `vu` | `trinh` | `cát`
+
+**uvw-5791**: `Kỳ_địa` + `quảng_nhi` + `quyết_thổ` + `cao_nhi` + `sảng` | `khải` → `Kỳ` | `địa` | `quảng` | `nhi` (×2) | `quyết` | `thổ` | `cao` | `sảng_khải`
+
+**uvw-6347**: `miệt_hôn_điếm` + `chi_khốn` + `chi_phong` → `miệt` | `hôn_điếm` | `chi` (×2) | `khốn` | `phong`
+
+**uvw-7555**: `Quân_Mỹ` → `Quân` | `Mỹ`
+
+**uvw-8056**: `Nigeria_Ibrahim` | `Baré_Maïnassara` → `Nigeria_Ibrahim_Baré_Maïnassara`
+
+**uvw-8082**: `Grand` | `Prix_Award` + `MTV_Video` | `Music` | `Awards_Japan` + `World_Music` | `Awards` + `Japan_Gold` | `Disc_Award` → `Grand_Prix_Award` | `MTV_Video_Music_Awards_Japan` | `World_Music_Awards` | `Japan_Gold_Disc_Award`
+
+### News (7/8 sentences with differences)
+
+**uvn-16886**: `đầm` | `đuôi` | `cá` + `thân` | `áo` + `BST_Love` → `đầm_đuôi_cá` | `thân_áo` | `BST` | `Love`
+
+**uvn-18693**: `niên_đại_mẫu` | `vật` → `niên_đại` | `mẫu_vật`
+
+**uvn-18725**: `phóng_vệ_tinh` → `phóng` | `vệ_tinh`
+
+**uvn-19673**: `Tăng_Hoàng_Vinh_giữ` → `Tăng_Hoàng_Vinh` | `giữ`
+
+**uvn-4740**: `lợi_nhuận_tích` | `lũy` → `lợi_nhuận` | `tích_lũy`
+
+**uvn-6121**: `of_refusing` + `Matt_really` + `new_leaf` + `and_has` → `of` | `refusing` | `Matt` | `really` | `new` | `leaf` | `and` | `has`
+
+**uvn-964**: `siêu` | `máy_tính` + `lượng` | `tử_lai` → `siêu_máy_tính` | `lượng_tử` | `lai`
+
+### Legal (3/4 sentences with differences)
+
+**vlc-10445**: `mi���n_thị_thực` → `miễn` | `thị_thực`
+
+**vlc-6220**: `Ủy` | `ban_thường_vụ` + `trình_bày_văn_bản` + `quy_phạm_pháp_luật` + `Chủ_tịch` | `nước` → `Ủy_ban` | `thường_vụ` | `trình_bày` | `văn_bản` | `quy_phạm` | `pháp_luật` | `Chủ_tịch_nước`
+
+**vlc-7243**: `phong_quân` (×2) + `hàm_sĩ_quan` (×2) + `tại` | `ngũ` → `phong` (×2) | `quân_hàm` (×2) | `sĩ_quan` (×2) | `tại_ngũ`
+
+## Key Takeaways
+
+1. **AL selection works**: The 92 gold sentences have Word F1 = 0.78, much lower than the 0.98 silver baseline, confirming that uncertainty scoring correctly identifies hard cases.
+
+2. **Cross-boundary merge is the #1 error**: CRF extends proper name spans into the next word, creating invalid compounds like `Nguyễn_Bính_Mưa_dầm`.
+
+3. **Wikipedia is hardest**: F1 = 0.707, driven by Hán-Việt classical text (single-syllable words mismerged) and foreign names.
+
+4. **Over-merge slightly > over-split**: 134 vs 119 errors. The CRF has a slight bias toward merging.
+
+5. **Legal domain has systematic issues**: 4-syllable compound over-merges (`quy_phạm_pháp_luật`, `trình_bày_văn_bản`) and boundary shifts on `quân_hàm`/`sĩ_quan`.
+
+6. **Foreign text always over-merged**: English words (`I_LOVE`, `DNA_replication`, `new_leaf`) are treated as Vietnamese compounds.
+
+## Recommendations for Cycle 2
+
+1. **Add proper name boundary features** to CRF: capitalization patterns, known name lists
+2. **Add foreign word detection**: Latin-script tokens should default to single-word segmentation
+3. **Increase legal domain sampling**: Only 4 sentences in Cycle 1, need more legal-specific gold data
+4. **Classical Hán-Việt**: Consider special handling or separate model for classical text passages
+5. **Retrain CRF** with the 92 gold corrections mixed into training data, then re-score remaining 19,900 sentences for Cycle 2 selection

PLAN_v1.1.md

@@ -4,7 +4,8 @@
 
 1. **Dataset**: 100K-sentence multi-domain WS dataset (BIO format) for CRF training
 2. **Guidelines**: Vietnamese annotation guidelines for WS, POS, DP
-3. **Gold annotation**: 2–5K gold sentences via active learning → in-domain LAS/UAS
+3. **Gold evaluation set**: Fixed gold dev/test (500 sentences) before AL begins
+4. **Gold annotation**: 2–5K gold sentences via active learning → in-domain LAS/UAS
 
 ## Status
 
@@ -14,25 +15,86 @@
 | 2 | Build BIO dataset + stratified split | Done | `udd-ws-v1.1-{train,dev,test}.txt` |
 | 3 | WS diagnostic checker (7 rules) | Done | `src/check_ws_errors.py`, `WS_CHECK_REPORT.md` |
 | 4 | WS annotation guidelines | Done | `guidelines/01. Word Segmentation/` |
-| 5 | WS error fix pipeline | Done | `src/fix_ws_errors.py`, `WS_FIX_REPORT.md` |
-| 5a | — Pass 1: Cross-boundary splits (19) | Done | |
-| 5b | — Pass 1.5: Long token splits (597) | Done | |
-| 5c | — Pass 2: Compound merges (22 terms) | Done | |
-| 5d | — Pass 3: BIO validation | Done | |
+| 5 | WS error fix pipeline (7 passes) | Done | `src/fix_ws_errors.py`, `WS_FIX_REPORT.md` |
+| 5a | — Pass 1: Cross-boundary splits (258) | Done | |
+| 5b | — Pass 1.5: Long token splits (0) | Done | |
+| 5c | — Pass 2: Compound merges (541, +25 from Cycle 1) | Done | |
+| 5d | — Pass 2.5: Foreign word splits (5,375) | Done | |
+| 5e | — Pass 2.75: Proper name boundary splits (816) | Done | |
+| 5f | — Pass 3: BIO validation (0 errors) | Done | |
 | 6 | Apply fixes + regenerate CoNLL-U | Done | `WS_FIX_REPORT.md` |
 | 7 | Re-run checker post-fix | Done | `WS_CHECK_REPORT.md` (updated) |
 | 8 | Sentence selection guidelines | Done | `guidelines/00. Sentence Selection/` |
 | 9 | AL literature survey (35 papers) | Done | `active_learning/references/` |
-| 10 | Train CRF baseline (80K silver) | Done | `tree-1/models/.../udd_ws_v1_1-20260211_034002/` |
+| 10 | Train CRF baseline (80K silver) | Done | `tree-1/models/.../udd_ws_v1_1-20260212_065135/` |
 | 11 | AL Cycle 0: uncertainty scoring (20K sent) | Done | `al_cycle0_ranked.tsv`, `al_cycle0_top500.tsv` |
-| 12 | AL Cycle 1: gold WS annotation (500 sent) | Planned | |
-| 13 | AL Cycle 2: retrain + annotate (500 sent) | Planned | |
-| 14 | AL Cycle 3: annotate (500 sent) | Planned | |
-| 15 | AL Cycle 4: annotate (500 sent) | Planned | |
-| 16 | POS annotation guidelines | Planned | `guidelines/02. POS Tagging/` |
-| 17 | DP annotation guidelines | Planned | `guidelines/03. Dependency Parsing/` |
-| 18 | Gold POS annotation (1K sentences) | Planned | AL Task 2 |
-| 19 | Gold DP annotation (800–1K sentences) | Planned | AL Task 3 |
+| 12 | Label Studio annotation tooling | Done | `src/ls_import_ws.py`, `src/ls_export_ws.py`, `src/ls_config_ws.xml` |
+| 13 | Dictionary plugin for annotation | Done | `src/build_dict_plugin.py`, `src/build_dict_search.py`, `src/dict_plugin.html` |
+| 14 | Generate Cycle 1 import + LS import | Done | `ls_import_cycle1.json` → LS project 4 (100 tasks) |
+| 15 | AL Cycle 1: gold WS annotation (100 sent) | Done | 95/100 annotated, 92 exported → `gold_ws_cycle1.txt` |
+| 15a | — Export gold BIO from Label Studio | Done | `ls_export_cycle1.json` → `gold_ws_cycle1.txt` (92 sent) |
+| 15b | — 3 sentences skipped (annotation gaps) | — | `uvn-3488`, `uvb-n-10657`, `uvb-f-10312` |
+| 16 | AL Cycle 1: evaluate CRF on gold | Done | Word F1=0.7808 on 92 gold sentences |
+| 16a | Post-Cycle 1: extend fix pipeline (3 new passes) | Done | Foreign splits, name boundary, +25 compounds |
+| 16b | Post-Cycle 1: re-run fixes on 100K silver | Done | 6,990 total fixes applied |
+| 16c | Post-Cycle 1: merge 92 gold into silver | Done | `src/merge_gold_silver.py`, 41 dev + 51 test replaced |
+| 16d | Post-Cycle 1: retrain CRF | Done | `udd_ws_v1_1-20260213_034134/`, Silver F1=0.9842 |
+| 16e | Post-Cycle 1: evaluate retrained CRF on gold | Done | Word F1=0.7988 (+0.018 from 0.7808) |
+| 16f | Post-Cycle 1: re-score for Cycle 2 | Done | New top-500 ranked, mean score 0.0130 |
+| **P0** | **Phase 0: Gold evaluation set (500 sentences)** | **Planned** | |
+| P0a | — Random sample 500 sentences (100/domain) from pool | Planned | `gold_eval_sample.tsv` |
+| P0b | — Annotate 500 gold WS in Label Studio | Planned | `gold_ws_eval.txt` |
+| P0c | — Split 250 dev + 250 test, freeze | Planned | `gold_ws_eval_dev.txt`, `gold_ws_eval_test.txt` |
+| P0d | — Eval baseline CRF on gold test (250 sent) | Planned | True baseline F1 on representative data |
+| P0e | — Remove 500 eval sentences from AL candidate pool | Planned | Pool: 20K → 19.5K |
+| 17 | AL Cycle 2: annotate (200 sent) | Planned | |
+| 18 | AL Cycle 3: annotate (200 sent) | Planned | |
+| 19 | AL Cycle 4: annotate (200 sent) | Planned | |
+| 19a | AL Cycle 5: annotate (200 sent) | Planned | |
+| 19b | AL Cycle 6: annotate (200 sent) | Planned | |
+| 19c | AL Cycle 7: annotate (200 sent) | Planned | |
+| 19d | AL Cycle 8: annotate (200 sent) | Planned | |
+| 19e | AL Cycle 9: annotate (200 sent) | Planned | |
+| 19f | AL Cycle 10: annotate (200 sent) | Planned | |
+| 20 | POS annotation guidelines | Planned | `guidelines/02. POS Tagging/` |
+| 21 | DP annotation guidelines | Planned | `guidelines/03. Dependency Parsing/` |
+| 22 | Gold POS annotation (1K sentences) | Planned | AL Task 2 |
+| 23 | Gold DP annotation (800–1K sentences) | Planned | AL Task 3 |
+
+## Phase 0: Gold Evaluation Set
+
+**Rationale**: The AL literature universally requires a **fixed, randomly-sampled gold evaluation set created before AL begins** (Zhang et al., EMNLP 2022; Shi et al., NAACL 2021; Settles & Craven, EMNLP 2008; Luth et al., NeurIPS 2023). Without this:
+- No representative F1 baseline (uncertainty-selected sentences are biased)
+- No valid learning curves across AL cycles
+- No reliable stopping criteria
+- Selection bias is undetectable
+
+**Why Cycle 1's 92 gold sentences cannot serve as eval set:**
+
+| Problem | Detail |
+|---------|--------|
+| Selection bias | Selected by **highest uncertainty**, not random → hardest cases only |
+| Domain skew | Fiction 30%, Wikipedia 30%, Non-fiction 26%, News 9%, Legal 4% (should be 20% each) |
+| Data leakage | Already merged into training data (41 dev + 51 test) |
+| Unrepresentative F1 | 0.78 on these vs ~0.98 on silver — overstates error rate |
+
+**Cycle 1's 92 gold sentences remain valuable as:**
+- Training data supplement (already merged into silver — correct usage)
+- Diagnostic set for error pattern analysis
+- NOT as evaluation set for stopping criteria
+
+**Phase 0 protocol:**
+
+1. **Random sample**: Draw 500 sentences uniformly from the 20K dev+test pool, stratified by domain (100 per domain: legal, news, Wikipedia, fiction, non-fiction)
+2. **Exclude from AL pool**: Remove these 500 from all future uncertainty scoring and AL selection
+3. **Annotate**: Gold WS annotation in Label Studio using the same guidelines as Cycle 1
+4. **Split**: 250 dev + 250 test (50 per domain in each split), random seed 42
+5. **Freeze**: These sets are **never modified** during AL cycles — no merging into training, no re-annotation, no selection bias
+6. **Baseline eval**: Evaluate current CRF on the 250 gold test sentences → this is the **true F1 baseline**
+
+**Expected annotation time**: 500 sentences ÷ 80 sent/day = ~6 days
+
+**Minimum size justification**: UD guidelines require ≥10K words per evaluation set. At ~26 syllables/sentence and ~20 words/sentence, 250 sentences ≈ 5,000 words per set. This is below the UD 10K threshold but sufficient for WS evaluation with reasonable confidence intervals. If budget allows, 400 dev + 400 test (800 total) would be ideal.
 
 ## Active Learning for Word Segmentation
 
@@ -42,75 +104,213 @@
 
 **Query strategy**: CRF token marginal uncertainty — `u_i = 1 - max(P(B|x,i), P(I|x,i))`. Sentence score = mean uncertainty × (1 + boundary_weight) + 0.01 × n_long_tokens. Also prioritize sentences with known error patterns (4+ syllable tokens, high-uncertainty zones).
 
+**Evaluation**: After each cycle, retrain CRF and evaluate on the **Phase 0 gold test set (250 sentences)**. Plot learning curve (F1 vs cumulative gold annotations).
+
 **Cycles**:
 
-| Cycle | Sentences | Cumulative | Status | Focus |
-|-------|----------:|-----------:|--------|-------|
-| 0 | 0 | 0 | Done | Train CRF on 80K silver, establish baseline |
-| 1 | 500 | 500 | Planned | Highest uncertainty + known error patterns |
-| 2 | 500 | 1,000 | Planned | Retrained CRF's new uncertain tokens |
-| 3 | 500 | 1,500 | Planned | Domain-specific compounds, remaining errors |
-| 4 | 500 | 2,000 | Planned | Diminishing returns check; stop if F1 plateaus |
+| Phase/Cycle | Sentences | Cumulative | Status | Focus |
+|-------------|----------:|-----------:|--------|-------|
+| **Phase 0** | **500** | **500** | **Planned** | **Gold eval set: 250 dev + 250 test (random, stratified)** |
+| Cycle 0 | 0 | 0 | Done | Train CRF on 80K silver, establish baseline |
+| Cycle 1 | 92 | 92 | Done | Top 100 uncertainty, 92 gold → diagnostic + training data |
+| Cycle 2 | 200 | ~290 | Planned | Retrained CRF's new uncertain tokens |
+| Cycle 3 | 200 | ~490 | Planned | Domain-specific compounds, remaining errors |
+| Cycle 4 | 200 | ~690 | Planned | Diminishing returns check |
+| Cycle 5 | 200 | ~890 | Planned | |
+| Cycle 6 | 200 | ~1,090 | Planned | |
+| Cycle 7 | 200 | ~1,290 | Planned | |
+| Cycle 8 | 200 | ~1,490 | Planned | |
+| Cycle 9 | 200 | ~1,690 | Planned | |
+| Cycle 10 | 200 | ~1,890 | Planned | Stop if F1 plateaus (<0.1% improvement) |
+
+**Total gold**: 500 (eval) + ~1,890 (AL) = ~2,390 sentences
 
-**Annotation**: Correct BIO tags using NIIVTB 9-rule framework (`guidelines/01. Word Segmentation/`). Flag ambiguous cases for guideline revision. Est. 80 sent/day → 500 sent ≈ 6 days/cycle.
+**Annotation**: Correct BIO tags in Label Studio using NIIVTB 9-rule framework (`guidelines/01. Word Segmentation/`). Dictionary plugin highlights multi-syllable words (green = in dict, red = not in dict) for faster review. Flag ambiguous cases for guideline revision. Est. 80 sent/day → 200 sent ≈ 2.5 days/cycle.
 
-**Stopping**: <0.1% F1 improvement on held-out gold test set (200 sentences from Cycle 1).
+**Stopping**: <0.1% F1 improvement on **Phase 0 gold test set** (250 sentences, fixed, never modified) over two consecutive cycles.
 
 **Quality**: Re-annotate 5% blind every 200 sentences → target >95% self-consistency.
 
-## Baseline Results (Cycle 0)
+## CRF Model Results
+
+### Cycle 0 (baseline: 80K silver)
 
 | Metric | Score |
 |--------|-------|
-| Syllable Accuracy | 0.9916 |
-| Word F1 | **0.9848** |
-| Word Precision | 0.9843 |
-| Word Recall | 0.9854 |
+| Syllable Accuracy | 0.9908 |
+| Word F1 | **0.9834** |
+| Word Precision | 0.9830 |
+| Word Recall | 0.9838 |
 
-Model: `tree-1/models/word_segmentation/udd_ws_v1_1-20260211_034002/`
+Model: `tree-1/models/word_segmentation/udd_ws_v1_1-20260212_065135/`
 Training: 80K silver sentences, pycrfsuite L-BFGS, c1=0.5, c2=0.001, 300 iterations.
 
 **Uncertainty scoring** (20K dev+test sentences):
-- Mean uncertainty: 0.009
-- 7,454 sentences (37.3%) with score ≥ 0.01
-- 696 sentences (3.5%) with score ≥ 0.05
-- 831 sentences have 4+ syllable tokens
-- Top 500 selected: 237 dev + 263 test
+- Mean uncertainty: 0.0103
+- Mean score: 0.0136
+- 8,636 sentences (43.2%) with score ≥ 0.01
+- 848 sentences (4.2%) with score ≥ 0.05
+- 1,068 sentences have 4+ syllable tokens
+- Top 500 selected: 253 dev + 247 test
+
+### Post-Cycle 1 (fixed silver + 92 gold)
+
+**Fix pipeline results** (3 new passes added):
+
+| Fix Type | Count | Description |
+|----------|------:|-------------|
+| Cross-boundary splits | 258 | `Mặt trận Tổ quốc`, `làm Chủ tịch`, etc. |
+| Long token splits | 0 | (already fixed in v1.1) |
+| Compound merges | 541 | `ủy ban` (371), `mái nhà` (41), `lượng tử` (24), +22 more |
+| Foreign word splits | 5,375 | Latin-only multi-syllable tokens split per syllable |
+| Name boundary splits | 816 | `Trần khởi nghiệp` → `Trần` + `khởi nghiệp` |
+| **Total fixes** | **6,990** | |
+
+**Silver test performance:**
+
+| Metric | Cycle 0 | Post-Cycle 1 | Δ |
+|--------|---------|-------------|---|
+| Syllable Accuracy | 0.9908 | 0.9913 | +0.0005 |
+| Word F1 | 0.9834 | **0.9842** | +0.0008 |
+
+Model: `tree-1/models/word_segmentation/udd_ws_v1_1-20260213_034134/`
+Training: 80K fixed silver + 92 gold sentences merged, pycrfsuite L-BFGS, 300 iterations.
+
+**Uncertainty scoring** (20K dev+test, retrained model):
+- Mean uncertainty: 0.0099 (was 0.0103)
+- Mean score: 0.0130 (was 0.0136)
+- 8,407 sentences (42.0%) with score ≥ 0.01 (was 8,636)
+- 773 sentences (3.9%) with score ≥ 0.05 (was 848)
+- 1,042 sentences have 4+ syllable tokens (was 1,068)
+- Top 500 selected: 261 dev + 239 test
+
+### Cycle 1 Gold Evaluation (92 uncertainty-selected sentences)
+
+**Important caveat**: These 92 sentences were selected by **highest CRF uncertainty**, not randomly sampled. The F1 scores below represent performance on the **hardest** sentences, not overall model performance. See Phase 0 for representative evaluation.
+
+**Overall metrics:**
+
+| Metric | Cycle 0 CRF | Post-Cycle 1 CRF | Δ |
+|--------|-------------|-------------------|---|
+| Syllable Accuracy | 0.8665 | 0.8758 | +0.0093 |
+| Word Precision | 0.7934 | 0.7977 | +0.0043 |
+| Word Recall | 0.7685 | 0.7998 | +0.0313 |
+| Word F1 | **0.7808** | **0.7988** | **+0.0180** |
+| Sentences with diffs | 75/92 | 74/92 | −1 |
+| Over-merge errors | 134 | 113 | −21 |
+| Over-split errors | 118 | 112 | −6 |
+
+**Per-domain breakdown (Post-Cycle 1 CRF):**
+
+| Domain | N | Syl Acc | P | R | F1 | Δ F1 |
+|--------|---|---------|---|---|-----|------|
+| News | 8 | 0.9152 | 0.8606 | 0.8497 | **0.8551** | +0.016 |
+| Legal | 4 | 0.8765 | 0.8298 | 0.8333 | **0.8316** | +0.119 |
+| Non-fiction | 24 | 0.9033 | 0.8283 | 0.8056 | 0.8168 | −0.007 |
+| Fiction | 28 | 0.8757 | 0.8132 | 0.8121 | 0.8127 | −0.002 |
+| Wikipedia | 28 | 0.8381 | 0.7340 | 0.7609 | **0.7472** | +0.040 |
+
+**Error analysis** (74/92 sentences with differences):
+
+| Error Type | Cycle 0 | Post-Cycle 1 | Δ |
+|------------|---------|-------------|---|
+| Over-merge | 134 | 113 | −21 |
+| Over-split | 118 | 112 | −6 |
+| B→I changes | 83 | 95 | +12 |
+| I→B changes | 118 | 92 | −26 |
+
+**Remaining dominant error patterns**:
+1. Cross-boundary merge at proper names: `Nguyễn_Bính_Mưa_dầm`, `Chu_Du_tính_tình`
+2. Hán-Việt boundary errors (wikipedia): `thiên_địa_lưu`, `Hải_quốc_văn_kiến`
+3. Foreign name segmentation: `Nigeria_Ibrahim_Baré_Maïnassara`
+4. Some compound splits remain: `lính | thú`, `ắc | quy`
+
+## Dataset Statistics (v1.1 regenerated)
+
+| | Train | Dev | Test |
+|---|---|---|---|
+| **Sentences** | 80,000 | 10,000 | 10,000 |
+| **Syllables** | 2,229,051 | 278,364 | 278,798 |
+| **Words** | 1,696,362 | 211,880 | 212,088 |
+| **Unique words** | 71,286 | 21,346 | 21,106 |
+| **Single-syl ratio** | 70.2% | 70.2% | 70.1% |
+| **Multi-syl ratio** | 29.8% | 29.8% | 29.9% |
+| **Avg syl/word** | 1.31 | 1.31 | 1.31 |
+| **OOV token rate** | — | 2.5% | 2.4% |
+
+Domain distribution: exactly 20% per domain (legal, news, wikipedia, fiction, non-fiction) in all splits.
+
+| Domain | Avg Syl/Sent | Syl/Word | Character |
+|--------|---:|---:|---|
+| Legal | 33.2 | 1.49 | Longest sentences, most compound words |
+| News | 32.9 | 1.33 | |
+| Wikipedia | 31.2 | 1.30 | |
+| Non-fiction | 22.1 | 1.22 | |
+| Fiction | 19.9 | 1.18 | Shortest sentences, fewest compounds |
+
+## Annotation Tooling
+
+| Script | Purpose |
+|--------|---------|
+| `src/ls_import_ws.py` | Convert AL top-500 TSV → Label Studio JSON with pre-annotations + CRF confidence scores |
+| `src/ls_export_ws.py` | Convert Label Studio JSON export → BIO format |
+| `src/ls_config_ws.xml` | Label Studio labeling config (W/WH/WM/WL labels) |
+| `src/build_dict_plugin.py` | Generate dictionary plugin HTML for Label Studio (highlight bar + search) |
+| `src/build_dict_search.py` | Generate standalone dictionary search HTML |
+| `ls_import_cycle1.json` | Label Studio import for Cycle 1 (500 sentences with pre-annotations) |
+
+## Cycle 1 Results
+
+- **Annotated**: 95/100 tasks in Label Studio (project 4)
+- **Exported**: 92 gold sentences → `gold_ws_cycle1.txt` (3 skipped: annotation gaps)
+- **Skipped sentences**: `uvn-3488` (missing spans for foreign words), `uvb-n-10657` (missing "là"), `uvb-f-10312` (missing "»")
 
 ## Next Steps
 
-1. AL Cycle 1: select 500 sentences from `al_cycle0_top500.tsv`, start gold WS annotation
-2. After Cycle 1: retrain CRF with gold corrections, re-score remaining sentences
-3. POS annotation guidelines
-4. DP annotation guidelines
+1. **Phase 0: Gold evaluation set** — random sample 500 sentences, annotate, split 250 dev + 250 test, freeze
+2. **Baseline eval** — evaluate CRF on Phase 0 gold test set → true F1 baseline
+3. **AL Cycle 2** — generate Label Studio import from new top-500 (excluding 500 eval + 92 Cycle 1)
+4. POS annotation guidelines
+5. DP annotation guidelines
 
 ## Timeline
 
 ```
-Week:   1    2    3    4    5    6    7    8    9   10   11   12
-       +----+----+----+----+----+----+----+----+----+----+----+
-Task 1: ====================
-         WS gold (2K sentences, 4 cycles)
-Task 2:           ==================
-                   POS gold (1K sentences, 4 cycles)
-Task 3:                     ============================
-                             DP gold (800-1K sentences, 6 cycles)
+Week:   1    2    3    4    5    6    7    8    9   10   11   12   13
+       +----+----+----+----+----+----+----+----+----+----+----+----+
+Phase0: =====
+         Gold eval set (500 sent, ~6 days)
+Task 1:      =====================
+              WS gold (2K sentences, AL cycles 2-10)
+Task 2:                ==================
+                        POS gold (1K sentences, 4 cycles)
+Task 3:                          ============================
+                                  DP gold (800-1K sentences, 6 cycles)
 Guide:  *-------*-------*-------*-------*-------*
         WS      POS     DP v0.1  DP v0.2  v1.0
 ```
 
-Est. total: ~100 annotator-days (~12 weeks with overlap).
+Est. total: ~106 annotator-days (~13 weeks with overlap, +1 week for Phase 0).
 
 ## Related Documents
 
 | Document | Content |
 |----------|---------|
 | `TECHNICAL_REPORT.md` | v1.0 technical report (10K UD treebank) |
-| `TECHNICAL_REPORT_1.1.md` | v1.1 technical report (WS dataset + AL framework) |
+| `TECHNICAL_REPORT_v1.1.md` | v1.1 technical report (WS dataset + AL framework) |
 | `WS_CHECK_REPORT.md` | Diagnostic checker output (7 rules) |
-| `WS_FIX_REPORT.md` | Fix pipeline output (examples, counts) |
+| `WS_FIX_REPORT.md` | Fix pipeline output (258 cross-boundary, 541 compound, 5,375 foreign, 816 name boundary) |
+| `AL_CYCLE1_REPORT.md` | Cycle 1 annotation report (92 gold sentences, error analysis) |
 | `al_cycle0_ranked.tsv` | All 20K sentences ranked by uncertainty |
 | `al_cycle0_top500.tsv` | Top 500 sentences for Cycle 1 annotation |
+| `ls_import_cycle1.json` | Label Studio import for Cycle 1 (500 pre-annotated sentences) |
+| `ls_export_cycle1.json` | Label Studio export for Cycle 1 (95 annotated tasks) |
+| `gold_ws_cycle1.txt` | Gold WS BIO for Cycle 1 (92 sentences) |
+| `src/eval_ws_gold.py` | Evaluate CRF vs gold (Word F1, per-domain, error analysis; `--model` for direct CRF prediction) |
+| `src/merge_gold_silver.py` | Merge gold annotations into silver BIO files for CRF retraining |
 | `src/al_score_ws.py` | AL uncertainty scoring script |
+| `src/ls_import_ws.py` | Label Studio import generator (BIO → JSON with CRF confidence) |
+| `src/ls_export_ws.py` | Label Studio export converter (JSON → BIO) |
+| `src/build_dict_plugin.py` | Dictionary plugin builder for Label Studio |
 | `guidelines/00. Sentence Selection/` | Sentence selection criteria |
 | `guidelines/01. Word Segmentation/` | WS annotation guidelines (NIIVTB 9 rules) |

TECHNICAL_REPORT_v1.1.md

@@ -1,56 +1,98 @@
-# UDD-1 v1.1: Toward a Gold-Standard Vietnamese Universal Dependencies Treebank via Active Learning
+# UDD-WS v1.1: A 100K-Sentence Multi-Domain Vietnamese Word Segmentation Dataset and Active Learning Framework for Gold-Standard UD Annotation
 
 **Underthesea NLP**
 
 ## Abstract
 
-UDD-1 v1.0 established a 10,000-sentence silver-standard Vietnamese UD treebank from the legal domain. Version 1.1 takes two steps toward gold-standard annotation. First, we scale the data foundation: a multi-domain word segmentation dataset of 100,000 sentences in BIO format (VLSP-compatible) across 5 domains (legal, news, Wikipedia, fiction, non-fiction), enabling training of robust word segmentation models that form the prerequisite for accurate dependency parsing. Second, we lay out a concrete active learning framework for constructing a gold-standard Vietnamese UD treebank under a **solo annotator** setting, with task-specific AL strategies for word segmentation (CRF marginal uncertainty), POS tagging (tag confusion targeting), and dependency parsing (head entropy with partial arc annotation). The pipeline is estimated at ~100 annotator-days to produce 2,000 gold WS, 1,000 gold POS, and 800--1,000 gold DP sentences.
+We present UDD-WS v1.1, a 100,000-sentence silver-standard Vietnamese word segmentation dataset in BIO format spanning five domains (legal, news, Wikipedia, fiction, non-fiction), and a three-task active learning framework for constructing gold-standard Universal Dependencies annotations under a solo annotator constraint. The dataset is built with a 14-rule quality filter pipeline, diversity-aware round-robin sampling across 4,291 source documents, and two-level deduplication (exact + digit-normalized). We train a CRF word segmenter on the silver data (Word F1 = 0.9842 on silver test) and conduct a first active learning cycle, annotating 92 gold sentences selected by token marginal uncertainty. Evaluation against gold reveals Word F1 = 0.7988 on these high-uncertainty sentences, with cross-boundary proper name merges and Sino-Vietnamese classical text as dominant error categories. Critically, we identify a methodological gap in our initial AL design --- the absence of a fixed, randomly-sampled gold evaluation set --- and propose Phase 0: a 500-sentence gold evaluation set (250 dev + 250 test) created before continuing AL cycles, following established AL evaluation methodology (Zhang et al., 2022; Shi et al., 2021). The AL framework targets 2,000 gold word segmentation, 1,000 gold POS, and 800--1,000 gold dependency sentences over approximately 106 annotator-days, with Vietnamese annotation guidelines co-developed through the process.
 
-Beyond gold annotation, a central goal of this work is the co-development of **standardized Vietnamese annotation guidelines** for the three core NLP tasks: word segmentation, POS tagging, and dependency parsing. Vietnamese currently lacks a unified, publicly available annotation standard that covers all three tasks consistently. Existing guidelines are fragmented --- the VLSP 2013 shared task defined word segmentation conventions, the NIIVTB project (Nguyen et al., 2018) established 9 rules for word boundary decisions, and UD_Vietnamese-VTB provides limited language-specific UD guidelines --- but no single resource integrates them into a coherent annotation framework. Through the active learning loop, where the most ambiguous and informative examples are systematically surfaced, we develop comprehensive Vietnamese annotation guidelines that address language-specific phenomena (copula *là*, passive markers *được/bị*, serial verb constructions, classifier phrases, topic-comment structure) with explicit decision procedures, worked examples, and cross-task consistency. These guidelines are intended as a reusable community resource for future Vietnamese treebank and dataset construction.
+## 1. Introduction
 
-This report documents the word segmentation dataset (udd-ws-v1.1), the active learning roadmap, and the guideline development methodology.
+### 1.1 Background
 
-## 1. Introduction
+Word segmentation is the foundational task in Vietnamese NLP. Unlike English, Vietnamese uses spaces to separate syllables rather than words, so multi-syllable words like *doanh nghiệp* ("enterprise") appear as two space-separated tokens. All downstream NLP tasks --- POS tagging, named entity recognition, dependency parsing --- depend on correct word segmentation, and segmentation errors propagate through the entire pipeline.
 
-### 1.1 Motivation
+Despite its importance, Vietnamese word segmentation resources are limited. The VLSP 2013 shared task (Nguyen et al., 2013) provided the primary benchmark dataset (~75K sentences, news domain), but no large-scale multi-domain BIO-format dataset exists for training robust segmenters. Meanwhile, Vietnamese Universal Dependencies treebanks remain small: UD_Vietnamese-VTB contains 3,323 sentences (Nguyen et al., 2014), and UDD-1 v1.0 provides 10,000 silver-standard sentences from the legal domain only.
+
+### 1.2 Motivation
 
 The TECHNICAL_REPORT_REVIEW of UDD-1 v1.0 identified a clear path forward:
 
-> *"Could you run the Underthesea parser on 50-100 sentences from the legal corpus that have been manually annotated by a Vietnamese linguist, to report in-domain LAS/UAS? This single addition would move the paper from borderline to solid accept."*
+> *"Could you run the Underthesea parser on 50--100 sentences from the legal corpus that have been manually annotated by a Vietnamese linguist, to report in-domain LAS/UAS?"*
 
-More broadly, the review highlighted three priorities: (1) gold-standard evaluation of annotation quality, (2) reduction of the 8.6% UPOS-forcing artifact rate, and (3) multi-domain coverage beyond the legal domain. Version 1.1 addresses all three through a two-pronged strategy:
+The review highlighted three priorities: (1) gold-standard evaluation of annotation quality, (2) reduction of the 8.6% UPOS-forcing artifact rate, and (3) multi-domain coverage beyond the legal domain.
 
-- **Word segmentation dataset (udd-ws-v1.1)**: A 100K-sentence, 5-domain BIO-tagged dataset that provides the training data for robust word segmentation --- the first step in the Vietnamese NLP pipeline (WS → POS → DP).
-- **Active learning framework**: A systematic plan to construct gold-standard UD annotations efficiently, using the silver-standard data as a starting point and selectively correcting the most informative errors.
+Version 1.1 addresses these through a two-pronged strategy: a 100K-sentence multi-domain word segmentation dataset providing the training data for robust tokenization, and an active learning framework for efficient gold-standard annotation across the full Vietnamese NLP pipeline (WS → POS → DP).
 
-### 1.2 Overview of Contributions
+### 1.3 Contributions
 
-| Contribution | Status | Description |
-|---|---|---|
-| udd-ws-v1.1 dataset | Complete | 100K sentences, 5 domains, BIO format, stratified splits |
-| Sentence quality pipeline | Complete | 14-rule quality filter + continuous quality scoring |
-| Diversity-aware sampling | Complete | Round-robin extraction with per-source caps |
-| Active learning literature survey | Complete | 35 papers covering AL for parsing, treebank construction, guidelines |
-| Vietnamese WS annotation guidelines | Planned | NIIVTB 9-rule framework adapted for UDD-1 |
-| AL framework design | This report | Concrete plan for gold-standard UD annotation |
-| Gold-standard annotation | Planned | Target: 2-5K sentences via active learning |
+Our contributions are threefold:
+
+1. **UDD-WS v1.1 dataset**: A 100,000-sentence, 5-domain word segmentation dataset in BIO format with a rigorous quality pipeline (14-rule filter, round-robin diversity sampling, two-level deduplication). To our knowledge, this is the largest multi-domain Vietnamese word segmentation dataset in BIO format.
+
+2. **Active learning framework with proper evaluation methodology**: A three-task sequential AL pipeline (WS → POS → DP) for gold-standard UD annotation under a solo annotator constraint, with Phase 0 (fixed gold evaluation set created before AL) following established AL methodology. We report results from the first AL cycle (92 gold WS sentences) including a detailed error taxonomy.
+
+3. **Silver-standard quality analysis**: A systematic assessment of word segmentation quality in silver-standard data, revealing 3,592 inconsistently segmented word forms (67,612 occurrences) and identifying dominant error categories that inform both the fix pipeline (6,990 corrections) and the AL query strategy.
+
+## 2. Related Work
+
+### 2.1 Vietnamese Word Segmentation
+
+Vietnamese word segmentation has been studied since the early 2000s. The VLSP 2013 shared task (Nguyen et al., 2013) established the primary benchmark, with CRF-based systems achieving 97--98% F1 on news text. VnCoreNLP (Nguyen et al., 2018a) reports 97.90% F1 on VLSP 2013 using a CRF with hand-crafted features and dictionary lookup. The Underthesea toolkit provides a word_tokenize function combining CRF and regex-based tokenization.
+
+Despite these advances, all existing benchmarks are news-domain. No Vietnamese WS dataset covers legal, literary, or encyclopedic text, where domain-specific vocabulary (legal compounds, Sino-Vietnamese classical terms, foreign names) poses distinct segmentation challenges.
+
+### 2.2 Vietnamese Treebanks
+
+Four Vietnamese treebanks are relevant to this work.
+
+**VTB / VnDT / UD_Vietnamese-VTB.** The Vietnamese Treebank (VTB; Nguyen et al., 2014) is a constituency treebank of ~10,000 news sentences. VnDT (Nguyen et al., 2014) was automatically converted from VTB's bracketing to dependency format (10,200 sentences), achieving UAS 80.7% / LAS 73.5% at the time. UD_Vietnamese-VTB is a further automatic conversion to Universal Dependencies format containing 3,323 sentences --- the only UD-compliant Vietnamese treebank currently in the UD repository.
+
+**BKTreebank.** Nguyen (2018) constructed a 6,909-sentence dependency treebank from Dantri news text, notable as the first Vietnamese dependency treebank **manually annotated from scratch** rather than converted from a constituency treebank. BKTreebank uses a Penn-style POS tagset with 23 tags including 6 Vietnamese-specific tags (CL for classifiers, PFN/NML for nominalizers, VA/AV for adjectival/verbal adjectives) and 26 Stanford dependency relations with 2 Vietnamese-specific additions (`case:pfn`, `mark:relcl`). Three annotators plus one reviewer achieved inter-annotator agreement of 94.5% for POS, 85.2% UAS, and 80.4% LAS. Baseline parsing results were UAS 84.4% / LAS 81.4% with MaltParser. However, BKTreebank uses Stanford Dependencies rather than Universal Dependencies, limiting interoperability with the UD ecosystem.
+
+**NIIVTB.** The NIIVTB project (Nguyen et al., 2018b) established 9 rules for Vietnamese word boundary decisions and achieved 97.28% inter-annotator consistency across ~40,000 sentences annotated for word segmentation, POS tagging, and bracketing. Their word segmentation guidelines form the basis for our annotation framework.
 
-## 2. Word Segmentation Dataset: udd-ws-v1.1
+**UDD-1 v1.0.** Underthesea NLP (2025) provides 10,000 silver-standard sentences from the legal domain, produced by neural parsing with rule-based post-processing. No gold-standard evaluation exists.
 
-### 2.1 Rationale
+| Treebank | Sentences | Domain | Annotation | UD-compliant | IAA |
+|----------|----------:|--------|------------|:---:|-----|
+| VnDT | 10,200 | News | Auto-converted | No | --- |
+| UD_Vietnamese-VTB | 3,323 | News | Auto-converted | Yes | --- |
+| BKTreebank | 6,909 | News | Manual | No | 94.5% POS, 80.4% LAS |
+| NIIVTB | ~40,000 | News | Manual | No | 97.28% WS |
+| UDD-1 v1.0 | 10,000 | Legal | Silver (neural) | Yes | --- |
 
-Word segmentation is the foundation of Vietnamese NLP. The dependency parser in UDD-1 v1.0 uses an implicit tokenizer that produces segmentation errors propagating to all downstream annotations (Section 5.6 of TECHNICAL_REPORT.md). Building a dedicated word segmentation dataset enables:
+**Table R1**: Vietnamese treebank comparison.
 
-1. Training domain-robust CRF word segmentation models (tree-1 pipeline)
-2. Evaluating segmentation quality across domains independently of the parser
-3. Providing a clean input pipeline for future gold-standard UD annotation
+No existing Vietnamese treebank provides gold-standard multi-domain annotations. All are limited to news text, and only UD_Vietnamese-VTB (auto-converted, 3,323 sentences) follows UD format. BKTreebank's manual annotation methodology and IAA reporting provide a valuable reference point for our solo-annotator AL approach, though our work targets UD compliance and multi-domain coverage.
 
-### 2.2 Data Collection
+### 2.3 Active Learning for NLP
 
-Sentences are drawn from 4 HuggingFace datasets, the same sources as UDD-1 v1.0 but scaled to 20,000 sentences per domain:
+Active learning reduces annotation cost by selecting the most informative examples. Zhang et al. (2022) provide a comprehensive survey of AL for NLP, covering query strategies, batch selection, stopping criteria, and evaluation methodology. For structured prediction tasks, key approaches include:
+
+- **Uncertainty sampling** for sequence labeling: CRF token marginal uncertainty (Settles and Craven, 2008) selects sentences where the model is least confident.
+- **Head entropy for dependency parsing**: Li et al. (2016) propose annotating only the arcs where the parser is uncertain (partial annotation), reducing effort by 40--60%.
+- **Batch diversity**: Shi et al. (2021) combine uncertainty with DPP-based diversity sampling for dependency parsing, achieving 20--30% cost reduction.
+- **Sequential task AL**: Zhang et al. (2023) demonstrate that partial annotation with self-training provides the best cost reduction across four structured prediction tasks.
+
+A critical but often underemphasized aspect of AL evaluation is the requirement for a **fixed, randomly-sampled evaluation set created before AL begins**. Zhang et al. (2022, Section 5.2) note that measuring performance on a development set "would be unstable if this set is too small," but having no evaluation set is far worse. Farquhar et al. (2021) formalize that AL introduces statistical bias in the training data, which is undetectable without an independent test set. Luth et al. (2023) identify five pitfalls in AL evaluation, with inconsistent evaluation methodology as the primary concern.
+
+### 2.4 Treebank Construction Methodology
+
+The standard practice in treebank construction is to define evaluation splits before or at the time of annotation. The Penn Treebank's sections 02--21/22/23 split (Marcus et al., 1993) has been the universal standard for 30 years. For AL-based treebank construction, all existing work --- including Shi et al. (2021), Li et al. (2016), and the Pomak-Philotis UD treebank (2023) --- assumes fixed evaluation sets exist before AL iterations begin.
+
+UD project guidelines (de Marneffe et al., 2021) specify minimum sizes for evaluation data: for treebanks of 20K--110K words, at least 10K words for test data and 10% of the remainder for dev data. Bouma and van Noord (2017) demonstrate increasing return on annotation investment through selective manual correction of automatically annotated data --- the same paradigm we follow.
+
+Brants and Skut (1998) showed that correcting pre-annotated treebank data is 3--5x faster than annotation from scratch, providing the throughput estimates that underpin our cost analysis.
+
+## 3. UDD-WS v1.1 Dataset
+
+### 3.1 Data Sources
+
+Sentences are drawn from four HuggingFace datasets across five domains, with 20,000 sentences per domain:
 
 | Domain | Source Dataset | Sentences | Sent ID Prefix |
-|--------|---------------|-----------|----------------|
+|--------|---------------|----------:|----------------|
 | Legal | `undertheseanlp/UTS_VLC` | 20,000 | `vlc-` |
 | News | `undertheseanlp/UVN-1` | 20,000 | `uvn-` |
 | Wikipedia | `undertheseanlp/UVW-2026` | 20,000 | `uvw-` |
@@ -58,124 +100,100 @@ Sentences are drawn from 4 HuggingFace datasets, the same sources as UDD-1 v1.0
 | Non-fiction | `undertheseanlp/UVB-v0.1` | 20,000 | `uvb-n-` |
 | **Total** | | **100,000** | |
 
-**Table 1**: Domain breakdown of udd-ws-v1.1.
-
-#### 2.2.1 Preprocessing
+**Table 1**: Domain breakdown of UDD-WS v1.1.
 
-All source texts undergo Unicode normalization (`underthesea.text_normalize()`), markdown removal, and whitespace normalization before sentence segmentation. Sentence splitting uses `safe_sent_tokenize()`, a wrapper around `underthesea.sent_tokenize()` that fixes incorrect splits inside Vietnamese thousands-separator numbers (e.g., "2.000" was split into "2." + "000..."). This affected ~1,700 sentences across the corpus.
+### 3.2 Quality Filter Pipeline
 
-#### 2.2.2 Quality Filters
+Every sentence must pass a 14-rule quality filter pipeline before inclusion. The full specification is available in `guidelines/00. Sentence Selection/`.
 
-Every sentence must pass a comprehensive quality filter pipeline (14 rules) before inclusion. The full specification is in `guidelines/00. Sentence Selection/Guideline.md`.
-
-**Common filters** (`base_valid()`, all domains):
+**Common filters** (all domains):
 
 | Category | Rule | Condition |
 |----------|------|-----------|
 | Structure | Length bounds | 20--300 characters |
 | Structure | Minimum words | >= 4 words |
 | Structure | Proper start | First character uppercase or digit |
-| Structure | Proper end | Terminal punctuation (`.!?…"»"'):`) |
+| Structure | Proper end | Terminal punctuation (`.!?...`) |
 | Language | Vietnamese content | At least one Vietnamese diacritical character |
 | Language | Not mostly uppercase | Uppercase characters <= 50% |
-| Language | Language detection | `underthesea.lang_detect()` == `"vi"` with fallback: accept if 3+ Vietnamese diacritical characters (prevents false rejection of sentences with foreign proper nouns) |
+| Language | Language detection | `lang_detect()` == `"vi"` with fallback |
 | Markup | No HTML/template | Reject `{`, `}`, `<`, `>`, `\|` |
-| Markup | No template key=value | Reject `\w+=\w+` (but allow spaced `=` like `GDP = 3.500 USD`) |
-| Markup | No file extensions | Reject `.jpg`, `.png`, `.gif`, `.svg`, `.webp` |
-| Brackets | Balanced | `(` count == `)` count; `[` count == `]` count |
-| Glued text | Multi-tone detection | Reject tokens with 2+ toned vowels (Vietnamese syllables carry at most one tone mark; multiple tones indicate OCR-like merge errors) |
-| Glued text | Digit-letter glue | Reject tokens with a toned vowel AND `\d[letter]` pattern |
-
-**Domain-specific filters** add further rules: legal header/article removal, news byline/caption filtering, Wikipedia stub/infobox/reference rejection, and books apply stricter thresholds (30--250 chars, 5--40 words, <= 30% uppercase, <= 15% digits, punctuation density, no mid-sentence ellipsis, limited dialogue).
+| Markup | No template key=value | Reject `\w+=\w+` (allow spaced `=`) |
+| Markup | No file extensions | Reject `.jpg`, `.png`, `.gif`, etc. |
+| Brackets | Balanced | `(` count == `)` count |
+| Glued text | Multi-tone detection | Reject tokens with 2+ toned vowels |
+| Glued text | Digit-letter glue | Reject toned vowel AND `\d[letter]` pattern |
 
-#### 2.2.3 Deduplication
+**Table 2**: Quality filter rules applied to all sentences.
 
-Two levels of deduplication are applied during extraction (not as post-processing):
+Domain-specific filters add further rules: legal header/article removal, news byline/caption filtering, Wikipedia stub/infobox/reference rejection, and stricter thresholds for book text (30--250 chars, 5--40 words, <= 30% uppercase, punctuation density limits).
 
-- **Exact deduplication**: Raw sentence strings are tracked in a `seen` set. Identical sentences from different documents are rejected.
-- **Near-duplicate detection**: All digit sequences are normalized to `#` (`normalize_for_dedup()`), then exact-matched. This catches formulaic sentences differing only in numbers (e.g., legal clauses referencing different article numbers, news reporting different statistics).
-
-#### 2.2.4 Diversity-Aware Sampling
+### 3.3 Diversity-Aware Sampling
 
 Initial sequential extraction produced low source diversity: fiction was dominated by 5 books (primarily Murakami's *1Q84*), and legal text came from only 106 documents. We replaced sequential extraction with **round-robin sampling** with per-source caps:
 
-1. **Phase 1 (Candidate collection)**: Scan all documents/books, collecting up to `MAX_PER_DOC=500` (or `MAX_PER_BOOK=500`) valid sentences per source. For non-book domains, scanning stops at 3x the target candidate count.
-2. **Phase 2 (Round-robin selection)**: Cycle through all sources, taking one sentence per source per round, with exact and near-duplicate deduplication. This ensures no single source dominates.
-
-Books are ranked by quality score (`goodreads_rating * min(num_ratings / 100, 10)`) and processed in descending order, so higher-quality books are scanned first.
+1. **Phase 1 (Candidate collection)**: Scan all documents, collecting up to 500 valid sentences per source.
+2. **Phase 2 (Round-robin selection)**: Cycle through all sources, taking one sentence per source per round, with exact and near-duplicate deduplication.
 
-**Source diversity achieved**:
+Books are ranked by quality score (Goodreads rating weighted by number of ratings) and processed in descending order.
 
-| Domain | Sequential (before) | Round-robin (after) |
+| Domain | Sequential | Round-robin |
 |--------|:---:|:---:|
 | Legal | 106 documents | 275 documents |
 | News | 1,030 documents | 3,091 documents |
 | Wikipedia | 180 articles | 785 articles |
 | Fiction | 5 books | 72 books |
 | Non-fiction | 20 books | 68 books |
+| **Total** | **1,341 sources** | **4,291 sources** |
 
-**Table 1b**: Source diversity improvement from round-robin sampling.
-
-#### 2.2.5 Sentence Quality Score
-
-Beyond binary pass/fail filters, each sentence is assigned a continuous quality score via `sentence_score()` returning a value in (0, 1). The score combines 6 sub-scores:
+**Table 3**: Source diversity improvement from round-robin sampling.
 
-| Sub-score | Weight | Description |
-|-----------|--------|-------------|
-| Length | 0.20 | Gaussian around ideal range [60, 200] chars |
-| Word count | 0.15 | Gaussian around ideal range [8, 35] words |
-| Structure | 0.20 | Proper start (+0.5) + proper end (+0.5) |
-| Cleanliness | 0.30 | Penalties for markup, unbalanced brackets, glued text, file extensions, excessive uppercase |
-| Completeness | 0.15 | Penalties for unbalanced quotes, excessive digit ratio |
-| Vietnamese density | multiplier | Ratio of Vietnamese diacritical characters to total letters (20%+ maps to 1.0). Applied as a **multiplier** so non-Vietnamese text scores near 0. |
+### 3.4 Deduplication
 
-Formula: `final_score = clamp(base_score * vietnamese_density, 0.01, 0.99)`
+Two levels of deduplication are applied during extraction:
 
-This score is used for quality reporting and will be used for active learning sentence selection (prioritizing sentences in the 0.7--0.95 range where annotation effort has the highest marginal value).
+- **Exact deduplication**: Raw sentence strings tracked in a `seen` set.
+- **Near-duplicate detection**: All digit sequences normalized to `#`, then exact-matched. This catches formulaic sentences differing only in numbers (e.g., legal clauses referencing different article numbers).
 
-### 2.3 BIO Annotation
+### 3.5 BIO Annotation
 
-Sentences are converted to syllable-level BIO format using:
+Sentences are converted to syllable-level BIO format:
 
-1. `underthesea.word_tokenize(sentence, format="text")` → compound tokens with underscores (e.g., `"Việt_Nam là một quốc_gia"`)
-2. `underthesea.pipeline.word_tokenize.regex_tokenize.tokenize()` → syllable splitting
-3. First syllable of each token → `B-W`, continuation syllables → `I-W`
+1. `underthesea.word_tokenize(sentence, format="text")` → compound tokens with underscores
+2. `regex_tokenize()` → syllable splitting
+3. First syllable → `B-W`, continuation syllables → `I-W`
 
 Output format (VLSP 2013 compatible, tab-separated with comment headers):
 
 ```
 # sent_id = vlc-1
-# text = Một doanh nghiệp lớn hoạt động hiệu quả .
-Một	B-W
-doanh	B-W
-nghiệp	I-W
-lớn	B-W
-hoạt	B-W
-động	I-W
-hiệu	B-W
-quả	I-W
-.	B-W
+# text = Mot doanh nghiep lon hoat dong hieu qua .
+Mot     B-W
+doanh   B-W
+nghiep  I-W
+lon     B-W
+hoat    B-W
+dong    I-W
+hieu    B-W
+qua     I-W
+.       B-W
 ```
 
-This format is directly loadable by tree-1's `load_data_vlsp2013()` function, which maps `B-W → B`, `I-W → I`.
-
-### 2.4 Dataset Statistics
+### 3.6 Dataset Statistics
 
 | | Train | Dev | Test | Total |
 |---|---:|---:|---:|---:|
 | Sentences | 80,000 | 10,000 | 10,000 | 100,000 |
-| Words | 1,592,531 | 201,565 | 197,832 | 1,991,928 |
-| Syllables | 2,087,103 | 263,953 | 259,338 | 2,610,394 |
-| Avg word/sent | 19.91 | 20.16 | 19.78 | 19.92 |
-| Avg syl/sent | 26.09 | 26.40 | 25.93 | 26.10 |
+| Words | 1,696,362 | 211,880 | 212,088 | 2,120,330 |
+| Syllables | 2,229,051 | 278,364 | 278,798 | 2,786,213 |
+| Avg word/sent | 21.20 | 21.19 | 21.21 | 21.20 |
+| Avg syl/sent | 27.86 | 27.84 | 27.88 | 27.86 |
 | Avg syl/word | 1.31 | 1.31 | 1.31 | 1.31 |
 
-**Table 2**: udd-ws-v1.1 split statistics.
+**Table 4**: UDD-WS v1.1 split statistics.
 
 Splits are stratified by domain (each domain contributes exactly 20% to every split) with random seed 42 for reproducibility.
 
-### 2.5 Word Length Distribution
-
 | Syllables per word | Count | Percentage |
 |:---:|---:|---:|
 | 1 | 1,403,963 | 70.48% |
@@ -184,417 +202,508 @@ Splits are stratified by domain (each domain contributes exactly 20% to every sp
 | 4 | 4,384 | 0.22% |
 | 5+ | 1,127 | 0.06% |
 
-**Table 3**: Word length distribution across the full dataset.
+**Table 5**: Word length distribution. ~70% single-syllable, ~28% two-syllable compounds, consistent with Vietnamese linguistics.
 
-The distribution is consistent with Vietnamese linguistics: ~70% single-syllable words, ~28% two-syllable compounds, and ~2% longer compounds. The 1.31 average syllables per word aligns with the 1.38 figure reported for legal text in SEGMENTATION_EVAL.md (the difference reflecting the inclusion of less formal fiction/non-fiction domains).
+| Domain | Avg Syl/Sent | Syl/Word | Character |
+|--------|---:|---:|---|
+| Legal | 33.2 | 1.49 | Longest sentences, most compounds |
+| News | 32.9 | 1.33 | |
+| Wikipedia | 31.2 | 1.30 | |
+| Non-fiction | 22.1 | 1.22 | |
+| Fiction | 19.9 | 1.18 | Shortest sentences, fewest compounds |
 
-### 2.6 Input Quality Summary
+**Table 6**: Per-domain characteristics.
 
-The 14-rule quality filter pipeline ensures that all 100,000 input sentences are well-formed Vietnamese text. Post-filtering validation confirms 0 duplicates, 0 near-duplicates, 0 unbalanced brackets, 0 markup remnants, 0 lowercase starts, 0 missing end punctuation, and 0 glued text across the entire dataset.
+### 3.7 Silver-Standard Quality Assessment
 
-### 2.7 Silver-Standard Caveat
+The BIO annotations are generated automatically by `underthesea.word_tokenize()` (v2.1.0) and inherit its segmentation biases. We assess silver quality using a 7-rule diagnostic checker (`check_ws_errors.py`):
 
-The BIO annotations are generated automatically by `underthesea.word_tokenize()` and inherit its segmentation biases. SEGMENTATION_EVAL.md documents known issues including 462 over-segmented dictionary words (7,373 occurrences) and 382 potentially under-segmented tokens. The dataset is a silver-standard resource suitable for CRF training but not for evaluation of segmentation quality. A gold-standard evaluation subset is planned as part of the active learning framework (Section 3).
+| Rule | Description | Unique Forms | Total Occurrences |
+|------|-------------|-------------:|------------------:|
+| 1 | Inconsistent segmentation | 3,592 | 67,612 |
+| 4 | Long tokens (4+ syllables) | 928 | 4,606 |
+| 5 | Punctuation boundary errors | 299 | 554 |
+| 6 | Number-word boundary errors | 150 | 175 |
+| 7 | Single-character anomalies | 13 | 972 |
 
-## 3. Active Learning Framework for Gold-Standard UD Annotation
+**Table 7**: Silver-standard quality diagnostics.
 
-### 3.1 Goal
+**Inconsistent segmentation** (Rule 1) is the dominant issue: 3,592 unique word forms appear both as single tokens and as split syllables across the dataset. For example, *suc khoe* ("health") appears 410 times as a single token and 383 times as split --- a near 50-50 split. This 67,612/2,786,213 = **2.4% inconsistency rate** at the token level characterizes the noise floor of the silver standard.
 
-Construct a **gold-standard** Vietnamese UD treebank of 2,000--5,000 sentences with verified word segmentation, POS tags, and dependency relations, sufficient to:
+Notable inconsistently segmented forms include:
 
-1. Report in-domain LAS/UAS on legal and multi-domain text
-2. Serve as evaluation data for parser development
-3. Serve as seed training data for active-learning-boosted parsers
-4. Establish Vietnamese-specific UD annotation guidelines through the annotation process
+| Form | As Single | As Split | Split Ratio |
+|------|----------:|---------:|---:|
+| suc khoe | 410 | 383 | 48.3% |
+| cong hoa | 782 | 163 | 17.2% |
+| goi la | 654 | 78 | 10.6% |
+| nguoi ta | 26 | 530 | 95.3% |
+| nhat la | 129 | 370 | 74.1% |
 
-### 3.2 Solo Annotator Setting
+**Table 8**: Top inconsistently segmented forms (simplified).
 
-This project operates under a **solo annotator** constraint: a single Vietnamese linguist performs all annotation. This is a realistic setting for under-resourced languages where trained annotators are scarce. The solo setting has specific implications:
+Based on gold annotation from AL Cycle 1, we developed a rule-based fix pipeline (`fix_ws_errors.py`) that corrects 6,990 errors across the 100K sentences:
 
-- **No inter-annotator agreement (IAA)**: Quality is ensured through consistency checks, model-based error detection, and guideline self-auditing rather than dual annotation.
-- **Consistency advantage**: A single annotator produces internally consistent annotations, avoiding the reconciliation overhead of multi-annotator setups.
-- **Anchoring risk**: The annotator corrects silver-standard pre-annotations, creating anchoring bias toward the parser's output. Mitigation: guidelines explicitly instruct the annotator to evaluate each decision independently; periodic blind re-annotation of 5% of sentences to measure self-consistency.
-- **Throughput**: One annotator correcting pre-annotated data can process 50--100 sentences/day for word segmentation, 30--50 for POS, and 15--30 for dependency parsing (estimates from Brants & Skut 1998, adjusted for Vietnamese complexity).
+| Fix Type | Count | Description |
+|----------|------:|-------------|
+| Cross-boundary splits | 258 | Proper names merged with following words |
+| Compound merges | 541 | Known compounds that were over-split |
+| Foreign word splits | 5,375 | Latin-script multi-syllable tokens |
+| Name boundary splits | 816 | Proper names extended across word boundaries |
+| **Total** | **6,990** | **7.0% of sentences affected** |
 
-**Quality assurance without IAA**:
+**Table 9**: Fix pipeline corrections applied to silver data.
 
-| Method | Purpose | Frequency |
-|--------|---------|-----------|
-| Self-consistency check | Re-annotate 5% of completed sentences blind | Every 200 sentences |
-| Model-based error detection | Flag arcs where retrained model disagrees with gold | After each AL cycle |
-| Dictionary validation | Cross-check WS against Viet74K dictionary | Continuous |
-| UD validator | Automated structural constraint checking | After each batch |
-| Guideline self-audit | Review decisions against written guidelines | Weekly |
+## 4. Active Learning Framework
+
+### 4.1 Design Principles
 
-**Table 4**: Quality assurance methods for solo annotator setting.
+Our AL framework is guided by four principles:
 
-### 3.3 Why Active Learning
+1. **Fixed evaluation before AL** (Phase 0): A randomly-sampled gold evaluation set must exist before any AL iteration begins, following Zhang et al. (2022), Shi et al. (2021), and standard practice in all AL literature.
+2. **Sequential task pipeline**: WS → POS → DP, where each task's gold output feeds the next task's annotation pool, preventing cascading errors.
+3. **Uncertainty + known error targeting**: Combining model uncertainty with known error pattern coverage ensures both model blind spots and systematic issues are addressed.
+4. **Solo annotator QA**: Quality assurance through self-consistency checks, model-based error detection, and guideline self-auditing, rather than inter-annotator agreement.
 
-Manual annotation of dependency treebanks is expensive. Brants & Skut (1998) showed that correcting pre-annotated data is 3--5x faster than annotation from scratch. Active learning further reduces cost by selecting the most informative examples for annotation.
+### 4.2 Phase 0: Gold Evaluation Set
 
-The literature (surveyed in `active_learning/references/research_active_learning_ud/`) demonstrates:
+**Rationale**: The standard AL evaluation protocol requires a fixed, randomly-sampled gold test set that exists before any AL iteration (Zhang et al., 2022, Section 5.2; Farquhar et al., 2021; Luth et al., 2023). Without this:
+- Learning curves (performance vs. annotation budget) cannot be plotted
+- Selection bias from uncertainty sampling is undetectable
+- Stopping criteria are unreliable
+- Model-data coupling cannot be assessed
 
-| Method | Cost Reduction | Reference |
-|--------|---------------|-----------|
-| Uncertainty sampling | ~50% fewer sentences | Hwa (2004) |
-| Head entropy + partial annotation | 40--60% less arc annotation | Li et al. (2016) |
-| DPP batch diversity + uncertainty | ~20--30% fewer sentences | Shi et al. (2021) |
-| Partial annotation + self-training | Best cost reduction across 4 tasks | Zhang et al. (2023) |
+Our initial Cycle 1 selected 100 sentences by **highest CRF uncertainty**, producing 92 gold sentences with a domain distribution skewed toward fiction (30%) and Wikipedia (30%) rather than the balanced 20% per domain. These sentences represent the hardest cases (Word F1 = 0.78 vs. 0.98 on silver) and cannot serve as a representative evaluation set.
 
-**Table 5**: Active learning cost reduction benchmarks from the literature.
+**Protocol**:
 
-UDD-1's situation is particularly favorable for AL: we already have 100K silver-standard sentences that can serve as the initial model's training data and as candidates for selective correction.
+1. **Random sample**: 500 sentences from the 20K dev+test pool, stratified by domain (100 per domain)
+2. **Annotate**: Gold WS following the NIIVTB 9-rule guidelines
+3. **Split**: 250 dev + 250 test (50 per domain per split)
+4. **Freeze**: Never modified during AL --- no merging into training, no re-selection
+5. **Exclude from AL pool**: These 500 sentences are removed from all future uncertainty scoring
 
-### 3.4 Three-Task Active Learning Pipeline
+**Size justification**: At ~26 syllables/sentence and ~20 words/sentence, 250 sentences provide ~5,000 words per evaluation set. While below the UD guideline of 10K words, this provides sufficient statistical power for word segmentation F1 evaluation. Comparable UD treebanks for under-resourced languages use similar sizes (Pomak: 635 test sentences; VTB: 1,123 test sentences for all three UD layers combined).
 
-The Vietnamese NLP pipeline is sequential: **Word Segmentation → POS Tagging → Dependency Parsing**. Each task depends on the output of the previous one, so errors cascade. We apply active learning independently to each task with task-specific strategies, proceeding in pipeline order.
+### 4.3 Three-Task Sequential Pipeline
+
+The Vietnamese NLP pipeline is sequential: **Word Segmentation → POS Tagging → Dependency Parsing**. Each task depends on the output of the previous one. We apply active learning independently to each task with task-specific query strategies:
 
 ```
-  ┌─────────────────────────────────────────────────────────────────┐
-  │                  Task 1: Word Segmentation                      │
-  │                                                                 │
-  │  Silver data ──► Train CRF ──► Score uncertainty ──► Annotate  │
-  │  (100K BIO)      (tree-1)      (token marginals)     (correct  │
-  │                                                       BIO tags) │
-  │                       ◄──── Retrain ◄──── Gold WS data         │
-  └───────────────────────────────┬─────────────────────────────────┘
-                                  │ Gold-segmented sentences
-  ┌───────────────────────────────▼─────────────────────────────────┐
-  │                  Task 2: POS Tagging                             │
-  │                                                                 │
-  │  Silver POS ──► Train CRF ──► Score uncertainty ──► Annotate   │
-  │  (auto-tagged)   (tree-1)     (tag marginals)      (correct    │
-  │                                                      UPOS tags) │
-  │                       ◄──── Retrain ◄──── Gold POS data        │
-  └───────────────────────────────┬─────────────────────────────────┘
-                                  │ Gold-segmented + Gold-POS sentences
-  ┌───────────────────────────────▼─────────────────────────────────┐
-  │                  Task 3: Dependency Parsing                      │
-  │                                                                 │
-  │  Silver DP ──► Train parser ──► Score uncertainty ──► Annotate  │
-  │  (auto-parsed)  (biaffine)     (head entropy)       (correct   │
-  │                                                      arcs +     │
-  │                                                      deprels)   │
-  │                       ◄──── Retrain ◄──── Gold DP data         │
-  └─────────────────────────────────────────────────────────────────┘
+  Task 1: Word Segmentation
+  Silver BIO → Train CRF → Score uncertainty → Annotate → Retrain
+                                                            ↓
+  Task 2: POS Tagging (on gold-segmented sentences)
+  Silver POS → Train CRF → Score uncertainty → Annotate → Retrain
+                                                            ↓
+  Task 3: Dependency Parsing (on gold-segmented + gold-POS sentences)
+  Silver DP → Train parser → Score head entropy → Annotate → Retrain
 ```
 
-### 3.5 Task 1: Active Learning for Word Segmentation
+### 4.4 Task 1: Word Segmentation
 
-**Objective**: Build a gold-standard word segmentation evaluation set and improve the CRF segmenter through targeted correction of silver BIO data.
+**Query strategy**: CRF token marginal uncertainty. For token at position $i$:
 
-**Starting point**: 100K silver BIO sentences (udd-ws-v1.1), CRF model trained on this data (tree-1, 98.90% syllable F1 on silver test set).
+$$u_i = 1 - \max(P(\text{B-W}|x, i),\ P(\text{I-W}|x, i))$$
 
-#### 3.5.1 Query Strategy: Token-Level Marginal Uncertainty
+**Sentence score**: mean uncertainty across all tokens, weighted by multi-syllable word boundary density, plus a bonus for sentences containing known error patterns (4+ syllable tokens, dictionary-flagged forms).
 
-CRF models produce marginal probabilities for each token's label. For word segmentation, the uncertainty of a token at position $i$ is:
+**Cycles**: 200 sentences per cycle after Phase 0, evaluated on the fixed gold test set (250 sentences). Target: 2,000 gold WS sentences total. Stopping criterion: <0.1% F1 improvement on gold test over two consecutive cycles.
 
-$$u_i = 1 - \max(P(\text{B-W}|x, i),\ P(\text{I-W}|x, i))$$
+### 4.5 Task 2: POS Tagging
 
-**Sentence-level score**: average uncertainty across all tokens in the sentence, weighted by the number of multi-syllable word boundaries (positions where a B-W/I-W decision is non-trivial).
+**Query strategy**: Tag marginal uncertainty + confusion-targeted selection. Two complementary criteria:
 
-**Selection protocol**:
-1. Train CRF on udd-ws-v1.1-train (80K silver sentences)
-2. Predict on udd-ws-v1.1-dev + udd-ws-v1.1-test (20K sentences) with marginal probabilities
-3. Rank sentences by aggregated token uncertainty
-4. Select top-N sentences stratified by domain (equal representation)
-5. Additionally select sentences containing known error patterns from SEGMENTATION_EVAL.md:
-   - Sentences with tokens matching the 462 over-segmented dictionary words
-   - Sentences with tokens matching the 382 under-segmented candidates
-   - Sentences with 4+ syllable tokens (potential under-segmentation)
+- **Token uncertainty**: $u_i = 1 - \max_t P(t|x, i)$ over 15 UPOS tags
+- **Confusion density**: proportion of tokens matching known confusion pairs (AUX/VERB, NOUN/VERB, ADJ/VERB, DET/PRON, ADP/SCONJ)
 
-#### 3.5.2 Annotation Procedure
+**Selection formula**: $\text{Score} = 0.6 \cdot \text{uncertainty} + 0.4 \cdot \text{confusion\_density}$
 
-The annotator corrects BIO tags in a text editor (or annotation tool), focusing on word boundaries:
+**Dependency**: Only sentences with verified gold WS from Task 1 enter the POS pool. Target: 1,000 gold POS sentences.
 
-1. **Review pre-annotated BIO output** with uncertain tokens highlighted
-2. **Correct boundaries** using the NIIVTB 9-rule framework (see `ANNOTATION_GUIDELINE_WORD_SEGMENTATION.md`):
-   - Rule 1 (Insertability test): Can another word be inserted between syllables?
-   - Rule 2 (Semantic opacity): Is the meaning compositional?
-   - Rules 3--9: Additional criteria for specific constructions
-3. **Flag ambiguous cases** for guideline development
+### 4.6 Task 3: Dependency Parsing
 
-**Batch size**: 500 sentences per AL cycle (solo annotator, ~1--2 days of work).
+**Query strategy**: Head entropy + DPP batch diversity + partial arc annotation (Li et al., 2016). Head entropy for token $i$:
 
-#### 3.5.3 AL Cycle for Word Segmentation
+$$H_i = -\sum_{h} P(h|x, i) \log P(h|x, i)$$
 
-| Cycle | Sentences | Cumulative Gold | Focus |
-|-------|-----------|----------------|-------|
-| 0 (Baseline) | 0 | 0 | Train CRF on 80K silver |
-| 1 | 500 | 500 | Highest uncertainty + known error patterns |
-| 2 | 500 | 1,000 | Retrained CRF's new uncertain tokens |
-| 3 | 500 | 1,500 | Domain-specific compounds, remaining errors |
-| 4 | 500 | 2,000 | Diminishing returns check; stop if F1 plateaus |
+Within selected sentences, the annotator only corrects arcs where $H_i > \tau$ (top 30--40% most uncertain arcs), accepting parser output for confident arcs. Batch selection uses DPP with PhoBERT sentence embeddings as the kernel, combining informativeness with structural diversity.
 
-**Stopping criterion**: Stop when the CRF retrained on silver+gold data achieves <0.1% F1 improvement on a held-out gold test set (200 sentences set aside from Cycle 1).
+**Dependency**: Only sentences with gold WS + gold POS from Tasks 1--2 enter the DP pool. Target: 800--1,000 gold DP sentences with partial arc annotation.
 
-**Expected outcome**: 2,000 gold WS sentences, CRF word F1 improvement from ~98.0% to >99%.
+### 4.7 Solo Annotator Quality Assurance
 
-### 3.6 Task 2: Active Learning for POS Tagging
+This project operates under a solo annotator constraint. Quality is ensured through five methods:
 
-**Objective**: Build gold POS annotations on the gold-segmented sentences from Task 1, and improve the CRF POS tagger.
+| Method | Purpose | Frequency |
+|--------|---------|-----------|
+| Self-consistency check | Re-annotate 5% blind | Every 200 sentences |
+| Model-based error detection | Flag disagreements between retrained model and gold | After each AL cycle |
+| Dictionary validation | Cross-check WS against Viet74K dictionary | Continuous |
+| UD validator | Automated structural constraint checking | After each batch |
+| Guideline self-audit | Review decisions against written guidelines | Weekly |
 
-**Starting point**: Gold-segmented sentences from Task 1. Silver POS tags from `underthesea.pos_tag()`, mapped to UPOS via `UPOS_MAP`. CRF POS tagger (tree-1, 95.89% accuracy on silver UDD-1).
+**Table 10**: Quality assurance framework for solo annotator.
 
-**Dependency on Task 1**: POS tagging operates on correctly segmented words. Only sentences with verified word segmentation from Task 1 enter the POS annotation pool.
+**Self-consistency target**: >95% agreement between original and re-annotation. This is comparable to the 97.28% inter-annotator agreement reported for the NIIVTB project (Nguyen et al., 2018b).
 
-#### 3.6.1 Query Strategy: Tag Marginal Uncertainty + Confusion-Targeted Selection
+**Anchoring bias mitigation**: The annotator corrects silver pre-annotations, creating anchoring bias. Guidelines explicitly instruct independent evaluation of each decision; periodic blind re-annotation measures the degree of anchoring.
 
-CRF POS taggers produce marginal probabilities over the 15 UPOS tags for each token. Two complementary selection criteria:
+### 4.8 Annotation Guideline Development
 
-**Criterion A — Token uncertainty**:
-$$u_i = 1 - \max_t P(t|x, i), \quad t \in \{\text{ADJ, ADP, ADV, ..., X}\}$$
+Vietnamese annotation guidelines are co-developed through the AL process, addressing language-specific phenomena:
 
-Sentence score = average of top-K most uncertain tokens (not all tokens, to avoid penalizing long easy sentences).
+1. **Word segmentation**: NIIVTB 9-rule framework adapted for UDD-1 domains (insertability test, semantic opacity, specific construction rules)
+2. **POS tagging**: AUX word list, confusion pair decision procedures (*co the* test for VERB, *rat* test for ADJ)
+3. **Dependency parsing**: Copula *la*, passive *duoc/bi*, serial verb constructions, classifier phrases, topic-comment structure
 
-**Criterion B — Confusion-targeted selection**: Prioritize sentences containing tokens from known confusion pairs identified in v1.0:
-- **AUX vs. VERB**: *được*, *bị*, *phải*, *có thể* (20 auxiliary words with dual function)
-- **NOUN vs. VERB**: Vietnamese words frequently function as both (e.g., *quy định* "regulation"/"to regulate")
-- **ADJ vs. VERB**: Stative verbs vs. adjectives (e.g., *đẹp* "beautiful/to be beautiful")
-- **DET vs. PRON**: *này*, *đó*, *nào* (deictic function)
-- **ADP vs. SCONJ**: *khi*, *vì*, *do* (preposition vs. subordinator)
+Guidelines evolve through a self-audit process: draft after Cycle 1, revision after Cycle 2 based on edge cases, consolidation after final cycle.
 
-**Selection formula**: Score = $\alpha \cdot \text{uncertainty} + (1-\alpha) \cdot \text{confusion\_density}$, where confusion density is the proportion of tokens matching known confusion pairs. $\alpha = 0.6$ to favor uncertainty while ensuring coverage of systematic errors.
+## 5. Experimental Setup
 
-#### 3.6.2 Annotation Procedure
+### 5.1 CRF Word Segmentation Model
 
-The annotator corrects UPOS tags on gold-segmented sentences:
+We train a CRF word segmenter using pycrfsuite with L-BFGS optimization:
 
-1. **Review silver UPOS tags** with uncertain and confusion-pair tokens highlighted
-2. **Correct tags** by examining each highlighted token in sentential context:
-   - For AUX/VERB: Apply the AUX word list + syntactic function test (does it modify another verb?)
-   - For NOUN/VERB: Apply the *có thể* test (can *có thể* "can" be inserted before it? → VERB)
-   - For ADJ/VERB: Apply the *rất* test (can *rất* "very" modify it? → ADJ)
-3. **Record XPOS** if the XPOS-UPOS mismatch is justified (functional reclassification) vs. an error
-4. **Do not consider dependency relations**: POS is annotated independently of deprel to break the UPOS-forcing cycle from v1.0
+| Parameter | Value |
+|-----------|-------|
+| Algorithm | L-BFGS |
+| c1 (L1 penalty) | 0.5 |
+| c2 (L2 penalty) | 0.001 |
+| Max iterations | 300 |
+| Training data | 80K silver sentences (udd-ws-v1.1-train) |
+| Feature template | Syllable n-grams, capitalization, character type |
 
-#### 3.6.3 AL Cycle for POS Tagging
+The model is implemented in tree-1's word segmentation pipeline and produces token marginal probabilities for AL uncertainty scoring.
 
-| Cycle | Sentences | Cumulative Gold | Focus |
-|-------|-----------|----------------|-------|
-| 0 (Baseline) | 0 | 0 | Evaluate CRF POS on gold-segmented sentences |
-| 1 | 300 | 300 | Highest uncertainty + AUX/VERB confusion |
-| 2 | 300 | 600 | Retrained model's new errors + NOUN/VERB confusion |
-| 3 | 200 | 800 | Remaining confusion pairs, domain-specific terms |
-| 4 | 200 | 1,000 | Diminishing returns check |
+### 5.2 Evaluation Metrics
 
-**Stopping criterion**: Stop when POS accuracy on held-out gold test set (100 sentences) improves <0.2% per cycle.
+- **Syllable Accuracy**: Fraction of syllables with correct B-W/I-W tag
+- **Word Precision**: Fraction of predicted words that match gold words
+- **Word Recall**: Fraction of gold words that match predicted words
+- **Word F1**: Harmonic mean of word precision and recall
+- **Error types**: Over-merge (silver merges across word boundaries), over-split (silver splits within words)
 
-**Expected outcome**: 1,000 gold POS sentences, UPOS accuracy improvement from ~95.9% to >97%, elimination of the UPOS-forcing artifact.
+### 5.3 AL Uncertainty Scoring
 
-### 3.7 Task 3: Active Learning for Dependency Parsing
+CRF token marginal uncertainty: $u_i = 1 - \max(P(\text{B-W}|x, i), P(\text{I-W}|x, i))$
 
-**Objective**: Build gold dependency annotations on sentences with verified WS and POS, producing the first in-domain LAS/UAS measurements for Vietnamese legal and multi-domain text.
+Sentence-level composite score:
 
-**Starting point**: Sentences with gold WS + gold POS from Tasks 1--2. Silver dependency arcs from `underthesea.dependency_parse()`. Biaffine parser (~76% LAS on VLSP 2020 news benchmark).
+$$\text{score}(s) = \bar{u}(s) \times (1 + w_{\text{boundary}}) + 0.01 \times n_{\text{long}}$$
 
-**Dependency on Tasks 1--2**: Only sentences with verified WS and POS enter the DP annotation pool. This ensures that measured LAS/UAS reflects parsing quality, not cascading tokenization or POS errors.
+where $\bar{u}(s)$ is mean token uncertainty, $w_{\text{boundary}}$ weights multi-syllable word boundaries, and $n_{\text{long}}$ counts 4+ syllable tokens in the sentence.
 
-#### 3.7.1 Query Strategy: Head Entropy + Partial Arc Annotation
+Scoring is performed on the 20K dev+test sentences (excluding Phase 0 eval sentences in future cycles). Top-N sentences are selected for annotation in each cycle.
 
-Dependency parsing has the highest annotation cost per sentence (each token requires a head and a deprel). Following Li et al. (2016), we use **partial annotation** to minimize effort: within each selected sentence, the annotator only corrects arcs that the parser is uncertain about.
+## 6. Results
 
-**Head entropy** for token $i$ with possible heads $h \in \{0, 1, ..., n\}$:
+### 6.1 CRF Baseline on Silver Data
 
-$$H_i = -\sum_{h} P(h|x, i) \log P(h|x, i)$$
+**Cycle 0** (CRF trained on 80K silver sentences, evaluated on 10K silver test):
 
-This requires a probabilistic parser that outputs head distributions. We use a biaffine parser (Dozat and Manning, 2017) trained on UDD-1 silver data + VLSP 2020, which naturally produces attention scores convertible to probabilities via softmax.
+| Metric | Score |
+|--------|-------|
+| Syllable Accuracy | 0.9908 |
+| Word F1 | **0.9834** |
+| Word Precision | 0.9830 |
+| Word Recall | 0.9838 |
 
-**Sentence selection**: DPP (Determinantal Point Process) batch selection (Shi et al., 2021) combining:
-- **Informativeness**: Sum of head entropy across tokens → selects uncertain sentences
-- **Diversity**: PhoBERT sentence embeddings as the DPP kernel → selects structurally diverse sentences
-- **Domain balance**: Equal quota per domain (legal, news, Wikipedia, fiction, non-fiction)
+**Table 11**: CRF baseline on silver test set.
 
-**Arc selection within sentences**: For each selected sentence, mark arcs where $H_i > \tau$ (threshold determined from validation set, typically keeping the top 30--40% most uncertain arcs). The annotator focuses on these arcs while accepting the parser's output for confident arcs.
+Note: Silver-on-silver evaluation inflates scores because the model and test data share the same systematic biases from `underthesea.word_tokenize()`. The true performance on gold data is substantially lower (see Section 6.2).
 
-#### 3.7.2 Annotation Procedure
+**Uncertainty scoring** on 20K dev+test sentences:
+- Mean uncertainty: 0.0103; mean composite score: 0.0136
+- 8,636 sentences (43.2%) with score >= 0.01
+- 848 sentences (4.2%) with score >= 0.05
+- 1,068 sentences contain 4+ syllable tokens
 
-The annotator corrects dependency arcs using a tree visualization tool (e.g., Arborator-Grew or ConlluEditor):
+### 6.2 AL Cycle 1: First Gold Evaluation
 
-1. **View the full parse tree** with uncertain arcs highlighted (red = high entropy, green = confident)
-2. **For each uncertain arc**, decide:
-   - **Head**: Which word does this token depend on? (click to reassign)
-   - **Deprel**: What is the relation? (select from dropdown of 37 UD base types)
-3. **Verify confident arcs only if they look wrong** — the annotator can skip green arcs but is encouraged to scan for obvious errors
-4. **Apply Vietnamese-specific decisions** using the evolving guideline document:
-   - Copula *là*: `cop` when linking subject to predicate nominal, `mark` in cleft constructions
-   - Passive *được/bị*: `aux:pass` when modifying another verb, `root`/`xcomp` when main verb
-   - Serial verbs: second verb as `xcomp` or `conj` depending on shared arguments
-   - Classifiers: `clf` (UD v2 subtype) for numeral-classifier-noun constructions
-   - Topic fronting: `dislocated` when the fronted NP is resumed by a pronoun, `nsubj` otherwise
+We selected the top 100 sentences by composite uncertainty score. After annotation, 92 sentences were exported as gold (3 skipped due to annotation gaps in Label Studio). These 92 sentences contain 1,506 syllables.
 
-#### 3.7.3 AL Cycle for Dependency Parsing
+**Important caveat**: These 92 sentences were selected by highest uncertainty, not randomly sampled. The results below represent CRF performance on the **hardest** sentences, not overall model performance.
 
-| Cycle | Sentences | Arcs Annotated (est.) | Cumulative Gold | Focus |
-|-------|-----------|----------------------|----------------|-------|
-| 0 (Baseline) | 0 | 0 | 0 | Evaluate parser on VLSP 2020 test |
-| 1 (Pilot) | 50 | ~500 (all arcs) | 50 | Full annotation; establish LAS/UAS baseline on legal text; bootstrap guidelines |
-| 2 | 200 | ~800 (partial) | 250 | High-entropy arcs only; focus on nsubj/obj/obl attachment errors |
-| 3 | 200 | ~800 (partial) | 450 | Coordination, subordination, relative clause errors |
-| 4 | 200 | ~800 (partial) | 650 | Multi-domain expansion (news, wiki, books) |
-| 5 | 150 | ~600 (partial) | 800 | Domain-specific constructions, long sentences |
-| 6+ | 100/cycle | ~400/cycle | 1,000+ | Continue until LAS improvement plateaus |
+| Metric | Cycle 0 CRF | Post-Cycle 1 CRF | Improvement |
+|--------|-------------|-------------------|-------------|
+| Syllable Accuracy | 0.8665 | 0.8758 | +0.0093 |
+| Word Precision | 0.7934 | 0.7977 | +0.0043 |
+| Word Recall | 0.7685 | 0.7998 | +0.0313 |
+| Word F1 | **0.7808** | **0.7988** | **+0.0180** |
+| Sentences with diffs | 75/92 | 74/92 | -1 |
 
-**Cycle 1 is special**: The first 50 sentences are fully annotated (all arcs, not just uncertain ones) to establish a reliable gold test set and measure the initial parser's in-domain LAS/UAS. This addresses the primary reviewer request.
+**Table 12**: CRF vs. gold on 92 uncertainty-selected sentences.
 
-**Stopping criterion**: Stop when retrained parser's LAS on the gold test set (50 sentences from Cycle 1) improves <0.5% per cycle, or when the annotator reports that most uncertain arcs are genuinely ambiguous (guideline-level issues rather than parser errors).
+The Post-Cycle 1 CRF was retrained on 80K fixed silver + 92 merged gold sentences, yielding +1.8% absolute F1 improvement on this diagnostic set.
 
-**Expected outcome**: 800--1,000 gold DP sentences with ~3,000--4,000 verified arcs. First reported in-domain LAS/UAS for Vietnamese legal text and multi-domain text.
+**Per-domain breakdown** (Post-Cycle 1 CRF):
 
-### 3.8 Cross-Task Annotation Flow
+| Domain | N | F1 | Improvement |
+|--------|---:|----:|----:|
+| News | 8 | **0.8551** | +0.016 |
+| Legal | 4 | **0.8316** | +0.119 |
+| Non-fiction | 24 | 0.8168 | -0.007 |
+| Fiction | 28 | 0.8127 | -0.002 |
+| Wikipedia | 28 | **0.7472** | +0.040 |
 
-The three tasks are executed sequentially with overlap. As Task 1 produces gold-segmented sentences, they flow into Task 2; as Task 2 produces gold-POS sentences, they flow into Task 3.
+**Table 13**: Per-domain F1 on uncertainty-selected gold sentences.
 
-```
-Week:   1    2    3    4    5    6    7    8    9   10   11   12
-       ├────┼────┼────┼────┼────┼────┼────┼────┼────┼────┼────┤
-Task 1: ████████████████████
-         WS Cycle 1-4 (2K sentences)
-Task 2:           ██████████████████
-                   POS Cycle 1-4 (1K sentences)
-Task 3:                     ████████████████████████████
-                             DP Cycle 1-6+ (800-1K sentences)
-Guidelines: ◆───────◆───────◆───────◆───────◆───────◆
-            v0.1    v0.2    v0.3    v0.4    v0.5    v1.0
-```
+Wikipedia is the hardest domain (F1 = 0.747), driven by Sino-Vietnamese classical text and foreign names. Legal shows the largest improvement (+11.9%), though with only 4 sentences.
 
-**Overlap**: Task 2 can begin once Task 1 has produced its first 500 gold sentences (Cycle 1). Task 3 can begin once Task 2 has produced its first 300 gold sentences. This overlap reduces total calendar time from ~18 weeks (sequential) to ~12 weeks.
+**Domain distribution bias**: The uncertainty-selected set is heavily skewed: Fiction 30%, Wikipedia 30%, Non-fiction 26%, News 9%, Legal 4% --- far from the balanced 20% per domain. This confirms why uncertainty-selected data cannot serve as a representative evaluation set (see Section 4.2).
 
-### 3.9 Guideline Development (Solo Annotator)
+### 6.3 Post-Cycle 1: Silver Data Improvement
 
-Without a second annotator for IAA, guidelines evolve through a **self-audit** process:
+After applying the fix pipeline (6,990 corrections) and merging 92 gold sentences into training data, the retrained CRF shows modest improvement on silver test:
 
-| Cycle | Guideline Action | Quality Check |
-|-------|-----------------|---------------|
-| WS Cycle 1 | Draft WS guidelines from NIIVTB 9 rules | Dictionary validation of all corrections |
-| WS Cycle 2 | Revise WS guidelines based on encountered edge cases | Re-annotate 25 sentences from Cycle 1 blind; measure self-consistency |
-| POS Cycle 1 | Draft POS guidelines (confusion pairs, AUX list) | Model-based error detection on corrected data |
-| POS Cycle 2 | Revise POS guidelines | Re-annotate 15 sentences from Cycle 1 |
-| DP Cycle 1 | Draft DP guidelines (pilot, 50 full sentences) | UD validator on all sentences |
-| DP Cycle 3 | Major guideline revision after 450 sentences | Re-annotate 25 sentences from Cycles 1--2 |
-| Final | Consolidate all guidelines into single document | Full self-consistency audit on 100 random sentences |
+| Metric | Cycle 0 | Post-Cycle 1 | Improvement |
+|--------|---------|-------------|-------------|
+| Syllable Accuracy | 0.9908 | 0.9913 | +0.0005 |
+| Word F1 | 0.9834 | **0.9842** | +0.0008 |
 
-**Self-consistency target**: >95% agreement between original and re-annotation on the same sentences. Disagreements indicate guideline ambiguities that need resolution.
+**Table 14**: CRF on silver test after fix pipeline + 92 gold.
 
-Vietnamese-specific challenges requiring guideline development:
-1. **Copula `là`**: Multiple syntactic functions (copula, focus marker, relative clause marker)
-2. **Passive markers `được/bị`**: AUX vs. main VERB distinction is context-dependent
-3. **Serial verb constructions**: Common in Vietnamese; requires explicit deprel convention
-4. **Classifier constructions**: Numeral-classifier-noun patterns need specific annotation rules
-5. **Topic-comment structure**: Vietnamese allows topic fronting; affects nsubj vs. dislocated
-6. **Legal domain vocabulary**: Terms like *điều*, *khoản*, *mục* have specific syntactic roles
+The small improvement on silver test (+0.08%) contrasts with the larger improvement on gold (+1.8%), suggesting the silver test set is a poor proxy for true quality improvement.
 
-### 3.10 Quality Targets
+### 6.4 Uncertainty Reduction After Cycle 1
 
-| Metric | v1.0 (Silver) | v1.1 Target (Gold Subset) |
-|--------|--------------|---------------------------|
-| Word segmentation F1 | Unknown (same as underthesea) | >98% (gold test) |
-| UPOS accuracy | ~91.4% (8.6% forced) | >97% (gold test) |
-| LAS | ~76% (news benchmark) | Measured on legal + multi-domain |
-| UAS | Unknown | Measured on legal + multi-domain |
-| Self-consistency | N/A | >95% on re-annotation |
-| Gold WS sentences | 0 | 2,000 |
-| Gold POS sentences | 0 | 1,000 |
-| Gold DP sentences | 0 | 800--1,000 |
+Re-scoring the 20K dev+test sentences with the retrained model:
 
-**Table 6**: Quality targets for v1.1 gold annotation.
+| Statistic | Cycle 0 | Post-Cycle 1 | Change |
+|-----------|---------|-------------|--------|
+| Mean uncertainty | 0.0103 | 0.0099 | -3.9% |
+| Mean composite score | 0.0136 | 0.0130 | -4.4% |
+| Sentences with score >= 0.01 | 8,636 (43.2%) | 8,407 (42.0%) | -229 |
+| Sentences with score >= 0.05 | 848 (4.2%) | 773 (3.9%) | -75 |
 
-### 3.11 Cost Estimation (Solo Annotator)
+**Table 15**: Uncertainty reduction after Cycle 1 retraining.
 
-| Task | Sentences | Est. Speed | Est. Days | Calendar Weeks |
-|------|-----------|-----------|-----------|----------------|
-| WS (4 cycles × 500) | 2,000 | 80 sent/day | 25 days | 5 weeks |
-| POS (4 cycles × 250) | 1,000 | 40 sent/day | 25 days | 5 weeks |
-| DP Pilot (full) | 50 | 15 sent/day | 3 days | 1 week |
-| DP Partial (5 cycles × 150--200) | 750--1,000 | 25 sent/day | 30--40 days | 6--8 weeks |
-| Guideline development | — | — | 10 days | distributed |
-| Self-consistency audits | ~200 re-annotated | — | 5 days | distributed |
-| **Total** | | | **~100 days** | **~12 weeks (with overlap)** |
+## 7. Analysis
 
-**Table 7**: Cost estimation for solo annotator.
+### 7.1 Error Taxonomy
 
-The partial annotation strategy for dependency parsing (annotating only ~40% of arcs per sentence) reduces DP annotation effort by approximately 60% compared to full annotation, consistent with Li et al. (2016). Total effort of ~100 annotator-days is feasible for a single linguist working full-time over 3 months, or part-time over 6 months.
+Analysis of the 92 gold sentences reveals 225 total errors (113 over-merge + 112 over-split). We identify six error categories:
 
-## 4. Data Format and Access
+**Category 1: Cross-boundary proper name merge** (most frequent). The CRF extends proper name spans into the following word:
+- `Nguyen_Binh_Mua_dam` → `Nguyen_Binh` | `Mua` | `dam`
+- `Tang_Hoang_Vinh_giu` → `Tang_Hoang_Vinh` | `giu`
 
-### 4.1 Word Segmentation Dataset (udd-ws-v1.1)
+**Root cause**: CRF continuation bias after proper name tokens.
 
-**Files** (local, UDD-1 repository):
-- `udd-ws-v1.1-train.txt` — 80,000 sentences (19 MB)
-- `udd-ws-v1.1-dev.txt` — 10,000 sentences (2.5 MB)
-- `udd-ws-v1.1-test.txt` — 10,000 sentences (2.4 MB)
-- `udd-ws-v1.1-{train,dev,test}.conllu` — CoNLL-U format (words as tokens)
+**Category 2: Sino-Vietnamese classical text** (Wikipedia-specific). Classical Chinese quotations use mostly single-syllable words, but the CRF aggressively merges:
+- `thien_dia_luu` → `thien_dia` | `luu`
+- `tat_huu_tung` → `tat` | `huu` | `tung`
 
-**Format**: BIO text with comment headers (VLSP 2013 compatible):
-```
-# sent_id = uvn-1234
-# text = Original sentence text here
-syllable1	B-W
-syllable2	I-W
-syllable3	B-W
-```
+**Root cause**: CRF trained on modern Vietnamese where multi-syllable words are common.
 
-### 4.2 Pipeline Scripts
+**Category 3: Foreign word merges**. Latin-script tokens merged as compounds:
+- `DNA_replication` → `DNA` | `replication`
+- `I_LOVE_YOU` → `I` | `LOVE` | `YOU`
 
-| Script | Purpose | Command |
-|--------|---------|---------|
-| `src/fetch_ws_sentences.py` | Fetch 100K sentences with quality filters and round-robin sampling | `uv run src/fetch_ws_sentences.py` |
-| `src/build_ws_dataset.py` | Convert to BIO + stratified split | `uv run src/build_ws_dataset.py` |
-| `src/fix_ws_errors.py` | Fix known WS errors (cross-boundary merges, always-split compounds) | `uv run src/fix_ws_errors.py` |
-| `src/check_ws_errors.py` | Rule-based WS error checker (7 rules) | `uv run src/check_ws_errors.py` |
-| `src/ws_statistics.py` | Convert to CoNLL-U + compute statistics | `uv run src/ws_statistics.py` |
+**Category 4: Compound boundary shifts**. CRF merges across true word boundaries:
+- `loi_nhuan_tich` + `luy` → `loi_nhuan` + `tich_luy` (boundary shift)
+- `phong_ve_tinh` → `phong` + `ve_tinh` (over-merge)
 
-### 4.3 Intermediate Files
+**Category 5: Compound under-merges** (over-splits). Known compounds split incorrectly:
+- `Uy` | `ban` → `Uy_ban` ("committee")
+- `linh` | `thu` → `linh_thu` ("soldier")
+- `mu` | `rua` → `mu_rua` ("turtle shell")
 
-- `ws_sentences_{vlc,uvn,uvw,uvb_f,uvb_n}.txt` — Raw sentences per domain (format: `idx\tsentence`)
+**Category 6: Proper name under-merges**:
+- `Truong_Han` | `Sieu` → `Truong_Han_Sieu`
+- `Nicolas` | `Fatio_de` | `Duillier` → `Nicolas_Fatio_de_Duillier`
 
-## 5. Relation to UDD-1 v1.0
+### 7.2 Error Distribution by Type
 
-UDD-1 v1.1 does not replace v1.0. The relationship is:
+| Error Type | Count | Percentage |
+|------------|------:|---:|
+| Cross-boundary name merge | ~40 | 17.8% |
+| Sino-Vietnamese boundary | ~35 | 15.6% |
+| Foreign word merge | ~20 | 8.9% |
+| Compound boundary shift | ~30 | 13.3% |
+| Compound under-merge | ~50 | 22.2% |
+| Proper name under-merge | ~20 | 8.9% |
+| Other | ~30 | 13.3% |
+| **Total** | **~225** | **100%** |
 
-| | v1.0 | v1.1 |
-|---|---|---|
-| **UD Treebank** | 10K sentences, legal domain, silver-standard CoNLL-U | Unchanged (v1.0 treebank remains) |
-| **WS Dataset** | N/A | 100K sentences, 5 domains, BIO format |
-| **Gold Annotation** | None | Planned: 2-5K sentences via active learning |
-| **Annotation Guidelines** | Implicit (parser behavior) | Explicit (co-developed through AL) |
-| **Domains** | Legal only | Legal, News, Wikipedia, Fiction, Non-fiction |
+**Table 16**: Error distribution by category.
+
+Over-merge (Categories 1--4, ~55%) slightly dominates over-split (Categories 5--6, ~31%), with the CRF showing a bias toward merging adjacent syllables.
+
+### 7.3 Comparison with VLSP 2013
+
+The VLSP 2013 shared task (Nguyen et al., 2013) established the Vietnamese WS benchmark with ~75K news sentences. Key differences with UDD-WS v1.1:
+
+| Aspect | VLSP 2013 | UDD-WS v1.1 |
+|--------|-----------|-------------|
+| Size | ~75K sentences | 100K sentences |
+| Domains | News only | 5 domains |
+| Format | Space-separated | BIO (syllable-level) |
+| Standard | Gold (human-annotated) | Silver (auto-annotated) |
+| SOTA F1 | 97.90% (VnCoreNLP) | 98.42% (CRF, silver-on-silver) |
+
+**Table 17**: Comparison with VLSP 2013 WS shared task.
+
+Our silver-on-silver F1 of 98.42% is not directly comparable to the VLSP 2013 gold-standard F1 of 97.90%, as the former benefits from shared biases between training and test data. The true performance on gold data, as measured by Phase 0, will provide a fair comparison point.
+
+## 8. Conclusion and Future Work
+
+We presented UDD-WS v1.1, a 100,000-sentence multi-domain Vietnamese word segmentation dataset, and a three-task active learning framework for gold-standard UD annotation. Our key findings from the first AL cycle are:
+
+1. **AL selection is effective**: CRF uncertainty scoring correctly identifies hard cases (F1 = 0.80 on uncertainty-selected gold vs. 0.98 on silver).
+2. **Targeted retraining helps**: 92 gold sentences yield +1.8% F1 improvement on the diagnostic set, demonstrating the value of selective correction.
+3. **Error analysis informs fixes**: The error taxonomy directly informed a fix pipeline correcting 6,990 silver errors, improving model robustness.
+4. **Fixed evaluation is essential**: Uncertainty-selected sentences are not representative (skewed domain distribution, artificially low F1). Phase 0 addresses this with a randomly-sampled gold evaluation set.
+
+**Immediate next steps**:
+1. Phase 0: Annotate 500 randomly-sampled gold sentences as fixed evaluation set (250 dev + 250 test)
+2. Establish true F1 baseline on representative gold data
+3. Continue AL Cycle 2+ with proper evaluation infrastructure
+4. Develop POS and DP annotation guidelines
+
+**Longer-term goals**:
+- Complete 2,000 gold WS sentences through AL Cycles 2--10
+- 1,000 gold POS sentences (Task 2)
+- 800--1,000 gold DP sentences with partial arc annotation (Task 3)
+- Consolidated Vietnamese annotation guidelines for WS, POS, and DP
+- First reported in-domain LAS/UAS for multi-domain Vietnamese text
+
+## Limitations
 
-The word segmentation dataset feeds into tree-1's CRF training pipeline. The gold annotations planned through active learning will eventually enable a v2.0 release with verified quality.
+1. **Silver-standard quality**: The 100K BIO annotations are generated by `underthesea.word_tokenize()` and inherit its biases. The 2.4% token-level inconsistency rate (3,592 forms, 67,612 occurrences) means the dataset is suitable for CRF training but not for evaluation of segmentation quality. Models trained on this data will learn the tokenizer's systematic errors.
 
-## 6. Conclusion
+2. **No gold evaluation set yet**: Phase 0 is planned but not yet executed. All current results are either on silver test data (inflated) or on uncertainty-selected gold data (biased). The true model performance on representative data is unknown.
 
-UDD-1 v1.1 establishes the data and methodological foundation for building a gold-standard Vietnamese UD treebank:
+3. **Single annotator**: All gold annotations come from a single annotator, precluding inter-annotator agreement measurement. Self-consistency checks (>95% target) provide a weaker quality guarantee than dual annotation with adjudication.
 
-1. **udd-ws-v1.1**: A 100,000-sentence, 5-domain word segmentation dataset in BIO format, built with a rigorous 14-rule quality filter pipeline, diversity-aware round-robin sampling (72 fiction books, 68 non-fiction books, 275 legal documents, 3,091 news documents, 785 Wikipedia articles), and two-level deduplication (exact + digit-normalized near-duplicate detection). The dataset provides training data for robust tokenization models that underpin all downstream annotations.
+4. **Anchoring bias**: The annotator corrects silver pre-annotations, creating anchoring bias toward the parser's output. While guidelines instruct independent evaluation and periodic blind re-annotation measures the degree of bias, some residual anchoring is expected.
 
-2. **Sentence quality infrastructure**: A continuous quality scoring function (`sentence_score()`) and comprehensive validation pipeline that ensures all input sentences are well-formed Vietnamese text, free of markup, OCR artifacts, and duplicates. The `safe_sent_tokenize()` wrapper fixes Vietnamese-specific sentence splitting errors in the upstream NLP toolkit.
+5. **CRF uncertainty miscalibration**: The CRF is trained on silver data with systematic errors. Its uncertainty estimates may be miscalibrated for exactly the errors that matter most: if the training data consistently segments a form in both directions (e.g., *suc khoe* at 48% split ratio), the CRF will be uncertain for the wrong reason (training noise vs. genuine ambiguity).
 
-3. **Three-task active learning pipeline**: Task-specific AL strategies for each layer of the Vietnamese NLP pipeline:
-   - **Word segmentation**: CRF token marginal uncertainty + dictionary-based error targeting → 2,000 gold sentences
-   - **POS tagging**: Tag marginal uncertainty + confusion-pair targeting (AUX/VERB, NOUN/VERB) → 1,000 gold sentences
-   - **Dependency parsing**: Head entropy + DPP batch diversity + partial arc annotation → 800--1,000 gold sentences
+6. **Cost estimates from outdated reference**: Throughput estimates (80 sent/day WS, 40 sent/day POS, 15--25 sent/day DP) are adapted from Brants and Skut (1998), a 28-year-old study on German constituency treebanks. No pilot timing study has been conducted for Vietnamese. Actual throughput may differ.
 
-4. **Solo annotator methodology**: Quality assurance through self-consistency checks (>95% target), model-based error detection, dictionary validation, and UD structural validation --- demonstrating that gold treebank construction is feasible without multiple annotators.
+7. **Unvalidated AL hyperparameters**: Several AL parameters are set without validation: the confusion-targeted POS selection weight (alpha = 0.6), the DP partial annotation threshold (tau), the DPP kernel construction. These will be tuned empirically as AL cycles proceed.
 
-5. **Vietnamese UD guidelines**: Co-developed through the annotation process, addressing Vietnamese-specific challenges (copula, passive markers, serial verbs, classifiers, topic-comment structure).
+8. **Domain difficulty imbalance**: The balanced 20K-per-domain design treats all domains equally in sentence count, but annotation difficulty varies substantially (legal text has more multi-syllable compounds; Wikipedia has Sino-Vietnamese classical text). Budget allocation should account for this.
 
-The immediate next step is Task 1, Cycle 1: selecting 500 sentences with highest CRF uncertainty from udd-ws-v1.1 for gold word segmentation annotation. In parallel, the DP pilot (50 fully annotated sentences) will produce the first in-domain LAS/UAS measurements for Vietnamese legal text. Total estimated effort: ~100 annotator-days over 12 weeks.
+## Ethics Statement
+
+**Data licensing**: The source datasets (`undertheseanlp/UTS_VLC`, `undertheseanlp/UVN-1`, `undertheseanlp/UVW-2026`, `undertheseanlp/UVB-v0.1`) are hosted on HuggingFace. Users should verify the licensing terms of each source dataset before commercial use of UDD-WS v1.1.
+
+**Annotation labor**: All annotation is performed by a single Vietnamese linguist who is also the project lead. No crowd-sourced annotation or underpaid labor is involved.
+
+**Potential for misuse**: The 100K silver-standard dataset, if used uncritically as gold-standard training data, could propagate systematic segmentation errors to downstream applications. Users should note the silver-standard caveat (Section 3.7) and the 2.4% inconsistency rate.
+
+**Bias considerations**: The dataset inherits biases from both the source texts (e.g., legal text from Vietnamese law, fiction from popular novels) and the `underthesea` tokenizer. The round-robin sampling mitigates source concentration bias but does not address content-level biases in the source corpora.
+
+## Acknowledgments
+
+We thank the anonymous reviewer of TECHNICAL_REPORT v1.0 for detailed feedback that motivated this work.
 
 ## References
 
-- Baldridge, J. and Osborne, M. (2004). Active Learning and the Total Cost of Annotation. In *Proceedings of EMNLP 2004*.
+- Baldridge, J. and Osborne, M. (2004). Active Learning and the Total Cost of Annotation. In *Proceedings of EMNLP 2004*, pp. 9--16.
 
-- Brants, T. and Skut, W. (1998). Automation of Treebank Annotation. In *Proceedings of CoNLL 1998*.
+- Bouma, G. and van Noord, G. (2017). Increasing Return on Annotation Investment: The Automatic Construction of a Universal Dependency Treebank for Dutch. In *Proceedings of the NoDaLiDa 2017 Workshop on Universal Dependencies (UDW 2017)*, pp. 19--26.
+
+- Brants, T. and Skut, W. (1998). Automation of Treebank Annotation. In *Proceedings of CoNLL 1998*, pp. 49--57.
 
 - de Marneffe, M.-C., Manning, C.D., Nivre, J., and Zeman, D. (2021). Universal Dependencies. *Computational Linguistics*, 47(2):255--308.
 
+- Farquhar, S., Gal, Y., and Rainforth, T. (2021). On Statistical Bias in Active Learning: How and When to Fix It. In *Proceedings of ICLR 2021*.
+
 - Hwa, R. (2004). Sample Selection for Statistical Parsing. *Computational Linguistics*, 30(3):253--276.
 
 - Li, Z., Zhang, M., Zhang, Y., Liu, Z., Chen, W., Wu, H., and Wang, H. (2016). Active Learning for Dependency Parsing with Partial Annotation. In *Proceedings of ACL 2016*, pp. 344--354.
 
-- Nguyen, Q., Vu, T., Nguyen, D., Nguyen, M., and Phan, T. (2018). Ensuring Annotation Consistency and Accuracy for Vietnamese Treebank. *Language Resources and Evaluation*, Springer.
+- Luth, L., Strater, T., Borth, D., and Meinhardt, T. (2023). Navigating the Pitfalls of Active Learning Evaluation: A Systematic Framework for Meaningful Assessment. In *Proceedings of NeurIPS 2023*.
+
+- Marcus, M.P., Santorini, B., and Marcinkiewicz, M.A. (1993). Building a Large Annotated Corpus of English: The Penn Treebank. *Computational Linguistics*, 19(2):313--330.
+
+- Nguyen, K.-H. (2018). BKTreebank: Building a Vietnamese Dependency Treebank. In *Proceedings of the Eleventh International Conference on Language Resources and Evaluation (LREC 2018)*, pp. 2772--2777. Miyazaki, Japan.
+
+- Nguyen, C.T., Nguyen, T.P., Phan, X.H., and Nguyen, T.T. (2013). Vietnamese Word Segmentation at VLSP 2013. In *Proceedings of VLSP 2013 Workshop*.
+
+- Nguyen, P.T., Vu, X.L., Nguyen, T.M.H., Nguyen, V.H., and Le, H.P. (2014). Building a Large Syntactically-Annotated Corpus of Vietnamese. In *Proceedings of LREC 2014*.
+
+- Nguyen, D.Q., Nguyen, D.Q., Vu, T., Dras, M., and Johnson, M. (2018a). A Fast and Accurate Vietnamese Word Segmenter. In *Proceedings of LREC 2018*, pp. 2582--2586.
+
+- Nguyen, Q.M., Vu, T.L., Nguyen, D.Q., Nguyen, M.Q., and Phan, T.H. (2018b). Ensuring Annotation Consistency and Accuracy for Vietnamese Treebank. *Language Resources and Evaluation*, 52:873--899. Springer.
+
+- Settles, B. and Craven, M. (2008). An Analysis of Active Learning Strategies for Sequence Labeling Tasks. In *Proceedings of EMNLP 2008*, pp. 1070--1079.
+
+- Shi, T., Benton, A., Malioutov, I., and Irsoy, O. (2021). Diversity-Aware Batch Active Learning for Dependency Parsing. In *Proceedings of NAACL 2021*, pp. 2616--2626.
 
-- Shi, T., Benton, A., Malioutov, I., and Irsoy, O. (2021). Diversity-Aware Batch Active Learning for Dependency Parsing. In *Proceedings of NAACL 2021*.
+- Zhang, Z., Strubell, E., and Hovy, E. (2022). A Survey of Active Learning for Natural Language Processing. In *Proceedings of EMNLP 2022*, pp. 6166--6190.
 
 - Zhang, Z., Strubell, E., and Hovy, E. (2023). Data-efficient Active Learning for Structured Prediction with Partial Annotation and Self-Training. In *Findings of EMNLP 2023*.
+
+## Appendix A: Sentence Quality Score
+
+Beyond binary pass/fail filters, each sentence is assigned a continuous quality score via `sentence_score()` returning a value in (0, 1):
+
+| Sub-score | Weight | Description |
+|-----------|--------|-------------|
+| Length | 0.20 | Gaussian around ideal range [60, 200] chars |
+| Word count | 0.15 | Gaussian around ideal range [8, 35] words |
+| Structure | 0.20 | Proper start (+0.5) + proper end (+0.5) |
+| Cleanliness | 0.30 | Penalties for markup, brackets, glued text |
+| Completeness | 0.15 | Penalties for unbalanced quotes, digit ratio |
+| Vietnamese density | multiplier | Ratio of Vietnamese diacritical chars to total |
+
+Formula: `final_score = clamp(base_score * vietnamese_density, 0.01, 0.99)`
+
+## Appendix B: Full Error Examples from AL Cycle 1
+
+### B.1 Cross-Boundary Proper Name Merges
+
+| Silver (wrong) | Gold (correct) | Domain |
+|----------------|----------------|--------|
+| `Nguyen_Binh_Mua_dam` | `Nguyen_Binh` \| `Mua` \| `dam` | Fiction |
+| `Chu_Du_tinh_tinh` | `Chu_Du` \| `tinh_tinh` | Non-fiction |
+| `Tang_Hoang_Vinh_giu` | `Tang_Hoang_Vinh` \| `giu` | News |
+| `Khong_Dinh_Dat_thoi` | `Khong_Dinh_Dat` \| `thoi` | Wikipedia |
+
+### B.2 Sino-Vietnamese Classical Text
+
+| Silver (wrong) | Gold (correct) | sent_id |
+|----------------|----------------|---------|
+| `thien_dia_luu` | `thien_dia` \| `luu` | uvw-11634 |
+| `tat_huu_tung` | `tat` \| `huu` \| `tung` | uvw-379 |
+| `am_thuc_chi` | `am_thuc` \| `chi` | uvw-379 |
+
+### B.3 Compound Boundary Shifts
+
+| Silver (wrong) | Gold (correct) | Type |
+|----------------|----------------|------|
+| `loi_nhuan_tich` + `luy` | `loi_nhuan` + `tich_luy` | Boundary shift |
+| `phong_ve_tinh` | `phong` + `ve_tinh` | Over-merge |
+| `quy_pham_phap_luat` | `quy_pham` + `phap_luat` | Over-merge |
+
+## Appendix C: Cost Estimation
+
+| Task | Sentences | Est. Speed | Est. Days | Calendar Weeks |
+|------|-----------|-----------|-----------|----------------|
+| Phase 0 (gold eval) | 500 | 80 sent/day | 6 days | 1 week |
+| WS AL (9 cycles x 200) | 1,800 | 80 sent/day | 23 days | 5 weeks |
+| POS (4 cycles x 250) | 1,000 | 40 sent/day | 25 days | 5 weeks |
+| DP Pilot (full) | 50 | 15 sent/day | 3 days | 1 week |
+| DP Partial (5 cycles x 150--200) | 750--1,000 | 25 sent/day | 30--40 days | 6--8 weeks |
+| Guideline development | -- | -- | 10 days | distributed |
+| Self-consistency audits | ~200 re-annotated | -- | 5 days | distributed |
+| **Total** | | | **~106 days** | **~13 weeks** |
+
+**Table C.1**: Cost estimation for solo annotator, including Phase 0.
+
+## Appendix D: Relation to UDD-1 v1.0
+
+UDD-WS v1.1 does not replace UDD-1 v1.0. The relationship is:
+
+| | v1.0 | v1.1 |
+|---|---|---|
+| **UD Treebank** | 10K sentences, legal domain, silver CoNLL-U | Unchanged |
+| **WS Dataset** | N/A | 100K sentences, 5 domains, BIO format |
+| **Gold Annotation** | None | Phase 0 (500 eval) + AL cycles (planned) |
+| **Annotation Guidelines** | Implicit (parser behavior) | Explicit (co-developed through AL) |
+| **Domains** | Legal only | Legal, News, Wikipedia, Fiction, Non-fiction |

TECHNICAL_REPORT_v1.1_REVIEW.md

@@ -0,0 +1,263 @@
+# Review: UDD-WS v1.1: A 100K-Sentence Multi-Domain Vietnamese Word Segmentation Dataset and Active Learning Framework for Gold-Standard UD Annotation
+
+**Reviewed**: TECHNICAL_REPORT_v1.1.md
+**Date**: 2026-02-13
+**Review format**: ACL Rolling Review (ARR)
+**Note**: This is a re-review following major revision from the initial v1.1 submission. The paper has been restructured to ACL Long Paper format and incorporates Phase 0 gold evaluation methodology.
+
+---
+
+## Paper Summary
+
+This paper presents UDD-WS v1.1, a 100,000-sentence silver-standard Vietnamese word segmentation dataset in BIO format spanning five domains (legal, news, Wikipedia, fiction, non-fiction), along with a three-task active learning framework for constructing gold-standard Universal Dependencies annotations. The paper reports results from a first AL cycle (92 gold-annotated sentences selected by CRF uncertainty), identifies the methodological necessity of a fixed gold evaluation set (Phase 0) before continuing AL, and provides a detailed error taxonomy of Vietnamese word segmentation errors across domains. The work is positioned as both a resource contribution (the multi-domain WS dataset) and a methodology contribution (the sequential AL pipeline with proper evaluation infrastructure).
+
+## Summary of Strengths
+
+1. **Excellent ACL-compliant structure with mandatory sections (full paper)**: The revision successfully restructures to ACL Long Paper format with all mandatory sections: Limitations (8 explicit items), Ethics Statement, proper Related Work with comprehensive treebank comparison (Table R1), and clear separation of Experimental Setup / Results / Analysis. The Limitations section is notably thorough and self-aware --- particularly items 2 (no gold eval yet), 5 (CRF miscalibration on inconsistent silver data), and 6 (cost estimates from 1998 study). This level of intellectual honesty is rare and strengthens the paper.
+
+2. **Phase 0 gold evaluation set methodology (Section 4.2)**: The most significant improvement over the initial draft. The paper now correctly identifies that uncertainty-selected sentences cannot serve as representative evaluation data (domain skew: Fiction 30%/Wikipedia 30% vs. target 20% each; F1 = 0.78 reflecting hardest cases, not overall performance). The Phase 0 protocol --- random stratified sampling of 500 sentences, 250 dev + 250 test, frozen before AL --- follows established methodology (Zhang et al., 2022; Farquhar et al., 2021; Luth et al., 2023). This transforms a fundamental methodological flaw into a design strength.
+
+3. **Comprehensive Vietnamese treebank comparison (Section 2.2, Table R1)**: The Related Work now covers all five relevant Vietnamese treebanks (VnDT, UD_Vietnamese-VTB, BKTreebank, NIIVTB, UDD-1 v1.0) with a clear comparison table. BKTreebank's inclusion is particularly valuable --- its IAA figures (94.5% POS, 80.4% LAS with 3 annotators + reviewer) provide the benchmark against which UDD-1's solo-annotator approach should be evaluated. The observation that "all existing Vietnamese treebanks are limited to news text" positions UDD-1's multi-domain contribution clearly.
+
+4. **Well-executed first AL cycle with actionable error taxonomy (Sections 6.2, 7.1)**: The 92 gold sentences produce concrete, useful results: a 6-category error taxonomy (cross-boundary name merge, Sino-Vietnamese boundaries, foreign words, compound shifts, compound under-merges, name under-merges), per-domain F1 breakdown revealing Wikipedia as hardest (0.747) and news as easiest (0.855), and +1.8% F1 improvement from targeted retraining. The error analysis directly informed the fix pipeline (6,990 corrections), demonstrating the practical value of even a small gold annotation effort.
+
+5. **Silver quality assessment with quantified inconsistency rate (Section 3.7)**: The paper now openly reports the 3,592 inconsistently segmented forms (67,612 occurrences, 2.4% token-level inconsistency) with concrete examples (e.g., *suc khoe* at 48.3% split ratio). This is a significant finding about the `underthesea` tokenizer's behavior and is essential context for anyone using the dataset.
+
+6. **Honest silver-vs-gold performance comparison (Section 6.3)**: The observation that silver-on-silver F1 improvement (+0.08%) vastly understates true quality improvement on gold (+1.8%) is an important methodological insight. The paper correctly concludes that "the silver test set is a poor proxy for true quality improvement."
+
+## Summary of Weaknesses
+
+1. **Phase 0 is still planned, not executed (Section 4.2)**: While the methodology is now correct, the fundamental weakness from the initial review persists: **no representative gold evaluation exists**. All reported F1 numbers are either on silver data (inflated by shared bias) or on uncertainty-selected data (deflated by selection toward hardest cases). The true model performance on representative data remains unknown. The paper cannot report a credible baseline F1 number. This limits the paper to a "methodology + plan" contribution rather than a resource contribution with demonstrated quality.
+
+2. **Missing critical related work on Vietnamese cross-domain parsing (DGDT, FDSE 2023)**: Two highly relevant recent publications are absent:
+   - **DGDT (2025)**: Huynh et al. published a Vietnamese Domain Generalization Dependency Treebank with train/dev/test from completely separated domains, showing 3.27% UAS / 5.09% LAS cross-domain performance drops. This directly validates UDD-1's multi-domain motivation and provides a benchmark for cross-domain Vietnamese parsing.
+   - **BERT-based sentence recommendation for Vietnamese UD (FDSE 2023)**: A Vietnamese-specific AL method for UD treebank construction using BERT to select informative sentences. This is the closest prior work to UDD-1's AL approach for Vietnamese and must be cited and discussed.
+   - **VLSP 2019/2020 dependency parsing shared tasks**: These established Vietnamese DP benchmarks (best LAS 76.27 in VLSP 2020) are not cited despite being directly relevant to Task 3 (dependency parsing).
+
+3. **No comparison with transformer-based WS systems (Section 7.3)**: The VLSP 2013 comparison (Table 17) lists VnCoreNLP at 97.90% as the WS SOTA. However, transformer-based systems now achieve 98.31--98.35% F1 (SpanSeg + XLM-RoBERTa; NlpHUST ELECTRA fine-tuning). The paper should acknowledge this and discuss why a CRF approach is appropriate for the AL loop (speed for uncertainty scoring, marginal probability access) despite lower ceiling accuracy. Additionally, ViWordFormer (AAAI 2025) challenges the entire syllable-to-word segmentation paradigm and warrants discussion.
+
+4. **Error distribution counts are approximate (Section 7.2, Table 16)**: All counts in the error taxonomy are prefixed with "~" (e.g., "~40", "~35", "~20"). For 92 gold sentences with 225 total errors, exact counts should be computable and reported. The approximation undermines the quantitative rigor of the analysis section.
+
+5. **Underspecified AL composite scoring formula (Section 5.3)**: The sentence-level score formula includes $w_{\text{boundary}}$ ("weights multi-syllable word boundaries") without defining this term precisely. How is boundary weight computed? Is it a count, a ratio, a learned parameter? The $0.01 \times n_{\text{long}}$ additive term is specified but the boundary weight is not. This makes the query strategy not fully reproducible.
+
+6. **Table numbering inconsistency**: Table R1 (treebank comparison in Related Work) breaks the sequential numbering (Tables 1--17 in the main body, then "Table R1" interspersed). This should be integrated into the main numbering sequence or clearly designated as an in-text table.
+
+## Scores
+
+### Soundness: 3
+
+The WS dataset is a real artifact with documented quality characteristics. The Phase 0 methodology is correct and well-justified. The AL Cycle 1 results (92 gold sentences) provide genuine empirical evidence. However, key metrics remain unavailable (no representative gold F1), error counts are approximate, and the AL scoring formula is underspecified. The paper makes no unsupported claims --- it is honest about what is planned vs. completed --- but the ratio of planned-to-completed work remains high.
+
+### Excitement: 3
+
+The paper addresses a genuine gap: no multi-domain Vietnamese WS dataset exists, and no prior work applies sequential AL across cascading NLP tasks (WS → POS → DP). The Phase 0 methodology and the self-critical identification of evaluation bias are intellectually valuable. However, excitement is moderated by the limited executed results (92 gold sentences out of a planned 2,390) and the absence of a representative gold baseline.
+
+### Overall Assessment: 3
+
+The paper has improved substantially from the initial v1.1 draft. The ACL structure is properly followed, the Related Work is comprehensive (with gaps noted above), the Limitations section is unusually thorough, and the Phase 0 methodology corrects a fundamental flaw. The paper now functions as a reasonable combined resource/methodology contribution. To move from borderline to solid accept, two additions would suffice: (1) execute Phase 0 and report a representative gold F1 baseline, and (2) address the missing Vietnamese cross-domain parsing references (DGDT, FDSE 2023, VLSP 2019/2020).
+
+### Reproducibility: 4
+
+The data pipeline is well-documented with scripts, quality filter specifications, random seed (42), and HuggingFace dataset identifiers. The CRF training configuration is specified (Table in Section 5.1). The BIO format is VLSP 2013-compatible. The main reproducibility gap is the AL scoring formula's underspecified boundary weight term. The `underthesea` version (v2.1.0) is now stated.
+
+### Confidence: 4
+
+I am familiar with the Vietnamese NLP, Universal Dependencies, active learning, and treebank construction literature. I have researched the current SOTA and verified claims against published results. I have reviewed the supporting materials (WS_CHECK_REPORT, AL_CYCLE1_REPORT, PLAN_v1.1) to contextualize the contribution.
+
+## Detailed Comments
+
+### Technical Soundness
+
+**Dataset (Section 3)**: The data collection pipeline is technically sound. The 14-rule quality filter is well-specified (Table 2) with appropriate domain-specific extensions. Round-robin sampling demonstrably improves source diversity (Table 3: 1,341 → 4,291 sources). The two-level deduplication (exact + digit-normalized) is appropriate for legal text. The BIO annotation procedure is standard and reproducible.
+
+The silver quality assessment (Section 3.7) is a significant addition. The 2.4% inconsistency rate (67,612/2,786,213 tokens) is clearly reported with concrete examples. However, the paper could distinguish between *genuine ambiguity* (forms like *nguoi ta* where both segmentations are linguistically defensible in different contexts) and *tokenizer errors* (forms like *suc khoe* where one segmentation is clearly correct). This distinction matters for AL: genuinely ambiguous forms should not necessarily be "fixed."
+
+**AL Framework (Section 4)**: The Phase 0 design is methodologically sound and well-grounded in literature. The four design principles (Section 4.1) are clearly stated and justified. The three-task sequential pipeline is correctly designed to prevent cascading errors.
+
+However, the paper does not address the **pool bias problem**: Phase 0 samples from the 20K dev+test silver pool, which already excludes 80K training sentences. If the dev+test pool has different characteristics from the training pool (e.g., domain distribution, sentence length, difficulty), the Phase 0 evaluation set may still not be representative of overall model performance. The paper should verify that the dev+test pool has the same distributional properties as the full 100K dataset.
+
+**Results (Section 6)**: The results are correctly presented with appropriate caveats. The distinction between silver-on-silver evaluation (inflated) and gold evaluation (biased by uncertainty selection) is clearly made. The +1.8% F1 improvement from 92 gold sentences is a genuine result, though measured on the diagnostic set rather than representative data.
+
+One concern: the paper reports Post-Cycle 1 F1 improvement on the **same 92 sentences** that were merged into training data. While this demonstrates that the fix pipeline and gold merge improved the model's predictions on hard cases, it is not a valid held-out evaluation. The paper should note this explicitly --- currently, the caveat about uncertainty selection is present but the data leakage issue (evaluation on training data) is not flagged.
+
+### Novelty and Contribution
+
+The paper's novelty lies in three areas:
+
+1. **First multi-domain Vietnamese WS dataset**: Research confirms no comparable resource exists. The closest analogs are Chinese PKUSEG (4 domains) and Thai BEST2010 (4 domains). UDD-WS v1.1 with 5 domains and 100K sentences fills a genuine gap.
+
+2. **Sequential AL across cascading NLP tasks**: The WS → POS → DP pipeline where each task's gold output feeds the next is novel. Prior AL work treats tasks independently or jointly, not sequentially with dependency constraints.
+
+3. **Self-critical evaluation methodology**: The paper's identification of its own evaluation bias (Phase 0 rationale in Section 4.2) and the detailed Limitations section demonstrate methodological maturity. This meta-contribution --- showing how to correctly set up AL evaluation for treebank construction --- is valuable for the community.
+
+The novelty is **moderate overall**: the individual techniques (CRF uncertainty, round-robin sampling, BIO format) are standard, but their combination for Vietnamese multi-domain treebank construction is new.
+
+### Clarity and Presentation
+
+The paper is well-structured and clearly written. The ACL format is properly followed with all mandatory sections. Key improvements from the initial draft:
+
+- Abstract accurately reflects both completed and planned work
+- Contributions (Section 1.3) are specific and verifiable
+- Tables are numbered consistently (with the exception of Table R1)
+- The distinction between completed and planned work is maintained throughout
+- Vietnamese examples use diacritics in running text but simplified ASCII in tables (acceptable for readability)
+
+Minor presentation issues:
+- Section 3.5: The BIO example uses simplified ASCII (*"Mot doanh nghiep..."*) but Section 3.7 uses diacritics (*"suc khoe"*). Consistency would help.
+- The Abstract is 198 words, within the ~200-word target.
+- "Table R1" breaks the numbering convention; should be "Table X" in sequence.
+
+### Reproducibility Assessment
+
+Strong points:
+- Pipeline scripts documented with exact commands (`uv run src/...`)
+- Quality filter rules specified with exact thresholds (Table 2)
+- CRF hyperparameters specified (Section 5.1)
+- Random seed (42) for split stratification
+- `underthesea` version (v2.1.0) stated
+- Data sources on HuggingFace with dataset identifiers
+
+Gaps:
+- AL composite score: $w_{\text{boundary}}$ undefined (Section 5.3)
+- Feature template described as "syllable n-grams, capitalization, character type" without exact specification
+- Label Studio configuration referenced but annotation interface not fully specified
+- Fix pipeline rules are documented in `WS_FIX_REPORT.md` but the paper only summarizes counts (Table 9), not the exact matching logic
+
+### Limitations and Ethics
+
+The Limitations section (8 items) is comprehensive and self-aware. Notable items:
+
+- **Limitation 2** (no gold eval yet) is the most important and is correctly identified as the primary gap.
+- **Limitation 5** (CRF miscalibration on inconsistent silver) identifies a subtle but important problem: uncertainty scoring conflates training noise with genuine ambiguity.
+- **Limitation 6** (1998 cost estimates) is honest --- even a 10-sentence pilot timing study would strengthen the paper.
+
+The Ethics Statement covers data licensing, annotation labor, misuse potential, and bias. The licensing caveat ("verify the licensing terms of each source dataset") is appropriate given that the paper aggregates from multiple sources.
+
+**Missing from Limitations**: The paper does not discuss the potential for the silver dataset to be used for **training Vietnamese LLMs** or **evaluation benchmarks**, where the 2.4% inconsistency rate could have outsized impact. A brief note on downstream misuse beyond CRF training would strengthen the Ethics section.
+
+## Related Work Research
+
+### Papers Found
+
+| Paper | Year | Method | Results | Relevance |
+|-------|------|--------|---------|-----------|
+| NlpHUST/vi-word-segmentation | 2022 | ELECTRA fine-tuned | F1 98.35% (VLSP 2013) | **Not cited** --- current Vietnamese WS SOTA |
+| SpanSeg + XLM-RoBERTa | 2021 | Span labeling | F1 98.31% (VLSP 2013) | **Not cited** --- transformer WS SOTA |
+| ViWordFormer | 2025 | Word formation model | Outperforms on ViHOS, PhoNER | **Not cited** --- challenges WS paradigm |
+| DGDT (Vietnamese cross-domain DP) | 2025 | Domain separation treebank | -5.09% LAS cross-domain | **Not cited** --- directly validates multi-domain motivation |
+| BERT sentence recommendation for Viet UD | 2023 | BERT-based AL for Vietnamese UD | Improved sentence selection | **Not cited** --- closest prior work to UDD-1 AL |
+| PhoNLP | 2021 | PhoBERT joint model | LAS 79.11 (VnDT) | **Not cited** --- Vietnamese DP SOTA |
+| VLSP 2020 DP shared task | 2020 | PhoBERT+ELMO biaffine | LAS 76.27 | **Not cited** --- Vietnamese DP benchmark |
+| VLSP 2019 DP shared task | 2019 | Biaffine + ensemble | LAS 61.28 | **Not cited** --- first Vietnamese UD shared task |
+| PKUSEG (Chinese multi-domain WS) | 2019 | Domain-adapted CRF/neural | +0.67 F1 OOD | **Not cited** --- closest resource comparison |
+| LLMs in AL loop (ECML-PKDD 2024) | 2024 | GPT-4 as annotator in AL | 42x cost savings | Not cited --- relevant to future directions |
+| Pomak UD AL treebank | 2025 | Active annotation | 6,351 sentences | Cited (Section 2.4) |
+| BKTreebank | 2018 | Manual dependency annotation | IAA 94.5% POS | Cited (Section 2.2) |
+
+### Missing Citations
+
+1. **DGDT (Huynh et al., 2025)**: Vietnamese Domain Generalization Dependency Treebank. Directly relevant --- first Vietnamese cross-domain DP evaluation, shows 5.09% LAS drop across domains. Essential for motivating UDD-1's multi-domain approach.
+
+2. **BERT-based sentence recommendation for Vietnamese UD (FDSE 2023)**: Vietnamese-specific AL for UD treebank construction. The closest prior work to UDD-1's methodology. Must be cited and discussed.
+
+3. **VLSP 2019 and 2020 dependency parsing shared tasks**: These established Vietnamese DP benchmarks with public data and leaderboards. The paper cites VLSP 2013 (WS) but omits VLSP 2019/2020 (DP), which are directly relevant to Task 3.
+
+4. **PhoNLP (Nguyen and Nguyen, 2021)**: Achieves LAS 79.11 on VnDT, the current Vietnamese DP SOTA. The paper cites ~76% LAS for underthesea but does not contextualize this against the SOTA.
+
+5. **PKUSEG (Luo et al., 2019)**: Chinese multi-domain word segmentation toolkit with 4 domains. The closest comparison to UDD-WS v1.1's multi-domain approach. Useful for cross-lingual context.
+
+6. **Transformer-based Vietnamese WS**: NlpHUST ELECTRA (F1 98.35%) and SpanSeg (F1 98.31%) surpass VnCoreNLP (97.90%). The comparison in Table 17 should include these.
+
+### SOTA Verification
+
+- **Claimed**: VnCoreNLP 97.90% F1 as WS SOTA (Section 2.1, Table 17)
+  **Actual**: NlpHUST ELECTRA 98.35% F1, SpanSeg 98.31% F1
+  **Assessment**: Outdated by ~0.45% F1. VnCoreNLP is the best CRF system but not overall SOTA.
+
+- **Claimed**: ~76% LAS on VLSP 2020 (Section 4.6, referencing v1.0)
+  **Actual**: Best VLSP 2020 system: 76.27% LAS (PhoBERT+ELMO). PhoNLP achieves 79.11% on VnDT.
+  **Assessment**: Consistent for VLSP 2020 but should note PhoNLP as current SOTA.
+
+- **Claimed**: Brants & Skut (1998) 3-5x speedup for pre-annotation correction (Section 2.4)
+  **Actual**: Verified in original paper.
+  **Assessment**: Accurate but noted as Limitation 6 (28-year-old study).
+
+- **Claimed**: No multi-domain Vietnamese WS dataset exists (Section 1.3)
+  **Actual**: Research confirms this claim. VLSP 2013 is news-only; no other Vietnamese WS dataset covers multiple domains.
+  **Assessment**: Verified --- UDD-WS v1.1 fills a genuine gap.
+
+## Questions for Authors
+
+1. **What is the exact definition of $w_{\text{boundary}}$ in the composite scoring formula (Section 5.3)?** The paper defines $\bar{u}(s)$ and $n_{\text{long}}$ but leaves the boundary weight undefined. Is it the fraction of I-W tokens? The number of B-W→I-W transitions? A learned parameter?
+
+2. **How many of the 3,592 inconsistently segmented forms represent genuine contextual ambiguity vs. tokenizer errors?** For instance, *nguoi ta* ("people/they") may genuinely function as either one word or two depending on context (pronoun vs. noun phrase). Can you estimate the proportion of "defensible inconsistency" vs. "error inconsistency"?
+
+3. **Have you verified that the dev+test pool (20K sentences) has the same distributional properties as the full 100K dataset?** If the stratified split introduces any systematic differences (e.g., different sentence lengths per domain), Phase 0 sampling from this pool may still not be fully representative.
+
+4. **What was the actual annotation throughput for Cycle 1?** The paper estimates 80 sent/day for WS based on Brants & Skut (1998). Cycle 1 annotated ~95 sentences --- how many calendar days did this take? Even rough timing data would validate the cost estimates.
+
+5. **Are you aware of the BERT-based sentence recommendation work for Vietnamese UD (FDSE 2023)?** This appears to be the closest prior work to your AL approach for Vietnamese treebank construction and should be discussed.
+
+## Minor Issues
+
+- Section 3.5: BIO example uses ASCII (*"Mot doanh nghiep"*) while running text uses diacritics. Use consistent representation.
+- Table R1 breaks the sequential numbering. Integrate as Table N or use a consistent convention for in-section tables.
+- Section 7.2: Error counts are approximate ("~40", "~35"). For 92 sentences, exact counts are feasible and expected.
+- Section 2.1: "The Underthesea toolkit provides a word_tokenize function combining CRF and regex-based tokenization" --- specify the version (v2.1.0, consistent with Section 3.7).
+- Section 4.5: The POS selection weight $\alpha = 0.6$ is stated without justification. A brief note on sensitivity or planned tuning would help.
+- References: Some entries use first-name initials inconsistently (e.g., "Nguyen, K.-H." vs. "Nguyen, C.T."). Standardize to ACL bibliography format.
+- The Acknowledgments section thanks "the anonymous reviewer of TECHNICAL_REPORT v1.0" --- this is unusual for ACL format where reviews are anonymous. Consider rephrasing.
+
+## Suggestions for Improvement
+
+1. **Execute Phase 0 before next submission (highest priority)**: Annotating 500 random sentences (~6 days) would provide the representative gold baseline F1, transforming the paper from "methodology + plan" to "resource + methodology + results." This single addition would move the score from 3 to 4.
+
+2. **Add missing Vietnamese cross-domain references**: Cite DGDT (2025), FDSE 2023 Vietnamese UD AL, VLSP 2019/2020 DP shared tasks, and PhoNLP. These directly validate the paper's motivation and contextualize the contribution.
+
+3. **Update WS SOTA comparison**: Add transformer-based WS results (NlpHUST 98.35%, SpanSeg 98.31%) to Table 17. Discuss why CRF is appropriate for the AL loop despite lower ceiling (speed, marginal probabilities, interpretability).
+
+4. **Report exact error counts**: Replace approximate counts in Table 16 with exact values. Manually categorize all 225 errors.
+
+5. **Define $w_{\text{boundary}}$ precisely**: Specify the boundary weight formula in Section 5.3 for full reproducibility.
+
+6. **Discuss PKUSEG as cross-lingual reference**: The Chinese multi-domain WS literature (PKUSEG, MCCWS) provides the closest comparison. Citing it strengthens the claim that multi-domain WS is a recognized challenge.
+
+7. **Add pilot timing data**: Report actual annotation time for Cycle 1 to validate or revise the 80 sent/day estimate.
+
+8. **Flag data leakage in Section 6.2**: The Post-Cycle 1 CRF evaluation on the 92 gold sentences that were merged into training data is not a held-out evaluation. Note this explicitly alongside the existing uncertainty-selection caveat.
+
+## Evaluation Checklist
+
+### Methodology
+- [x] Research questions clearly stated
+- [x] Methods appropriate for research questions
+- [ ] Baselines appropriate and fairly compared (WS SOTA outdated)
+- [ ] Statistical significance properly addressed (no confidence intervals on F1)
+- [x] Limitations of approach acknowledged (8 items, comprehensive)
+
+### Experiments
+- [x] Datasets properly described (source, size, splits, preprocessing)
+- [x] Evaluation metrics appropriate for the task
+- [x] Training details sufficient for reproduction (CRF hyperparameters in Table, Section 5.1)
+- [ ] Ablation studies or analysis provided (error taxonomy yes, ablation no)
+- [x] Results support the claims made (with appropriate caveats)
+
+### Presentation
+- [x] Abstract accurately summarizes contributions
+- [x] Introduction motivates the problem
+- [ ] Related work comprehensive and fair (missing DGDT, VLSP 2019/2020, PhoNLP, transformer WS)
+- [x] Figures/tables readable and informative
+- [x] Conclusion matches actual contributions
+
+### Related Work Verification
+- [ ] Key prior work on same task is cited (DGDT, FDSE 2023 missing)
+- [ ] Baseline comparisons use current methods (VnCoreNLP is CRF SOTA, not overall SOTA)
+- [ ] SOTA claims are accurate and up-to-date (WS SOTA is 98.35%, not 97.90%)
+- [ ] No significant missing references (6 important papers missing)
+- [x] Fair characterization of competing approaches
+
+### Responsible NLP
+- [x] Limitations section present and substantive (8 items)
+- [x] Potential negative impacts discussed (Ethics Statement)
+- [x] Data collection ethics addressed (licensing caveat)
+- [x] Bias considerations mentioned (source bias + tokenizer bias)

WS_FIX_REPORT.md

@@ -4,212 +4,320 @@ Fixes applied by `src/fix_ws_errors.py` to UDD-1.1 word segmentation BIO files.
 
 ## Summary
 
-| File | Cross-boundary splits | Long token splits | Compound merges | Validation errors |
-|------|----------------------:|------------------:|----------------:|------------------:|
-| udd-ws-v1.1-train.txt | 4,969 | 725 | 1,870 | 0 |
-| udd-ws-v1.1-dev.txt | 598 | 87 | 222 | 0 |
-| udd-ws-v1.1-test.txt | 611 | 79 | 233 | 0 |
-| **TOTAL** | **6,178** | **891** | **2,325** | **0** |
+| File | Cross-boundary | Long token | Compound merges | Foreign splits | Name boundary | Validation errors |
+|------|---------------:|-----------:|----------------:|---------------:|--------------:|------------------:|
+| udd-ws-v1.1-train.txt | 207 | 0 | 409 | 4,287 | 668 | 0 |
+| udd-ws-v1.1-dev.txt | 26 | 0 | 61 | 561 | 68 | 0 |
+| udd-ws-v1.1-test.txt | 25 | 0 | 71 | 527 | 80 | 0 |
+| **TOTAL** | **258** | **0** | **541** | **5,375** | **816** | **0** |
 
 ## Merge Frequency by Term
 
 | Term | Count |
 |:-----|------:|
-| hàng hóa | 759 |
-| chữa bệnh | 326 |
-| khám bệnh | 255 |
-| vụ án | 209 |
-| hòa giải | 138 |
-| kèm theo | 107 |
-| trả lại | 86 |
-| làm chủ | 82 |
-| phạt tù | 82 |
-| hủy bỏ | 81 |
-| phiên tòa | 44 |
-| bị hại | 37 |
-| ghi rõ | 35 |
-| rút gọn | 34 |
-| tiền công | 27 |
-| quá hạn | 9 |
-| chủ tọa | 7 |
-| lời khai | 5 |
-| thuê khoán | 1 |
-| bốc hàng | 1 |
+| ủy ban | 371 |
+| mái nhà | 41 |
+| làm chủ | 27 |
+| lượng tử | 24 |
+| tại ngũ | 16 |
+| người thương | 16 |
+| xua tay | 9 |
+| nghiến răng | 7 |
+| như điên | 7 |
+| tích lũy | 5 |
+| trêu ghẹo | 2 |
+| chăn lợn | 2 |
+| khay trà | 2 |
+| sương mai | 2 |
+| đồng tự | 2 |
+| đường thẳng | 2 |
+| mu rùa | 1 |
+| nói gở | 1 |
+| lính thú | 1 |
+| đầm đuôi cá | 1 |
+| hành chánh | 1 |
+| bơi chó | 1 |
 
 ## Cross-Boundary Split Examples
 
 ### udd-ws-v1.1-train.txt
 
-- [uvw-12640] split "nhà Minh" → 'nhà' + 'Minh'
-- [vlc-18105] split "ban Thường vụ Quốc hội" → 'ban' + 'Thường vụ' + 'Quốc hội'
-- [uvn-5192] split "nam MC" → 'nam' + 'MC'
-- [uvw-13051] split "diễn văn Gettysburg" → 'diễn văn' + 'Gettysburg'
-- [uvn-11996] split "sàn London" → 'sàn' + 'London'
-- [uvb-f-11285] split "Tiên phong Thiên Giới" → 'Tiên phong' + 'Thiên Giới'
-- [uvn-1652] split "giải Vietstock" → 'giải' + 'Vietstock'
-- [uvn-1652] split "Nông nghiệp và Môi trường" → 'Nông nghiệp và' + 'Môi trường'
-- [uvn-15129] split "chuyên Sinh" → 'chuyên' + 'Sinh'
-- [uvb-n-18150] split "vua Kiệt" → 'vua' + 'Kiệt'
-- [uvn-11110] split "bắc Tử" → 'bắc' + 'Tử'
-- [uvb-n-7068] split "phái Duy Vật" → 'phái' + 'Duy Vật'
-- [uvb-n-10940] split "quân Anh" → 'quân' + 'Anh'
-- [uvn-3318] split "chuẩn Anh" → 'chuẩn' + 'Anh'
-- [uvw-11565] split "nhà Thương" → 'nhà' + 'Thương'
-- [uvw-14254] split "châu Á" → 'châu' + 'Á'
-- [uvn-6648] split "hình X-quang" → 'hình' + 'X-quang'
-- [uvn-3392] split "bảng G ." → 'bảng' + 'G .'
-- [uvn-3392] split "bảng H" → 'bảng' + 'H'
-- [uvn-8106] split "châu Âu" → 'châu' + 'Âu'
-- ... and 4949 more
+- [uvn-4610] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-9598] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvn-14855] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-15939] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-4115] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-16369] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-16369] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-17534] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-4739] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-896] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvw-8925] split "Xã hội Chủ nghĩa" → 'Xã hội' + 'Chủ nghĩa'
+- [vlc-15109] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-3849] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-999] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-16713] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-10256] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-15677] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-4321] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-8038] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-2887] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- ... and 187 more
 
 ### udd-ws-v1.1-dev.txt
 
-- [uvb-n-730] split "mừng Lễ" → 'mừng' + 'Lễ'
-- [uvn-7093] split "chữ T" → 'chữ' + 'T'
-- [uvn-3234] split "môn Văn cấp" → 'môn' + 'Văn cấp'
-- [uvw-10521] split "of Plants" → 'of' + 'Plants'
-- [uvb-n-5230] split "châu Âu" → 'châu' + 'Âu'
-- [uvw-380] split "quẻ Sư 師" → 'quẻ' + 'Sư 師'
-- [uvb-f-19862] split "bán TV" → 'bán' + 'TV'
-- [uvn-207] split "9.0 IELTS" → '9.0' + 'IELTS'
-- [uvw-9988] split "Đế quốc Đông" → 'Đế quốc' + 'Đông'
-- [uvw-8703] split "phân Riemann" → 'phân' + 'Riemann'
-- [vlc-16923] split "và Phát triển nông thôn" → 'và' + 'Phát triển nông thôn'
-- [uvw-7715] split "châu Phi Francophone" → 'châu' + 'Phi Francophone'
-- [vlc-19028] split "và Phát triển nông thôn" → 'và' + 'Phát triển nông thôn'
-- [uvw-18298] split ":: Một" → '::' + 'Một'
-- [uvn-6630] split "hành Perseverance" → 'hành' + 'Perseverance'
-- [uvn-2592] split "da Luz" → 'da' + 'Luz'
-- [uvn-15632] split "phi hành gia NASA" → 'phi hành gia' + 'NASA'
-- [vlc-8981] split "thường vụ Quốc hội" → 'thường vụ' + 'Quốc hội'
-- [uvw-7003] split "châu Âu" → 'châu' + 'Âu'
-- [uvn-115] split "Giáo dục và Đào tạo" → 'Giáo dục và' + 'Đào tạo'
-- ... and 578 more
+- [uvw-18563] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-3635] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-10450] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-11201] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-310] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-12944] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-6630] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-6630] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvw-7217] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-7765] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-17628] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-15712] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvw-7424] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-6567] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-17311] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-17818] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvw-15253] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-4949] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvn-1143] split "làm Chủ tịch" → 'làm' + 'Chủ tịch'
+- [vlc-8360] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- ... and 6 more
 
 ### udd-ws-v1.1-test.txt
 
-- [uvw-10170] split "phương Nam" → 'phương' + 'Nam'
-- [uvw-10170] split "thời Trung Cổ" → 'thời' + 'Trung Cổ'
-- [vlc-6223] split "thường vụ Quốc hội" → 'thường vụ' + 'Quốc hội'
+- [vlc-17267] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
 - [vlc-6223] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
-- [uvn-2071] split "tây Nam Cực" → 'tây' + 'Nam Cực'
-- [uvn-19983] split "fanpage Facebook" → 'fanpage' + 'Facebook'
-- [uvn-19983] split "nhắn SMS" → 'nhắn' + 'SMS'
-- [uvw-4834] split "châu Phi" → 'châu' + 'Phi'
-- [uvw-4834] split "châu Phi" → 'châu' + 'Phi'
-- [uvb-n-1229] split "matières Le" → 'matières' + 'Le'
-- [uvb-f-17918] split "gả Loan" → 'gả' + 'Loan'
-- [uvb-f-17918] split "phán Lợi" → 'phán' + 'Lợi'
-- [uvb-f-11852] split "‘ Người" → '‘' + 'Người'
-- [uvb-f-4210] split "quân Urgals" → 'quân' + 'Urgals'
-- [uvw-18184] split "quân H'mong" → 'quân' + "H'mong"
-- [vlc-13904] split "thường vụ Quốc hội" → 'thường vụ' + 'Quốc hội'
-- [uvn-17235] split "giải Jackpot" → 'giải' + 'Jackpot'
-- [uvn-3787] split "Thế chiến Ii" → 'Thế chiến' + 'Ii'
-- [uvw-14649] split "nhà Hồ" → 'nhà' + 'Hồ'
-- [uvw-14649] split "quân Minh" → 'quân' + 'Minh'
-- ... and 591 more
+- [uvn-10658] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-1254] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-3581] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-1319] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-2238] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-14701] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-14701] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-16929] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-10389] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-16627] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-6090] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-7684] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-1544] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [uvn-16619] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-18829] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-15826] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-9744] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- [vlc-19549] split "Mặt trận Tổ quốc" → 'Mặt trận' + 'Tổ quốc'
+- ... and 5 more
 
 ## Long Token Split Examples
 
+## Foreign Word Split Examples
+
+### udd-ws-v1.1-train.txt
+
+- [uvw-7593] split-foreign "Max Planck" → 'Max' + 'Planck'
+- [uvw-7593] split-foreign "Werner Heisenberg" → 'Werner' + 'Heisenberg'
+- [uvw-7593] split-foreign "Erwin Schrödinger" → 'Erwin' + 'Schrödinger'
+- [uvw-7593] split-foreign "Paul Dirac" → 'Paul' + 'Dirac'
+- [uvn-19074] split-foreign "Gazprom Neft" → 'Gazprom' + 'Neft'
+- [uvn-1654] split-foreign "Elon Musk" → 'Elon' + 'Musk'
+- [uvb-n-12310] split-foreign "lon ton" → 'lon' + 'ton'
+- [uvb-n-1071] split-foreign "Gallimard Jeunesse" → 'Gallimard' + 'Jeunesse'
+- [uvb-n-1071] split-foreign "Revault d’Allones" → 'Revault' + 'd’Allones'
+- [uvb-n-1071] split-foreign "Revault d’Allones" → 'Revault' + 'd’Allones'
+- [uvn-588] split-foreign "Sony Music" → 'Sony' + 'Music'
+- [uvn-7078] split-foreign "Platinum Victory" → 'Platinum' + 'Victory'
+- [uvn-7078] split-foreign "Jardine Cycle" → 'Jardine' + 'Cycle'
+- [uvn-7078] split-foreign "Jardine Matheson" → 'Jardine' + 'Matheson'
+- [uvw-7071] split-foreign "Bartholomen Velho" → 'Bartholomen' + 'Velho'
+- [uvw-7071] split-foreign "Vaz Dourado" → 'Vaz' + 'Dourado'
+- [uvw-7071] split-foreign "Van Langren" → 'Van' + 'Langren'
+- [uvw-13401] split-foreign "Ishpatina Ridge" → 'Ishpatina' + 'Ridge'
+- [uvn-18582] split-foreign "Rhincodon typus" → 'Rhincodon' + 'typus'
+- [uvn-18582] split-foreign "Chiloscyllium plagiosum" → 'Chiloscyllium' + 'plagiosum'
+- [uvn-11577] split-foreign "Vahid Kardhani" → 'Vahid' + 'Kardhani'
+- [uvb-n-1000] split-foreign "Samaneri Sudhamma" → 'Samaneri' + 'Sudhamma'
+- [uvn-5523] split-foreign "phanh phui" → 'phanh' + 'phui'
+- [uvn-5523] split-foreign "HACKER NEWS" → 'HACKER' + 'NEWS'
+- [uvn-9592] split-foreign "Smart Workstation" → 'Smart' + 'Workstation'
+- [uvn-12958] split-foreign "Research with" → 'Research' + 'with'
+- [uvw-251] split-foreign "Philosophiæ Naturalis" → 'Philosophiæ' + 'Naturalis'
+- [uvw-251] split-foreign "Principia Mathematica" → 'Principia' + 'Mathematica'
+- [uvn-179] split-foreign "Jung Geun-sik" → 'Jung' + 'Geun-sik'
+- [uvn-15685] split-foreign "Brian Yuliarto" → 'Brian' + 'Yuliarto'
+- ... and 4257 more
+
+### udd-ws-v1.1-dev.txt
+
+- [uvw-16942] split-foreign "Fyodorovich Romanov" → 'Fyodorovich' + 'Romanov'
+- [uvn-6675] split-foreign "Notre Dame" → 'Notre' + 'Dame'
+- [uvn-6675] split-foreign "Pathfinding Robotic" → 'Pathfinding' + 'Robotic'
+- [uvn-6675] split-foreign "Observation Unit" → 'Observation' + 'Unit'
+- [uvn-16075] split-foreign "Guo Yungao" → 'Guo' + 'Yungao'
+- [uvw-11029] split-foreign "Robert Recorde" → 'Robert' + 'Recorde'
+- [uvw-4504] split-foreign "Kremser S33" → 'Kremser' + 'S33'
+- [uvw-4504] split-foreign "Vienna B1" → 'Vienna' + 'B1'
+- [uvw-12819] split-foreign "Francis Asbury" → 'Francis' + 'Asbury'
+- [uvn-16644] split-foreign "Jerome Powell" → 'Jerome' + 'Powell'
+- [uvw-710] split-foreign "endoplasmic reticulum" → 'endoplasmic' + 'reticulum'
+- [uvn-1745] split-foreign "Scotsman All-Carbon" → 'Scotsman' + 'All-Carbon'
+- [uvn-1745] split-foreign "Fiber Scooter" → 'Fiber' + 'Scooter'
+- [uvn-1745] split-foreign "Superstrata Bike" → 'Superstrata' + 'Bike'
+- [uvb-n-515] split-foreign "Anthony Weston" → 'Anthony' + 'Weston'
+- [uvb-n-515] split-foreign "Adam Smith" → 'Adam' + 'Smith'
+- [uvb-n-515] split-foreign "Madsen Pirie" → 'Madsen' + 'Pirie'
+- [uvn-4460] split-foreign "Patek Philippe" → 'Patek' + 'Philippe'
+- [uvw-19975] split-foreign "tetrachloride titan" → 'tetrachloride' + 'titan'
+- [uvw-4144] split-foreign "Unicode Consortium" → 'Unicode' + 'Consortium'
+- [uvw-16494] split-foreign "Rey Juan" → 'Rey' + 'Juan'
+- [uvb-f-3265] split-foreign "Sam Roffe" → 'Sam' + 'Roffe'
+- [uvn-2757] split-foreign "Jens-Frederik Nielsen" → 'Jens-Frederik' + 'Nielsen'
+- [uvn-2757] split-foreign "Mette Frederiksen" → 'Mette' + 'Frederiksen'
+- [uvn-2757] split-foreign "Emmanuel Macron" → 'Emmanuel' + 'Macron'
+- [uvb-n-16036] split-foreign "Klemens von" → 'Klemens' + 'von'
+- [uvw-9683] split-foreign "Ibn al-Haytham" → 'Ibn' + 'al-Haytham'
+- [uvw-9683] split-foreign "Kitab al-manazir" → 'Kitab' + 'al-manazir'
+- [uvn-12996] split-foreign "V-Business Pro" → 'V-Business' + 'Pro'
+- [uvn-12996] split-foreign "V-Business Advance" → 'V-Business' + 'Advance'
+- ... and 531 more
+
+### udd-ws-v1.1-test.txt
+
+- [uvb-n-19086] split-foreign "Ailic Mayenkh" → 'Ailic' + 'Mayenkh'
+- [uvn-4221] split-foreign "Rice Platform" → 'Rice' + 'Platform'
+- [uvn-2798] split-foreign "HR Asia" → 'HR' + 'Asia'
+- [uvw-1174] split-foreign "Cthulhu Macula" → 'Cthulhu' + 'Macula'
+- [uvb-n-3361] split-foreign "Groce Malo" → 'Groce' + 'Malo'
+- [uvn-6814] split-foreign "Peng Zhihui" → 'Peng' + 'Zhihui'
+- [uvn-812] split-foreign "Engineered Arts" → 'Engineered' + 'Arts'
+- [uvw-17796] split-foreign "Swords Society" → 'Swords' + 'Society'
+- [uvw-11767] split-foreign "Lech Kaczyński" → 'Lech' + 'Kaczyński'
+- [uvn-10632] split-foreign "FTSE Russell" → 'FTSE' + 'Russell'
+- [uvw-14690] split-foreign "Christian Friedrich" → 'Christian' + 'Friedrich'
+- [uvw-6352] split-foreign "Mein Kampf" → 'Mein' + 'Kampf'
+- [uvw-4709] split-foreign "Los Angeles" → 'Los' + 'Angeles'
+- [uvw-4709] split-foreign "Los Angeles" → 'Los' + 'Angeles'
+- [uvn-15627] split-foreign "Nick Hague" → 'Nick' + 'Hague'
+- [uvn-15627] split-foreign "Aleksandr Gorbunov" → 'Aleksandr' + 'Gorbunov'
+- [uvb-n-2364] split-foreign "Abdul Razak" → 'Abdul' + 'Razak'
+- [uvw-8985] split-foreign "MediaWiki wiki" → 'MediaWiki' + 'wiki'
+- [uvw-8985] split-foreign "WebPlatform Docs" → 'WebPlatform' + 'Docs'
+- [uvn-13054] split-foreign "Shark Tank" → 'Shark' + 'Tank'
+- [uvn-6301] split-foreign "Carina Hong" → 'Carina' + 'Hong'
+- [uvw-10066] split-foreign "Nicolaus Copernicus" → 'Nicolaus' + 'Copernicus'
+- [uvw-15154] split-foreign "Khong Chiam" → 'Khong' + 'Chiam'
+- [uvw-15154] split-foreign "Ubon Ratchathani" → 'Ubon' + 'Ratchathani'
+- [uvn-17184] split-foreign "Marie van" → 'Marie' + 'van'
+- [uvw-14969] split-foreign "Beniamino Gigli" → 'Beniamino' + 'Gigli'
+- [uvw-11704] split-foreign "Black Swamp" → 'Black' + 'Swamp'
+- [uvw-4837] split-foreign "Coffee Tea" → 'Coffee' + 'Tea'
+- [uvw-4837] split-foreign "Black tea" → 'Black' + 'tea'
+- [uvw-4837] split-foreign "Green tea" → 'Green' + 'tea'
+- ... and 497 more
+
+## Name Boundary Split Examples
+
 ### udd-ws-v1.1-train.txt
 
-- [vlc-16418] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-366] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-18939] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvw-3718] split "lực lượng vũ trang giáo phái" → 'lực lượng vũ trang' + 'giáo phái'
-- [vlc-19939] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvn-5647] split "phương tiện thông tin đại chúng" → 'phương tiện' + 'thông tin đại chúng'
-- [vlc-4038] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-9598] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-121] split "luật tố tụng hình sự" → 'luật' + 'tố tụng hình sự'
-- [vlc-121] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-2994] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-2994] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [uvw-3692] split "Thanh Tâm Tài Nhân thảo" → 'Thanh Tâm' + 'Tài Nhân' + 'thảo'
-- [uvn-11014] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-7080] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-1414] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [uvn-15822] split "Chúng sở hữu cơ thể" → 'Chúng' + 'sở hữu' + 'cơ thể'
-- [vlc-14193] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-15939] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-17799] split "Bộ Tài nguyên và Môi trường" → 'Bộ' + 'Tài nguyên' + 'và' + 'Môi trường'
-- [vlc-17799] split "Bộ Tài nguyên và Môi trường" → 'Bộ' + 'Tài nguyên' + 'và' + 'Môi trường'
-- [vlc-7359] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvn-12995] split "Tổng giám đốc điều hành" → 'Tổng giám đốc' + 'điều hành'
-- [vlc-4115] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-16369] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-16369] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-7736] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-7433] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvn-2956] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-131] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- ... and 695 more
+- [uvw-6789] split-name-boundary "A0 biểu thị" → 'A0' + 'biểu thị'
+- [uvw-6789] split-name-boundary "A9 biểu thị" → 'A9' + 'biểu thị'
+- [uvn-6189] split-name-boundary "Tân tiến sĩ" → 'Tân' + 'tiến sĩ'
+- [uvw-10682] split-name-boundary "Âu châu lục địa" → 'Âu' + 'châu lục địa'
+- [uvn-13266] split-name-boundary "EVNSPC nhiệm kỳ" → 'EVNSPC' + 'nhiệm kỳ'
+- [uvw-4500] split-name-boundary "Thánh đại công" → 'Thánh' + 'đại công'
+- [uvw-16965] split-name-boundary "Đức thống nhất" → 'Đức' + 'thống nhất'
+- [uvb-n-818] split-name-boundary "Trần khởi nghiệp" → 'Trần' + 'khởi nghiệp'
+- [uvb-n-15660] split-name-boundary "Chiến Quốc dân chúng" → 'Chiến Quốc' + 'dân chúng'
+- [uvb-f-15644] split-name-boundary "Quỳnh cảm thấy" → 'Quỳnh' + 'cảm thấy'
+- [uvb-n-12580] split-name-boundary "Văn chậm rãi" → 'Văn' + 'chậm rãi'
+- [uvb-f-8769] split-name-boundary "Trần thản nhiên" → 'Trần' + 'thản nhiên'
+- [uvb-n-9642] split-name-boundary "Chị em gái" → 'Chị' + 'em gái'
+- [uvb-f-16955] split-name-boundary "Bầu không khí" → 'Bầu' + 'không khí'
+- [uvb-n-8707] split-name-boundary "Syrie đối xử" → 'Syrie' + 'đối xử'
+- [uvn-521] split-name-boundary "Trịnh Hà thừa hưởng" → 'Trịnh Hà' + 'thừa hưởng'
+- [uvw-19476] split-name-boundary "Cơ sinh học" → 'Cơ' + 'sinh học'
+- [uvw-16946] split-name-boundary "Einstein chứng minh" → 'Einstein' + 'chứng minh'
+- [uvn-8273] split-name-boundary "PVF-CANDvới lợi thế" → 'PVF-CANDvới' + 'lợi thế'
+- [uvn-4990] split-name-boundary "Cựu chiến binh" → 'Cựu' + 'chiến binh'
+- [uvn-12163] split-name-boundary "Khoa học tập" → 'Khoa' + 'học tập'
+- [uvb-f-2975] split-name-boundary "Thế nghĩa là" → 'Thế' + 'nghĩa là'
+- [uvw-11278] split-name-boundary "GDP danh nghĩa" → 'GDP' + 'danh nghĩa'
+- [uvn-2671] split-name-boundary "Tập thể dục" → 'Tập' + 'thể dục'
+- [uvb-n-10229] split-name-boundary "Ăn chênh lệch" → 'Ăn' + 'chênh lệch'
+- [vlc-19523] split-name-boundary "Chủ sở hữu quyền" → 'Chủ' + 'sở hữu quyền'
+- [uvb-n-7589] split-name-boundary "Để ý nghĩa" → 'Để' + 'ý nghĩa'
+- [uvn-8775] split-name-boundary "Mỹ cáo buộc" → 'Mỹ' + 'cáo buộc'
+- [uvb-n-17047] split-name-boundary "Valentinianl bất tài" → 'Valentinianl' + 'bất tài'
+- [uvw-3030] split-name-boundary "Thái học sinh" → 'Thái' + 'học sinh'
+- ... and 638 more
 
 ### udd-ws-v1.1-dev.txt
 
-- [vlc-18372] split "phương tiện thông tin đại chúng" → 'phương tiện' + 'thông tin đại chúng'
-- [vlc-7298] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-14549] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-3635] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-210] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-210] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-17769] split "Tòa án nhân dân tối cao" → 'Tòa án nhân dân' + 'tối cao'
-- [vlc-10450] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-8980] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-18762] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvw-13677] split "thức ăn biến đổi gen" → 'thức ăn' + 'biến đổi gen'
-- [vlc-8382] split "Viện kiểm sát nhân dân tối cao" → 'Viện kiểm sát' + 'nhân dân' + 'tối cao'
-- [vlc-562] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvn-1191] split "quyền sở hữu toàn dân" → 'quyền sở hữu' + 'toàn' + 'dân'
-- [uvw-5987] split "nhà thiên văn nghiệp dư" → 'nhà thiên văn' + 'nghiệp' + 'dư'
-- [uvw-5987] split "nhà thiên văn nghiệp dư" → 'nhà thiên văn' + 'nghiệp' + 'dư'
-- [vlc-11201] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [uvb-f-16068] split "Dương dở khóc dở cười" → 'Dương' + 'dở' + 'khóc' + 'dở' + 'cười'
-- [vlc-11811] split "Viện kiểm sát nhân dân" → 'Viện kiểm sát' + 'nhân dân'
-- [vlc-11811] split "Viện kiểm sát nhân dân" → 'Viện kiểm sát' + 'nhân dân'
-- [uvw-6333] split "phóng vệ tinh nhân tạo" → 'phóng vệ tinh' + 'nhân tạo'
-- [vlc-13552] split "phương tiện thông tin đại chúng" → 'phương tiện' + 'thông tin đại chúng'
-- [vlc-17367] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-784] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-375] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-310] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-8628] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvb-n-18761] split "lý thuyết vi trùng học" → 'lý thuyết' + 'vi trùng' + 'học'
-- [vlc-12944] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [uvb-n-8034] split "phong vũ biểu tâm hồn" → 'phong' + 'vũ' + 'biểu' + 'tâm hồn'
-- ... and 57 more
+- [uvn-15053] split-name-boundary "Lâm chủ động" → 'Lâm' + 'chủ động'
+- [uvn-18170] split-name-boundary "T&T thúc đẩy" → 'T&T' + 'thúc đẩy'
+- [uvb-f-12120] split-name-boundary "Lê cau mày" → 'Lê' + 'cau mày'
+- [uvw-1745] split-name-boundary "Newton thiết lập" → 'Newton' + 'thiết lập'
+- [uvb-n-3769] split-name-boundary "Chu Du tính tình" → 'Chu Du' + 'tính tình'
+- [uvw-14536] split-name-boundary "Lương khởi nghĩa" → 'Lương' + 'khởi nghĩa'
+- [uvw-754] split-name-boundary "Nguyễn bắt đầu" → 'Nguyễn' + 'bắt đầu'
+- [uvb-n-18390] split-name-boundary "Văn sửng sốt" → 'Văn' + 'sửng sốt'
+- [uvb-f-6739] split-name-boundary "Ngẫm nghĩ lại" → 'Ngẫm' + 'nghĩ lại'
+- [uvn-16785] split-name-boundary "Biên bản ghi nhớ" → 'Biên' + 'bản ghi nhớ'
+- [uvb-f-16752] split-name-boundary "Nữ minh tinh" → 'Nữ' + 'minh tinh'
+- [uvw-5598] split-name-boundary "Texas biến đổi" → 'Texas' + 'biến đổi'
+- [uvn-2059] split-name-boundary "Hạ nghị sĩ" → 'Hạ' + 'nghị sĩ'
+- [uvn-5151] split-name-boundary "Masanđạt doanh thu" → 'Masanđạt' + 'doanh thu'
+- [uvw-19553] split-name-boundary "Khu tự trị" → 'Khu' + 'tự trị'
+- [uvb-n-5806] split-name-boundary "Đáng lẽ ra" → 'Đáng' + 'lẽ ra'
+- [uvb-n-1302] split-name-boundary "Huyền tương đối" → 'Huyền' + 'tương đối'
+- [uvw-16476] split-name-boundary "Mạng khoa học" → 'Mạng' + 'khoa học'
+- [uvw-2471] split-name-boundary "Lê biên soạn" → 'Lê' + 'biên soạn'
+- [uvb-n-3787] split-name-boundary "Đài phát thanh" → 'Đài' + 'phát thanh'
+- [uvb-n-15532] split-name-boundary "La Mã thần thánh" → 'La Mã' + 'thần thánh'
+- [uvn-47] split-name-boundary "Xét chứng chỉ" → 'Xét' + 'chứng chỉ'
+- [uvw-16315] split-name-boundary "Mendeleev chỉnh lý" → 'Mendeleev' + 'chỉnh lý'
+- [uvn-12038] split-name-boundary "Cơ quan chức năng" → 'Cơ' + 'quan chức năng'
+- [uvw-3820] split-name-boundary "Âm tức là" → 'Âm' + 'tức là'
+- [uvn-9015] split-name-boundary "Cơ quan chức năng" → 'Cơ' + 'quan chức năng'
+- [uvb-n-14607] split-name-boundary "Kỵ ghi nhớ" → 'Kỵ' + 'ghi nhớ'
+- [uvb-f-3703] split-name-boundary "Duy cố tình" → 'Duy' + 'cố tình'
+- [vlc-19841] split-name-boundary "Chủ sở hữu quyền" → 'Chủ' + 'sở hữu quyền'
+- [uvb-n-1525] split-name-boundary "Cao Lãnh hưởng ứng" → 'Cao Lãnh' + 'hưởng ứng'
+- ... and 38 more
 
 ### udd-ws-v1.1-test.txt
 
-- [uvb-n-19086] split "Phù duy bất khả thức" → 'Phù' + 'duy' + 'bất' + 'khả' + 'thức'
-- [vlc-17267] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-6223] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-9785] split "Viện kiểm sát nhân dân tối cao" → 'Viện kiểm sát' + 'nhân dân' + 'tối cao'
-- [vlc-16181] split "Bộ Tài nguyên và Môi trường" → 'Bộ' + 'Tài nguyên' + 'và' + 'Môi trường'
-- [vlc-12025] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-3574] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-10711] split "Viện kiểm sát nhân dân tối cao" → 'Viện kiểm sát' + 'nhân dân' + 'tối cao'
-- [uvn-11938] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-4389] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-1254] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-5002] split "luật tố tụng hình sự" → 'luật' + 'tố tụng hình sự'
-- [uvw-14472] split "nhà thiên văn vật lý" → 'nhà thiên văn' + 'vật lý'
-- [vlc-1994] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-8312] split "Tổng liên đoàn Lao động" → 'Tổng liên đoàn' + 'Lao động'
-- [uvw-15118] split "Đông Hoài quốc hoàng đế" → 'Đông' + 'Hoài' + 'quốc' + 'hoàng đế'
-- [uvw-15118] split "Hoàn Nhan A Cốt Đả" → 'Hoàn' + 'Nhan' + 'A' + 'Cốt' + 'Đả'
-- [vlc-4206] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [uvw-19136] split "sinh vật biến đổi gen" → 'sinh vật' + 'biến đổi gen'
-- [vlc-3581] split "Mặt trận T�� quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-1319] split "Đảng Cộng sản Việt Nam" → 'Đảng' + 'Cộng sản' + 'Việt Nam'
-- [vlc-1319] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-1319] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [uvw-156] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-6908] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- [vlc-2238] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-14701] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-14701] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-16929] split "Mặt trận Tổ quốc Việt Nam" → 'Mặt trận Tổ quốc' + 'Việt Nam'
-- [vlc-870] split "xã hội chủ nghĩa Việt Nam" → 'xã hội chủ nghĩa' + 'Việt Nam'
-- ... and 49 more
+- [uvw-12971] split-name-boundary "California nhân danh" → 'California' + 'nhân danh'
+- [uvn-6100] split-name-boundary "Hoàng đầu thú" → 'Hoàng' + 'đầu thú'
+- [uvw-16039] split-name-boundary "Thông hoảng hốt" → 'Thông' + 'hoảng hốt'
+- [uvw-9799] split-name-boundary "Môn-Khmer nguyên thủy" → 'Môn-Khmer' + 'nguyên thủy'
+- [uvw-8566] split-name-boundary "Hoàng thái tử" → 'Hoàng' + 'thái tử'
+- [uvb-n-17470] split-name-boundary "Văn mệt mỏi" → 'Văn' + 'mệt mỏi'
+- [uvw-8360] split-name-boundary "A di chuyển" → 'A' + 'di chuyển'
+- [uvn-350] split-name-boundary "Ba Đình đối đầu" → 'Ba Đình' + 'đối đầu'
+- [uvw-6709] split-name-boundary "Hàm tán xạ" → 'Hàm' + 'tán xạ'
+- [uvw-13962] split-name-boundary "Đông Âu vệ tinh" → 'Đông Âu' + 'vệ tinh'
+- [uvn-9573] split-name-boundary "Mangilal sở hữu" → 'Mangilal' + 'sở hữu'
+- [uvb-f-14820] split-name-boundary "Arya lặng lẽ" → 'Arya' + 'lặng lẽ'
+- [uvw-8943] split-name-boundary "Mỹ hứng chịu" → 'Mỹ' + 'hứng chịu'
+- [uvw-8943] split-name-boundary "Tennessee hứng chịu" → 'Tennessee' + 'hứng chịu'
+- [uvn-17072] split-name-boundary "Kim tiết lộ" → 'Kim' + 'tiết lộ'
+- [vlc-2624] split-name-boundary "Mạch tích hợp" → 'Mạch' + 'tích hợp'
+- [uvw-1478] split-name-boundary "Phản vật chất" → 'Phản' + 'vật chất'
+- [vlc-9115] split-name-boundary "Chủ sở hữu" → 'Chủ' + 'sở hữu'
+- [uvw-4004] split-name-boundary "Chu kỳ tính" → 'Chu' + 'kỳ tính'
+- [vlc-842] split-name-boundary "Hạ sĩ quan" → 'Hạ' + 'sĩ quan'
+- [uvb-f-15773] split-name-boundary "Wamba cố tình" → 'Wamba' + 'cố tình'
+- [uvn-12071] split-name-boundary "Kinh doanh số" → 'Kinh' + 'doanh số'
+- [uvw-14537] split-name-boundary "Đỗ Nguyễn trông nom" → 'Đỗ Nguyễn' + 'trông nom'
+- [vlc-12093] split-name-boundary "Chủ sở hữu quyền" → 'Chủ' + 'sở hữu quyền'
+- [uvn-2062] split-name-boundary "Ukraine nhượng bộ" → 'Ukraine' + 'nhượng bộ'
+- [uvb-f-2026] split-name-boundary "Lữ hoan hỉ" → 'Lữ' + 'hoan hỉ'
+- [uvn-19730] split-name-boundary "Jeong-hee cáo buộc" → 'Jeong-hee' + 'cáo buộc'
+- [uvb-f-12536] split-name-boundary "Lê chua chát" → 'Lê' + 'chua chát'
+- [uvb-n-6581] split-name-boundary "Tuy âm dương" → 'Tuy' + 'âm dương'
+- [uvb-f-10929] split-name-boundary "Vĩ cảm nhận" → 'Vĩ' + 'cảm nhận'
+- ... and 50 more

al_cycle2_top200.tsv

@@ -0,0 +1,201 @@
+rank	file	sent_idx	score	n_syllables	uncertainty	boundary_weight	n_long_tokens	max_uncertainty	text
+1	test	5312	0.127511	15	0.109295	0.1667	0	0.442459	Các hóa thạch Neanderthal 1 được cho là có niên đại 40.000 năm tuổi .
+2	test	4226	0.118466	19	0.076564	0.4167	1	0.438001	Dù vậy , thầy trò HLV Hoàng Tuấn Anh vấp phải hàng thủ được tổ chức tốt củaU .
+3	test	4592	0.117090	14	0.083636	0.4000	0	0.487245	Nó thậm chí vừa kỷ niệm 67 năm bay quanh hành tinh xanh .
+4	dev	1986	0.114202	10	0.091362	0.2500	0	0.497709	Văn sửng sốt : - Làm sao lại bỏ ?
+5	test	7993	0.107452	15	0.073089	0.3333	1	0.444471	Các Chúa Nguyễn về danh nghĩa là quan lại của Nhà Lê Trung hưng .
+6	test	2938	0.103645	15	0.089826	0.1538	0	0.466537	Nữ minh tinh nhíu mày : - Ra ngoài hồ sao lâu quá vậy ?
+7	test	4770	0.102785	22	0.070080	0.4667	0	0.485898	Người Nữ Chân đã trả thù cho vương quốc Bột Hải từng bị người Khiết Đan tiêu diệt vào năm xưa .
+8	dev	103	0.102390	18	0.090345	0.1333	0	0.453747	Ngài đã trao truyền cho Ma Ha Ca Diếp cái gì khi trao cho ngài cành hoa ?
+9	test	5515	0.102103	9	0.079413	0.2857	0	0.492459	Thế David đối xử tốt với chị chứ ?
+10	dev	9698	0.101365	23	0.082935	0.2222	0	0.385680	Các axit béo thường bão hòa hoặc không bão hòa đơn với độ dài chuỗi từ 16 đến 26 nguyên tử carbon .
+11	dev	1395	0.099803	8	0.087327	0.1429	0	0.362241	Mặt Trời có hạng quang phổ G2V .
+12	dev	4864	0.098360	17	0.063645	0.5455	0	0.385769	Thường sử dân vô tri vô dục , sử phù trí giả bất cảm vi dã .
+13	test	7109	0.098148	6	0.065432	0.5000	0	0.392534	Giải nghĩa : Duyệt dã .
+14	dev	3698	0.096770	12	0.072577	0.3333	0	0.486716	Chẳng lẽ cứ vua Lê , chúa Trịnh mãi thế này .
+15	test	6806	0.096738	11	0.061561	0.5714	0	0.480032	Nữ thần tới để đôn đốc tướng quân lựa chọn .
+16	dev	3159	0.096030	26	0.061459	0.5625	0	0.462741	Công nghệ nhận diện căn cước công dân gắn chip thông quaNFCgiúp quá trình xác thực danh tính diễn ra nhanh chóng và chính xác .
+17	test	6613	0.095808	17	0.078901	0.2143	0	0.463867	Nữ minh tinh bĩu môi : - Nhưng tôi lại không thích nói chuyện với anh !
+18	test	7577	0.095739	15	0.082973	0.1538	0	0.469382	Họ chửi rủa ỏm tỏi : - Tây vô đây làm gì đông quá !
+19	test	3804	0.095277	23	0.082284	0.1579	0	0.415389	Nhất Linh ĐÔI BẠN Chương 2 Phần I Bên ông tuần có mở tiệc thọ mừng cụ Bang , bà nội Dũng .
+20	test	7391	0.095185	13	0.073219	0.3000	0	0.398074	VN30 vượt 1.800 điểm ( Ảnh minh họa : Đăng Đức ) .
+21	test	6337	0.095036	15	0.082365	0.1538	0	0.470823	Đây là 5/6 con sông của Lục Đầu Giang cùng với sông Kinh Thầy .
+22	dev	8081	0.094295	8	0.070721	0.3333	0	0.461954	Tựa đề và phương pháp dịch 4.5 .
+23	dev	4386	0.092331	13	0.067361	0.2222	1	0.473245	Gần 30.500 du học sinh và hơn 110 trường đã phản hồi .
+24	dev	8691	0.092190	13	0.085098	0.0833	0	0.469826	Nhưng lúc San về , y bảo : - Anh liệu đấy !
+25	test	2469	0.092038	24	0.072030	0.2778	0	0.416939	Khi Nhất Tiếu Hồng Trần nhìn thấy con số bạo thương này , kinh ngạc tới mức không thốt ra được nửa câu .
+26	dev	7211	0.091950	17	0.059497	0.5455	0	0.442904	Huân chương Tự do hạng Nhất của Nhà nước Cộng hòa Dân chủ Nhân dân Lào .
+27	dev	6989	0.091855	11	0.066804	0.3750	0	0.397094	Y thành thực yêu nghề và yêu các trẻ em .
+28	dev	7942	0.091527	14	0.084990	0.0769	0	0.408987	Gã đang tán gì không biết nmà Kỳ Hương cười tươi như hoa .
+29	test	1577	0.091421	16	0.073137	0.2500	0	0.427447	Hoa tủi còn đâu duyên tác hợp , Mây bay rồi nữa giấc chiêm bao !
+30	dev	6168	0.090475	34	0.071848	0.2593	0	0.454712	Nghị Hách , mặc lòng mặc bộ áo trào vào ngày dạ tiệc , cũng chấp tay vái dài lưng cúi thật khom , mà rằng : - Bẩm lạy cụ lớn ạ .
+31	dev	5426	0.090424	12	0.067818	0.3333	0	0.451500	Holly lúc nào cũng ngưỡng mộ tài nấu nướng của mẹ .
+32	dev	7059	0.090244	12	0.075203	0.2000	0	0.467545	Văn vừa nói vừa ấn cái đồng hồ vào tay Minh .
+33	test	8734	0.090198	9	0.070154	0.2857	0	0.447857	Và tình ái là sợi dây vấn vít .
+34	test	1833	0.090018	11	0.057284	0.5714	0	0.451056	Định nghĩa Thông dịch Dịch đuổi Dịch song hành 7.2 .
+35	test	2251	0.089105	8	0.077967	0.1429	0	0.477627	Liêm rất ghét những kẻ vong bản .
+36	dev	8295	0.088628	17	0.062561	0.4167	0	0.457358	Cô ưu tiên những trang phục trẻ trung , quyến rũ tôn vòng eo thon gọn .
+37	dev	832	0.088105	17	0.072557	0.2143	0	0.487539	Một anh bạn tôi thôi học 20 năm rồi mà còn oán môn Địa chất học .
+38	test	7451	0.087897	39	0.050528	0.5417	1	0.428335	CỬA HÀNG TIỆN LỢI “ TIỆN LỢI ” NHỜ HỆ TH���NG LUÂN CHUYỂN HÀNG HÓA Chúng ta hãy cùng phân tích hoạt động mua hàng tại cửa hàng tiện lợi từ khía cạnh phí giao dịch liên quan .
+39	test	4041	0.087631	12	0.071698	0.2222	0	0.470650	Phía sau , Thái Mạo dẫn binh lính đuổi đến rồi .
+40	dev	796	0.087543	10	0.070034	0.2500	0	0.458788	Gió đêm mùa hạ sao lại lạnh thế này ?
+41	test	4435	0.087460	13	0.067277	0.3000	0	0.498480	Đương đọc , Quỳnh bỗng phải dừng lại để tay áp ngực .
+42	test	1926	0.086980	26	0.073598	0.1818	0	0.493773	Thủy Tiễn » « Vèo » , « vèo » - Vô ích thôi nhóc ạ , như thế không làm gì được ta đâu .
+43	test	8174	0.086778	17	0.076569	0.1333	0	0.446461	San cười phì một cái ra đằng mũi , rồi lại chúi đầu vào sách học .
+44	dev	6779	0.086491	19	0.077387	0.1176	0	0.474255	Tao mà cố tình hại mày , lát nữa xuống bơi , hà bá rút cẳng tao liền !
+45	test	8864	0.086474	16	0.054625	0.4000	1	0.426044	Triết gia đi đi lại lại trong phòng , bực mình vì bị bác bỏ .
+46	dev	1644	0.086101	10	0.047834	0.8000	0	0.468189	Viện trưởng Viện kiểm sát nhân dân tối cao .
+47	test	5290	0.085898	22	0.066376	0.2941	0	0.457848	Đội bóng xứ sen hồng có hành trình chinh phục tấm vé chính thức dự VCK toàn quốc rất thuyết phục .
+48	test	5925	0.085605	29	0.058089	0.4737	0	0.426702	Lý Công Uẩn lên ngôi Thế thứ các triều đại từ thời Hùng Vương đến hết TK X I. Thế thứ các triều đại thời sơ sử II .
+49	dev	2947	0.085583	12	0.057055	0.5000	0	0.342904	Niết Bàn là lý tưởng cao đẹp nhất của chúng sanh .
+50	test	5429	0.084998	18	0.056665	0.5000	0	0.460121	Thị dĩ thánh nhân xử thượng nhi dân bất trọng , xử tiền nhi dân bất hại .
+51	test	1210	0.084340	10	0.075906	0.1111	0	0.457209	Ngữ đoạn và cấu trúc của ngữ đoạn 2.1 .
+52	test	1107	0.084321	31	0.067457	0.2500	0	0.499115	Christopher Paolini Eragon - Cậu bé cưỡi rồng Tập 1 Thung lũng Palancar Sáng hôm sau , vầng dương ló dạng với những màu hồng và vàng tươi rực rỡ .
+53	test	4422	0.084027	21	0.071423	0.1765	0	0.433664	Mấy anh lính khố xanh toan giơ cao cái roi mây thì quan công sứ ra hiệu ngăn lại hỏi ...
+54	test	7130	0.083636	23	0.069090	0.2105	0	0.495195	Quẻ Thuần Khôn còn gọi là quẻ Khôn ( 坤 kūn ) , tức Đất là quẻ số 2 trong Kinh Dịch .
+55	dev	5861	0.083627	17	0.063950	0.3077	0	0.491722	Trong lùng bùng áo mưa , Khang nói , Khang muốn về nhà thay quần áo .
+56	test	9950	0.083516	48	0.058813	0.2500	1	0.483714	Nguyễn Đăng Tiến tước Quản Vũ hầu , tức Cai Gia ( theo Hoàng Lê nhất thống chí ) một tay " giặc già " Trung Quốc gốc người Việt Đông , Quảng Tây sang đầu quân làm thuộc hạ , tân khách dưới trướng Nguyễn Khản .
+57	test	838	0.083323	7	0.071420	0.1667	0	0.497610	Cay đắng chạy vào Malaysia 31 .
+58	test	4666	0.082789	21	0.051381	0.4167	1	0.465111	Ông từng giữ chức Chủ tịch và Tổng Thư ký của Tổng Liên đoàn Vovinam Việt Võ Đạo Thế giới .
+59	test	8882	0.082225	9	0.063953	0.2857	0	0.490249	Sao bà Hoàng lại nói như thế nhỉ ?
+60	test	8524	0.082209	31	0.059530	0.3810	0	0.384412	Ngẫm nghĩ giây lát , thị nữ bèn tâu : - Trình bệ hạ , quả là quan điện tiền Trần Thủ Độ có năng lui tới cung hoàng hậu .
+61	test	1619	0.081681	30	0.059900	0.3636	0	0.483023	Mọi người có thể tích lũy những vật thặng thừa và từ đó sự phân biệt giàu nghèo do sự lớn nhỏ mạnh yếu về lao động xuất hiện .
+62	dev	9085	0.081666	37	0.064009	0.2759	0	0.495617	Ứng dụng của ắc quy Các phương pháp phóng và nạp Nạp ắc quy lần thứ nhất Ắc quy mới lắp hoặc sau khi sửa chữa thay thế bản cực xong , phải nạp hình thành .
+63	dev	9841	0.081625	15	0.081625	0.0000	0	0.478545	Như người ta vẫn nói : “ Cho con mèo đồng xu Koban ” .
+64	dev	5095	0.081587	31	0.063164	0.2917	0	0.406860	Bản hồ địa giả thân hạ 本乎地者親下 vật nào gốc ở đất thì thân thuộc với cõi dưới mệnh người trông Thượng cửu Thượng cửu : kháng long hữu hối .
+65	test	2511	0.081424	15	0.069792	0.1667	0	0.415228	Cửu tam Hào từ : Lao khiêm , quân tử hữu chung , cát .
+66	test	9833	0.081417	23	0.066614	0.2222	0	0.445216	Đối quark là các hạt như " đối đỏ " , " đối xanh lam " , " đối xanh lơ " .
+67	dev	8121	0.081388	44	0.055582	0.4643	0	0.459157	Tuy nhiên , Khang vương Triệu Cấu của Bắc Tống tránh được nạn , đồng thời xưng đế ở Nam Kinh Quy Đức phủ ( nay thuộc Thương Khâu , Hà Nam ) , kiến quốc Nam Tống , tức Tống Cao Tông .
+68	dev	3069	0.081200	31	0.056000	0.4500	0	0.405153	Công Tử Phù Tô không chạy theo nàng , mà lại cùng Công Tử Liên Thành song chiến với con Linh Sơn Tuyết Hầu mắt đỏ lông trắng như tuyết .
+69	test	9339	0.081184	23	0.066424	0.2222	0	0.454714	Con người ốm yếu thế , ốm yếu từ những bệnh tật tích lũy trong năm , sáu , mười nghìn năm .
+70	test	7548	0.081175	20	0.064940	0.2500	0	0.336833	Nãy giờ ngồi làm thinh , Kỳ Nam bỗng tủm tim cưới khiến Kỳ Hương cau mày khó chịu .
+71	dev	4246	0.081055	8	0.060791	0.3333	0	0.221456	Truyện vị thần núi Hồng Lĩnh 31 .
+72	dev	4780	0.081043	31	0.072938	0.1111	0	0.482829	Ngô Phù Sai không giết con mình nhưng cất quân đi đánh Tề và hội chư hầu lần nữa , nước Việt đánh úp nước Ngô và giết chết Hữu !
+73	dev	2317	0.080775	9	0.053850	0.5000	0	0.279596	Đặc trưng của dịch thuật thính thị 6.2 .
+74	dev	2306	0.080758	9	0.060569	0.3333	0	0.384998	Lee Nguyễn Cầu thủ Hoa Kỳ gốc Việt .
+75	test	976	0.080685	12	0.060514	0.3333	0	0.263441	Thái độ trào lộng nhường chỗ cho châm biếm chua chát .
+76	test	3445	0.080647	15	0.069894	0.1538	0	0.474387	Trần căn phòng không có vết ẩm như người ta thấy khắp tòa nhà .
+77	dev	460	0.080567	38	0.063147	0.2759	0	0.485083	Cho nên trong tác phẩm “ Cung oán ngâm khúc ” của tác giả Ôn Như Hầu đã có câu : “ Phải duyên hương lửa cùng nhau , Xe dê lọ rắc lá dâu mới vào .
+78	test	6026	0.080384	21	0.064307	0.2500	0	0.438402	Diệt Thiên Tru Ma : " Đám chó Khải Hoàn kia , ta lập tức tới giết sạch bọn ngươi !
+79	dev	4155	0.079644	9	0.070794	0.1250	0	0.299169	Một ngày mùa đông tuyết phủ trắng xóa .
+80	test	5672	0.079348	12	0.066123	0.2000	0	0.488454	Nhờ công Lê Lai cứu chúa này mà Okada thoát nạn .
+81	dev	2368	0.079296	14	0.079296	0.0000	0	0.428298	Cau tiến sát tấm cửa liếp , cởi giỏ cá đeo bên hông .
+82	test	3631	0.079109	36	0.061027	0.2963	0	0.466549	Anh trai Đường Minh nằm trên giường , vẻ mặt từa tựa Trác Mộc Cường Ba lúc nhác thấy tấm ảnh , cứ ngây ngây dại dại nhìn lên bức tường dán chi chít ảnh .
+83	dev	1610	0.078465	30	0.062231	0.2609	0	0.461420	Thời niên thiếu Năm Quý Mão ( 1783 ) , Nguyễn Du thi Hương ở trường Sơn Nam , đậu Tam trường ( Sinh đồ ) lúc 18 tuổi .
+84	test	7826	0.078035	37	0.050540	0.3462	1	0.462731	Không một ông nào trả lời vì có lẽ các ông đại diện ấy đến lúc ấy đều bán tín bán nghi không biết ăn chuột bao tử như thế là văn minh hay man dã .
+85	dev	148	0.078000	25	0.058500	0.3333	0	0.359546	Bước vào đại học , nền tảng tích lũy từ các kỳ thi học sinh giỏi giúp Lâm chủ động sắp xếp bài vở .
+86	dev	1030	0.077849	14	0.066727	0.1667	0	0.458599	Anh tập ngậm đũa để ăn uống , lật sách và gõ phím .
+87	dev	5952	0.077848	9	0.051899	0.5000	0	0.433390	Tác giả chịu ơn độc giả nhiều lắm .
+88	dev	4228	0.077757	14	0.066649	0.1667	0	0.371469	Qua một cánh đồng ngót hai cây số mới thấy đường cái quan .
+89	dev	3394	0.077748	9	0.060471	0.2857	0	0.401418	Tao chơi bời không phải vô mục đích .
+90	test	7668	0.077733	20	0.058300	0.3333	0	0.334984	Khi ở nhà chồng , Thị Kính giữ phận làm dâu , tôn kính phụng dưỡng bố mẹ chồng .
+91	dev	1011	0.077532	36	0.060302	0.2857	0	0.427724	Tiêu Tiêu lườm Tĩnh Chi bằng ánh mắt bất lực rồi đáp với giọng chưa hết tức giận , “ này , cậu ngốc thật hay cố ý giả bộ trong sáng ngây thơ thế ?
+92	dev	6341	0.077531	33	0.060571	0.2800	0	0.448857	Đà Lạt chiều em gió gác trăng thềm Một chút sương mơ ảo huyền mộng mị Một chút tơ vương rối hồn thi sĩ Gỡ giùm ta giăng mắc , hỡi em ..
+93	dev	9723	0.077383	11	0.056279	0.3750	0	0.372973	Suy luận gièm pha Đánh rắn phải đánh vào đầu .
+94	dev	269	0.077355	21	0.050281	0.5385	0	0.387546	Tôi bị xử phạt vi phạm hành chính liên quan đến an toàn vệ sinhthực phẩmkhi kinh doanh quán ăn .
+95	test	4780	0.077185	9	0.060033	0.2857	0	0.491836	Như vậy mà cậu không thấy mệt sao ?
+96	dev	3745	0.077038	10	0.061630	0.2500	0	0.280040	Đề thi Đẫm Máu Tác giả : Lôi Mễ .
+97	dev	959	0.076946	16	0.056427	0.3636	0	0.454889	Cộng hòa Séc là quốc gia đa đảng theo chế độ cộng hòa đại nghị .
+98	test	3777	0.076745	22	0.041861	0.8333	0	0.375912	Ngày nay chúng ta tạo nhân ngày mai chúng ta thọ quả , thực là rõ ràng rành mạch biết bao .
+99	dev	4329	0.076731	19	0.068205	0.1250	0	0.490920	Dưới ánh đèn măng-sông sáng lóe , lúc ấy tôi mới ngắm kỹ dung mạo ông Cai Móm .
+100	test	3819	0.076712	14	0.059009	0.3000	0	0.438582	Tình yêu của họ là một thứ tình yêu lệ thuộc kẻ khác .
+101	test	9664	0.076650	15	0.060225	0.2727	0	0.449106	Thời đó ai ai cũng rõ Chu Du khôi ngô “ tuấn tú ” .
+102	dev	7839	0.076608	15	0.071500	0.0714	0	0.442286	Kì lạ là học sinh mới vẩn tỉnh bơ ăn phần trưa của mình .
+103	test	4187	0.076496	8	0.066934	0.1429	0	0.322865	Maurice và Adolphe sợ quá phát khóc ...
+104	test	5805	0.076324	9	0.047702	0.6000	0	0.396684	A. Định nghĩa B . Phân loại 1 .
+105	dev	2378	0.076312	18	0.059353	0.2857	0	0.362455	Mỗi làm gió thoảng qua đều được ướp hương hoa loa kèn thơm ngát và dịu ngọt .
+106	dev	6866	0.076010	10	0.068409	0.1111	0	0.428811	Khổng Tử bèn rời nước Sở về nước Vệ .
+107	dev	9325	0.075954	12	0.050636	0.5000	0	0.370414	Nghiên cứu mới công bố trên tạp chí quốc tếNature Communication .
+108	dev	4434	0.075877	9	0.066393	0.1429	0	0.343580	Cái gì là lạ nhoi nhoi trong tôi ?
+109	test	6429	0.075794	8	0.066320	0.1429	0	0.495409	Em nào giặt đồ mà chẳng cần ...
+110	dev	2874	0.075443	30	0.057233	0.3182	0	0.439693	Trương Phúc Phan dùng kế trá hàng , tuyển mộ 15 người Chà Và ( có gốc gác từ Malacca ) ra Côn Đảo làm thuê cho quân Anh .
+111	dev	9803	0.075373	26	0.060879	0.2381	0	0.356382	Nhà vua cầm sừng tê văn dài bảy tấc mà đi xuống biển , đại để cũng như nói cầm sừng tê đi xuống nước .
+112	dev	7431	0.075365	31	0.060778	0.2400	0	0.480767	Tên gọi Sào Nam ( 巢南 ) được lấy từ câu " 越鳥巢南枝 [ Việt điểu sào nam chi , nghĩa là Chim Việt làm tổ cành Nam ] " .
+113	dev	3905	0.075328	10	0.052729	0.4286	0	0.292166	Kể ra công sở tiêu tiền rất phí phạm .
+114	dev	2765	0.075259	26	0.054186	0.3889	0	0.354216	Trác Mộc Cường Ba kéo tay người phụ nữ , chỉ vào Phương Tân đứng xa xa nói : ‘ A ma , đô na !
+115	dev	8323	0.075191	18	0.058482	0.2857	0	0.430440	Mạnh Phàm Triết kêu lên thất thanh , rồi cùng Phương Mộc ngã nhào xuống sân thượng .
+116	dev	7668	0.075161	52	0.037174	0.4839	2	0.471260	Một số di tích tiêu biểu tại khu vực Óc Eo – Ba Thê đã được khai quật và bảo tồn bao gồm : Nam Linh Sơn Tự , Gò Cây Me , Gò Út Trạnh , Gò Óc Eo , Gò Cây Thị A và B , Gò Giồng Cát ...
+117	test	3268	0.075130	16	0.070434	0.0667	0	0.407115	Sự thực , em cũng đã yêu vụng giấu thầm anh trong bao nhiêu lâu !
+118	test	8948	0.074990	30	0.059992	0.2500	0	0.498176	Vị tướng già đứng nhìn người chép sách trẻ tuổi và hóm hỉnh nheo mắt : - Cháu sẽ thu xếp khăn gói lên đường trong tuần trăng này .
+119	test	1651	0.074973	13	0.068725	0.0909	0	0.381690	Phương Tân nghi hoặc trong lòng , lộ ra cả nét mặt .
+120	dev	8449	0.074818	37	0.058192	0.2857	0	0.462078	Ngoài khối hiệp thông chính , còn tồn tại các giáo hội ly khai theo lịch cũ ( Chính thống giáo Cựu lịch ) , hoặc nghi thức cũ ( Tín hữu Cựu nghi thức ) .
+121	dev	6243	0.074749	50	0.049595	0.3056	1	0.443688	Đến lúc này , Liêu Thiên Tộ Đế mới xem trọng sự việc , đồng thời hạ lệnh thân chinh , song quân Liêu bị quân Kim đánh bại , còn nội bộ triều Liêu lại xảy ra việc Da Luật Chương Nô và Cao Vĩnh Xương làm phản .
+122	dev	9296	0.074705	15	0.059764	0.2500	0	0.490440	Kỳ Nam bật cười : - Có cái chuyên môn như vậy nữa sao ?
+123	dev	9421	0.074615	20	0.059692	0.2500	0	0.482721	Mặc dầu vậy , chúng vẫn tông cửa xông vào , rất ít gia đình Do Thái tránh được .
+124	test	7378	0.074589	20	0.063401	0.1765	0	0.482855	Đọc dò và đọc lướt theo mục đích Đọc dò Bài luyện tập Đọc lướt Bài luyện tập 7 .
+125	dev	5276	0.074495	20	0.063321	0.1765	0	0.456219	Kinh sách thường dạy : Phật và chúng sanh tánh thường rỗng lặng , chính là nghĩa như vậy .
+126	test	9085	0.074444	16	0.051180	0.4545	0	0.463774	Một thể chất bạc nhược không bao giờ có đặng một tinh thần dũng mãnh .
+127	dev	6163	0.074312	34	0.063383	0.1724	0	0.488757	Tỉ dụ như câu này , ở chương 2 : “ Hữu vô tương SANH , “ Nan dị tương THÀNH , “ Trường đoản tương HÌNH , “ Cao hạ tương KHUYNH .
+128	test	6627	0.074230	12	0.055673	0.3333	0	0.433737	Kẻ tả hữu thưa : “ Ấy là Dương Tu ” .
+129	test	6221	0.074173	36	0.055629	0.3333	0	0.429212	Sự kiện 908 – Một năm sau khi tiếm vị , Hậu Lương Thái Tổ cho hạ độc giết chết Lý Chúc , tức Đường Ai Đế , hoàng đế cuối cùng của triều Đường .
+130	dev	5220	0.074140	14	0.063549	0.1667	0	0.453948	Trong tự nhiên và văn hóa Tập tin : Color icon pink v2 .
+131	dev	1481	0.074068	10	0.057608	0.2857	0	0.388282	Khả trinh 可貞 – có thể giữ bền được .
+132	dev	8916	0.074041	8	0.064786	0.1429	0	0.446816	Hương già mà bày đặt trông mẹ .
+133	test	9576	0.073958	53	0.055817	0.3250	0	0.444774	Làm như thế cốt để mưu nghiệp lớn , chọn ở chỗ giữa , làm kế cho con cháu muôn vạn đời , trên kính mệnh trời , dưới theo ý dân , nếu có chỗ tiện thì dời đổi , cho nên vận nước lâu dài , phong tục giàu thịnh .
+134	test	7450	0.073856	13	0.051131	0.4444	0	0.374198	Tuy thế , thỉnh thoảng chàng cũng làm được việc giá trị .
+135	test	4585	0.073824	17	0.056454	0.3077	0	0.444112	Thế Lữ hoan hỉ tuyên bố : Từ nay , chúng ta đã có Xuân Diệu .
+136	dev	5341	0.073682	29	0.063156	0.1667	0	0.410964	Đại Bình nguyên nằm ở trung bộ Texas , trải dài từ Vùng Cán xoong của tiểu bang và Llano Estacado cho đến Vùng Đồi Texas gần Austin .
+137	dev	4170	0.073531	10	0.051472	0.4286	0	0.481334	Nguyên phệ , nguyên vĩnh trinh , vô cữu .
+138	test	8058	0.073486	15	0.063688	0.1538	0	0.379196	Trước mắt anh , những luống cúc mâm xôi vàng rực đã bung nở .
+139	dev	947	0.073358	24	0.051962	0.4118	0	0.431265	Nó ra đời như một liệu pháp y tế để kiểm soát các cơn động kinh ở bệnh nhân động kinh kháng thuốc .
+140	dev	6672	0.073331	8	0.054998	0.3333	0	0.437901	Ba chàng trai ngơ ngác nhìn nhau .
+141	dev	9005	0.073292	13	0.061077	0.2000	0	0.441073	Khi ông tạ thế , vua Minh Mạng phong Lương năng bá .
+142	test	6982	0.073233	12	0.067130	0.0909	0	0.466531	Chị Nga đứng im , tròn xoe mắt , không hiểu .
+143	test	195	0.073208	14	0.052292	0.4000	0	0.451739	Đành phải phó mặc mọi sự , chống đỡ bây giờ vô hiệu .
+144	test	7397	0.073016	33	0.042047	0.2609	2	0.414529	Đế Lai bèn để ái nữ là Âu Cơ và các thị tỳ ở lại nơi hành tại rồi đi chu du thiên hạ , ngắm các nơi danh lam thắng cảnh .
+145	dev	8652	0.072991	13	0.061762	0.1818	0	0.432060	Không bao giờ sự vật hé ra một tý cho họ thấy .
+146	dev	344	0.072935	39	0.049903	0.4615	0	0.475066	Nơi nguyên quán Võ Tánh là Gò Công , ấp Gò Tre có hai câu đối ghi công ơn họ Võ như sau : “ Khổng Tước kỳ , khẳng khái Cần vương , tam hùng thủ liệt .
+147	dev	4097	0.072916	17	0.060049	0.2143	0	0.418667	Tôi không bao giờ có ý đùa cụ , nhưng đã có lần dò ý cụ .
+148	test	3199	0.072866	19	0.069031	0.0556	0	0.379981	Con chó đi rồi , tríêt gia duỗi thẳng chân về phía lò sưởi đang cháy tí tách .
+149	test	4693	0.072819	16	0.054615	0.3333	0	0.411808	Sách có câu : " Tinh thần minh mẩn trong thân thể tráng kiện " .
+150	test	1162	0.072775	9	0.064689	0.1250	0	0.346305	Tiếng lọach xọach kéo chốt rồi cửa mở .
+151	test	211	0.072616	12	0.059413	0.2222	0	0.330733	Cây lương thực chính vùng này là lúa nước và ngô .
+152	test	2715	0.072512	12	0.060427	0.2000	0	0.407835	Hương thơm ngọt ngào của món cranachan thoảng vào từ bếp .
+153	test	427	0.072379	18	0.060315	0.2000	0	0.475350	Lão Sloan cúi nhìn chăm chú , miệng lẩm bẩm : - Chắc đồ chôm chỉa đây .
+154	dev	17	0.072159	9	0.054119	0.3333	0	0.461577	Triệu Việt Vương đã chết như thế nào ?
+155	test	7331	0.072123	35	0.053577	0.3462	0	0.488467	Theo Hoài Thanh trong Thi nhân Việt Nam : “ Độ ấy Thơ Mới vừa ra đời , Thế Lữ như vừng sao đột hiện ánh sáng chói khắp cả trời thơ Việt Nam .
+156	test	7856	0.072076	16	0.063066	0.1429	0	0.445630	Còn gì bằng được ông Chánh mật thám gọi đến chữa cho con gái nữa .
+157	dev	4300	0.071919	18	0.051941	0.3846	0	0.492143	Rồi những ngày tháng cứ Mưa dầm thấm từ từ Nàng rằng anh quyết tử Yêu yêu ...
+158	dev	8322	0.071744	39	0.058867	0.2188	0	0.457048	Đao Đao Kiến Huyết nhào lên điên cuồng tấn công , trong nháy mắt thò tay vào bọc lấy hồng dược uống thêm , nhưng vẫn không chịu nổi một chiêu bạo kích của Ma Tiêu Dao kia .
+159	dev	2786	0.071588	13	0.060575	0.1818	0	0.477334	Thủ cựu đợi thời chi tượng : giữ mức cũ đợi thời .
+160	dev	6588	0.071568	13	0.071568	0.0000	0	0.398594	Sao mày đốn thế , mày để tao gào rát cả cổ !
+161	dev	1281	0.071485	9	0.055599	0.2857	0	0.343125	Đồng hồ báo xăng chỉ ba phần tư .
+162	test	3149	0.071397	8	0.071397	0.0000	0	0.290680	Để đầu trần để tránh lửa bén .
+163	test	6835	0.071345	16	0.057968	0.2308	0	0.475611	Có thể vì Tuyền thích gương mặt Khang , với mũi dài và mắt xếch .
+164	test	7572	0.071192	15	0.060239	0.1818	0	0.381069	Trong Doanh hoàn chí lược ( 1849 ) , Từ Kế Dư viếtː " ...
+165	test	7554	0.071139	11	0.071139	0.0000	0	0.493656	Súng nổ rền trời … tôi nói dóc làm gì .
+166	dev	6613	0.071066	36	0.048731	0.4583	0	0.386043	Đến thế kỷ 16 , đại lãnh chúa Hideoshi phái một hạm đội mở cuộc viễn chinh đầu tiên tiến chiếm Cao Ly , nhưng hạm đội Nhật bị hải quân Cao Ly đánh tan .
+167	dev	429	0.071032	12	0.064575	0.1000	0	0.382149	Một cơn gió mạnh đem theo hương thơm ngào ngạt hơn .
+168	test	4080	0.070979	23	0.049376	0.4375	0	0.489624	Xét nghiệm định lượng lactat cao 6,58 cho thấy người bệnh bị sốc nhiễm khuẩn khiến huyết áp tụt sâu 78/45 mmHg .
+169	dev	2851	0.070935	29	0.044823	0.1364	2	0.464640	Một người khác xưng mình là “ Tề Thiên Đại Thánh ” và thường gây hấn với một người khác xưng là “ Nam Hải Quán Âm ” .
+170	dev	6536	0.070934	30	0.056258	0.2609	0	0.399172	Tinh thần trí não họ như của đứa trẻ thơ , mừng giận thương vui đều phát hiện liền bằng những cử động sát một bên với xúc động .
+171	dev	9147	0.070888	9	0.055135	0.2857	0	0.402720	Bắc Việt của một ngày xa xưa ơi !
+172	dev	623	0.070861	21	0.055943	0.2667	0	0.481031	Những thiên Hà Cừ Thư , Bình Chuẩn Thư , viết với nhãn lực của một nhà kinh tế học .
+173	test	1974	0.070815	12	0.064913	0.0909	0	0.424932	Nếu thế mà vẫn không được thì đành chịu vậy ” .
+174	dev	6195	0.070735	67	0.049300	0.4348	0	0.477989	Nguồn gốc của Kim Vân Kiều truyện Theo Trần Đình Sử trong Thi pháp Truyện Kiều , Kim Vân Kiều truyện đã qua sáu lần biên soạn lại : Bản ghi chép sớm nhất về sự tích Thúy Kiều - Từ Hải là Kỷ tiễu trừ Từ Hải bản mạt ( Ghi chép đầu đuôi chuyện dẹp trừ Từ Hải ) của Mao Khôn đời Minh .
+175	test	4908	0.070677	10	0.056541	0.2500	0	0.421126	Ngoài xấp tiền giấy còn một đống tiền hào .
+176	dev	7978	0.070414	15	0.061026	0.1538	0	0.470716	Cúc Hương nháy mắt : - Nói vậy chứ anh để tụi này trả .
+177	test	7782	0.070395	9	0.062573	0.1250	0	0.311407	Nam nhíu mày : - Ông ngoại nào ?
+178	test	4248	0.070362	14	0.050259	0.4000	0	0.394092	Làm sao thần có thể an toàn được nếu định lừa bệ hạ ?
+179	test	2394	0.070320	18	0.066413	0.0588	0	0.432113	Dù phải lội sình và bị muỗi cắn , đỉa bám thì cũng nên đi cho biết .
+180	test	8544	0.070090	23	0.057346	0.2222	0	0.497318	Nền Đệ Ngũ cộng hòa dưới quyền Charles de Gaulle được thành lập vào năm 1958 và tồn tại cho đến nay .
+181	dev	4679	0.069849	13	0.059103	0.1818	0	0.496467	Cúc Hương lên tiếng : - Con Thục nói vậy không được .
+182	test	6695	0.069797	9	0.054287	0.2857	0	0.208383	Hành lang hôi mùi cải luộc chiếu nát .
+183	test	4783	0.069674	19	0.047671	0.4615	0	0.263335	Ba hiệp đầu chứng kiến thế trận cân bằng , hai đội liên tục ăn miếng trả miếng .
+184	dev	4754	0.069618	23	0.051457	0.3529	0	0.409701	Ngày mai đây , người ta sẽ lấy đi sản nghiệp suốt mấy mươi năm khó nhọc gầy dựng của ba mẹ .
+185	dev	912	0.069618	17	0.049142	0.4167	0	0.470716	Cúc Hương nháy mắt : - Mày để ý anh chàng Phong Khê kia chứ gì ?
+186	dev	3578	0.069538	44	0.052153	0.3333	0	0.447925	Năm 1533 , Nguyễn Kim đón con trai của Lê Chiêu Tông tên Lê Ninh , lập làm vua tức vua Lê Trang Tông , nhờ công ấy ông được phong làm Thượng phụ thái sư Hưng quốc công chưởng nội ngoại sự .
+187	test	5033	0.069520	21	0.066209	0.0500	0	0.408108	Đêm nay chim ngủ đâu Mà đêm nay hồn anh ở đâu Này chú chim xanh Bao giờ báo tiệp ?
+188	dev	5163	0.069380	13	0.063598	0.0909	0	0.412497	Cô Tư Hạnh tốc nóp , dụi mắt : - Nó đâu ?
+189	dev	8998	0.069294	21	0.051970	0.3333	0	0.472209	Tổng quan Ngân Hà bản chất là một thiên hà xoắn ốc chặn ngang kiểu SBbc theo phân loại Hubble .
+190	dev	5623	0.069076	22	0.059657	0.1579	0	0.441546	Tiến yêu Lãng như em ruột , vợ Tiến còn bé tí cũng bắt chước chồng , coi Lãng như em .
+191	dev	8855	0.069027	28	0.049189	0.2000	1	0.458044	Bài chia sẻ của Đình Bắc ẢNH CHỤP MÀN HÌNH Đình Bắc vừa trải qua giải đấu chói sáng tại VCK U . 23 châu Á 2026 .
+192	dev	6700	0.068847	30	0.054094	0.2727	0	0.414839	CAO VĂN LUẬN Viện trưởng viện Đại Học Huế Chu Tử SỐNG Phần I Chương 1 Huyền hốt hoảng gọi với theo người thanh niên : - Anh Thịnh !
+193	dev	3915	0.068588	24	0.059642	0.1500	0	0.425539	Tống Mẫn Công nói : “ Ta với Nam Cung Trường Vạn là chỗ thân nhau lắm , cần gì điều ấy ” .
+194	test	2710	0.068587	18	0.060967	0.1250	0	0.384732	U mày đau bụng , kêu rối rít lên , làm tao cũng quên bẵng đi mất ...
+195	test	29	0.068497	13	0.047421	0.4444	0	0.338639	Lời tâu của Lý Tiến , Lý Cầm và Trương Trọng 13 .
+196	dev	8974	0.068337	18	0.040107	0.4545	1	0.467712	Ngày nay , Thụy Điển là một nước quân chủ lập hiến với thể chế đại nghị .
+197	test	8657	0.068301	20	0.048212	0.4167	0	0.487623	Chiêu Quốc vương Trần Ích Tắc , trấn thủ lộ này , mời ông đến làm chủ hội vật .
+198	dev	6791	0.068133	36	0.056777	0.2000	0	0.459344	Hệ Mặt Trời nằm ở mặt trong của Cánh tay Orion - một cấu trúc hình xoắn ốc chứa đầy bụi và khí gas , cách tâm quay Galactic Center khoảng 26,000 năm ánh sáng .
+199	test	4537	0.068118	13	0.052399	0.3000	0	0.403474	Chiều chiều nghe vượn hú , Hoa lá rụng , buồn buồn .
+200	dev	3912	0.067944	32	0.054355	0.2500	0	0.426858	Căn cứ năm 627 , sứ giả Phù Nam còn đến tiến cống nhà Đường , nên có thể suy ra nước Phù Nam bị tiêu diệt phải sau năm này .

gold_ws_cycle1.txt

@@ -0,0 +1,1782 @@
+# sent_id = uvb-f-10324
+# text = Tiếng nước đập ì ầm vang trong không gian .
+Tiếng	B-W
+nước	B-W
+đập	B-W
+ì	B-W
+ầm	I-W
+vang	B-W
+trong	B-W
+không	B-W
+gian	I-W
+.	B-W
+
+# sent_id = uvb-f-10494
+# text = Văng vẳng đâu đây dạo khúc Ánh Trăng của Beethoven .
+Văng	B-W
+vẳng	I-W
+đâu	B-W
+đây	I-W
+dạo	B-W
+khúc	B-W
+Ánh	B-W
+Trăng	I-W
+của	B-W
+Beethoven	B-W
+.	B-W
+
+# sent_id = uvb-f-1078
+# text = Mặt Kỳ Nam xụ xuống mất hứng .
+Mặt	B-W
+Kỳ	B-W
+Nam	I-W
+xụ	B-W
+xuống	B-W
+mất	B-W
+hứng	I-W
+.	B-W
+
+# sent_id = uvb-f-11528
+# text = Trong chuyện của Homer thần linh giữ vai trò đáng kể .
+Trong	B-W
+chuyện	B-W
+của	B-W
+Homer	B-W
+thần	B-W
+linh	I-W
+giữ	B-W
+vai	B-W
+trò	I-W
+đáng	B-W
+kể	I-W
+.	B-W
+
+# sent_id = uvb-f-116
+# text = Thân không là lính thú Sao không về cố hương ?
+Thân	B-W
+không	B-W
+là	B-W
+lính	B-W
+thú	I-W
+Sao	B-W
+không	B-W
+về	B-W
+cố	B-W
+hương	I-W
+?	B-W
+
+# sent_id = uvb-f-12571
+# text = Anh không nỡ tẩy chay em chút nào .
+Anh	B-W
+không	B-W
+nỡ	B-W
+tẩy	B-W
+chay	I-W
+em	B-W
+chút	B-W
+nào	B-W
+.	B-W
+
+# sent_id = uvb-f-13184
+# text = Ta thấy Odysseus phải ra tay đối phó với tình thế một mất một còn .
+Ta	B-W
+thấy	B-W
+Odysseus	B-W
+phải	B-W
+ra	B-W
+tay	I-W
+đối	B-W
+phó	I-W
+với	B-W
+tình	B-W
+thế	I-W
+một	B-W
+mất	I-W
+một	I-W
+còn	I-W
+.	B-W
+
+# sent_id = uvb-f-14791
+# text = Vậy mà tao phóng xe như điên sợ trễ .
+Vậy	B-W
+mà	I-W
+tao	B-W
+phóng	B-W
+xe	B-W
+như	B-W
+điên	I-W
+sợ	B-W
+trễ	B-W
+.	B-W
+
+# sent_id = uvb-f-15743
+# text = Say sưa tự tôn con người coi mình như thần linh .
+Say	B-W
+sưa	I-W
+tự	B-W
+tôn	I-W
+con	B-W
+người	I-W
+coi	B-W
+mình	B-W
+như	B-W
+thần	B-W
+linh	I-W
+.	B-W
+
+# sent_id = uvb-f-16030
+# text = Quỳnh nghiến răng , rủa thầm .
+Quỳnh	B-W
+nghiến	B-W
+răng	I-W
+,	B-W
+rủa	B-W
+thầm	I-W
+.	B-W
+
+# sent_id = uvb-f-16202
+# text = Nữ minh tinh khịt mũi : - Con ễnh ương chứ ai !
+Nữ	B-W
+minh	I-W
+tinh	I-W
+khịt	B-W
+mũi	I-W
+:	B-W
+-	B-W
+Con	B-W
+ễnh	B-W
+ương	I-W
+chứ	B-W
+ai	B-W
+!	B-W
+
+# sent_id = uvb-f-16922
+# text = Cecelia Ahern PS I LOVE YOU Chương 4 PS I LOVE YOU - SỨC MẠNH TÌNH YÊU Vào hôm sinh nhật Holly đứng trước chiếc gương lớn ngắm mình .
+Cecelia	B-W
+Ahern	B-W
+PS	B-W
+I	B-W
+LOVE	B-W
+YOU	B-W
+Chương	B-W
+4	B-W
+PS	B-W
+I	B-W
+LOVE	B-W
+YOU	I-W
+-	B-W
+SỨC	B-W
+MẠNH	I-W
+TÌNH	I-W
+YÊU	I-W
+Vào	B-W
+hôm	B-W
+sinh	B-W
+nhật	I-W
+Holly	B-W
+đứng	B-W
+trước	B-W
+chiếc	B-W
+gương	B-W
+lớn	B-W
+ngắm	B-W
+mình	B-W
+.	B-W
+
+# sent_id = uvb-f-17901
+# text = Bọn gia nhân túc trực rất đông , người vận quần áo khá sang trọng , kẻ bận sơ sài tầm thường như anh chăn lợn Gurth .
+Bọn	B-W
+gia	B-W
+nhân	I-W
+túc	B-W
+trực	I-W
+rất	B-W
+đông	B-W
+,	B-W
+người	B-W
+vận	B-W
+quần	B-W
+áo	I-W
+khá	B-W
+sang	B-W
+trọng	I-W
+,	B-W
+kẻ	B-W
+bận	B-W
+sơ	B-W
+sài	I-W
+tầm	B-W
+thường	B-W
+như	B-W
+anh	B-W
+chăn	B-W
+lợn	I-W
+Gurth	B-W
+.	B-W
+
+# sent_id = uvb-f-18098
+# text = Khải nheo nheo mắt : - Nhớ nghen !
+Khải	B-W
+nheo	B-W
+nheo	I-W
+mắt	B-W
+:	B-W
+-	B-W
+Nhớ	B-W
+nghen	B-W
+!	B-W
+
+# sent_id = uvb-f-19929
+# text = Anh ta chỉ là một sự xao lãng thôi .
+Anh	B-W
+ta	B-W
+chỉ	B-W
+là	B-W
+một	B-W
+sự	B-W
+xao	B-W
+lãng	I-W
+thôi	B-W
+.	B-W
+
+# sent_id = uvb-f-2195
+# text = Em nhớ về hởi người thương Bờ xe nước tạc thơi gian đó Ai cắt cớ nói chi bao điều khó Đá La Hà vẫn trỗi mọc còn kia .
+Em	B-W
+nhớ	B-W
+về	B-W
+hởi	B-W
+người	B-W
+thương	I-W
+Bờ	B-W
+xe	I-W
+nước	I-W
+tạc	B-W
+thơi	B-W
+gian	I-W
+đó	B-W
+Ai	B-W
+cắt	B-W
+cớ	I-W
+nói	B-W
+chi	B-W
+bao	B-W
+điều	B-W
+khó	B-W
+Đá	B-W
+La	I-W
+Hà	I-W
+vẫn	B-W
+trỗi	B-W
+mọc	B-W
+còn	B-W
+kia	B-W
+.	B-W
+
+# sent_id = uvb-f-3070
+# text = Xin chào sĩ quan quân kỳ Iago !
+Xin	B-W
+chào	I-W
+sĩ	B-W
+quan	I-W
+quân	B-W
+kỳ	I-W
+Iago	B-W
+!	B-W
+
+# sent_id = uvb-f-5207
+# text = Nhiệm bán tín bán nghi : - Tâm lý gì lạ vậy ?
+Nhiệm	B-W
+bán	B-W
+tín	I-W
+bán	I-W
+nghi	I-W
+:	B-W
+-	B-W
+Tâm	B-W
+lý	I-W
+gì	B-W
+lạ	B-W
+vậy	B-W
+?	B-W
+
+# sent_id = uvb-f-6015
+# text = Liên xua tay mắng yêu : − Mình chỉ nói gở !
+Liên	B-W
+xua	B-W
+tay	I-W
+mắng	B-W
+yêu	I-W
+:	B-W
+−	B-W
+Mình	B-W
+chỉ	B-W
+nói	B-W
+gở	I-W
+!	B-W
+
+# sent_id = uvb-f-6292
+# text = Trương Hán Siêu bưng một khay trà đặt lên án .
+Trương	B-W
+Hán	I-W
+Siêu	I-W
+bưng	B-W
+một	B-W
+khay	B-W
+trà	I-W
+đặt	B-W
+lên	B-W
+án	B-W
+.	B-W
+
+# sent_id = uvb-f-6893
+# text = GỬI CỐ NHÂN Nguyễn Bính Mưa dầm gió bấc cố nhân ơi !
+GỬI	B-W
+CỐ	B-W
+NHÂN	I-W
+Nguyễn	B-W
+Bính	I-W
+Mưa	B-W
+dầm	B-W
+gió	B-W
+bấc	I-W
+cố	B-W
+nhân	I-W
+ơi	B-W
+!	B-W
+
+# sent_id = uvb-f-7016
+# text = Nhỡ Nhàng Nguyễn Bính Công tôi xe chỉ vót nan Phất diều mướn gió nơi nàng thả chơi Nỡ nào tắt gió nàng ơi !
+Nhỡ	B-W
+Nhàng	I-W
+Nguyễn	B-W
+Bính	I-W
+Công	B-W
+tôi	B-W
+xe	B-W
+chỉ	B-W
+vót	B-W
+nan	B-W
+Phất	B-W
+diều	B-W
+mướn	B-W
+gió	B-W
+nơi	B-W
+nàng	B-W
+thả	B-W
+chơi	B-W
+Nỡ	B-W
+nào	B-W
+tắt	B-W
+gió	B-W
+nàng	B-W
+ơi	B-W
+!	B-W
+
+# sent_id = uvb-f-7435
+# text = Mặt San đung đưa và tiếng y dơn dớt .
+Mặt	B-W
+San	B-W
+đung	B-W
+đưa	I-W
+và	B-W
+tiếng	B-W
+y	B-W
+dơn	B-W
+dớt	I-W
+.	B-W
+
+# sent_id = uvb-f-7705
+# text = Phó hương quản sao lại chứa chấp hắn ?
+Phó	B-W
+hương	B-W
+quản	I-W
+sao	B-W
+lại	B-W
+chứa	B-W
+chấp	B-W
+hắn	B-W
+?	B-W
+
+# sent_id = uvb-f-8631
+# text = Chém ruồi ai dùng gươm vàng làm chi .
+Chém	B-W
+ruồi	B-W
+ai	B-W
+dùng	B-W
+gươm	B-W
+vàng	B-W
+làm	B-W
+chi	I-W
+.	B-W
+
+# sent_id = uvb-f-8734
+# text = Trà đá thì chỉ có học môn bơi chó !
+Trà	B-W
+đá	I-W
+thì	B-W
+chỉ	B-W
+có	B-W
+học	B-W
+môn	B-W
+bơi	B-W
+chó	I-W
+!	B-W
+
+# sent_id = uvb-f-9174
+# text = Tự chúng ta tạo nghiệp lành hoặc nghiệp dữ , tạo càng lâu thì sức mạnh càng lắm .
+Tự	B-W
+chúng	B-W
+ta	I-W
+tạo	B-W
+nghiệp	B-W
+lành	I-W
+hoặc	B-W
+nghiệp	B-W
+dữ	I-W
+,	B-W
+tạo	B-W
+càng	B-W
+lâu	B-W
+thì	B-W
+sức	B-W
+mạnh	I-W
+càng	B-W
+lắm	B-W
+.	B-W
+
+# sent_id = uvb-f-9389
+# text = Thái Vĩ ngập ngừng giây lát , rồi cũng đi đến đó ngồi .
+Thái	B-W
+Vĩ	I-W
+ngập	B-W
+ngừng	I-W
+giây	B-W
+lát	I-W
+,	B-W
+rồi	B-W
+cũng	B-W
+đi	B-W
+đến	B-W
+đó	B-W
+ngồi	B-W
+.	B-W
+
+# sent_id = uvb-n-10024
+# text = Phải biến Quấy , Nên biến Hư , Sống biến Chết , Lạnh biến Nóng , Vinh biến Nhục , Thiện biến Ác ...
+Phải	B-W
+biến	B-W
+Quấy	B-W
+,	B-W
+Nên	B-W
+biến	B-W
+Hư	B-W
+,	B-W
+Sống	B-W
+biến	B-W
+Chết	B-W
+,	B-W
+Lạnh	B-W
+biến	B-W
+Nóng	B-W
+,	B-W
+Vinh	B-W
+biến	B-W
+Nhục	B-W
+,	B-W
+Thiện	B-W
+biến	I-W
+Ác	B-W
+...	B-W
+
+# sent_id = uvb-n-11434
+# text = Truyện trích từ Sử ký Tư Mã Thiên .
+Truyện	B-W
+trích	B-W
+từ	B-W
+Sử	B-W
+ký	I-W
+Tư	B-W
+Mã	I-W
+Thiên	I-W
+.	B-W
+
+# sent_id = uvb-n-11587
+# text = Tất nhiên tiên sinh phải biết cầm âm ?
+Tất	B-W
+nhiên	I-W
+tiên	B-W
+sinh	I-W
+phải	B-W
+biết	B-W
+cầm	B-W
+âm	I-W
+?	B-W
+
+# sent_id = uvb-n-11976
+# text = Những vị kỳ lão ở xứ Gò Công còn nhớ câu : Lệ thủy trình tường thoại , Quy khâu trúc phước cơ .
+Những	B-W
+vị	B-W
+kỳ	B-W
+lão	I-W
+ở	B-W
+xứ	B-W
+Gò	B-W
+Công	I-W
+còn	B-W
+nhớ	B-W
+câu	B-W
+:	B-W
+Lệ	B-W
+thủy	I-W
+trình	B-W
+tường	B-W
+thoại	I-W
+,	B-W
+Quy	B-W
+khâu	I-W
+trúc	B-W
+phước	B-W
+cơ	I-W
+.	B-W
+
+# sent_id = uvb-n-12321
+# text = Trời hôm nay , sương mai phớt nhẹ lên mái nhà .
+Trời	B-W
+hôm	B-W
+nay	I-W
+,	B-W
+sương	B-W
+mai	I-W
+phớt	B-W
+nhẹ	B-W
+lên	B-W
+mái	B-W
+nhà	I-W
+.	B-W
+
+# sent_id = uvb-n-13233
+# text = Khoa học dạy ta : nước lạnh đặc lại thì bớt thể vóc của nó đi .
+Khoa	B-W
+học	I-W
+dạy	B-W
+ta	B-W
+:	B-W
+nước	B-W
+lạnh	B-W
+đặc	B-W
+lại	B-W
+thì	B-W
+bớt	B-W
+thể	B-W
+vóc	I-W
+của	B-W
+nó	B-W
+đi	B-W
+.	B-W
+
+# sent_id = uvb-n-13288
+# text = Bộ Mạnh Tử dày gấp rưỡi bộ Luận Ngữ , gồm 14 chương : I. Lương Huệ vương , thượng .
+Bộ	B-W
+Mạnh	B-W
+Tử	I-W
+dày	B-W
+gấp	B-W
+rưỡi	B-W
+bộ	B-W
+Luận	B-W
+Ngữ	I-W
+,	B-W
+gồm	B-W
+14	B-W
+chương	B-W
+:	B-W
+I.	B-W
+Lương	B-W
+Huệ	I-W
+vương	I-W
+,	B-W
+thượng	B-W
+.	B-W
+
+# sent_id = uvb-n-15114
+# text = Chuông điện thoại thật phiền .
+Chuông	B-W
+điện	B-W
+thoại	I-W
+thật	B-W
+phiền	B-W
+.	B-W
+
+# sent_id = uvb-n-16553
+# text = Một ngày kia Tháo đương ngủ ngày , rớt mền .
+Một	B-W
+ngày	B-W
+kia	I-W
+Tháo	B-W
+đương	B-W
+ngủ	B-W
+ngày	I-W
+,	B-W
+rớt	B-W
+mền	B-W
+.	B-W
+
+# sent_id = uvb-n-17014
+# text = Mẹ con con chịu trọng ân của đức ngài .
+Mẹ	B-W
+con	I-W
+con	B-W
+chịu	B-W
+trọng	B-W
+ân	I-W
+của	B-W
+đức	B-W
+ngài	I-W
+.	B-W
+
+# sent_id = uvb-n-18226
+# text = Ngữ đoạn và cấu trúc của ngữ đoạn Khái niệm ngữ đoạn ( hay ngữ ) .
+Ngữ	B-W
+đoạn	I-W
+và	B-W
+cấu	B-W
+trúc	I-W
+của	B-W
+ngữ	B-W
+đoạn	I-W
+Khái	B-W
+niệm	I-W
+ngữ	B-W
+đoạn	I-W
+(	B-W
+hay	B-W
+ngữ	I-W
+)	B-W
+.	B-W
+
+# sent_id = uvb-n-18736
+# text = Lục Giả đáp : - Vương hơn chứ .
+Lục	B-W
+Giả	I-W
+đáp	B-W
+:	B-W
+-	B-W
+Vương	B-W
+hơn	B-W
+chứ	B-W
+.	B-W
+
+# sent_id = uvb-n-19309
+# text = Cái thằng " nhãi " sinh viên trường Lạc Hồng định trêu ghẹo anh chắc ?
+Cái	B-W
+thằng	B-W
+"	B-W
+nhãi	B-W
+"	B-W
+sinh	B-W
+viên	I-W
+trường	B-W
+Lạc	B-W
+Hồng	I-W
+định	B-W
+trêu	B-W
+ghẹo	I-W
+anh	B-W
+chắc	B-W
+?	B-W
+
+# sent_id = uvb-n-2201
+# text = Hideoshi quyết phục thù một lần nữa .
+Hideoshi	B-W
+quyết	B-W
+phục	B-W
+thù	I-W
+một	B-W
+lần	B-W
+nữa	B-W
+.	B-W
+
+# sent_id = uvb-n-3769
+# text = Bởi vì Chu Du tính tình mạnh mẽ , hào phóng .
+Bởi	B-W
+vì	I-W
+Chu	B-W
+Du	I-W
+tính	B-W
+tình	I-W
+mạnh	B-W
+mẽ	I-W
+,	B-W
+hào	B-W
+phóng	I-W
+.	B-W
+
+# sent_id = uvb-n-3847
+# text = Thấy những chiếc răng nanh sắc nhọn , triết gia lùi lại .
+Thấy	B-W
+những	B-W
+chiếc	B-W
+răng	B-W
+nanh	I-W
+sắc	B-W
+nhọn	I-W
+,	B-W
+triết	B-W
+gia	I-W
+lùi	B-W
+lại	B-W
+.	B-W
+
+# sent_id = uvb-n-3874
+# text = Kính dâng Tam Bảo chứng minh , và thành thực cầu mong các bậc cao minh phủ chính cho những điều sơ lậu .
+Kính	B-W
+dâng	B-W
+Tam	B-W
+Bảo	I-W
+chứng	B-W
+minh	I-W
+,	B-W
+và	B-W
+thành	B-W
+thực	I-W
+cầu	B-W
+mong	I-W
+các	B-W
+bậc	B-W
+cao	B-W
+minh	I-W
+phủ	B-W
+chính	I-W
+cho	B-W
+những	B-W
+điều	B-W
+sơ	B-W
+lậu	I-W
+.	B-W
+
+# sent_id = uvb-n-3968
+# text = Lại có khi dùng đồng tự làm âm .
+Lại	B-W
+có	B-W
+khi	I-W
+dùng	B-W
+đồng	B-W
+tự	I-W
+làm	B-W
+âm	B-W
+.	B-W
+
+# sent_id = uvb-n-511
+# text = Bà Nam Phương nhận số tiền trên , gởi lời cảm ơn Cụ Hồ và Ủy ban Hành chánh Huế .
+Bà	B-W
+Nam	B-W
+Phương	I-W
+nhận	B-W
+số	B-W
+tiền	B-W
+trên	B-W
+,	B-W
+gởi	B-W
+lời	B-W
+cảm	B-W
+ơn	I-W
+Cụ	B-W
+Hồ	I-W
+và	B-W
+Ủy	B-W
+ban	I-W
+Hành	B-W
+chánh	I-W
+Huế	B-W
+.	B-W
+
+# sent_id = uvb-n-567
+# text = NHẬP TỊCH Trà nhập tịch Lễ mộc dục Đại tế Xướng ca Giao hiếu Cơm quả , cơm quan viên Khoản đãi IV .
+NHẬP	B-W
+TỊCH	I-W
+Trà	B-W
+nhập	I-W
+tịch	I-W
+Lễ	B-W
+mộc	B-W
+dục	I-W
+Đại	B-W
+tế	I-W
+Xướng	B-W
+ca	I-W
+Giao	B-W
+hiếu	I-W
+Cơm	B-W
+quả	I-W
+,	B-W
+cơm	B-W
+quan	B-W
+viên	I-W
+Khoản	B-W
+đãi	B-W
+IV	B-W
+.	B-W
+
+# sent_id = uvb-n-5687
+# text = Với tấm hân hoan ấy , tôi thành thực cám ơn tác giả và xin trân trọng giới thiệu Dịch Thuật và Tự Do với bạn đọc .
+Với	B-W
+tấm	B-W
+hân	B-W
+hoan	I-W
+ấy	B-W
+,	B-W
+tôi	B-W
+thành	B-W
+thực	I-W
+cám	B-W
+ơn	I-W
+tác	B-W
+giả	I-W
+và	B-W
+xin	B-W
+trân	B-W
+trọng	I-W
+giới	B-W
+thiệu	I-W
+Dịch	B-W
+Thuật	I-W
+và	B-W
+Tự	B-W
+Do	I-W
+với	B-W
+bạn	B-W
+đọc	I-W
+.	B-W
+
+# sent_id = uvb-n-6010
+# text = Ví dụ như chơi trò thổi bong bóng .
+Ví	B-W
+dụ	I-W
+như	B-W
+chơi	B-W
+trò	B-W
+thổi	B-W
+bong	B-W
+bóng	I-W
+.	B-W
+
+# sent_id = uvb-n-7402
+# text = Đôi khi họ đốt mu rùa để đoán điềm mộng mị .
+Đôi	B-W
+khi	I-W
+họ	B-W
+đốt	B-W
+mu	B-W
+rùa	I-W
+để	B-W
+đoán	B-W
+điềm	B-W
+mộng	B-W
+mị	I-W
+.	B-W
+
+# sent_id = uvb-n-7669
+# text = Cung Khôn Thái được thiết lập ở gần điện Cần Thánh chỗ vua ở .
+Cung	B-W
+Khôn	I-W
+Thái	I-W
+được	B-W
+thiết	B-W
+lập	I-W
+ở	B-W
+gần	B-W
+điện	B-W
+Cần	B-W
+Thánh	I-W
+chỗ	B-W
+vua	B-W
+ở	B-W
+.	B-W
+
+# sent_id = uvn-14241
+# text = Sau khi cho giáo viên nhập điểm nhà trường còn hậu kiểm nữa .
+Sau	B-W
+khi	B-W
+cho	B-W
+giáo	B-W
+viên	I-W
+nhập	B-W
+điểm	B-W
+nhà	B-W
+trường	I-W
+còn	B-W
+hậu	B-W
+kiểm	I-W
+nữa	B-W
+.	B-W
+
+# sent_id = uvn-16886
+# text = Trong thiết kế đầm đuôi cá cùng phần thân áo bó sát từ thương hiệu Giambattista Valli đến từ BST Love của nhà mốt .
+Trong	B-W
+thiết	B-W
+kế	I-W
+đầm	B-W
+đuôi	I-W
+cá	I-W
+cùng	B-W
+phần	B-W
+thân	B-W
+áo	I-W
+bó	B-W
+sát	I-W
+từ	B-W
+thương	B-W
+hiệu	I-W
+Giambattista	B-W
+Valli	I-W
+đến	B-W
+từ	B-W
+BST	B-W
+Love	B-W
+của	B-W
+nhà	B-W
+mốt	I-W
+.	B-W
+
+# sent_id = uvn-18693
+# text = Để tìm niên đại mẫu vật lâu đời hơn , họ cần sử dụng đồng vị khác như beryllium-10 .
+Để	B-W
+tìm	B-W
+niên	B-W
+đại	I-W
+mẫu	B-W
+vật	I-W
+lâu	B-W
+đời	I-W
+hơn	B-W
+,	B-W
+họ	B-W
+cần	B-W
+sử	B-W
+dụng	I-W
+đồng	B-W
+vị	I-W
+khác	B-W
+như	B-W
+beryllium-10	B-W
+.	B-W
+
+# sent_id = uvn-18725
+# text = Trường Chinh 8A sẽ chủ yếu phóng vệ tinh lên các quỹ đạo đồng bộ Mặt Trời .
+Trường	B-W
+Chinh	I-W
+8A	I-W
+sẽ	B-W
+chủ	B-W
+yếu	I-W
+phóng	B-W
+vệ	B-W
+tinh	I-W
+lên	B-W
+các	B-W
+quỹ	B-W
+đạo	I-W
+đồng	B-W
+bộ	I-W
+Mặt	B-W
+Trời	I-W
+.	B-W
+
+# sent_id = uvn-19673
+# text = Bị can Tăng Hoàng Vinh giữ vai trò giúp sức cho Hòa trong việc nhận và tiêu thụ hàng .
+Bị	B-W
+can	I-W
+Tăng	B-W
+Hoàng	I-W
+Vinh	I-W
+giữ	B-W
+vai	B-W
+trò	I-W
+giúp	B-W
+sức	I-W
+cho	B-W
+Hòa	B-W
+trong	B-W
+việc	B-W
+nhận	B-W
+và	B-W
+tiêu	B-W
+thụ	I-W
+hàng	B-W
+.	B-W
+
+# sent_id = uvn-4740
+# text = Đây là phần lợi nhuận tích lũy nhiều năm .
+Đây	B-W
+là	B-W
+phần	B-W
+lợi	B-W
+nhuận	I-W
+tích	B-W
+lũy	I-W
+nhiều	B-W
+năm	B-W
+.	B-W
+
+# sent_id = uvn-6121
+# text = Ví dụ : After years of refusing to take life seriously , Matt really seems to have turned over a new leaf and has started to find work ( Sau nhiều năm từ chối sống một cách nghiêm túc , Matt dường như đã bước sang trang mới và bắt đầu tìm việc ) .
+Ví	B-W
+dụ	I-W
+:	B-W
+After	B-W
+years	B-W
+of	B-W
+refusing	B-W
+to	B-W
+take	B-W
+life	B-W
+seriously	B-W
+,	B-W
+Matt	B-W
+really	B-W
+seems	B-W
+to	B-W
+have	B-W
+turned	B-W
+over	B-W
+a	B-W
+new	B-W
+leaf	B-W
+and	B-W
+has	B-W
+started	B-W
+to	B-W
+find	B-W
+work	I-W
+(	B-W
+Sau	B-W
+nhiều	B-W
+năm	B-W
+từ	B-W
+chối	I-W
+sống	B-W
+một	B-W
+cách	I-W
+nghiêm	B-W
+túc	I-W
+,	B-W
+Matt	B-W
+dường	B-W
+như	I-W
+đã	B-W
+bước	B-W
+sang	B-W
+trang	B-W
+mới	B-W
+và	B-W
+bắt	B-W
+đầu	I-W
+tìm	B-W
+việc	B-W
+)	B-W
+.	B-W
+
+# sent_id = uvn-964
+# text = Các kỹ sư Nhật Bản khởi động siêu máy tính lượng tử lai đầu tiên trên thế giới .
+Các	B-W
+kỹ	B-W
+sư	I-W
+Nhật	B-W
+Bản	I-W
+khởi	B-W
+động	I-W
+siêu	B-W
+máy	I-W
+tính	I-W
+lượng	B-W
+tử	I-W
+lai	B-W
+đầu	B-W
+tiên	I-W
+trên	B-W
+thế	B-W
+giới	I-W
+.	B-W
+
+# sent_id = uvw-10198
+# text = Hữu thưởng vu đại quốc .
+Hữu	B-W
+thưởng	B-W
+vu	B-W
+đại	B-W
+quốc	I-W
+.	B-W
+
+# sent_id = uvw-11634
+# text = Có người nghĩa sĩ viếng câu đối rằng : :: Thệ tâm thiên địa lưu trường xích , :: Thiết xỉ giang sơn thổ thiệt hồng .
+Có	B-W
+người	B-W
+nghĩa	B-W
+sĩ	I-W
+viếng	B-W
+câu	B-W
+đối	I-W
+rằng	B-W
+:	B-W
+::	B-W
+Thệ	B-W
+tâm	B-W
+thiên	B-W
+địa	I-W
+lưu	B-W
+trường	B-W
+xích	B-W
+,	B-W
+::	B-W
+Thiết	B-W
+xỉ	B-W
+giang	B-W
+sơn	I-W
+thổ	B-W
+thiệt	B-W
+hồng	B-W
+.	B-W
+
+# sent_id = uvw-13194
+# text = Quân xướng thần họa .
+Quân	B-W
+xướng	B-W
+thần	B-W
+họa	B-W
+.	B-W
+
+# sent_id = uvw-13430
+# text = Đường thẳng chứa ba điểm đó được gọi là đường thẳng Euler .
+Đường	B-W
+thẳng	I-W
+chứa	B-W
+ba	B-W
+điểm	B-W
+đó	B-W
+được	B-W
+gọi	B-W
+là	I-W
+đường	B-W
+thẳng	I-W
+Euler	B-W
+.	B-W
+
+# sent_id = uvw-14316
+# text = Hệ tiên đề Hilbert Hệ tiên đề số học Lý thuyết tập hợp Frankael-Zermelo Tiên đề chọn Tiên đề trong vật lý Tiên đề Bohr Các tiên đề Bohr là các tiên đề của mô hình Bohr , được sử dụng để giải thích các hiện tượng vật lý , ví dụ như công thức Rydberg về các vạch quang phổ của nguyên tử hydro .
+Hệ	B-W
+tiên	I-W
+đề	I-W
+Hilbert	B-W
+Hệ	B-W
+tiên	I-W
+đề	I-W
+số	B-W
+học	I-W
+Lý	B-W
+thuyết	I-W
+tập	B-W
+hợp	I-W
+Frankael-Zermelo	B-W
+Tiên	B-W
+đề	I-W
+chọn	B-W
+Tiên	B-W
+đề	I-W
+trong	B-W
+vật	B-W
+lý	I-W
+Tiên	B-W
+đề	I-W
+Bohr	B-W
+Các	B-W
+tiên	B-W
+đề	I-W
+Bohr	B-W
+là	B-W
+các	B-W
+tiên	B-W
+đề	I-W
+của	B-W
+mô	B-W
+hình	I-W
+Bohr	B-W
+,	B-W
+được	B-W
+sử	B-W
+dụng	I-W
+để	B-W
+giải	B-W
+thích	I-W
+các	B-W
+hiện	B-W
+tượng	I-W
+vật	B-W
+lý	I-W
+,	B-W
+ví	B-W
+dụ	I-W
+như	B-W
+công	B-W
+thức	I-W
+Rydberg	B-W
+về	B-W
+các	B-W
+vạch	B-W
+quang	B-W
+phổ	I-W
+của	B-W
+nguyên	B-W
+tử	I-W
+hydro	B-W
+.	B-W
+
+# sent_id = uvw-1451
+# text = Trong thế giới Quan Âm gồm có Quan Âm Nam Hải và Quan Âm Diệu Thiện .
+Trong	B-W
+thế	B-W
+giới	I-W
+Quan	B-W
+Âm	I-W
+gồm	B-W
+có	B-W
+Quan	B-W
+Âm	I-W
+Nam	B-W
+Hải	I-W
+và	B-W
+Quan	B-W
+Âm	I-W
+Diệu	B-W
+Thiện	I-W
+.	B-W
+
+# sent_id = uvw-15121
+# text = Nhà Habsburg cũng bắt đầu tích lũy lãnh thổ cách xa các vùng đất cha truyền con nối .
+Nhà	B-W
+Habsburg	B-W
+cũng	B-W
+bắt	B-W
+đầu	I-W
+tích	B-W
+lũy	I-W
+lãnh	B-W
+thổ	I-W
+cách	B-W
+xa	B-W
+các	B-W
+vùng	B-W
+đất	I-W
+cha	B-W
+truyền	I-W
+con	I-W
+nối	I-W
+.	B-W
+
+# sent_id = uvw-15306
+# text = Nhà Tống tiêu diệt quốc gia của Trí Cao năm 1055 .
+Nhà	B-W
+Tống	B-W
+tiêu	B-W
+diệt	I-W
+quốc	B-W
+gia	I-W
+của	B-W
+Trí	B-W
+Cao	I-W
+năm	B-W
+1055	B-W
+.	B-W
+
+# sent_id = uvw-15392
+# text = Trần Tư Trai 《 Hải quốc văn kiến lục 》 nói : 《 An Nam lấy Giao Chỉ làm Đông Kinh , lấy Quảng Nam làm Tây kinh .
+Trần	B-W
+Tư	I-W
+Trai	I-W
+《	B-W
+Hải	B-W
+quốc	I-W
+văn	I-W
+kiến	I-W
+lục	I-W
+》	B-W
+nói	B-W
+:	B-W
+《	B-W
+An	B-W
+Nam	I-W
+lấy	B-W
+Giao	B-W
+Chỉ	I-W
+làm	B-W
+Đông	B-W
+Kinh	I-W
+,	B-W
+lấy	B-W
+Quảng	B-W
+Nam	I-W
+làm	B-W
+Tây	B-W
+kinh	I-W
+.	B-W
+
+# sent_id = uvw-15619
+# text = Long Hồ dinh đổi tên thành Hà Tiên trấn .
+Long	B-W
+Hồ	I-W
+dinh	B-W
+đổi	B-W
+tên	B-W
+thành	B-W
+Hà	B-W
+Tiên	I-W
+trấn	B-W
+.	B-W
+
+# sent_id = uvw-16915
+# text = Hình ảnh Tập tin : DNA replication .
+Hình	B-W
+ảnh	I-W
+Tập	B-W
+tin	I-W
+:	B-W
+DNA	B-W
+replication	B-W
+.	B-W
+
+# sent_id = uvw-17218
+# text = Điểm nóng chảy và điểm sôi Hợp chất hữu cơ rất dễ nóng chảy hay sôi .
+Điểm	B-W
+nóng	B-W
+chảy	I-W
+và	B-W
+điểm	B-W
+sôi	B-W
+Hợp	B-W
+chất	I-W
+hữu	B-W
+cơ	I-W
+rất	B-W
+dễ	B-W
+nóng	B-W
+chảy	I-W
+hay	B-W
+sôi	B-W
+.	B-W
+
+# sent_id = uvw-17235
+# text = Khổng Dĩnh Đạt thời Đường thì nói " Hoa Hạ vi Trung Quốc dã " .
+Khổng	B-W
+Dĩnh	I-W
+Đạt	I-W
+thời	B-W
+Đường	B-W
+thì	B-W
+nói	B-W
+"	B-W
+Hoa	B-W
+Hạ	I-W
+vi	B-W
+Trung	B-W
+Quốc	I-W
+dã	B-W
+"	B-W
+.	B-W
+
+# sent_id = uvw-18245
+# text = Ngừng nạp 1 giờ cho các ngăn ắc quy ổn định .
+Ngừng	B-W
+nạp	B-W
+1	B-W
+giờ	B-W
+cho	B-W
+các	B-W
+ngăn	B-W
+ắc	B-W
+quy	I-W
+ổn	B-W
+định	I-W
+.	B-W
+
+# sent_id = uvw-19010
+# text = Ông thân thiết với nhà toán học Thụy Sĩ Nicolas Fatio de Duillier .
+Ông	B-W
+thân	B-W
+thiết	I-W
+với	B-W
+nhà	B-W
+toán	B-W
+học	I-W
+Thụy	B-W
+Sĩ	I-W
+Nicolas	B-W
+Fatio	I-W
+de	I-W
+Duillier	I-W
+.	B-W
+
+# sent_id = uvw-2178
+# text = Đồng nguyên chất mềm và dễ uốn ; bề mặt đồng tươi có màu cam đỏ .
+Đồng	B-W
+nguyên	B-W
+chất	I-W
+mềm	B-W
+và	B-W
+dễ	B-W
+uốn	B-W
+;	B-W
+bề	B-W
+mặt	I-W
+đồng	B-W
+tươi	B-W
+có	B-W
+màu	B-W
+cam	B-W
+đỏ	I-W
+.	B-W
+
+# sent_id = uvw-3214
+# text = Thiên địa hòa xướng chi tượng : tượng trời đất giao hòa .
+Thiên	B-W
+địa	I-W
+hòa	B-W
+xướng	I-W
+chi	B-W
+tượng	B-W
+:	B-W
+tượng	B-W
+trời	B-W
+đất	I-W
+giao	B-W
+hòa	I-W
+.	B-W
+
+# sent_id = uvw-379
+# text = Phục Hy ghi : Nhu giả ẩm thực chi đạo dã , ẩm thực tất hữu tụng , cố thụ chi dĩ tụng .
+Phục	B-W
+Hy	I-W
+ghi	B-W
+:	B-W
+Nhu	B-W
+giả	B-W
+ẩm	B-W
+thực	I-W
+chi	B-W
+đạo	B-W
+dã	I-W
+,	B-W
+ẩm	B-W
+thực	I-W
+tất	B-W
+hữu	B-W
+tụng	B-W
+,	B-W
+cố	B-W
+thụ	B-W
+chi	B-W
+dĩ	B-W
+tụng	B-W
+.	B-W
+
+# sent_id = uvw-4040
+# text = Nhi Đinh Lê nhị gia , nãi tuẫn kỷ tư , hốt thiên mệnh , võng đạo Thương Chu chi tích , thường an quyết ấp vu tư , trí thế đại phất trường , toán số đoản xúc , bách tính hao tổn , vạn vật thất nghi .
+Nhi	B-W
+Đinh	B-W
+Lê	I-W
+nhị	B-W
+gia	I-W
+,	B-W
+nãi	B-W
+tuẫn	B-W
+kỷ	B-W
+tư	I-W
+,	B-W
+hốt	B-W
+thiên	B-W
+mệnh	I-W
+,	B-W
+võng	B-W
+đạo	B-W
+Thương	B-W
+Chu	I-W
+chi	B-W
+tích	I-W
+,	B-W
+thường	B-W
+an	B-W
+quyết	B-W
+ấp	B-W
+vu	B-W
+tư	I-W
+,	B-W
+trí	B-W
+thế	B-W
+đại	I-W
+phất	B-W
+trường	B-W
+,	B-W
+toán	B-W
+số	I-W
+đoản	B-W
+xúc	I-W
+,	B-W
+bách	B-W
+tính	I-W
+hao	B-W
+tổn	I-W
+,	B-W
+vạn	B-W
+vật	I-W
+thất	B-W
+nghi	I-W
+.	B-W
+
+# sent_id = uvw-415
+# text = Giải nghĩa : Chỉ dã .
+Giải	B-W
+nghĩa	I-W
+:	B-W
+Chỉ	B-W
+dã	I-W
+.	B-W
+
+# sent_id = uvw-4428
+# text = Cửu ngũ : Nhu vu tửu thực , trinh cát .
+Cửu	B-W
+ngũ	I-W
+:	B-W
+Nhu	B-W
+vu	B-W
+tửu	B-W
+thực	I-W
+,	B-W
+trinh	B-W
+cát	B-W
+.	B-W
+
+# sent_id = uvw-5791
+# text = Kỳ địa quảng nhi thản bình , quyết thổ cao nhi sảng khải .
+Kỳ	B-W
+địa	B-W
+quảng	B-W
+nhi	B-W
+thản	B-W
+bình	I-W
+,	B-W
+quyết	B-W
+thổ	B-W
+cao	B-W
+nhi	B-W
+sảng	B-W
+khải	I-W
+.	B-W
+
+# sent_id = uvw-6347
+# text = Dân cư miệt hôn điếm chi khốn ; vạn vật cực phồn phụ chi phong .
+Dân	B-W
+cư	I-W
+miệt	B-W
+hôn	B-W
+điếm	I-W
+chi	B-W
+khốn	B-W
+;	B-W
+vạn	B-W
+vật	I-W
+cực	B-W
+phồn	B-W
+phụ	I-W
+chi	B-W
+phong	B-W
+.	B-W
+
+# sent_id = uvw-6691
+# text = Họ Cariamidae : Chim mào bắt rắn .
+Họ	B-W
+Cariamidae	B-W
+:	B-W
+Chim	B-W
+mào	I-W
+bắt	B-W
+rắn	I-W
+.	B-W
+
+# sent_id = uvw-7555
+# text = Holocaust : Quân Mỹ giải phóng trại tập trung Dachau .
+Holocaust	B-W
+:	B-W
+Quân	B-W
+Mỹ	B-W
+giải	B-W
+phóng	I-W
+trại	B-W
+tập	B-W
+trung	I-W
+Dachau	B-W
+.	B-W
+
+# sent_id = uvw-8056
+# text = 1999 – Tổng thống Nigeria Ibrahim Baré Maïnassara bị ám sát .
+1999	B-W
+–	B-W
+Tổng	B-W
+thống	I-W
+Nigeria	B-W
+Ibrahim	I-W
+Baré	I-W
+Maïnassara	I-W
+bị	B-W
+ám	B-W
+sát	I-W
+.	B-W
+
+# sent_id = uvw-8082
+# text = Cô cũng 2 lần thắng giải Grand Prix Award cùng nhiều giải thưởng danh giá như MTV Video Music Awards Japan , World Music Awards , Japan Gold Disc Award .
+Cô	B-W
+cũng	B-W
+2	B-W
+lần	B-W
+thắng	B-W
+giải	B-W
+Grand	B-W
+Prix	I-W
+Award	I-W
+cùng	B-W
+nhiều	B-W
+giải	B-W
+thưởng	I-W
+danh	B-W
+giá	I-W
+như	B-W
+MTV	B-W
+Video	I-W
+Music	I-W
+Awards	I-W
+Japan	I-W
+,	B-W
+World	B-W
+Music	I-W
+Awards	I-W
+,	B-W
+Japan	B-W
+Gold	I-W
+Disc	I-W
+Award	I-W
+.	B-W
+
+# sent_id = uvw-998
+# text = Công trình về hiệu ứng quang điện của ông mang tính bước ngoặt khai sinh ra lý thuyết lượng tử .
+Công	B-W
+trình	I-W
+về	B-W
+hiệu	B-W
+ứng	I-W
+quang	B-W
+điện	I-W
+của	B-W
+ông	B-W
+mang	B-W
+tính	B-W
+bước	B-W
+ngoặt	I-W
+khai	B-W
+sinh	I-W
+ra	B-W
+lý	B-W
+thuyết	I-W
+lượng	B-W
+tử	I-W
+.	B-W
+
+# sent_id = vlc-10445
+# text = Đơn phương miễn thị thực 1 .
+Đơn	B-W
+phương	I-W
+miễn	B-W
+thị	B-W
+thực	I-W
+1	B-W
+.	B-W
+
+# sent_id = vlc-4443
+# text = Sơ yếu lý lịch và Phiếu lý lịch tư pháp .
+Sơ	B-W
+yếu	I-W
+lý	B-W
+lịch	I-W
+và	B-W
+Phiếu	B-W
+lý	B-W
+lịch	I-W
+tư	B-W
+pháp	I-W
+.	B-W
+
+# sent_id = vlc-6220
+# text = Ủy ban thường vụ Quốc hội quy định về thể thức và kỹ thuật trình bày văn bản quy phạm pháp luật của Quốc hội , Ủy ban thường vụ Quốc hội , Chủ tịch nước .
+Ủy	B-W
+ban	I-W
+thường	B-W
+vụ	I-W
+Quốc	B-W
+hội	I-W
+quy	B-W
+định	I-W
+về	B-W
+thể	B-W
+thức	I-W
+và	B-W
+kỹ	B-W
+thuật	I-W
+trình	B-W
+bày	I-W
+văn	B-W
+bản	I-W
+quy	B-W
+phạm	I-W
+pháp	B-W
+luật	I-W
+của	B-W
+Quốc	B-W
+hội	I-W
+,	B-W
+Ủy	B-W
+ban	I-W
+thường	B-W
+vụ	I-W
+Quốc	B-W
+hội	I-W
+,	B-W
+Chủ	B-W
+tịch	I-W
+nước	I-W
+.	B-W
+
+# sent_id = vlc-7243
+# text = Đối tượng được phong quân hàm sĩ quan tại ngũ Những người sau đây được xét phong quân hàm sĩ quan tại ngũ : 1 .
+Đối	B-W
+tượng	I-W
+được	B-W
+phong	B-W
+quân	B-W
+hàm	I-W
+sĩ	B-W
+quan	I-W
+tại	B-W
+ngũ	I-W
+Những	B-W
+người	B-W
+sau	B-W
+đây	B-W
+được	B-W
+xét	B-W
+phong	B-W
+quân	B-W
+hàm	I-W
+sĩ	B-W
+quan	I-W
+tại	B-W
+ngũ	I-W
+:	B-W
+1	B-W
+.	B-W
+

guidelines/01. Word Segmentation/Annotation Guideline v1.1.md

@@ -0,0 +1,336 @@
+# Annotation Guideline v1.1: Word Segmentation
+
+Phiên bản cập nhật sau AL Cycle 1 (92 câu gold, 5 domains).
+
+## 1. Tổng quan
+
+Guideline này bổ sung các quy tắc thực hành từ kinh nghiệm gán nhãn Cycle 1. Các quy tắc nền tảng (NIIVTB 9 rules, VLSP) vẫn giữ nguyên — xem `Annotation Guideline Summary.md`.
+
+### 1.1 Kết quả Cycle 1
+
+| Metric | Score |
+|--------|-------|
+| Câu gán nhãn | 92 (từ top 100 uncertainty) |
+| Word F1 (silver vs gold) | 0.7808 |
+| Over-merge errors | 134 |
+| Over-split errors | 119 |
+| Câu có lỗi | 75/92 (81.5%) |
+
+Domain khó nhất: Wikipedia (F1=0.707), Legal (F1=0.713).
+
+## 2. Quy tắc tên riêng
+
+### 2.1 Tên người Việt
+
+Tên người Việt gồm các thành phần: Họ + Tên đệm + Tên. Mỗi thành phần viết hoa chữ cái đầu.
+
+**Quy tắc**: Gộp toàn bộ tên người thành một từ.
+
+| Đúng | Sai |
+|------|-----|
+| `Nguyễn_Bính` | `Nguyễn` \| `Bính` |
+| `Trương_Hán_Siêu` | `Trương_Hán` \| `Siêu` |
+| `Tăng_Hoàng_Vinh` | `Tăng_Hoàng` \| `Vinh` |
+| `Chu_Du` | `Chu` \| `Du` |
+| `Kỳ_Nam` | `Kỳ` \| `Nam` |
+
+**Lưu ý ranh giới**: Không gộp tên người với từ kế tiếp.
+
+| Đúng | Sai (CRF hay mắc) |
+|------|-----|
+| `Nguyễn_Bính` \| `Mưa` \| `dầm` | `Nguyễn_Bính_Mưa_dầm` |
+| `Chu_Du` \| `tính_tình` | `Chu_Du_tính_tình` |
+| `Tăng_Hoàng_Vinh` \| `giữ` | `Tăng_Hoàng_Vinh_giữ` |
+| `Quỳnh` \| `nghiến_răng` | `Quỳnh_nghiến` + `răng` |
+| `Khải` \| `nheo_nheo` | `Khải_nheo` + `nheo` |
+
+**Dấu hiệu nhận biết ranh giới**: Từ sau tên riêng thường là động từ, tính từ, hoặc danh từ chung (viết thường). Nếu syllable kế tiếp viết thường → tách.
+
+### 2.2 Tên người Hán-Việt / Lịch sử
+
+Tên nhân vật lịch sử Hán-Việt: gộp toàn bộ hiệu/tên.
+
+| Đúng | Sai |
+|------|-----|
+| `Lương_Huệ_vương` | `Lương` \| `Huệ_vương` |
+| `Khổng_Dĩnh_Đạt` | `Khổng_Dĩnh` \| `Đạt` |
+| `Lục_Giả` | `Lục` \| `Giả` |
+| `Phục_Hy` | `Phục` \| `Hy` |
+
+### 2.3 Tên người nước ngoài
+
+Gộp toàn bộ tên (kể cả giới từ "de", "von"):
+
+| Đúng | Sai |
+|------|-----|
+| `Nicolas_Fatio_de_Duillier` | `Nicolas` \| `Fatio_de` \| `Duillier` |
+| `Nigeria_Ibrahim_Baré_Maïnassara` | `Nigeria_Ibrahim` \| `Baré_Maïnassara` |
+| `Beethoven` | (đúng, 1 syllable) |
+
+### 2.4 Tên địa danh
+
+| Đúng | Sai |
+|------|-----|
+| `Hà_Tiên` \| `trấn` | `Hà_Tiên_trấn` |
+| `Long_Hồ` \| `dinh` | `Long` \| `Hồ_dinh` |
+| `Lạc_Hồng` | `Lạc` \| `Hồng` |
+| `Cần_Thánh` | `Cần` \| `Thánh` |
+
+**Quy tắc**: Tên địa danh riêng gộp thành một từ. Danh từ chung chỉ đơn vị hành chính (trấn, dinh, huyện, tỉnh...) tách riêng.
+
+### 2.5 Tên tổ chức / Giải thưởng
+
+Gộp toàn bộ tên chính thức:
+
+| Đúng | Sai |
+|------|-----|
+| `World_Music_Awards` | `World_Music` \| `Awards` |
+| `MTV_Video_Music_Awards_Japan` | `MTV_Video` \| `Music` \| `Awards_Japan` |
+| `Grand_Prix_Award` | `Grand` \| `Prix_Award` |
+| `Japan_Gold_Disc_Award` | `Japan_Gold` \| `Disc_Award` |
+
+## 3. Quy tắc từ nước ngoài (Latin script)
+
+### 3.1 Nguyên tắc chung
+
+Mỗi từ tiếng Anh/Latin là một từ riêng biệt. **Không gộp** các từ tiếng Anh liền nhau thành một cụm (trừ tên riêng — xem mục 2.5).
+
+| Đúng | Sai |
+|------|-----|
+| `I` \| `LOVE` \| `YOU` | `I_LOVE` \| `YOU` |
+| `DNA` \| `replication` | `DNA_replication` |
+| `BST` \| `Love` | `BST_Love` |
+| `Matt` \| `really` | `Matt_really` |
+| `new` \| `leaf` | `new_leaf` |
+| `of` \| `refusing` | `of_refusing` |
+
+### 3.2 Viết tắt
+
+Viết tắt (BST, DNA, PS, I.) là từ riêng biệt:
+
+| Đúng | Sai |
+|------|-----|
+| `Ahern` \| `PS` | `Ahern_PS` |
+| `I.` \| `Lương_Huệ_vương` | `I._Lương` |
+
+## 4. Quy tắc Hán-Việt cổ văn
+
+Nhiều câu Wikipedia chứa trích dẫn cổ văn Hán-Việt. Trong cổ văn, phần lớn từ là **đơn âm tiết**.
+
+### 4.1 Nguyên tắc
+
+Hư từ Hán-Việt cổ (chi 之, dĩ 以, hữu 有, vô 無, tất 必, nhi 而...) luôn là từ riêng biệt:
+
+| Đúng | Sai |
+|------|-----|
+| `tất` \| `hữu` \| `tụng` | `tất_hữu_tụng` |
+| `chi` \| `dĩ` \| `tụng` | `chi_dĩ_tụng` |
+| `ẩm_thực` \| `chi` | `ẩm_thực_chi` |
+| `thiên_địa` \| `lưu` | `thiên_địa_lưu` |
+
+### 4.2 Từ ghép Hán-Việt vẫn gộp
+
+Từ ghép Hán-Việt 2+ âm tiết có nghĩa cố định vẫn gộp:
+
+| Gộp (đúng) | Lý do |
+|-------------|-------|
+| `thiên_địa` (天地) | Từ ghép đẳng lập: trời + đất |
+| `giang_sơn` (江山) | Từ ghép đẳng lập: sông + núi |
+| `ẩm_thực` (飲食) | Từ ghép đẳng lập: ăn + uống |
+| `đại_quốc` (大國) | Từ ghép chính phụ: nước + lớn |
+| `hòa_xướng` (和唱) | Từ ghép chính phụ: hòa + hát |
+| `hôn_điếm` (昏店) | Từ ghép cố định |
+| `Đinh_Lê` (丁黎) | Tên triều đại |
+| `nhị_gia` (二家) | Cụm cố định |
+
+### 4.3 Phân biệt cổ văn vs hiện đại
+
+Cùng một chuỗi syllable có thể gộp khác nhau tùy ngữ cảnh:
+
+| Ngữ cảnh | Tách | Ví dụ |
+|-----------|------|-------|
+| Cổ văn: `高 而 廣` | `cao` \| `nhi` \| `quảng` | 3 từ đơn |
+| Hiện đại: `cao nhi` | Không tồn tại | — |
+| Cổ văn: `決 土` | `quyết` \| `thổ` | 2 từ đơn |
+| Hiện đại: `quyết_định` | `quyết_định` | 1 từ ghép |
+
+## 5. Quy tắc từ ghép boundary shift
+
+Đây là lỗi phổ biến nhất: CRF gộp đúng số syllable nhưng sai ranh giới.
+
+### 5.1 Pattern: A_B_C → A_B | C (tách sai ở giữa)
+
+| Sai | Đúng | Ghi chú |
+|-----|------|---------|
+| `lợi_nhuận_tích` \| `lũy` | `lợi_nhuận` \| `tích_lũy` | boundary shift |
+| `niên_đại_mẫu` \| `vật` | `niên_đại` \| `mẫu_vật` | boundary shift |
+| `ngăn_ắc` \| `quy_ổn_định` | `ngăn` \| `ắc_quy` \| `ổn_định` | boundary shift |
+| `nanh_sắc` \| `nhọn` | `răng_nanh` \| `sắc_nhọn` | boundary shift |
+| `thực_cầu` \| `mong` | `thành_thực` \| `cầu_mong` | boundary shift |
+| `Hải_quốc_văn_kiến` \| `lục_》` | `Hải_quốc_văn_kiến_lục` \| `》` | tách dấu |
+
+**Cách nhận biết**: Kiểm tra xem từ ghép có nghĩa hợp lý không. Nếu `A_B` không phải từ trong từ điển mà `B_C` mới đúng → boundary shift.
+
+### 5.2 Pattern: Tên riêng + Từ kế tiếp
+
+| Sai | Đúng |
+|-----|------|
+| `Nữ_minh` \| `tinh_khịt` | `Nữ_minh_tinh` \| `khịt_mũi` |
+| `Liên_xua` \| `tay` | `Liên` \| `xua_tay` |
+| `Hồ_dinh_đổi` | `Long_Hồ` \| `dinh` \| `đổi` |
+
+## 6. Quy tắc từ ghép thường gặp
+
+### 6.1 Từ ghép phải gộp (CRF hay tách sai)
+
+| Từ | Loại | Ví dụ sai |
+|----|------|-----------|
+| `Ủy_ban` | Danh từ ghép | `Ủy` \| `ban` |
+| `Chủ_tịch_nước` | Chức danh | `Chủ_tịch` \| `nước` |
+| `lính_thú` | Danh từ ghép | `lính` \| `thú` |
+| `mu_rùa` | Danh từ ghép | `mu` \| `rùa` |
+| `trêu_ghẹo` | Động từ ghép đẳng lập | `trêu` \| `ghẹo` |
+| `sương_mai` | Danh từ ghép | `sương` \| `mai` |
+| `mái_nhà` | Danh từ ghép chính phụ | `mái` \| `nhà` |
+| `siêu_máy_tính` | Danh từ ghép | `siêu` \| `máy_tính` |
+| `lượng_tử` | Danh từ Hán-Việt | `lượng` \| `tử` |
+| `tích_lũy` | Động từ Hán-Việt | `tích` \| `lũy` |
+| `tại_ngũ` | Tính từ Hán-Việt | `tại` \| `ngũ` |
+| `Hành_chánh` | Danh từ Hán-Việt | `Hành` \| `chánh` |
+| `nghiến_răng` | Động từ ghép | `nghiến` \| `răng` |
+| `nheo_nheo` | Từ láy | `nheo` \| `nheo` |
+| `dơn_dớt` | Từ láy | `dơn` \| `dớt` |
+| `đầm_đuôi_cá` | Danh từ ghép (thời trang) | `đầm` \| `đuôi` \| `cá` |
+| `cha_truyền_con_nối` | Thành ngữ | `cha` \| `truyền_con` \| `nối` |
+| `Đường_thẳng` | Thuật ngữ toán học | `Đường` \| `thẳng` |
+| `nói_gở` | Động từ ghép | `nói` \| `gở` |
+| `xua_tay` | Động từ ghép | `xua` \| `tay` |
+| `bơi_chó` | Danh từ ghép | `bơi` \| `chó` |
+| `người_thương` | Danh từ ghép | `người` \| `thương` |
+| `chăn_lợn` | Động từ ghép | `chăn` \| `lợn` |
+| `khay_trà` | Danh từ ghép | `khay` \| `trà` |
+| `đồng_tự` | Thuật ngữ | `đồng` \| `tự` |
+
+### 6.2 Từ ghép phải tách (CRF hay gộp sai)
+
+| Từ (tách) | Loại | Ví dụ sai |
+|------------|------|-----------|
+| `phải` \| `biết` | Phụ từ + Động từ | `phải_biết` |
+| `có` \| `học` | Phụ từ + Động từ | `có_học` |
+| `lạnh` \| `đặc` | Tính từ + Bổ ngữ | `lạnh_đặc` |
+| `chơi` \| `trò` | Động từ + Danh từ | `chơi_trò` |
+| `làm` \| `âm` | Động từ + Danh từ | `làm_âm` |
+| `dạo` \| `khúc` | Động từ + Danh từ | `dạo_khúc` |
+| `miễn` \| `thị_thực` | Động từ + Danh từ | `miễn_thị_thực` |
+| `phóng` \| `vệ_tinh` | Động từ + Danh từ | `phóng_vệ_tinh` |
+
+**Nguyên tắc**: Nếu có thể chèn từ vào giữa (Rule 1) → tách. Ví dụ: `phải biết` → `phải nên biết` ✓ → tách.
+
+### 6.3 Từ ghép pháp luật
+
+| Gộp | Tách |
+|-----|------|
+| `quy_phạm` + `pháp_luật` (2 từ) | ~~`quy_phạm_pháp_luật`~~ (sai) |
+| `trình_bày` + `văn_bản` (2 từ) | ~~`trình_bày_văn_bản`~~ (sai) |
+| `thường_vụ` (1 từ) | ~~`ban_thường_vụ`~~ (sai) |
+| `quân_hàm` (1 từ) | ~~`hàm_sĩ_quan`~~ (sai, boundary shift) |
+| `sĩ_quan` (1 từ) | ~~`phong_quân`~~ (sai, boundary shift) |
+| `thị_thực` (1 từ) | ~~`miễn_thị_thực`~~ (sai) |
+
+**Nguyên tắc**: Thuật ngữ pháp luật 2 âm tiết → gộp (`quy_phạm`, `pháp_luật`, `sĩ_quan`, `quân_hàm`). Cụm 4 âm tiết → kiểm tra có phải 2 từ ghép 2 âm tiết ghép lại không.
+
+## 7. Quy tắc đặc biệt
+
+### 7.1 "Nhà" + Nghề nghiệp
+
+| Pattern | Tách | Ghi chú |
+|---------|------|---------|
+| `nhà` \| `toán_học` | `nhà` = danh từ chung, `toán_học` = danh từ ghép | Tách |
+| `nhà` \| `cầm_quyền` | | Tách |
+
+Nhưng: `nhà_máy` (factory) → gộp (không thể chèn).
+
+### 7.2 Tên tác phẩm / Tiêu đề
+
+Tên tác phẩm tách riêng khỏi tên tác giả:
+
+| Đúng | Sai |
+|------|-----|
+| `Nguyễn_Bính` \| `Mưa` \| `dầm` | `Nguyễn_Bính_Mưa_dầm` |
+| `Nguyễn_Bính` \| `Công` | `Nguyễn_Bính_Công` |
+| `Hải_quốc_văn_kiến_lục` | `Hải_quốc_văn_kiến` \| `lục` |
+
+Lưu ý: Tên sách Hán-Việt gộp toàn bộ (`Hải_quốc_văn_kiến_lục` = 1 từ).
+
+### 7.3 Tên cung điện / Địa danh lịch sử
+
+| Đúng | Sai |
+|------|-----|
+| `Cung_Khôn_Thái` | `Cung` \| `Khôn_Thái` |
+| `Bờ_xe_nước` | `Bờ` \| `xe` \| `nước` |
+
+### 7.4 Thành ngữ / Quán ngữ
+
+Gộp nếu nghĩa không suy ra từ thành phần:
+
+| Gộp | Loại |
+|-----|------|
+| `cha_truyền_con_nối` | Thành ngữ 4 âm tiết |
+| `như_điên` | Quán ngữ so sánh |
+| `tẩy_chay` | Từ ghép (từ boycott) |
+
+### 7.5 Dấu câu đặc biệt
+
+Dấu câu luôn là từ riêng biệt:
+
+| Đúng | Sai |
+|------|-----|
+| `》` (riêng) | `lục_》` |
+| `»` (riêng) | gộp với từ trước |
+
+## 8. Label Studio Workflow
+
+### 8.1 Labels
+
+| Label | Ý nghĩa | Dùng khi |
+|-------|----------|----------|
+| **W** | Word (từ thường) | Tên riêng, từ nước ngoài, từ ít gặp |
+| **WH** | Word High confidence | Từ phổ biến, CRF confidence > 0.95 |
+| **WM** | Word Medium confidence | CRF confidence 0.80–0.95 |
+| **WL** | Word Low confidence | CRF confidence < 0.80, cần kiểm tra kỹ |
+
+### 8.2 Quy trình gán nhãn
+
+1. Đọc toàn bộ câu trước khi sửa
+2. Kiểm tra các span WL (low confidence) trước — đây thường là lỗi
+3. Kiểm tra ranh giới tên riêng (viết hoa)
+4. Kiểm tra từ nước ngoài (Latin script)
+5. Sử dụng dictionary plugin (xanh = có trong từ điển, đỏ = không có)
+6. Đảm bảo **tất cả syllable đều được gán nhãn** (tránh bỏ sót)
+
+### 8.3 Lỗi thường gặp khi gán nhãn
+
+| Lỗi | Hậu quả |
+|-----|---------|
+| Bỏ sót syllable (không gán nhãn) | Câu bị loại khi export |
+| Gộp dấu câu với từ trước | BIO sai |
+| Bỏ qua từ nước ngoài | Câu bị loại |
+
+## 9. Tóm tắt quy tắc ưu tiên
+
+Khi không chắc chắn, áp dụng theo thứ tự:
+
+1. **Tra từ điển** (Rule 7): Nếu có trong từ điển → gộp
+2. **Test chèn** (Rule 1): Chèn được từ vào giữa → tách
+3. **Hình vị ràng buộc** (Rule 6): Thành phần không đứng riêng được → gộp
+4. **Nghĩa tổng hợp** (Rule 2–5): Nghĩa khác tổng thành phần → gộp
+5. **Tên riêng**: Gộp toàn bộ, tách khỏi từ tiếp theo
+6. **Cổ văn Hán-Việt**: Mặc định đơn âm tiết, trừ từ ghép cố định
+7. **Từ nước ngoài**: Mỗi từ Latin là một từ riêng (trừ tên riêng)
+
+## 10. Changelog
+
+| Version | Date | Changes |
+|---------|------|---------|
+| v1.0 | 2026-02 | Initial guidelines (NIIVTB 9 rules + legal domain) |
+| v1.1 | 2026-02 | Added: proper name rules, foreign word rules, Hán-Việt classical text rules, boundary shift patterns, compound word lists from Cycle 1 gold annotation (92 sentences) |

references/1998.conll.brants/paper.md

@@ -0,0 +1,1260 @@
+---
+title: "Automation of Treebank Annotation"
+authors:
+  - "Thorsten Brants"
+  - "Wojciech Skut"
+year: 1998
+venue: "CoNLL 1998"
+url: "https://aclanthology.org/W98-1207/"
+---
+
+**Automation** **of Treebank** **Annotation**
+
+
+Thorsten Brants and Wojciech Skut
+
+Universit£t des Saarlandes
+Computational Linguistics
+D-66041 Saarbrficken, Germany
+```
+            {brant s, skut }Qcoli. uni-sb, de
+
+```
+
+**•..** **`Abstract`**
+
+
+This paper describes applications of stochastic
+and symbolic NLP methods to treebank annotation. In paxticular we focus on (1) the automation of treebank annotation, (2) the comparison of conflicting annotations for the same
+sentence and (3) the automatic detection of inconsistencies. These techniques are currently
+employed for building a German treebank.
+
+
+
+rnu's nachgedacht werden
+VMFIN VVPP VAINF
+
+must thought-over be
+
+_'it has to be thought over'_
+
+
+
+**1** Introduction
+
+The emergence of new statistical NLP methods increases the demand for corpora annotated with syntactic structures. The construction of such a corpus (a _treebank)_ is a time-consuming task that can
+hardly be carried out unless some annotation work is
+automated. Purely automatic annotation, however,
+is not reliable enough to be employed without some
+form of human supervision and hand-correction.
+This interactive annotation strategy requires tools
+for error detection and consistency checking.
+
+The present paper reviews our experience with the
+development of automatic annotation tools which
+are currently used for building a corpus of German
+newspaper text.
+
+The next section gives an overview of the annotation format. Section 3 describes three applications
+of statistical NLP methods to treebank annotation.
+Finally, section 4 discusses mechanisms for comparing structures assigned by different annotators.
+
+**2** Annotating Argument Structure
+
+2.1 Annotation Scheme
+
+Unlike most treebanks of English, our corpus is annotated with _predicate-argumenl s~ructures_ and not
+phrase-structure trees. The reason is the free word
+order in German, a feature seriously affecting the
+transparency of traditional phrase structures. Thus
+local and non-local dependencies are represented in
+
+
+
+Dar'uber
+
+**PROAV**
+```
+ about-it
+
+```
+
+
+$.
+
+
+
+Figure h Sample structure from the Treebank
+
+
+the same way, at the cost of allowing crossing tree
+branches, as shown in figure 11 .
+
+Such a direct representation of the predicateargument relation makes annotation easier than it
+would be if additional trace-filler co-references were
+used for encoding discontinuous constituents. Furthermore, our scheme facilitates automatic extraction of valence frames and the construction of semantic representations.
+
+On the other hand, the predicate-argument structures used for annotating our corpus can still be converted automatically into phrase-structure trees if
+necessary, cf. (Skut et al., 1997a). For more details
+on the annotation scheme v. (Skut et al., 1997b).
+
+
+2.2 The Annotation Mode
+
+In order to make annotation more reliable, each sentence is annotated independently by two annotators.
+Afterwards, the results are compared, and both annotators have to agree on a unique structure. In
+
+
+1 The nodes and edges are labeled with category and
+function symbols, respectively (see appendix A).
+
+
+
+_Brants and Skut_ 49 _Automation of Treebank Annotation_
+
+
+Thorsten Brants and Wojciech Skut (1998) **Automation of Treebank Annotation.** In D.M.W. _Powers_ (ed.)
+_NeMLaP3/CoNLL98: New Methods in Language Processing and Computational Natural Language Learning,_ ACL, pp 49-57.
+
+
+case of persistent disagreement or uncertainty, the
+grammarian supervising the annotation work is consulted.
+
+It has turned out that comparing annotations
+involves significantly more effort than annotation
+proper. As we do not want to abandon the annotateand-compare strategy, additional effort has been put
+into the development of tools supporting the comparison of annotated structures (see section 4).
+
+
+3 Automation
+
+
+
+(1)
+
+
+**Selbst** **besucht**
+
+ADV VVPP
+
+
+himself **visited**
+
+
+
+**hat** **Peter** **Sabine**
+
+**VAFIN** **NE** **NE**
+
+
+**has** **Peter** **Sabine**
+
+
+
+**never**
+
+
+
+**hie**
+## **ADV l**
+
+
+
+**hie**
+
+
+
+**ADV**
+
+
+
+The efficiency of annotation can be significantly increased by using automatic annotation tools. Nevertheless, some form of human supervision and handcorrection is necessary to ensure sufficient reliability. As pointed out by (Marcus, Santorini, and
+Marcinkiewicz, 1994), such a _semi-automatic_ annotation strategy turns out to be superior to purely
+manual annotation in terms of accuracy and efficiency. Thus in most treebank projects, the task of
+the annotators consists in correcting the output of a
+parser, cf. (Marcus, Santorini, and Marcinkiewicz,
+
+
+
+1994), (Black et al., 1996).
+
+
+
+As for our project, the unavailability of
+broad-coverage argument-structure and dependency
+parsers made us adopt a bootstrapping strategy.
+Having started with completely manual annotation,
+we axe gradually increasing the degree of automation. The corpus annotated so far serves as training material for annotation tools based on statistical NLP methods, and the degree of automation
+increases with the amount of annotated sentences.
+
+Automatic processing and manual input are combined interactively: the annotator specifies some information, another piece of information is added automatically, the annotator adds new information or
+corrects parts of the structure, new parts are added
+automatically, and so on. The size and type of such
+annotation increments depends on the size of the
+training corpus. Currently, manual annotation consists in specifying the hierarchical structure, whereas
+category and function labels as well as simple substructures are assigned automatically. These automation steps are described in the following sections.
+
+
+
+**3.1** **Tagging Grammatical Functions**
+
+Assigning grammatical functions to a given hierarchical structure is based on a generalization of standard paxt-of-speech tagging techniques.
+
+In contrast to a standard probabilistic POS tagger (e.g. (Cutting et al., 1992; Feldweg, 1995)), the
+tagger for grammatical functions works with lexical
+
+
+
+_'Peter never visited Sabine himself'_
+
+
+Figure 2: Example sentence
+
+
+and contextual probability measures _PO.(')_ depending on the category of a mother node (Q). This additional parameter is necessary since the sequence of
+grammatical functions depends heavily on the type
+of phrase in which it occurs. Thus each category (S,
+VP, NP, PP etc.) defines a separate Markov model.
+
+Under this perspective, categories of daughter
+nodes correspond to the outputs of a Markov model
+(i.e., like words in POS tagging). Grammatical functions can be viewed as states of the model, analogously to tags in a standard part-of-speech tagger.
+
+Given a sequence of word and phrase categories
+T = T1...Tk and a parent category Q, we calculate the sequence of grammatical functions G =
+G1... Gk that link T and Q as
+
+
+_axgmax PQ( a[T)_ (1)
+G
+
+
+_PQ(G). Po(TIG)_
+= axgmax
+_a_ _PQ(T)_
+
+= axgmaxPo(a ) • _Pq(TIG)_
+
+G
+
+
+Assuming the Markov property we have
+
+
+k
+_PQ(TIG) = I"I PQ(T~IG,)_ _(2)_
+
+i----1
+and (using a trigram model)
+
+
+k
+_**po(G) : IX Po(V, IGi-2, G,-1)**_ _**(a)**_
+
+i=1
+
+The contexts are smoothed by linear interpolation
+of unigrams, bigrams, and trigrams. Their weights
+are calculated by deleted interpolation (Brown et al.,
+1992).
+
+
+
+_Brants and Skut_ 50 _Automation of Treebank.4nnotation_
+
+
+The structure of a sample sentence is shown jn
+figure 2. Here, the probability of the S node having
+this particular sequence of children is calculated as
+
+
+_Ps(G,T)_ _=_ Ps(OCl$,$)-Ps(VPlOC)
+
+         - Ps(HDI$, OC). Ps(VAFINIHD)
+
+          - Ps(SBIOC, HD)-Ps(NEISB )
+
+-Ps(NGIHD, SB)- Ps(ADVING )
+
+($ indicates the start of the sequence).
+
+The predictions of the tagger are correct in approx. 94% of all cases. During the annotation process this is further increased by exploiting a precision/recall trade-off (cf. section 3.5).
+
+3.2 Tagging Phrasal Categories
+
+The second level of automation is the recognition of
+phrasal categories, which frees the annotator from
+typing phrase labels. The task is performed by an
+extension of the grammatical function tagger presented in the previous section.
+
+Recall that each phrasal category defines a different Markov model. Given the categories of the
+children nodes in a phrase, we can run these models
+in parallel. The model that assigns the most probable sequence of grammatical functions determines
+the category label to be assigned to the parent node.
+
+Formally, we calculate the phrase category Q (and
+at the same time the sequence of grammatical functions G = G1 ... Gk) on the basis of the sequence of
+daughters 7" = T1 ... Tk with
+
+
+
+E~
+
+
+em in Tel Aviv **lebender** Dichter
+ART APPIq NE NE ADJA NN
+
+**1** ---- -- 0 4--1. **+**
+
+a in Tel Aviv living poet
+
+_'a poet living in Tel Aviv'_
+
+
+Figure 3: Structural tags
+
+
+(Ratnaparkhi, 1997) uses an iterative procedure to
+assign two types of tags _(start X_ and _join X,_ where
+X denotes the type of the phrase) combined with a
+process to build trees.
+
+We go one step further and assign simple structures in one pass. Furthermore, the nodes and
+branches of these tree chunks have to be assigned
+category and function labels.
+
+The basic idea is to encode structures of limited depth using a finite number of tags. Given a
+sequence of words (w0, wl .... wn/, we consider the
+structural relation ri holding between _wi_ and _wi-1_
+for 1 < i < n. For the recognition of NPs and PPs,
+it is sufficient to distinguish the following seven values of rl which uniquely identify sub-structures of
+limited depth.
+
+
+
+argmax _maxPQ(GIT)._
+Q 6
+
+
+This procedure can also be performed using one
+large {combined) Markov model that enables a very
+efficient calculation of the maximum.
+
+The overall accuracy of this approach is 95%.
+
+3.3 Tagging Hierarchical **Structure**
+
+The next automation step is the recognition of syntactic structures. In general, this task is much more
+difficult than assigning category and function labels,
+and requires a significantly larger training corpus
+than the one currently available. What can be done
+at the present stage is the recognition of relatively
+simple structures such as NPs and PPs.
+
+(Church, 1988) used a simple mechanism to mark
+the boundaries of NPs. He used part-of-speech tagging and added two flags to the part-of-speech tags
+to mark the beginning and the end of an NP.
+
+Our goal is more ambitious in that we mark not
+only the phrase boundaries of NPs but also the complete structure of a wider class of phrases, starting
+with APs, NPs and PPs.
+
+
+
+If more than one of the conditions above are met,
+the first of the corresponding tags in the list is assigned. A structure tagged with these symbols is
+shown in figure 3.
+
+In addition, we encode the POS tag ti assigned to
+w~. On the basis of these two pieces of information
+we define _structural tags as_ pairs Si = (ri, _ti)._ Such
+
+
+
+ri -
+
+
+0 if _parent(wi) =parent(wi_l)_
++ if _parent(wi) =parent2(wi_l)_
+++ if _parent(wi) = parentZ(wi_a)_
+
+  - if _parent2(wi) =parent(wi_l)_
+-- if _parentZ(wi) = parent(wi_l)_
+= if _parent2(wi) = parentg-(wi_l)_
+1 else
+
+
+
+_Brants and Skut_ _51_ _Automation of Treebank Annotation_
+
+
+tags constitute a finite alphabet of symbols describing the structure and syntactic category of phrases
+of depth < 3.
+
+The task is to assign the most probable sequence
+of structural tags ((So, $1, ..., Sn)) to a sequence of
+part-of-speech tags _(To, T1, ..., Tn)._
+
+Given a sequence of part-of-speech tags T =
+T1 ... T~, we calculate the sequence of structural tags
+S = $1 ... Sk such that
+
+
+_argmax P( S]T)_ (4)
+s
+
+
+_P(S) - P(TIS)_
+= argmax
+_s_ _P(T)_
+
+= argmaxP(S). _P(TIS)_
+
+S
+
+
+The part-of-speech tags are encoded in the structural tag (t), so S uniquely determines T. Therefore,
+we have _P(T[S)_ = 1 ifTi = ii and 0 otherwise, which
+simplifies calculations:
+
+
+argmax _P(S). P(T[S)_ (5)
+s
+
+
+= argmax H _P(SiISi-2, Si-1)P(TdSO_
+
+S i=1
+
+
+As in the previous models, the contexts are
+smoothed by linear interpolation of unigrams, bigrams, and trigrams. Their weights are calculated
+by deleted interpolation.
+
+This _chunk tagging_ technique can be applied to
+treebank annotation in two ways. Firstly, we could
+use it as a preprocessor; the annotator would then
+complete and correct the output of the chunk tagger.
+
+The second alternative is to combine this chunking
+with manual input in an interactive way. Then the
+annotator has to determine the boundaries of the
+sub-structure that is to be build by the program.
+
+Obviously, the second solution is favorable since
+the user supplies information about chunk boundaries, while in the preprocessing mode the tagger has
+to find both the boundaries and the internal structure of the chunks.
+
+The assignment of structural tags is correct in
+more than 94% of the cases. For detailed results
+see section 3.6.3.
+
+
+**3.4** **Interaction and Alternation**
+
+To illustrate the interaction of manual input and the
+automatic annotation techniques described above,
+we show the way in which the structure in figure
+
+
+
+3 is constructed. The current version of the annotation tool supports automatic assignment of category
+and phrase labels, so the user has to specify the hierarchical structure step by step 2.
+
+The starting point is the plain string of words together with their part-of-speech tags. The annotator first selects the words _Tel Aviv_ and executes the
+command "group" (this is all done with the mouse).
+Then the program inserts the category label MPN
+(multi-lexeme proper noun) and assigns the grammatical function PNC (proper noun component) to
+both words (cf. sections 3.2 and 3.1).
+
+Having completed the first sub-structure, the annotator selects the newly created MPN and the
+preposition _in,_ and creates a new phrase. The
+tool automatically inserts the phrase label PP and
+the grammatical functions AC (adpositional case
+marker) and NK (noun kernel component). The following two steps are to determine the components
+of the AP and, finally, those of the NP.
+
+At any time, the annotator has the opportunity
+to change and correct entries made by the program.
+
+This interactive annotation mode is favorable
+from the point of view of consistency checking. The
+first reason is that the annotation increments are
+rather small, so the annotator corrects not an entire
+parse tree, but a fairly simple local structure. The
+automatic assignment of phrase and function labels
+is generally more reliable than manual input because
+it is free of typically human errors (see the precision
+results in (Brants, Skut, and Krenn, 1997)). Thus
+the annotator can concentrate on the more difficult
+task, i.e., building complex syntactic structures.
+
+The second reason is that errors corrected at
+lower levels in the structure facilitate the recognition of structures at higher levels, thus many wrong
+readings are excluded by confirming or correcting a
+choice at a lower level.
+
+The partial automation of the annotation process
+(automatic regocnition of phrase labels and grammatical functions) has reduced the average annotation time from about 10 to 1.5 - 2 minutes per
+sentence, i.e. 600 - 800 tokens per minute, which
+is comparable to the figures published by the creators of the Penn Treebank in (Marcus, Santorini,
+and Marcinkiewicz, 1994).
+
+The test version of the annotation tool using the
+statistical chunking technique described in section
+3.3 permits even larger annotation increments and
+we expect a further increase in annotation speed.
+The user just has to select the words constituting an
+
+
+~The chunk tagger has not yet been fully integrated
+into the annotation **tool.**
+
+
+
+_Brants and Skut_ 52 _Automation of Treebank Annotation_
+
+
+NP or PP. The program assigns a sequence of structural tags to them; these tags are then converted to
+a tree structure and all labels are inserted.
+
+
+3.5 tteliability
+
+To make automatic annotation more reliable, the
+program assigning labels performs an additional reliability check. We do not only calculate the best
+assignment, but also the second-best alternative and
+its probability. If the probability of the alternative
+comes very close to that of the best sequence of labels, we regard the choice as unreliable, and the annotator is asked for confirmation.
+
+Currently, we employ three reliability levels, expressed by quotients of probabilities _Pbest/Psecond-_
+If this quotient is close to one (i.e., smaller than
+some threshold 01), the decision counts as unreliable, and annotation is left to the annotator. If the
+quotient is very large (i.e., greater than some threshold 02 > 91), the decision is regarded as reliable and
+the respective annotation is made by the program.
+
+If the quotient fails between 91 and 02, the decision
+is tagged as "almost reliable". The annotation is
+inserted by the program, but has to be confirmed by
+the annotator.
+
+This method enables the detection of a number of
+errors that are likely to be missed if the annotator
+is not asked for confirmation.
+
+The results of using these reliability levels are reported in the experiments section below.
+
+**3.6** **Experiments**
+
+This section reports on the accuracy achieved by the
+methods described in the previous sections.
+
+At present, our corpus contains approx. 6300 sentences (115,000 tokens) of German newspaper text
+(Frankfurter Rundschan). Results of tagging grammatical functions and phrase categories have improved slightly compared to those reported for a
+smaller corpus of approx. 1200 sentences (Brants,
+Skut, and Krenn, 1997). Accuracy figures for tagging the hierarchical structure are published for the
+first time.
+
+For each experiment, the corpus was divided into
+two disjoint parts: 90% training data and 10% test
+data. This procedure was repeated ten times, and
+the results were averaged.
+
+The thresholds 01 and 02 determining the reliability levels were set to 91 = 5 and 02 = 100.
+```
+3.6.1 Grammatical Functions
+ We employ the technique described in section 3.1
+to assign grammatical functions to a structure defined by an annotator. Grammatical functions are
+
+```
+
+
+Table 1: Levels of reliability and the percentage of
+cases in which the tagger assigned a correct grammatical function (or would have assigned ifa decision
+had been forced).
+
+
+
+_grammatical_
+_function_
+
+reliable
+marked
+unreliable
+overall
+
+
+```
+cases correct
+```
+
+88% 97.0%
+8% 85.0%
+4% 59.5%
+100% 94.6%
+
+
+
+Table 2: Levels of reliability and the percentage of
+cases in which the tagger assigned a correct phrase
+category (or would have assigned it if a decision had
+been forced).
+
+
+
+_phrase_
+
+_category_
+
+reliable
+marked
+unreliable
+overall
+
+
+```
+cases correct
+```
+
+76% 99.0%
+19% 91.5%
+5% 56.7%
+100% 95.4%
+
+
+
+represented by edge labels. Additionally, we exploit
+the recall/accuracy tradeoff as described in section
+3.5. The tagset of grammatical functions consists of
+45 tags.
+
+Tagging results are shown in table 1. Overall accuracy is 94.6%. 88% of all predictions are classified
+as reliable, which is the most important class for
+the actual annotation task. Accuracy in this class
+is 97.0%. It depends on the category of the phrase,
+e.g. accuracy for reliable cases reaches 99% for 51Ps
+and PPs.
+
+**3.6.2** **Phrasal Categories**
+
+Now the task is to assign phrasal categories to a
+structure specified by the annotator, i.e., only the
+hierarchical structure is given. We employ the technique of competing Markov models as described in
+```
+section 3.2 to assign phrase categories to the structure. Additionally, we compute alternatives to assign one of the three reliability levels to each decision
+as described in section 3.5. The tagset for phrasal
+categories consists of 25 tags.
+ As can be seen from table 2, the results of assigning phrasal categories are even better than those of
+assigning grammatical functions. Overall accuracy
+is 95.4%. Tags that are regarded as reliable (76% of
+all cases) have an accuracy of 99.0%, which results
+
+```
+
+
+_Brants and Skut_ 53 _Automation of Treebank Annotation_
+
+
+Table 3: Chunk tagger accuracy with respect to hierarchical structure.
+
+
+
+proc compare(A, B)
+```
+  for each non-terminal node X in A:
+   search node Y in B
+    such that yield(X) = yield(Y)
+```
+
+if `Y exists:`
+```
+    emit different labels if any
+   if Y does not exist:
+    emit X and its yield
+  end
+end
+
+```
+
+Figure 4: Basic asymmetric algorithm to compare
+annotation A with annotation B of the same sentence
+
+
+(Calder, 1997) presents a method of comparing
+the structure of context free trees found in different annotations. This section presents an extension
+of this algorithm that compares predicate-argument
+structures possibly containing crossing branches (cf.
+figure 2). Node and edge labels, representing phrasal
+categories and grammatical functions, are also taken
+into account.
+
+Phrasal (non-terminal) nodes are compared on
+the basis of their yields: the yield of a nonterminal node X in an annotation A is the ordered set
+of terminals that are (directly or indirectly) dominated by X. The yield need not be contiguous
+since predicate-argument structures allow discontinuous constituents.
+
+If both annotations contain nonterminal nodes
+that cover the same terminal nodes, the labels of the
+nonterminal nodes and their edges are compared.
+
+This results in a combined measure of structural
+and labeling differences, which is very useful in
+cleaning the corpus and keeping track of the development of the treebank.
+
+We use the basic algorithm shown in figure 4 to
+determine the differences in two annotations A and
+B. The basic form is asymmetric. Therefore, a complete comparison consists of two runs, one for each
+direction, and the outputs of both runs are combined.
+
+Figures 5 and 6 show examples of the output of
+the algorithm. These outputs can be directly used
+to mark the corresponding nodes and edges.
+
+The yield is sufficient to uniquely determine corresponding nodes since the annotations used here do
+not contain unary branching nodes. If unary branching occurs, both the parent and the child have the
+same terminal yield and further mechanism to determine corresponding nodes are needed. (Calder,
+
+1997) points out possible solutions to this problem.
+
+
+
+_structural_
+_tags_
+
+
+reliable
+marked
+unreliable
+overall
+
+
+
+cases correct
+
+
+86% 95.8%
+11% 93.2%
+3% 67.0%
+100% 94.4%
+
+
+
+in very reliable annotations.
+
+
+3.6.3 Chunk Tagger
+
+The chunk tagger described in section 3.3 assigns
+tags encoding structural information to a sequence
+of words and tags. The accuracy figures presented
+here refer to the correct assignments of these tags
+(see table 3).
+
+The assignment of structural tags allows us to construct a tree; the labels are afterwards assigned in
+a bottom-up fashion by the function/category label
+tagger described in earlier sections.
+
+Overall accuracy is 94.4% and reaches 95.8% in
+the reliable cases.
+
+A different measure of the chunker's correctness is
+the percentage of _complete phrases_ recognized correctly. In order to determine this percentage, we
+extracted all chunks of the maximal depth recognizable by the chunker. In a cross evaluation, 87.3% of
+these chunks were recognized correctly as far as the
+hierarchical structure is concerned.
+
+
+4 Comparing Trees
+
+
+Annotations produced by different annotators are
+compared automatically and differences are marked.
+The output of the comparison is given to the annotators. First, each of the annotators goes through
+the differences on his own and corrects obvious errors. Then remaining differences are resolved in a
+discussion of the annotators.
+
+Additionally, the program calculates the probabilities of the two different annotations. This is intended to be a first step towards resolving conflicting
+annotations automatically. Both parts, tree matching and the calculation of probabilities for complete
+trees are described in the following sections.
+
+
+4.1 Tree Matching
+
+The problem addressed here is the comparison of
+two syntactic structures that share identical terminal nodes (the words of the annotated sentence).
+
+
+
+_Brants and Skut_ 54 _Automation of Treebank Annotation_
+
+
+# **/2 t- t**
+
+**Selbst** **besucht** **hat** **Peter**
+0 1 2 3
+ADV VVPP VAFIN NE
+
+
+### ++
+
+**Sabine** **nie**
+4 5
+NE ADV
+
+
+
+4.2 Probabilities
+
+The probabilities of each sub-structure of depth one
+are calculated separately according to the model described in sections 3.1 and 3.2. Subsequently, the
+product of these probabilities is used as a scoring
+function for the complete structure• This method is
+based on the assumption that productions at different levels in a structure are independent, which is
+inherent to context free rules.
+
+Using the formulas from sections 3.1 and 3.2, the
+probability _P(A)_ of an annotation A is evaluated as
+
+
+_P(A) = HP(Qi)_
+
+_i=l_
+_nnt_
+**=** **H PQ,(T,, G,)**
+i--1
+
+_rtnt ki_
+_**= 1] 1]**_ _Poi(gi,ylg,,~-2,_ g/,~-l)
+
+i=lj=l
+
+**• PQ(ti,y I gij)**
+
+
+
+himself visited has Peter **Sabine**
+
+_'Peter never visited Sabine himself'_
+
+```
+sentence I errors I
+(1) structure: 500 VP lOCI 0 1 4
+
+```
+
+
+never
+
+
+```
+          (Selbst besucht Sabine)
+ (2) structure: 500 VP lOCI 0 I 4 S
+          (Selbst besucht Sabine hie)
+
+```
+
+Figure 5: Erroneous annotation (2) of the example
+sentence in figure 2 _(hie_ should be attached to S
+instead of VP), together with the output of the tree
+comparison algorithm• All nodes are numbered to
+enable identification. Additionally, this output can
+be used to highlight the corresponding nodes and
+edges.
+
+
+
+(3)
+
+
+**Selbst**
+
+0
+ADV
+
+
+himself
+
+
+
+**besucht** **hat** **Peter** **Sabine**
+1 2 3 4
+VVPP VAFIN NE NE
+
+
+**visited** has Peter Sabine
+
+_'Peter never visited Sabine himself'_
+
+
+# **?**
+
+
+
+**nie**
+
+
+
+5
+ADV
+
+
+
+A annotation (structure) for a sentence
+_nnt_ number of nonterminal nodes in A
+_nt_ number of terminal nodes in A
+n number of nodes = _nnt + nt_
+Qi ith phrase in A
+T/ sequence of tags in _Qi_
+_Gi_ sequence of gramm, func. in Qi
+ki number of elements in Qi
+_tij_ tag of jth child in Qi
+_gl,i_ grammatical function of jth
+child in _Qi_
+Probabilities computed in this way cannot be used
+directly to compare two annotations since they favor
+annotations with fewer nodes. Each new nonterminal node introduces a new element in the product
+and makes it smaller.
+
+Therefore, we normalize the probabilities w.r.t.
+the number of nodes in the annotation, which yields
+the perplexity _PP(A)_ of an annotation A:
+
+
+
+never
+
+
+
+**`sentence`** **1** **`errors`** **1**
+```
+(1) edge: 5 (ADV) [NG] nie
+(3) edge: 5 (ADV) [MO] hie
+
+```
+
+Figure 6: Erroneous annotation (3) of the example
+sentence in figure 2 _(nie_ should have grammatical
+function NG instead of MO), together with the output of the tree comparison algorithm.
+
+
+
+_PP(A)=~'p~A)_ (6)
+
+
+4.3 Application to a Corpus
+
+The procedures of tree matching and probability calculation were applied to our corpus, which currently
+consists of approx. 6300 sentences (115,000 tokens)
+of German newspaper text, each sentence annotated
+at least twice.
+
+We measured the agreement of independent annotations after first annotation but before correction
+
+
+
+_Brants and Skut_ 55 _Automation of Treebank Annotation_
+
+
+Table 4: Comparison of independent semi-automatic
+annotations (1) after first, independent annotation
+and (2) after comparison but before the final discussion (current stage).
+
+
+
+Table 5: Using model perplexities to compare different annotations: Accuracy of using the hypothesis that a correct annotation has a lower perplexity
+than a wrong annotation.
+
+recall precision
+30% 95.3%
+45% 92.2%
+60% 88.6%
+85% 81.4%
+100% 65.8%
+
+
+the whole annotation to be wrong and the sentence
+counts as an error.
+
+If we make the assumption that a correct annotation always has a lower perplexity than a wrong
+annotation for the same sentence, the system would
+make a correct decision for 65.8% of the sentences
+(see table 5, last row).
+
+For approx. 70% of all sentences, at least one
+of the initial annotations was completely correct.
+This means that the two initial annotations and the
+automatic comparison yield a corpus with approx.
+65.8% × 70% = 46% completely correct annotations
+(complete structure and all tags).
+
+One can further increase precision at the cost of
+recall by requiring the difference in perplexity to exceed some minimum distance. This precision/recall
+tradeoff is also shown in table 5.
+
+
+5 Conclusion
+
+The techniques and automatic tools described in this
+paper are designed to support annotation proper,
+online/offline consistency checking and the comparison of independent annotations of the same sentences. Most of the techniques employ stochastic
+processing methods, which guarantee high accuracy
+and robustness.
+
+The bootstrapping approach adopted in our
+project makes the degree of automation a function of
+available training data. Easier processing tasks are
+automated first. Experience gained and data annotated at a lower level allow to increase the level of
+automation step by step. The current size of our
+corpus (approx. 6300 sentences) enables reliable automatic assignment of category and function labels
+as well as simple structures.
+
+Future work will be concerned with developing
+automatic annotation methods handling complex
+structures, which should ultimately lead to the development of a parser for predicate-argument trees
+containing crossing branches.
+
+
+
+word level:
+(1) ident, parent node
+(2) ident, gram. func.
+node level:
+(3) identical nodes
+(4) identical nodes/labels
+(5) ident, node/gram, func.
+**sentence** level:
+(6) identical structure
+(7) identical annotation
+
+
+
+92.3% 98.7%
+93.8% 99.1%
+
+
+87.6% 98.1%
+84.2% 97.4%
+76.6% 96.3%
+
+
+48.6% 90.8%
+34.6% 87.9%
+
+
+
+**<1>** **(2>**
+
+
+
+(1), and after correction but before the final discussion (2), which is the current stage of the corpus.
+The results are shown in table 4.
+
+As for measuring differences, we can count them
+at word, node and sentence level.
+
+
+
+At the word level, we are interested in (1) the
+number of correctly assigned parent categories (does
+a word belong to a PP, NP, etc.?), and (2) the number of correctly assigned grammatical functions (is
+a word a head, modifier, subject, etc.?).
+
+At the node level (non-terminals, phrases) we
+measure (3) the number of identical nodes, i.e., if
+there is a node in one annotation, we check whether
+it corresponds to a node in the other annotation having the same yield. Additionally, we count (4) the
+number of identical nodes having the same phrasal
+category, and (5) the number of identical nodes having the same phrasal category and the same grammatical function within its parent phrase.
+
+
+
+At the sentence level, we measure (6) the number
+of annotated sentences having the same structure,
+and, which is the strictest measure, (7) the number
+of sentences having the same structure and the same
+labels (i.e., exactly the same annotation).
+
+At the node level, we find 87.6% agreement in independent annotations. A large amount of the differences come from misinterpretation of the annotation guidelines by the annotators and are eliminated
+after comparison, which results in 98.1% agreement.
+This kind of comparison is the one most frequently
+used in the statistical parsing community for comparing parser output.
+
+The sentence level is the strictest measure, and
+the agreement is low (34.6% identical annotations
+after first annotation, 87.9% after comparison). But
+at this level, one error (e.g. a wrong label) renders
+
+
+
+_Brants and Skut_ 56 _Automation of Treebank Annotation_
+
+
+6 Acknowledgements
+
+This work is part of the DFG Sonderforschungsbereich 378 _Resource-Adaptive Cognitive Processes,_
+Project C3 _Concurrent'Grammar Processing._
+
+We wish to thank the universities of Stuttgart
+and Tiibingen for kindly providing us with a handcorrected part-of-speech tagged corpus. We also
+wish to thank Oliver Plaehn, who did a great job
+in implementing the annotation tool, and Peter
+Sch~ifer, who built the tree comparison tool. Special
+thanks go to the five annotators continually increasing the size and the quality of our corpus. And finally, we thank Sabine Kramp for proof-reading this
+paper.
+
+
+**References**
+
+Black, Ezra, Stephen Eubank, Hideki Kashioka,
+
+David Magerman, Roger Garside, and Geoffrey
+Leech. 1996. Beyond skeleton parsing: Producing
+a comprehensive large-scale general-English treebank with full grammatical analysis. In _Proc. of_
+
+_COLING-96,_ pages 107-113, Kopenhagen, Denmark.
+
+
+```
+Appendix A: Tagsets
+
+This section contains descriptions of tags used in this
+paper. These are not complete lists.
+```
+
+A.1 Part-of-Speech Tags
+
+We use the Stuttgart-T/ibingen-Tagset. The complete set is described in (Thielen and Schiller, 1995).
+
+ADJA `attributive adjective`
+```
+ ADV adverb "
+ APPR preposition
+ ART article
+```
+
+NE proper noun
+NN `common noun`
+```
+ PROAV pronominal adverb
+ VAFIN finite auxiliary
+ VAINF infinite auxiliary
+ VMFIN finite modal verb
+ VVPP past participle of main verb
+
+A.2 Phrasal Categories
+ AP adjective phrase
+ MPN multi-word proper noun
+ NP noun phrase
+ PP prepositional phrase
+```
+
+S `sentence`
+`VP` verb phrase
+
+A.3 Grammatical Functions
+
+AC adpositional case marker
+HD head
+MO modifier
+NG negation
+NK noun kernel
+OA accusative object
+OC dausal object
+PNC proper noun component
+SB `subject`
+
+
+
+Ratnaparkhi, Adwait. 1997. `A` linear observed
+time statistical parser based on maximum entropy
+models." In _Proceedings of EMNLP.97,_ Providence, RI, USA.
+
+Skut, Wojciech, Thorsten Brants, Brigitte Krenn,
+
+and Hans Uszkoreit. 1997a. Annotating unrestricted German text. In _FacMagung der Sektion_
+
+_Computerlinguistik der Deutschen Gesellschaft_
+_fffr Sprachwissenschafl,_ Heidelberg, Germany.
+
+
+
+Skut, Wojciech, Brigitte Krenn, Thorsten Brants,
+
+
+
+and Hans Uszkoreit. 1997b. An annotation
+scheme for free word order languages. In _Proceed-_
+
+
+
+_ings of ANLP-97,_ Washington, DC.
+
+
+
+Thielen, Christine and Anne Schiller. 1995. Ein
+
+kleines und erweitertes Tagset f/its Deutsche.
+In _Tagungsberichte des Arbeitstreffens Lezikon_
+
+_+ Text 17./18. Februar 1994, Schlofl Hohen-_
+_t(~bingen. Lezicographica Series Maior,_ Tfibingen.
+Niemeyer.
+
+
+
+Brants, Thorsten, Wojciech Skut, and Brigitte
+
+Krenn. 1997. Tagging grammatical functions. In
+_Proceedings of EMNLP-97,_ Providence, RI, USA.
+
+
+
+Brown, P. F., V. J. Della Pietra, Peter V. deSouza,
+
+Jenifer C. Lai, and Robert L. Mercer. 1992. Classbased n-gram models of natural language. _Com-_
+_putational Linguistics,_ 18(4):467-479.
+
+
+
+Calder, Jo. 1997. On aligning trees. In _Proc. of_
+
+_EMNLP-97,_ Providence, RI, USA.
+
+
+
+Church, Kenneth Ward. 1988. A stochastic parts
+
+program and noun phrase parser for unrestricted
+text. In _Proc. Second Conference on Applied Nat-_
+_ural Language Processing,_ pages 136-143, Austin,
+Texas, USA.
+
+
+
+Cutting, Doug, Julian Kupiec, Jan Pedersen, and
+
+Penelope Sibun. 1992. A practical part-of-speech
+tagger. In _Proceedings of the 3rd Conference_
+_on Applied Natural Language Processing (ACL),_
+pages 133-140.
+
+
+
+Feldweg, Helmut. 1995. Implementation and eval
+uation of a german hmm for pos disambiguation.
+In _Proceedings of EACL-SIGDAT-95 Workshop,_
+Dublin, Ireland.
+
+
+
+Marcus, Mitchell, Beatrice Santorini, and Mary Ann
+
+Marcinkiewicz. 1994. Building a large annotated
+corpus of English: the Penn Treebank. In Susan Armstrong, editor, _Using Large Corpora._ MIT
+Press.
+
+
+
+_Bran ts and Skut_ _5 7_ _Automation of Treebank,4 nnotation_
+
+

references/2004.cl.hwa/paper.md

@@ -0,0 +1,1954 @@
+---
+title: "Sample Selection for Statistical Parsing"
+authors:
+  - "Rebecca Hwa"
+year: 2004
+venue: "Computational Linguistics 30(3)"
+url: "https://aclanthology.org/J04-3001/"
+---
+
+# **Sample Selection for Statistical Parsing**
+
+Rebecca Hwa _[∗]_
+
+University of Pittsburgh
+
+
+_Corpus-based_ _statistical_ _parsing_ _relies_ _on_ _using_ _large_ _quantities_ _of_ _annotated_ _text_ _as_ _training_
+_examples. Building this kind of resource is expensive and labor-intensive. This work proposes to_
+_usesampleselectiontofindhelpfultrainingexamplesandreducehumaneffortspentonannotating_
+_less informative ones. We consider several criteria for predicting whether unlabeled data might_
+_be_ _a_ _helpful_ _training_ _example._ _Experiments_ _are_ _performed_ _across_ _two_ _syntactic_ _learning_ _tasks_
+_and within the single task of parsing across two learning models to compare the effect of different_
+_predictive_ _criteria._ _We_ _find_ _that_ _sample_ _selection_ _can_ _significantly_ _reduce_ _the_ _size_ _of_ _annotated_
+_training corpora and that uncertainty is a robust predictive criterion that can be easily applied to_
+_different learning models._
+
+
+**1.** **Introduction**
+
+
+Many learning tasks for natural language processing require supervised training; that
+is, the system successfully learns a concept only if it has been given annotated training data. For example, while it is difficult to induce a grammar with raw text alone,
+the task is tractable when the syntactic analysis for each sentence is provided as a
+part of the training data (Pereira and Schabes 1992). Current state-of-the-art statistical parsers (Collins 1999; Charniak 2000) are all trained on large annotated corpora
+such as the Penn Treebank (Marcus, Santorini, and Marcinkiewicz 1993). However,
+supervised training data are difficult to obtain; existing corpora might not contain the
+relevant type of annotation, and the data might not be in the domain of interest. For
+example, one might need lexical-semantic analyses in addition to the syntactic analyses in the treebank, or one might be interested in processing languages, domains,
+or genres for which there are no annotated corpora. Because supervised training demands significant human involvement (e.g., annotating the syntactic structure of each
+sentence by hand), creating a new corpus is a labor-intensive and time-consuming endeavor. The goal of this work is to minimize a system’s reliance on annotated training
+data.
+One promising approach to mitigating the annotation bottleneck problem is to
+use **sample** **selection**, a variant of active learning. Sample selection is an interactive
+learning method in which the machine takes the initiative in selecting unlabeled data
+for the human to annotate. Under this framework, the system has access to a large pool
+of unlabeled data, and it has to predict how much it can learn from each candidate in
+the pool if that candidate is labeled. More quantitatively, we associate each candidate
+in the pool with a **training** **utility** **value** (TUV). If the system can accurately identify
+the subset of examples with the highest TUV, it will have located the most beneficial
+
+
+_∗_ Computer Science Department, Pittsburgh, PA 15260. E-mail: hwa@cs.pitt.edu.
+
+
+Submission received: 14 October 2002; Revised submission received: 30 September 2003; Accepted for
+publication: 22 December 2003
+
+
+_⃝_ c 2004 Association for Computational Linguistics
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+training examples, thus freeing the annotators from having to label less informative
+examples.
+In this article, we apply sample selection to two syntactic learning tasks: training
+a prepositional-phrase attachment (PP-attachment) model and training a statistical
+parsing model. We are interested in addressing two main questions. First, what are
+good predictors of a candidate’s training utility? We propose several predictive criteria
+and define evaluation functions based on them to rank the candidates’ utility. We
+have performed experiments comparing the effect of these evaluation functions on
+the size of the training corpus. We find that, with a judiciously chosen evaluation
+function, sample selection can significantly reduce the size of the training corpus. The
+second main question is: Are the predictors consistently effective for different types of
+learners? We compare the predictive criteria both across tasks (between PP-attachment
+and parsing) and within a single task (applying the criteria to two parsing models:
+an expectation-maximization-trained parser and a count-based parser). We find that
+the learner’s uncertainty is a robust predictive criterion that can be easily applied to
+different learning models.
+
+
+**2.** **Learning** **with** **Sample** **Selection**
+
+
+Unlike traditional learning systems that receive training examples indiscriminately,
+a sample selection learning system actively influences its own progress by choosing
+new examples to incorporate into its training set. There are two types of selection algorithms: **committee-based** and **single** **learner** . A committee-based selection algorithm
+works with multiple learners, each maintaining a different hypothesis (perhaps pertaining to different aspects of the problem). The candidate examples that lead to the
+most disagreements among the different learners are considered to have the highest
+TUV (Cohn, Atlas, and Ladner 1994; Freund et al. 1997). For computationally intensive
+problems, such as parsing, keeping multiple learners may be impractical.
+In this work, we focus on sample selection using a single learner that keeps one
+working hypothesis. Without access to multiple hypotheses, the selection algorithm
+can nonetheless estimate the TUV of a candidate. We identify the following three
+classes of predictive criteria:
+
+
+1. _Problem-space:_ Knowledge about the problem space may provide
+information about the type of candidates that are particularly plentiful or
+difficult to learn. This criterion focuses on the general attributes of the
+learning problem, such as the distribution of the input data and
+properties of the learning algorithm, but it ignores the current state of
+the hypothesis.
+
+
+2. _Performance of the hypothesis:_ Testing the candidates on the current
+working hypothesis shows the type of input data on which the
+hypothesis may perform weakly. That is, if the current hypothesis is
+unable to label a candidate or is uncertain about it, then the candidate
+might be a good training example (Lewis and Catlett 1994). The
+underlying assumption is that an uncertain output is likely to be wrong.
+
+
+3. _Parameters of the hypothesis:_ Estimating the potential impact that the
+candidates will have on the parameters of the current working
+hypothesis locates those examples that will change the current
+hypothesis the most.
+
+
+254
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+_U_ is a set of unlabeled candidates.
+_L_ is a set of labeled training examples.
+_C_ is the current hypothesis.
+**Initialize:**
+_C ←_ _Train_ ( _L_ ).
+**Repeat**
+_N_ _←_ _Select_ ( _n_, _U_, _C_, _f_ ).
+_U ←_ _U −_ _N_ .
+_L ←_ _L ∪_ _Label_ ( _N_ ).
+_C ←_ _Train_ ( _L_ ).
+**Until** ( _C_ is good enough) **or** ( _U_ = _∅_ ) **or** (cutoff).
+
+
+**Figure** **1**
+Pseudo code for the sample selection learning algorithm.
+
+
+Figure 1 outlines the single-learner sample selection training loop in pseudocode.
+Initially, the training set, _L_, consists of a small number of labeled examples, based on
+which the learner proposes its first hypothesis of the target concept, _C_ . Also available
+to the learner is a large pool of unlabeled training candidates, _U_ . In each training
+iteration, the selection algorithm, _Select_ ( _n_, _U_, _C_, _f_ ), ranks the candidates of _U_ according
+to their expected TUVs and returns the _n_ candidates with the highest values. The
+algorithm predicts the TUV of each candidate, _u_ _∈_ _U_, with an evaluation function,
+_f_ ( _u_, _C_ ). This function may rely on the hypothesis concept _C_ to estimate the utility of a
+candidate _u_ . The _n_ chosen candidates are then labeled by human experts and added
+to the existing training set. Running the learning algorithm, _Train_ ( _L_ ), on the updated
+training set, the system proposes a new hypothesis regarding the target concept that
+is the most compatible with the examples seen thus far. The loop continues until one
+of three stopping conditions is met: The hypothesis is considered to perform well
+enough, all candidates are labeled, or an absolute cutoff point is reached (e.g., no
+more resources).
+
+
+**3.** **Sample** **Selection** **for** **Prepositional-Phrase** **Attachment**
+
+
+One common source of structural ambiguities arises from syntactic constructs in which
+a prepositional phrase might be equally likely to modify the verb or the noun preceding it. Researchers have proposed many computational models for resolving PPattachment ambiguities. Some well-known approaches include rule-based models (Brill
+and Resnik 1994), backed-off models (Collins and Brooks 1995), and a maximumentropy model (Ratnaparkhi 1998). Following the tradition of using learning PPattachment as a way to gain insight into the parsing problem, we first apply sample
+selection to reduce the amount of annotation used in training a PP-attachment model.
+We use the Collins-Brooks model as the basic learning algorithm and experiment with
+several evaluation functions based on the types of predictive criteria described earlier.
+Our experiments show that the best evaluation function can reduce the number of
+labeled examples by nearly half without loss of accuracy.
+
+
+**3.1** **A** **Summary** **of** **the** **Collins-Brooks** **Model**
+The Collins-Brooks model takes prepositional phrases and their attachment classifications as training examples: each is represented as a quintuple of the form ( _v_, _n_, _p_, _n_ 2, _a_ ),
+where _v_, _n_, _p_, and _n_ 2 are the head words of the verb phrase, the object noun phrase, the
+
+
+255
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+
+
+
+
+
+
+
+
+
+
+
+**Figure** **2**
+The Collins-Brooks PP-attachment classification algorithm.
+
+
+preposition, and the prepositional noun phrase, respectively, and _a_ specifies the attachment classification. For example, ( _wrote a book in three days, attach-verb_ ) would be annotated as ( _wrote_, _book_, _in_, _days_, _verb_ ). The head words can be automatically extracted using
+a heuristic table lookup in the manner described by Magerman (1994). For this learning
+problem, the supervision is the one-bit information of whether _p_ should attach to _v_ or
+to _n_ . In order to learn the attachment preferences of prepositional phrases, the system
+builds attachment statistics for each the **characteristic tuple** of all training examples. A
+characteristic tuple is some subset of the four head words in the example, with the condition that one of the elements must be the preposition. Each training example forms
+eight characteristic tuples: ( _v_, _n_, _p_, _n_ 2), ( _v_, _n_, _p_ ), ( _v_, _p_, _n_ 2), ( _n_, _p_, _n_ 2), ( _v_, _p_ ), ( _n_, _p_ ), ( _p_, _n_ 2), ( _p_ ).
+The attachment statistics are a collection of the occurrence frequencies for all the characteristic tuples in the training set and the occurrence frequencies for the characteristic
+tuples of those examples determined to attach to nouns. For some characteristic tuple
+_t_, _Count_ ( _t_ ) denotes the former and _Count_ NP( _t_ ) denotes the latter. In terms of the sample
+selection algorithm, the collection of counts represents the learner’s current hypothesis
+( _C_ in Figure 1). Figure 2 provides the pseudocode for the _Train_ routine.
+Once trained, the system can be used to classify test cases based on the statistics of
+the most similar training examples and back off as necessary. For instance, to determine
+the PP-attachment for a test case, the classifier would first consider the ratio of the
+two frequency counts for the four-word characteristic tuple of the test case. If the tuple
+
+
+256
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+**Figure** **3**
+In this example, the classification of the test case preposition is backed off to the
+two-word-tuple level. In the diagram, each circle represents a characteristic tuple. A filled
+circle denotes that the tuple has occurred in the training set. The dashed rectangular box
+indicates the back-off level on which the classification is made.
+
+
+never occurred in the training example, the classifier would then back off to look at the
+test case’s three three-word characteristic tuples. It would continue to back off further,
+if necessary. In the case that the model has no information on any of the characteristic
+tuples of the test case, it would, by default, classify the test case as an instance of noun
+attachment. Figure 3 shows using the back-off scheme on a test case. We describe in
+the _Test_ pseudocode routine in Figure 2 the model’s classification procedure for each
+back-off level.
+
+
+**3.2** **Evaluation** **Functions**
+Based on the three classes of predictive criteria discussed in Section 2, we propose
+several evaluation functions for the Collins-Brooks model.
+
+
+**3.2.1 The Problem Space.** One source of knowledge to exploit is our understanding of
+the PP-attachment model and properties of English prepositional phrases. For instance,
+we know that the most problematic test cases for the PP-attachment model are those
+for which it has no statistics at all. Therefore, those data that the system has not
+yet encountered might be good candidates. The first evaluation function we define,
+_f_ novel( _u_, _C_ ), equates the TUV of a candidate _u_ with its degree of novelty, the number
+of its characteristic tuples that currently have zero counts: [1]
+
+
+
+
+   _f_ novel( _u_, _C_ ) =
+
+_t∈Tuples_ ( _u_ )
+
+
+
+
+1 : _Count_ ( _t_ ) = 0
+0 : otherwise
+
+
+
+This evaluation function has some blatant defects. It may distort the data distribution
+so much that the system will not be able to build up a reliable collection of statistics.
+The function does not take into account the intuition that those data that rarely occur,
+no matter how novel, probably have overall low training utility. Moreover, the scoring
+scheme does not make any distinction between the characteristic tuples of a candidate.
+
+
+1 Note that the current hypothesis _C_ is ignored in evaluation functions of this class because they depend
+only on the knowledge about the problem space.
+
+
+257
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+|had|Col2|
+|---|---|
+|idea<br>on,<br>on,<br>on,<br>on,<br>on,<br>**topic**<br><br>idea<br>on,<br>topic<br>topic<br>idea<br>had<br>on,<br>had<br>idea<br>had<br>topic|idea<br>on,<br>on,<br>on,<br>on,<br>on,<br>**topic**<br><br>idea<br>on,<br>topic<br>topic<br>idea<br>had<br>on,<br>had<br>idea<br>had<br>topic|
+|on,|on,|
+|||
+|noun|noun|
+
+
+
+u1, u2 u3 u4 u5
+
+
+**Figure** **4**
+If candidate _u_ 1 is selected, a total of 22 tuples can be ignored. The dashed rectangles show the
+classification level before training, and the solid rectangles show the classification level after
+the statistics of _u_ 1 have been taken. The obviated tuples are represented by the filled black
+circles.
+
+
+We know, however, that the PP-attachment classifier is a back-off model that makes
+its decision based first on statistics of the characteristic tuple with the most words. A
+more sophisticated sampling of the data domain should consider not only the novelty
+of the data, but also the frequency of its occurrence, as well as the quality of its characteristic tuples. We define a back-off-model-based evaluation function, _f_ backoff( _u_, _C_ ), that
+scores a candidate _u_ by counting the number of characteristic tuples that would be
+obviated in all candidates if _u_ were included in the training set. For example, suppose
+we have a small pool of five candidates, and we are about to pick the first training
+example:
+
+
+_u_ 1 = (put, book, on, shelf)
+_u_ 2 = (put, book, on, shelf)
+_u_ 3 = (put, idea, on, shelf)
+_u_ 4 = (wrote, book, on, shelf)
+_u_ 5 = (had, idea, on, topic)
+
+
+According to _f_ backoff, either _u_ 1 or _u_ 2 would be the best choice. By selecting either as
+the first training example, we could ignore all but the four-word characteristic tuple
+for both _u_ 1 and _u_ 2 (a saving of seven tuples each); since _u_ 3 and _u_ 4 each have three
+words in common with the first two candidates, they would no longer depend on
+their lower four tuples; and although we would also improve the statistics for one
+of _u_ 5’s tuples ( _on_ ), nothing could be pruned from _u_ 5’s characteristic tuples. Thus,
+_f_ backoff( _u_ 1, _C_ ) = _f_ backoff( _u_ 2, _C_ ) = 7 + 7 + 4 + 4 = 22 (see Figure 4).
+Under _f_ backoff, if _u_ 1 were chosen as the first example, _u_ 2 would lose all its utility,
+because we could not prune any extra characteristic tuples by using _u_ 2. That is, in
+the next round of selection, _f_ backoff( _u_ 2, _C_ ) = 0. Candidate _u_ 5 would be the best second
+example because it would now have the most tuples to prune (7 tuples).
+The evaluation function _f_ backoff improves upon _f_ novel in two ways. First, novel candidates that occur frequently are favored over those that rarely come up. As we have
+seen in the above example, a candidate that is similar to other candidates can eliminate more characteristic tuples all at once. Second, the evaluation strategy follows the
+working principle of the back-off model and discounts lower-level characteristic tuples
+that do not affect the classification process, even if they were “novel.” For instance,
+
+
+258
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+after selecting _u_ 1 as the first training example, we would no longer care about the
+two-word tuples of _u_ 4 such as ( _wrote, on_ ), even though we have no statistics for them.
+A potential problem with _f_ backoff is that after all the obvious candidates have been
+selected, the function is not very good at differentiating between the remaining candidates that have about the same level of novelty and occur infrequently.
+
+
+**3.2.2** **The** **Performance** **of** **the** **Hypothesis.** The evaluation functions discussed in the
+previous section score candidates based on prior knowledge alone, independent of the
+current state of the learner’s hypothesis and the annotation of the selected training
+examples. To attune the selection of training examples to the learner’s progress, an
+evaluation function might factor in its current hypothesis in predicting a candidate’s
+TUV.
+One way to incorporate the current hypothesis into the evaluation function is
+to score each candidate using the current model, assuming its hypothesis is right.
+An error-driven evaluation function, _f_ err, equates the TUV of a candidate with the
+hypothesis’ estimate of its likelihood to misclassify that candidate (i.e., one minus the
+probability of the most-likely class). If the hypothesis predicts that the likelihood of a
+prepositional phrase to attach to the noun is 80%, and if the hypothesis is accurate,
+then there is a 20% chance that it has misclassified.
+A related evaluation function is one that measures the hypothesis’s **uncertainty**
+across all classes, rather than focusing on only the most likely class. Intuitively, if the
+hypothesis classifies a candidate as equally likely to attach to the verb as to the noun, it
+is the most uncertain of its answer. If the hypothesis assigns a candidate to a class with
+a probability of one, then it is the most certain of its answer. For the binary-class case,
+the uncertainty-based evaluation function, _f_ unc, can be expressed in the same way as
+the error-driven function, as a function that is symmetric about 0.5 and monotonically
+decreases if the hypothesis prefers one class over another: [2]
+
+
+_f_ unc( _u_, _C_ ) = _f_ �err( _u_, _C_ )
+
+1 _−_ _P_ ( _noun_ _| u_, _C_ ) : _P_ ( _noun_ _| u_, _C_ ) _≥_ 0 _._ 5
+=
+_P_ ( _noun_ _| u_, _C_ ) : otherwise
+
+= 0 _._ 5 _−_ _abs_ (0 _._ 5 _−_ _P_ ( _noun_ _| u_, _C_ )) (1)
+
+
+In the general case of choosing between multiple classes, _f_ err and _f_ unc are different from
+one another. We shall return to this point in Section 4.1.2 when we consider training
+parsers.
+The potential drawback of the performance-based evaluation functions is that they
+assume that the hypothesis is correct. Selecting training examples based on a poor hypothesis is prone to pitfalls. On the one hand, the hypothesis may be overly confident
+about the certainty of its decisions. For example, the hypothesis may assign _noun_ to
+a candidate with a probability of one based on parameter estimates computed from a
+single previous observation in which a similar example was labeled as _noun_ . Despite
+the unreliable statistics, this candidate would not be selected, since the hypothesis
+considers this a known case. Conversely, the hypothesis may also direct the selection algorithm to chase after undecidable cases. For example, consider prepositional
+phrases (PPs) with _in_ as the head. These PPs occur frequently, and about half of them
+should attach to the object noun. Even though training on more labeled _in_ examples
+
+
+2 As long as it adheres to these criteria, the specific form of the function is irrelevant, since the selection
+is not determined by the absolute scores of the candidates, but by their scores relative to each other.
+
+
+259
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+does not improve the model’s performance on future _in_ PPs, the selection algorithm
+will keep on requesting more _in_ training examples because the hypothesis remains uncertain about this preposition. [3] With an unlucky starting hypothesis, these evaluation
+functions may select uninformative candidates initially.
+
+
+**3.2.3** **The** **Parameters** **of** **the** **Hypothesis.** The potential problems with performancebased evaluation function stems from their trust in the model’s diagnosis of its own
+progress. Another way to incorporate the current hypothesis is to determine how good
+it is and what type of examples will improve it the most. In this section we propose
+an evaluation function that scores candidates based on their utilities in increasing the
+confidence about the parameters of the hypothesis (i.e., the collection of statistics over
+the characteristic tuples of the training examples).
+Training the parameters of the PP-attachment model is similar to empirically determining the bias of a coin. We measure the coin’s bias by repeatedly tossing it and
+keeping track of the percentage of times it lands on heads. The more trials we perform,
+the more confident we become about our estimation of the bias. Similarly, in estimating
+_p_, the likelihood of a PP’s attaching to its object noun, we are more confident about the
+classification decision based on statistics with higher counts than based on statistics
+with lower counts. A quantitative measurement of our confidence in a statistic is the
+**confidence** **interval** . This is a region around the measured statistic, bounding the area
+within which the true statistic may lie. More specifically, the confidence interval for _p_,
+a binomial parameter, is defined as
+
+
+
+
+
+
+
+
+
+
+
+
+¯ _p_ (1 _−_ ¯ _p_ )
+
+
+
+1
+_conf_ ~~_i_~~ _nt_ (¯ _p_, _n_ ) =
+
+
+
+_−_ ¯ _p_ )
+
++ _[t]_ [2]
+_n_ 4 _n_
+
+
+
+4 _n_ [2]
+
+
+
+1 + _[t]_ [2]
+
+
+
+¯ _p_ + _[t]_ [2]
+
+
+
+_n_
+
+
+
+2 _n_ _[±][ t]_
+
+
+
+where ¯ _p_ is the expected value of _p_ based on _n_ trials, and _t_ is a **threshold** **value** that
+depends on the number of trials and the level of confidence we desire. For instance,
+if we want to be 90% confident that the true statistic _p_ lies within the interval, and ¯ _p_
+is based on _n_ = 30 trials, then we set _t_ to be 1.697. [4] Applying the confidence interval
+concept to evaluating candidates for the back-off PP-attachment model, we define
+a function _f_ conf that scores a candidate by taking the average of the lengths of the
+confidence interval of each back-off level. That is,
+
+
+
+_f_ conf( _u_, _C_ ) =
+
+
+
+�4
+
+_|conf_ _int_ (¯ _pl_ ( _u_, _C_ ), _nl_ ( _u_, _C_ )) _|_
+_l_ =1
+
+4
+
+
+
+where ¯ _pl_ ( _u_, _C_ ) is the probability that model _C_ will attach _u_ to _noun_ at back-off level _l_,
+and _nl_ ( _u_, _C_ ) is the number of training examples upon which this classification is based.
+The confidence-based evaluation function has several potential problems. One of
+its flaws is similar to that of _f_ novel. In the early stage, _f_ conf picks the same examples
+
+
+3 This phenomenon is particularly acute in the early stages of refining the hypothesis because most
+decisions are based on statistics of the head preposition alone; in the later stages, the hypothesis can
+usually rely on higher-ordered characteristic tuples that tend to be better classifiers.
+4 For _n ≤_ 120, the values of _t_ can be found in standard statistic textbooks; for _n ≥_ 120, _t_ = 1 _._ 6576.
+Because the derivation for the confidence interval equation makes a normality assumption, the
+equation does not hold for small values of _n_ (cf Larsen and Marx [1986], pp. 277–278). When _n_ is large,
+
+
+
+the contributions from the terms in _[t]_ [2]
+
+
+
+large _n_, ¯ _p ± t_ [�]
+
+
+
+_n_ [are] [negligible.] [Dropping] [these] [terms,] [we] [have] [the] _[t]_ [statistic] [for]
+
+
+
+¯ _p_ (1 _−_ ¯ _p_ ) _/n_ .
+
+
+
+260
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+as _f_ novel, because we have no confidence in the statistics of novel examples. Therefore,
+_f_ conf is also prone to chase after examples that rarely occur to build up the confidence
+of some unimportant parameters. A second problem is that _f_ conf ignores the output of
+the model. Thus, if candidate A has a confidence interval around [0 _._ 6, 1] and candidate
+B has a confidence interval around [0 _._ 4, 0 _._ 7], then _f_ conf will prefer candidate A, even
+though training on A will not change the hypothesis’s performance, since the entire
+confidence interval is already in the _noun_ zone.
+
+
+**3.2.4** **Hybrid** **Function.** The three categories of predictive criteria discussed above are
+complementary, each focusing on a different aspect of the learner’s weakness. Therefore, it may be beneficial to combine these criteria into one evaluation function. For
+instance, the deficiency of the confidence-based evaluation function described in the
+previous section can be avoided if the confidence interval covering the region around
+the uncertainty boundary (candidate B in the example just discussed) is weighed more
+heavily than one around the end points (candidate A).
+In this section, we introduce a new function that tries to factor in both the uncertainty of the model performance and the confidence of the model parameters. First, we
+define a function, called _area_ (¯ _p_, _n_ ), that computes the area under a Gaussian function
+_N_ ( _x_, _µ_, _σ_ ) with a mean of 0.5 and a standard deviation of 0.1 that is bounded by the
+confidence interval as computed by _conf_ _int_ (¯ _p_, _n_ ) (see Figure 5). [5] That is, suppose ¯ _p_
+has a confidence interval of [ _a_, _b_ ]; then
+
+
+               - _b_
+_area_ (¯ _p_, _n_ ) = _N_ ( _x_, 0 _._ 5, 0 _._ 1) _dx_
+
+_a_
+
+
+Computing _area_ for each back-off level, we define an evaluation function, _f_ area( _u_, _C_ ),
+as their average. This function can be viewed as a product of _f_ conf and _f_ unc. [6]
+
+
+**3.3** **Experimental** **Comparison**
+To determine the relative merits of the proposed evaluation functions, we compare the
+learning curve of training with sample selection according to each function against a
+baseline of random selection in an empirical study. The corpus for this comparison is
+a collection of phrases extracted from the Wall Street Journal (WSJ) Treebank. We use
+Section 00 as the development set and Sections 2-23 as the training and test sets. We
+perform 10-fold cross-validation to ensure the statistical significance of the results. For
+each fold, the training candidate pool contains about 21,000 phrases, and the test set
+contained about 2,000 phrases.
+As shown in Figure 1, the learner generates an initial hypothesis based on a small
+set of training examples, _L_ . These examples are randomly selected from the pool of
+unlabeled candidates and annotated by a human. Random sampling ensures that the
+initial trained set reflects the distribution of the candidate pool and thus that the
+initial hypothesis is unbiased. Starting with an unbiased hypothesis is important for
+those evaluation functions whose scoring metrics are affected by the accuracy of the
+hypothesis. In these experiments, _L_ initially contains 500 randomly selected examples.
+In each selection iteration, all the candidates are scored by the evaluation function,
+and _n_ examples with the highest TUVs are picked out from _U_ to be labeled and added
+
+
+5 The standard deviation value for the Gaussian is chosen so that more than 98% of the mass of the
+distribution is between 0.25 and 0.75.
+6 Note that we can replace the function in equation (1) with the _N_ ( _x_, 0 _._ 5, _σ_ ) without affecting _f_ unc,
+because it is also symmetric about 0.5 and monotonically decreasing as the input value moves further.
+
+
+261
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+0 0.4 1.0
+
+
+Likelihood of Attach NP
+
+
+**Figure** **5**
+An example: Suppose that the candidate has a likelihood of 0.4 for noun attachment and a
+confidence interval of width 0.1. Then _area_ computes the area bounded by the confidence
+interval and the Gaussian curve.
+
+
+to _L_ . Ideally, we would like to have _n_ = 1 for each iteration. In practice, however, it is
+often more convenient for the human annotator to label data in larger batches rather
+than one at a time. In these experiments, we use a batch size of _n_ = 500 examples.
+We make note of one caveat to this kind of _n_ -best batch selection. Under a
+hypothesis-dependent evaluation function, identical examples will receive identical
+scores. Because identical (or very similar) examples tend to address the same deficiency in the hypothesis, adding _n_ very similar examples to the training set is unlikely
+to lead to big improvements in the hypothesis. To diversify the examples in each batch,
+we simulate single-example selection (whenever possible) by reestimating the scores
+of the candidates after each selection. Suppose we have just chosen to add candidate
+_x_ to the batch. Then, before selecting the next candidate, we estimate the potential
+decrease in scores of candidates similar to _x_ once it belongs to the annotated training
+set. The estimation is based entirely on the knowledge that _x_ is chosen, but not on
+the classification of _x_ . Thus, only certain types of evaluation functions are amenable
+to the reestimation process. For example, if scores have been assigned by _f_ conf, then
+we know that the confidence intervals of the candidates similar to _x_ must decrease
+slightly after learning _x_ . On the other hand, if scores have been assigned by _f_ unc, then
+we cannot perceive any changes in the scores of similar candidates without knowing
+the true classification of _x_ .
+
+
+**3.3.1** **Results** **and** **Discussion.** This section presents the empirical measurements of
+the model’s performances using training examples selected by different evaluation
+functions. We compare each proposed function with the baseline of random selection
+( _f_ rand). The results are graphically depicted from two perspectives. One (e.g., Figure
+6(a)–6(c)) plots the learning curves of the functions, showing the relationship between
+the number of training examples ( _x_ -axis) and the performance of the model on test
+data ( _y_ -axis). We deem one evaluation function to be better than another if its learning
+curve envelopes the other’s. An alternative way to interpret the results is to focus on
+the reduction in training size offered by one evaluation function over another for some
+particular performance level. Figure 6(d) is a bar graph comparing all the evaluation
+
+
+262
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+76
+
+
+74
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+76
+
+
+74
+
+
+
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+76
+
+
+74
+
+
+
+
+
+0 5000 10000 15000 20000
+
+
+Number of examples in the training set
+
+
+
+0 5000 10000 15000 20000
+
+
+Number of examples in the training set
+
+
+
+(a) (b)
+
+
+
+
+
+25,000
+
+
+20,000
+
+
+15,000
+
+
+10,000
+
+
+5,000
+
+
+0
+
+
+
+0 5000 10000 15000 20000
+
+
+Number of examples in the training set
+
+
+
+Evaluation Functions
+
+
+
+(c) (d)
+
+
+**Figure** **6**
+A comparison of the performance of different evaluation functions: (a) compares the learning
+curves of the functions that use knowledge about the problem space ( _f_ novel and _f_ backoff) with
+that of the baseline; (b) compares the learning curves of performance-based function ( _func_ and
+_fconf_ ) with the baseline; (c) compares the learning curve of _f_ area, which combines uncertainty
+and confidence, with _f_ unc, _f_ conf, and the baseline; (d) compares all the functions for the number
+of training examples selected at the final performance level (83.8%).
+
+
+functions at the highest performance level. The graph shows that in order to train a
+model that attaches PPs with an accuracy rate of 83.8%, sample selection with _f_ novel
+requires 2,500 fewer examples than the baseline.
+Compared to _f_ novel, _f_ backoff selects more helpful training examples in the early stage.
+As shown in Figure 6(a), the improvement rate of the model under _f_ backoff is always
+at least as fast that for as _f_ novel. However, the differences between these two functions
+become smaller for higher performance levels. This outcome validates our predictions.
+Scoring candidates by a combination of their novelty, occurrence frequencies, and the
+qualities of their characteristic tuples, _f_ backoff selects helpful early (the first 4,000 or so)
+training examples. Then, just as in _f_ novel, the learning rate remains stagnant for the
+next 2,000 poorly selected examples. Finally, when the remaining candidates all have
+similar novelty values and contain mostly characteristic tuples that occur infrequently,
+the selection becomes random.
+Figure 6(b) compares the two evaluation functions that score candidates based on
+the current state of the hypothesis. Although both functions suffer a slow start, they
+are more effective than _f_ backoff at reducing the training set when learning high-quality
+models. Initially, because all the unknown statistics are initialized to 0.5, selection based
+on _f_ unc is essentially random sampling. Only after the hypothesis becomes sufficiently
+accurate (after training on about 5,000 annotated examples) does it begin to make
+
+
+263
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+informed selections. Following a similar but more exaggerated pattern, the confidencebased function, _f_ conf, also improves slowly at the beginning before finally overtaking
+the baseline. As we noted earlier, because the hypothesis is not confident about novel
+candidates, _f_ conf and _f_ novel tend to select the same early examples. Therefore, the early
+learning rate of _f_ conf is as poor as that of _f_ novel. In the later stage, while _f_ novel continues
+to flounder, _f_ conf can select better candidates based on a more reliable hypothesis.
+Finally, the best-performing evaluation function is the hybrid approach. Figure
+6(c) shows that the learning curve of _f_ area combines the earlier success of _f_ unc and the
+later success of _f_ conf to always outperform the other functions. As shown in Figure
+6(d), it requires the least number of examples to achieve the highest performance level
+of 83.8%. Compared to the baseline, _f_ area requires 47% fewer examples to achieve this
+performance level. From these comparison studies, we conclude that involving the
+hypothesis in the selection process is a key factor in reducing the size of the training
+set.
+
+
+**4.** **Sample** **Selecting** **for** **Statistical** **Parsing**
+
+
+In applying sample selection to training a PP-attachment model, we have observed
+that all effective evaluation functions make use of the model’s current hypothesis in
+estimating the training utility of the candidates. Although knowledge about the problem space seems to help sharpening the learning curve initially, overall, it is not a good
+predictor. In this section, we investigate whether these observations hold true for training statistical parsing models as well. Moreover, in order to determine whether the
+performances of the predictive criteria are consistent across different learning models
+within the same domain, we have performed the study on two parsing models: one
+based on a context-free variant of tree-adjoining grammars (Joshi, Levy, and Takahashi 1975), the Probabilistic Lexicalized Tree Insertion Grammar (PLTIG) formalism
+(Schabes and Waters 1993; Hwa 1998), and Collins’s Model 2 parser (1997). Although
+both models are lexicalized, statistical parsers, their learning algorithms are different.
+The Collins Parser is a fully supervised, history-based learner that models the parameters of the parser by taking statistics directly from the training data. In contrast,
+PLTIG’s expectation-maximization-based induction algorithm is partially supervised;
+the model’s parameters are estimated indirectly from the training data.
+As a superset of the PP-attachment task, parsing is a more challenging learning
+problem. Whereas a trained PP-attachment model is a binary classifier, a parser must
+identify the correct syntactic analysis out of all possible parses for a sentence. This
+classification task is more difficult than PP-attachment, since the number of possible
+parses for a sentence grows exponentially with respect to its length. Consequently,
+the annotator’s task is more complex. Whereas the person labeling the training data
+for PP-attachment reveals one unit of information (always choosing between _noun_ or
+_verb_ ), the annotation needed for parser training is usually greater than one unit, [7] and
+the type of labels varies from sentence to sentence. Because the annotation complexity
+differs from sentence to sentence, the evaluation functions must strike a balance between maximizing potential informational gain and minimizing the expected amount
+
+
+7 We consider each pair of brackets in the training sentence to be one unit of supervised information,
+assuming that the number of brackets correlates linearly with the amount of effort spent by the human
+annotator. This correlation is an approximation, however; in real life, adding one pair of brackets to a
+longer sentence may require more effort than adding a pair of brackets to a shorter one. To capture
+bracketing interdependencies at this level, we would need to develop a model of the annotation
+decision process and incorporate it as an additional factor in the evaluation functions.
+
+
+264
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+of annotation exerted. We propose a set of evaluation functions similar in spirit to those
+for the PP-attachment learner, but extended to accommodate the parsing domain.
+
+
+
+**4.1** **Evaluation** **Functions**
+**4.1.1** **Problem** **Space.** Similarly to scoring a PP candidate based on the novelty and
+frequencies of its characteristic tuples, we define an evaluation function, _f_ lex( **w**, _G_ ) that
+scores a sentence candidate, **w**, based on the novelty and frequencies of word pair
+co-occurrences: 
+
+
+_f_ lex( **w**, _G_ ) =
+
+
+
+
+_wi_, _wj∈_ **w** _[new]_ [(] _[w][i]_ [,] _[ w][j]_ [)] _[ ×][ coocc]_ [(] _[w][i]_ [,] _[ w][j]_ [)]
+
+
+
+_length_ ( **w** )
+
+
+
+where **w** is the unlabeled sentence candidate, _G_ is the current parsing model (which
+is ignored by problem-space-based evaluation functions), _new_ ( _wi_, _wj_ ) is an indicator
+function that returns one if we have not yet selected any sentence in which _wi_ and _wj_
+co-occurred, and _coocc_ ( _wi_, _wj_ ) is a function that returns the number of times that _wi_ cooccurs [8] with _wj_ in the candidate pool. We expect these evaluation functions to be less
+relevant for the parsing domain than for the PP-attachment domain for two reasons.
+First, because we do not have the actual parses, the extraction of lexical relationships
+is based on co-occurrence statistics, not syntactic relationships. Second, because the
+distribution of words that form lexical relationships is wider and more uniform than
+that of words that form PP characteristic tuples, most word pairs will be novel and
+appear only once.
+Another simple evaluation function based on the problem space is one that estimates the TUV of a candidate from its sentence length:
+
+
+_f_ len( **w**, _G_ ) = _length_ ( **w** )
+
+
+The intuition behind this function is based on the general observation that longer
+sentences tend to have complex structures and introduce more opportunities for ambiguous parses. Although these evaluation functions may seem simplistic, they have
+one major advantage: They are easy to compute and require little processing time.
+Because inducing parsing models demands significantly more time than inducing PPattachment models, it becomes more important that the evaluation functions for parsing models be as efficient as possible.
+
+
+**4.1.2** **The** **Performance** **of** **the** **Hypothesis.** We previously defined two performancebased evaluation functions: _f_ err, the model’s estimate of the likelihood that is has
+made a classification error, and _func_, the model’s estimate of its uncertainty in making
+the classification. We have shown the two functions to have similar performance for
+the PP-attachment task. This is not the case for statistical parsing, however, because
+the number of possible classes (parse trees) differs from sentence to sentence. For
+example, suppose we wish to compare one candidate for which the current parsing
+model generated four equally likely parses with another candidate for which the model
+generated 1 parse with probability of 0.2 and 99 other parses with a probability of
+0.01 (such that they sum to 0.98). The error-driven function, _f_ err, would score the latter
+candidate higher because its most likely parse has a lower probability than that of the
+most likely parse of the former candidate; the uncertainty-based function, _f_ unc, would
+score the former candidate higher because the model does not have a strong preference
+
+
+8 We consider two words to be co-occuring if their log-likelihood ratio is greater than some threshold
+value determined with held-out data.
+
+
+265
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+for one parse over any other. In this section, we provide a formal definition for both
+functions.
+Suppose that a parsing model _G_ generates a candidate sentence **w** with probability
+_P_ ( **w** _| G_ ), and that the set _V_ contains all possible parses that _G_ generated for **w** . Then,
+we denote the probability of _G_ ’s generating a single parse, _v ∈V_, as _P_ ( _v | G_ ) such that
+
+
+_P_ ( _v | G_ ) = _P_ ( **w** _| G_ ). The parse chosen for **w** is the most likely parse in _V_, denoted
+_v∈V_
+
+as _v_ max, where
+_v_ max = argmax _v∈V_ _P_ ( _v | G_ )
+
+
+Note that _P_ ( _v_ _|_ _G_ ) reflects the probability of one particular parse tree, _v_, out of all
+possible parse trees for all possible sentences that _G_ can generate. To compute the
+likelihood of a parse’s being the correct parse out of the possible parses of **w** according
+to _G_, denoted as _P_ ( _v |_ **w**, _G_ ), we need to normalize the tree probability by the sentence
+probability. So according to _G_, the likelihood that _v_ max is the correct parse for **w** is [9]
+
+
+_P_ ( _v_ max _| G_ )
+_P_ ( _v_ max _|_ **w**, _G_ ) =
+
+_P_ ( **w** _| G_ )
+
+= ~~�~~ _P_ ( _v_ max _| G_ ) (2)
+
+_v∈V_ _[P]_ [(] _[v][ |][ G]_ [)] _[.]_
+
+
+Therefore, the error-driven evaluation function is defined as
+
+
+_f_ err( **w**, _G_ ) = 1 _−_ _P_ ( _v_ max _|_ **w**, _G_ )
+
+
+
+Unlike the error-driven function, which focuses on the most likely parse, the
+uncertainty-based function takes the probability distribution of all parses into account.
+To quantitatively characterize its distribution, we compute the entropy of the distribution. That is,
+
+
+
+
+  _H_ ( _V_ ) = _−_
+
+
+
+_p_ ( _v_ ) lg( _p_ ( _v_ )) (3)
+_v∈V_
+
+
+
+where _V_ is a random variable that can take any possible outcome in set _V_, and _p_ ( _v_ ) =
+Pr( _V_ = _v_ ) is the density function. Further details about the properties of entropy can
+be found in textbooks on information theory (e.g., Cover and Thomas 1991).
+Determining the parse tree for a sentence from a set of possible parses can be
+viewed as assigning a value to a random variable. Thus, a direct application of the
+entropy definition to the probability distribution of the parses for sentence **w** in _G_
+computes its **tree** **entropy**, _TE_ ( **w**, _G_ ), the expected number of bits needed to encode
+the distribution of possible parses for **w** . However, we may not wish to compare
+two sentences with different numbers of parses by their entropy directly. If the parse
+probability distributions for both sentences are uniform, the sentence with more parses
+will have a higher entropy. Because longer sentences typically have more parses, using
+entropy directly would result in a bias toward selecting long sentences. To normalize
+for the number of parses, the uncertainty-based evaluation function, _f_ unc, is defined as
+a measurement of similarity between the actual probability distribution of the parses
+and a hypothetical uniform distribution for that set of parses. In particular, we divide
+
+
+9 Note that _P_ ( **w** _|v_, _G_ ) = 1 for any _v ∈V_, where _V_ is the set of all possible parses for **w**, because _v_ exists
+only when **w** is observed.
+
+
+266
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+the tree entropy by the log of the number of parses: [10]
+
+
+_f_ unc( **w**, _G_ ) = _[TE]_ [(] **[w]** [,] _[ G]_ [)]
+
+lg( _∥V∥_ )
+
+
+We now derive the expression for _TE_ ( **w**, _G_ ). Recall from equation (2) that if _G_
+produces a set of parses, _V_, for sentence **w**, the set of probabilities _P_ ( _v |_ **w**, _G_ ) (for all
+_v ∈V_ ) defines the distribution of parsing likelihoods for sentence **w** :
+
+
+     
+_P_ ( _v |_ **w**, _G_ ) = 1
+_v∈V_
+
+
+Note that _P_ ( _v |_ **w**, _G_ ) can be viewed as a density function _p_ ( _v_ ) (i.e., the probability of
+assigning _v_ to a random variable _V_ ). Mapping it back into the entropy definition from
+equation (3), we derive the tree entropy of **w** as follows:
+
+
+
+_TE_ ( **w**, _G_ ) = _H_ ( _V_ )
+
+
+
+
+ = _−_
+
+
+
+_p_ ( _v_ ) lg( _p_ ( _v_ ))
+_v∈V_
+
+
+
+
+ = _−_
+
+
+_v∈V_
+
+ = _−_
+
+
+_v∈V_
+
+
+
+_P_ ( _v | G_ ) _P_ ( **w** _| G_ ) [lg][(] _[P]_ [(] _[v][ |][ G]_ [)) +]
+
+_v∈V_
+
+
+
+_P_ ( _v | G_ ) _[P]_ [(] _[v][ |][ G]_ [)]
+_P_ ( **w** _| G_ ) [lg][(] _P_ ( **w** _| G_ ) [)]
+
+
+
+_P_ ( _v | G_ )
+_P_ ( **w** _| G_ ) [lg][(] _[P]_ [(] **[w]** _[ |][ G]_ [))]
+
+
+
+
+
+_P_ ( _v | G_ )
+_v∈V_
+
+
+
+1
+= _−_
+_P_ ( **w** _| G_ )
+
+
+1
+= _−_
+_P_ ( **w** _| G_ )
+
+
+
+
+
+
+
+_P_ ( _v | G_ ) lg( _P_ ( _v | G_ )) + [lg][(] _[P]_ [(] **[w]** _[ |][ G]_ [))]
+
+_P_ ( **w** _| G_ )
+
+_v∈V_
+
+
+
+_P_ ( _v | G_ ) lg( _P_ ( _v | G_ )) + lg( _P_ ( **w** _| G_ ))
+_v∈V_
+
+
+
+_P_ ( **w** _| G_ )
+
+
+
+
+
+
+
+Using the bottom-up, dynamic programming technique (see the appendix for details) of computing inside probabilities (Lari and Young 1990), we can efficiently compute the probability of the sentence, _P_ ( **w** _| G_ ). Similarly, the algorithm can be modified
+to compute the quantity [�] _P_ ( _v | G_ ) lg( _P_ ( _v | G_ )).
+
+_v∈V_
+
+
+**4.1.3** **The** **Parameters** **of** **the** **Hypothesis.** Although the confidence-based function
+gives good TUV estimates to candidates for training PP-attachment models, it is not
+clear how a similar technique can be applied to training parsers. Whereas binary
+classification tasks can be described by binomial distributions, for which the confidence interval is well defined, a parsing model is made up of many multinomial
+classification decisions. We therefore need a way to characterize the confidence for
+each decision as well as a way to combine them into an overall confidence. Another
+difficulty is that the complexity of the induction algorithm deters us from reestimating the TUVs of the remaining candidates after selecting each new candidate. As we
+
+
+10 When _f_ unc( **w**, _G_ ) = 1, the parser is considered to be the most uncertain about a particular sentence.
+Instead of dividing tree entropies, one could have computed the Kullback-Leibler distance between the
+two distributions (in which case a score of zero would indicate the highest level of uncertainty).
+Because the selection is based on relative scores, as long as the function is monotonic, the exact form of
+the function should not have much impact on the outcome.
+
+
+267
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+discussed in Section 3.3, reestimation is important for batched annotation. Without
+some means of updating the TUVs after each selection, the learner will not realize that
+it has already selected a candidate to train some parameter with low confidence until
+the retraining phase, which occurs only at the end of the batch selection; therefore, it
+may continue to select very similar candidates to train the same parameter. Even if we
+assume that the statistics can be updated, reestimating the TUVs is a computationally
+expensive operation. Essentially, all the remaining candidates that share some parameters with the selected candidate will need to be re-parsed. For these practical reasons, we do not include an evaluation function measuring confidence for the parsing
+experiment.
+
+
+**4.2** **Experiments** **and** **Results**
+We compare the effectiveness of sample selection using the proposed evaluation functions against a baseline of random selection ( _f_ rand( **w**, _G_ ) = _rand_ ()). Similarly to previous
+experimental designs, the learner is given a small set of annotated seed data from the
+WSJ Treebank and a large set of unlabeled data (also from the WSJ Treebank but with
+the labels removed) from which to select new training examples. All training data are
+from Sections 2–21 of the treebank. We monitor the learning progress of the parser by
+testing it on unseen test sentences. We use Section 00 for development and Section
+23 for testing. This study is repeated for two different models, the PLTIG parser and
+Collins’s Model 2 parser.
+
+
+**4.2.1** **An** **Expectation-Maximization-Based** **Learner.** In the first experiment, we use
+an induction algorithm (Hwa 2001a) based on the expectation-maximization (EM)
+principle that induces parsers for PLTIGs. The algorithm performs heuristic search
+through an iterative reestimation procedure to find local optima: sets of values for
+the grammar parameters that maximizes the grammar’s likelihood of generating the
+training data. In principle, the algorithm supports unsupervised learning; however,
+because the search space has too many local optima, the algorithm tends to converge
+on a model that is unsuitable for parsing. Here, we consider a partially supervised
+variant in which we assume that the learner is given the phrasal boundaries of the
+training sentences but not the label of the constituent units. For example, the sentence
+_Several_ _fund_ _managers_ _expect_ _a_ _rough_ _market_ _this_ _morning_ _before_ _prices_ _stabilize._ would be
+labeled as “((Several fund managers) (expect ((a rough market) (this morning)) (before
+(prices stabilize))).)” Our algorithm is similar to the approach taken by Pereira and
+Schabes (1992) for inducing PCFG parsers.
+Because the EM algorithm itself is an iterative procedure, performing sample selection on top of an EM-based learner is an extremely computational-intensive process.
+Here, we restrict the experiments for the PLTIG parsers to a smaller-scale study in the
+following two aspects. First, the lexical anchors of the grammar rules are backed off to
+part-of-speech tags; this restricts the size of the grammar vocabulary to 48. Second, the
+unlabeled candidate pool is set to contain 3,600 sentences, which is sufficiently large
+for inducing a grammar of this size. The initial model is trained on 500 labeled seed
+sentences. For each selection iteration, an additional 100 sentences are moved from
+the unlabeled pool to be labeled and added to the training set. After training, the
+updated parser is then tested on unseen sentences (backed off to their part-of-speech
+tags) and compared to the gold standard. Because the induced PLTIG produces binarybranching parse trees, which have more layers than the gold standard, we measure
+parsing accuracy in terms of the crossing-bracket metric. The study is repeated for
+10 trials, each using a different portion of the full training set, to ensure statistical
+significance (using pairwise _t_ -test at 95% confidence).
+
+
+268
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+81
+
+
+
+80
+
+
+79
+
+
+78
+
+
+77
+
+
+76
+
+
+
+
+
+5000 10000 15000 20000 25000 30000 35000 40000 45000
+
+
+Number of labeled brackets in the training set
+
+(a)
+
+
+40,000
+
+
+35,000
+
+
+30,000
+
+
+25,000
+
+
+20,000
+
+
+15,000
+
+
+10,000
+
+
+5,000
+
+
+0
+
+
+Evaluation functions
+
+
+(b)
+
+
+**Figure** **7**
+PLTIG parser: (a) A comparison of the evaluation functions’ learning curves. (b) A comparison
+of the evaluation functions for a test performance score of 80%.
+
+
+The results of the experiment are graphically shown in Figure 7. As with the
+PP-attachment studies, Figure 7(a) compares the learning curves of the proposed evaluation functions to that of the baseline. Note that even though these functions select
+examples in terms of entire sentences, the amount of annotation is measured in the
+graphs ( _x_ -axis) in terms of the number of _brackets_ rather than sentences. Unlike in
+the PP-attachment case, the amount of effort from the annotators varies significantly
+from example to example. A short and simple sentence takes much less time to annotate than a long and complex sentence. We address this effect by approximating
+the amount of effort as the number of brackets the annotator needs to label. Thus,
+we deem one evaluation function more effective than another if, for the desired level
+of performance, the smallest set of sentences selected by the function contains fewer
+brackets than that of the other function. Figure 7(b) compares the evaluation functions
+at the final test performance level of 80%.
+
+
+269
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+Qualitatively comparing the learning curves in the figure, we see that with the appropriate evaluation function, sample selection does reduce the amount of annotation.
+Similarly to our findings in the PP-attachment study, the simple problem-space-based
+evaluation function, _f_ len, offers only little savings; its performance is nearly indistinguishable from that of the baseline, for the most part. [11] The evaluation functions based
+on hypothesis performances, on the other hand, do reduce the amount of annotation in
+the training data. Of the two that we proposed for this category, the tree entropy evaluation function, _f_ unc, has a slight edge over the error-driven evaluation function, _f_ err.
+For a quantitative comparison, let us consider the set of grammars that achieve
+an average parsing accuracy of 80% on the test sentences. We consider a grammar to
+be comparable to that of the baseline if its mean test score is at least as high as that
+of the baseline and if the difference between the means is not statistically significant.
+The baseline case requires an average of about 38,000 brackets in the training data. In
+contrast, to induce a grammar that reaches the same 80% parsing accuracy with the
+examples selected by _f_ unc, the learner requires, on average, 19,000 training brackets.
+Although the learning rate of _f_ err is slower than that of _f_ unc overall, it seems to have
+caught up in the end; it needs 21,000 training brackets, slightly more than _f_ unc. While
+the simplistic sentence length evaluation function, _f_ len, is less helpful, its learning rate
+still improves slightly faster than that of the baseline. A grammar of comparable quality
+can be induced from a set of training examples selected by _f_ len containing an average
+of 28,000 brackets. [12]
+
+
+**4.2.2** **A** **History-Based** **Learner.** In the second experiment, the basic learning model
+is Collins’s (1997) Model 2 parser, which uses a history-based learning algorithm that
+takes statistics directly over the treebank. As a fully supervised algorithm, it does not
+have to iteratively reestimate its parameters and is computationally efficient enough
+for us to carry out a large-scale experiment. For this set of studies, the unlabeled
+candidate pool consists of around 39,000 sentences. The initial model is trained on
+500 labeled seed sentences, and at each selection iteration, an additional 100 sentences
+are moved from the unlabeled pool into the training set. The parsing performance on
+the test sentences is measured in terms of the parser’s F-score, the harmonic average
+of the labeled precision and labeled recall rates over the constituents (Van Rijsbergen
+1979). [13]
+
+We plot the comparisons between different evaluation functions and the baseline
+for the history-based parser in Figure 8. The examples selected by the problem-spacebased functions do not seem to be helpful. Their learning curves are, for the most part,
+slightly worse than the baseline. In contrast, the parsers trained on data selected by
+the error-driven and uncertainty-based functions learn faster than the baseline; and as
+before, _f_ unc performs slightly better than _f_ err.
+For the final parsing performance of 88%, the parser requires a baseline training set
+of 30,500 sentences annotated with about 695,000 constituents. The same performance
+can be achieved with a training set of 20,500 sentences selected by _f_ err, which contains
+about 577,000 annotated constituents; or with a training set of 17,500 sentences selected
+by _f_ unc, which contains about 505,000 annotated constituents, reducing the number of
+annotated constituents by 27%. Comparing the outcome of this experiment with that of
+
+
+11 In this experiment, we have omitted the evaluation function for selecting novel lexical relationships,
+_f_ lex, because the grammar does not use actual lexical anchors.
+12 In terms of the number of sentences, the baseline _f_ rand selected 2,600 sentences; _f_ len selected 1,300
+sentences; and _f_ err and _f_ unc each selected 900 sentences.
+13 _F_ = [2] _[×]_ _LR_ _[LR]_ + _[×]_ _LP_ _[LP]_ [,] [where] _[LR]_ [is] [the] [labeled] [recall] [score] [and] _[LP]_ [is] [the] [labeled] [precision] [score.]
+
+
+270
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+88
+
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+
+
+
+100000 200000 300000 400000 500000 600000 700000 800000 900000
+
+
+Number of labeled constituents in the training set
+
+(a)
+
+
+800,000
+
+
+700,000
+
+
+600,000
+
+
+500,000
+
+
+400,000
+
+
+300,000
+
+
+200,000
+
+
+100,000
+
+
+0
+
+
+Evaluation Functions
+
+
+(b)
+
+
+**Figure** **8**
+Model 2 parser: (a) A comparison of the learning curves of the evaluation functions. (b) A
+comparison of all the evaluation functions at the test performance level of 88%.
+
+
+the experiment involving the EM-based learner, we see that the training data reduction
+rates are less dramatic than before. This may be because both _f_ unc and _f_ err ignore lexical
+items and chase after sentences containing words that rarely occur. Recent work by
+Tang, Luo, and Roukos (2002) suggests that a hybrid approach that combines features
+of the problem space and the uncertainty of the parser may result in better performance
+for lexicalized parsers.
+
+
+**5.** **Related** **Work**
+
+
+Sample selection benefits problems in which the cost of acquiring raw data is cheap but
+the cost of annotating them is high, as is certainly the case for many supervised learning tasks in natural language processing. In addition to PP-attachment, as discussed
+in this article, sample selection has been successfully applied to other classification
+
+
+271
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+applications. Some examples include text categorization (Lewis and Catlett 1994), base
+noun phrase chunking (Ngai and Yarowsky 2000), part-of-speech tagging (Engelson
+Dagan 1996), spelling confusion set disambiguation (Banko and Brill 2001), and word
+sense disambiguation (Fujii et al. 1998).
+More challenging are learning problems whose objective is not classification, but
+generation of complex structures. One example in this direction is applying sample
+selection to semantic parsing (Thompson, Califf, and Mooney 1999), in which sentences
+are paired with their semantic representation using a deterministic shift-reduce parser.
+A recent effort that focuses on statistical syntactic parsing is the work by Tang, Lou,
+and Roukos (2002). Their results suggest that the number of training examples can be
+further reduced by using a hybrid evaluation function that combines a hypothesisperformance-based metric such as tree entropy (“word entropy” in their terminology)
+with a problem-space-based metric such as sentence clusters.
+Aside from active learning, researchers have applied other learning techniques
+to combat the annotation bottleneck problem in parsing. For example, Henderson
+and Brill (2002) consider the case in which acquiring additional human-annotated
+training data is not possible. They show that parser performance can be improved by
+using boosting and bagging techniques with multiple parsers. This approach assumes
+that there are enough existing labeled data to train the individual parsers. Another
+technique for making better use of unlabeled data is cotraining (Blum and Mitchell
+1998), in which two sufficiently different learners help each other learn by labeling
+training data for one another. The work of Sarkar (2001) and Steedman, Osborne, et
+al. (2003) suggests that co-training can be helpful for statistical parsing. Pierce and
+Cardie (2001) have shown, in the context of base noun identification, that combining
+sample selection and cotraining can be an effective learning framework for large-scale
+training. Similar approaches are being explored for parsing (Steedman, Hwa, et al.
+2003; Hwa et al. 2003).
+
+
+**6.** **Conclusion**
+
+
+In this article, we have argued that sample selection is a powerful learning technique
+for reducing the amount of human-labeled training data. Our empirical studies suggest
+that sample selection is helpful not only for binary classification tasks such as PPattachment, but also for applications that generate complex outputs such as syntactic
+parsing.
+We have proposed several criteria for predicting the training utility of the unlabeled candidates and developed evaluation functions to rank them. We have conducted
+experiments to compare the functions’ ability to select the most helpful training examples. We have found that the uncertainty criterion is a good predictor that consistently
+finds helpful examples. In our experiments, evaluation functions that factor in the
+uncertainty criterion consistently outperform the baseline of random selection across
+different tasks and learning algorithms. For learning a PP-attachment model, the most
+helpful evaluation function is a hybrid that factors in the prediction performance of the
+hypothesis and the confidence for the values of the parameters of the hypothesis. For
+training a parser, we found that uncertainty-based evaluation functions that use tree
+entropy were the most helpful for both the EM-based learner and the history-based
+learner.
+The current work points us in several future directions. First, we shall continue
+to develop alternative formulations of evaluation functions to improve the learning rates of parsers. Under the current framework, we did not experiment with any
+hypothesis-parameter-based evaluation functions for the parser induction task; how
+
+272
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+ever, hypothesis-parameter-based functions may be feasible under a multilearner setting, using parallel machines. Second, while in this work we focused on selecting
+entire sentences as training examples, we believe that further reduction in the amount
+of annotated training data might be possible if the system could ask the annotators
+more-specific questions. For example, if the learner is unsure only of a local decision
+within a sentence (such as a PP-attachment ambiguity), the annotator should not have
+to label the entire sentence.
+In order to allow for finer-grained interactions between the system and the annotators, we have to address some new challenges. To begin with, we must weigh
+in other factors in addition to the amount of annotations. For instance, the learner
+may ask about multiple substrings in one sentence. Even if the total number of labels were fewer, the same sentence would still need to be mentally processed by the
+annotators multiple times. This situation is particularly problematic when there are
+very few annotators, as it becomes much more likely that a person will encounter the
+same sentence many times. Moreover, we must ensure that the questions asked by the
+learner are well-formed. If the learner were simply to present the annotator with some
+substring that it could not process, the substring might not form a proper linguistic
+constituent for the annotator to label. Additionally, we are interested in exploring the
+interaction between sample selection and other semisupervised approaches such as
+boosting, reranking, and cotraining. Finally, based on our experience with parsing, we
+believe that active-learning techniques may be applicable to other tasks that produce
+complex outputs such as machine translation.
+
+
+**Appendix:** **Efficient** **Computation** **of** **Tree** **Entropy**
+
+
+
+As discussed in Section 4.1.2, for learning tasks such as parsing, the number of possible classifications is so large that it may not be computationally efficient to calculate
+the degree of uncertainty using the tree entropy definition. In the equation for the tree
+entropy of **w** ( _TE_ ( **w**, _G_ )) presented in Section 4.1.2, the computation requires summing
+over all possible parses, but the number of possible parses for a sentence grows exponentially with respect to the sentence length. In this appendix, we show that tree
+entropy can be efficiently computed using dynamic programming.
+For illustrative purposes, we describe the computation process using a PCFG expressed in Chomsky normal form. [14] The basic idea is to compose the tree entropy of
+the entire sentence from the tree entropy of the subtrees. The process is similar to
+that for computing the probability of the entire sentence from the probabilities of substrings (called **Inside** **Probabilities** ). We follow the notation convention of Lari and
+Young (1990).
+The inside probability of a nonterminal _X_ generating the substring _wi . . . wj_ is
+denoted as _e_ ( _X_, _i_, _j_ ); it is the sum of the probabilities of all possible subtrees that have
+_X_ as the root and _wi . . . wj_ as the leaf nodes. We define a new function _h_ ( _X_, _i_, _j_ ) to
+represent the corresponding entropy for the substring:
+
+
+
+
+   _h_ ( _X_, _i_, _j_ ) = _−_
+
+
+
+_P_ ( **x** _| G_ ) lg( _P_ ( **x** _| G_ ))
+
+
+
+_∗_
+**x** _∈X⇒wi...wj_
+
+
+where _G_ is the current model. Under this notation, the tree entropy of a sentence,
+
+
+_P_ ( _v | G_ ) lg _P_ ( _v | G_ ), is denoted as _h_ ( _S_, 1, _n_ ).
+_v∈V_
+
+
+14 That is, every production rule must be in one of two forms: a nonterminal expands into two more
+nonterminals, or a nonterminal expands into a terminal.
+
+
+273
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+Analogously to the computation of inside probabilities, we compute _h_ ( _X_, _i_, _j_ ) recursively. The base case is when the nonterminal _X_ generates a single token substring
+_wi_ . The only possible tree has _X_ at the root, immediately dominating the leaf node _wi_ .
+Therefore, the tree entropy is
+
+
+_h_ ( _X_, _i_, _i_ ) = _e_ ( _X_, _i_, _i_ ) lg( _e_ ( _X_, _i_, _i_ ))
+
+
+For the general case, _h_ ( _X_, _i_, _j_ ), we must find all rules of the form _X →_ _YZ_, where _Y_ and
+_∗_
+_Z_ are nonterminals, that have contributed toward _X_ _⇒_ _wi . . . wj_ . To do so, we consider
+_∗_
+all possible ways dividing up _wi . . . wj_ into two pieces such that _Y_ _⇒_ _wi . . . wk_ and
+_∗_
+_Z_ _⇒_ _wk_ +1 _. . . wj_ :
+
+
+
+
+
+_hY_, _Z_, _k_ ( _X_, _i_, _j_ )
+( _X→YZ_ )
+
+
+
+_h_ ( _X_, _i_, _j_ ) =
+
+
+
+
+- _j−_ 1
+
+
+_k_ = _i_
+
+
+
+The function _hY_, _Z_, _k_ ( _X_, _i_, _j_ ) is a portion of _h_ ( _X_, _i_, _j_ ) that accounts for those parses in which
+the rule _X →_ _YZ_ is used and the division point is at word _wk_ . The nonterminals _Y_ and
+_Z_ may, in turn, generate their substrings with multiple parses. Let _Y_ represent the set
+_∗_ _∗_
+of parses for _Y_ _⇒_ _wi . . . wk_ ; let _Z_ represent the set of parses for _Z_ _⇒_ _wk_ +1 _. . . wj_ ; and
+let _x_ represent the parse step of _X →_ _YZ_ . Then, there are a total of _∥Y∥× ∥Z∥_ parses,
+and the probability of each parse is _P_ ( _x_ ) _P_ ( **y** ) _P_ ( **z** ), where **y** _∈Y_ and **z** _∈Z_ . To compute
+_hY_, _Z_, _k_, we need to sum over all possible parses:
+
+
+
+
+    _hY_, _Z_, _k_ ( _X_, _i_, _j_ ) = _−_
+
+
+
+_P_ ( _x_ ) _P_ ( **y** ) _P_ ( **z** ) lg( _P_ ( _x_ ) _P_ ( **y** ) _P_ ( **z** ))
+**y** _∈Y_, **z** _∈Z_
+
+
+
+
+ = _−_
+
+
+
+_P_ ( _x_ ) _P_ ( **y** ) _P_ ( **z** )(lg _P_ ( _x_ ) + lg _P_ ( **y** ) + lg _P_ ( **z** ))
+**y** _∈Y_, **z** _∈Z_
+
+
+
+= _−P_ ( _x_ ) lg( _P_ ( _x_ )) _e_ ( _Y_, _i_, _k_ ) _e_ ( _Z_, _k_ + 1, _j_ ) + _P_ ( _x_ ) _h_ ( _Y_, _i_, _k_ ) _e_ ( _Z_, _k_ + 1, _j_ )
+
++ _P_ ( _x_ ) _e_ ( _Y_, _i_, _k_ ) _h_ ( _Z_, _k_ + 1, _j_ )
+
+
+Thus, the tree entropy of the entire sentence can be recursively computed from the
+entropy values of the substrings.
+
+
+
+**Acknowledgments**
+We thank Joshua Goodman, Lillian Lee,
+Wheeler Ruml, and Stuart Shieber for
+helpful discussions, and Ric Crabbe, Philip
+Resnik, and the reviewers for their
+constructive comments on this article.
+Portions of this work have appeared
+previously (Hwa 2000, 2001b); we thank the
+reviewers of those papers for their helpful
+comments. Parts of this work was carried
+out while the author was a graduate student
+at Harvard University, supported by the
+National Science Foundation under Grant
+No. IRI 9712068. The work is also supported
+by the Department of Defense contract
+RD-02-5700, and ONR MURI Contract
+FCPO.810548265.
+
+
+274
+
+
+
+**References**
+Banko, Michele and Eric Brill. 2001. Scaling
+to very very large corpora for natural
+language disambiguation. In _Proceedings of_
+_the 39th Annual Meeting of the Association for_
+_Computational Linguistics_, Toulouse,
+France, pages 26–33.
+Blum, Avrim and Tom Mitchell. 1998.
+Combining labeled and unlabeled data
+with co-training. In _Proceedings of the 1998_
+_Conference on Computational Learning_
+_Theory_, pages 92–100, Madison, WI.
+Brill, Eric and Philip S. Resnik. 1994. A rule
+based approach to PP attachment
+disambiguation. In _Proceedings of the 15th_
+_International Conference on Computational_
+_Linguistics (COLING)_, Kyoto, Japan, pages
+1198–1204.
+
+
+Hwa Sample Selection for Statistical Parsing
+
+
+
+Charniak, Eugene. 2000. A
+maximum-entropy-inspired parser. In
+_Proceedings of the First Meeting of the North_
+_American Association for Computational_
+_Linguistics_, Seattle.
+Cohn, David, Les Atlas, and Richard
+Ladner. 1994. Improving generalization
+with active learning. _Machine Learning_,
+15(2):201–221.
+Collins, Michael. 1997. Three generative,
+lexicalised models for statistical parsing.
+In _Proceedings of the 35th Annual Meeting of_
+_the Association for Computational Linguistics_,
+pages 16–23, Madrid.
+Collins, Michael. 1999. _Head-Driven Statistical_
+_Models for Natural Language Parsing_ . Ph.D.
+thesis, University of Pennsylvania,
+Philadelphia.
+Collins, Michael and James Brooks. 1995.
+Prepositional phrase attachment through
+a backed-off model. In _Proceedings of the_
+_Third Workshop on Very Large Corpora_,
+Cambridge, MA, pages 27–38.
+Cover, Thomas M. and Joy A. Thomas.
+1991. _Elements of Information Theory_ . John
+Wiley, New York.
+Engelson, Sean P. and Ido Dagan. 1996.
+Minimizing manual annotation cost in
+supervised training from corpora. In
+_Proceedings of the 34th Annual Meeting of the_
+_Association for Computational Linguistics_,
+Santa Cruz, CA, pages 319–326.
+Freund, Yoav, H. Sebastian Seung, Eli
+Shamir, and Naftali Tishby. 1997. Selective
+sampling using the query by committee
+algorithm. _Machine Learning_,
+28(2–3):133–168.
+Fujii, Atsushi, Kentaro Inui, Takenobu
+Tokunaga, and Hozumi Tanaka. 1998.
+Selective sampling for example-based
+word sense disambiguation. _Computational_
+_Linguistics_, 24(4):573–598.
+Henderson, John C. and Eric Brill. 2000.
+Bagging and boosting a treebank parser.
+In _Proceedings of the First Meeting of the_
+_North American Association for Computational_
+_Linguistics_, Seattle, pages 34–41.
+Hwa, Rebecca. 1998. An empirical
+evaluation of probabilistic lexicalized tree
+insertion grammars. In _Proceedings of the_
+_36th Annual Meeting of the Association for_
+_Computational Linguistics and 17th_
+_International Conference on Computational_
+_Linguistics_, Montreal, volume 1, pages
+557–563.
+Hwa, Rebecca. 2000. Sample selection for
+statistical grammar induction. In
+_Proceedings of 2000 Joint SIGDAT Conference_
+_on Empirical Methods in Natural Language_
+_Processing and Very Large Corpora_, pages
+45–52, Hong Kong, October.
+
+
+
+Hwa, Rebecca. 2001a. _Learning Probabilistic_
+_Lexicalized Grammars for Natural Language_
+_Processing_ . Ph.D. thesis, Harvard
+University, Cambridge, MA.
+Hwa, Rebecca. 2001b. On minimizing
+training corpus for parser acquisition. In
+_Proceedings of the ACL 2001 Workshop on_
+_Computational Natural Language Learning_
+_(ConLL-2001)_, Toulouse, France, pages
+84–89.
+Hwa, Rebecca, Miles Osborne, Anoop
+Sarkar, and Mark Steedman. 2003.
+Corrected co-training for statistical
+parsers. In _Proceedings of the ICML_
+_Workshop on the Continuum from Labeled to_
+_Unlabeled Data in Machine Learning and Data_
+_Mining at the 20th International Conference of_
+_Machine Learning (ICML-2003)_,
+Washington, DC, pages 95–102, August.
+Joshi, Aravind K., Leon S. Levy, and
+Masako Takahashi. 1975. Tree adjunction
+grammars. _Journal of Computer and System_
+_Sciences_, 10(1): 136–163.
+Lari, Karim A. and Steve J. Young. 1990.
+The estimation of stochastic context-free
+grammars using the inside-outside
+algorithm. _Computer Speech and Language_,
+4:35–56.
+Larsen, Richard J. and Morris L. Marx. 1986.
+_An Introduction to Mathematical Statistics_
+_and Its Applications_ . Prentice-Hall,
+Englewood Cliffs, NJ.
+Lewis, David D. and Jason Catlett. 1994.
+Heterogeneous uncertainty sampling for
+supervised learning. In _Proceedings of the_
+_Eleventh International Conference on Machine_
+_Learning_, San Francisco, pages 148–156.
+Magerman, David. 1994. _Natural Language_
+_Parsing as Statistical Pattern Recognition_ .
+Ph.D. thesis, Stanford University,
+Stanford, CA.
+Marcus, Mitchell, Beatrice Santorini, and
+Mary Ann Marcinkiewicz. 1993. Building
+a large annotated corpus of English: The
+Penn Treebank. _Computational Linguistics_,
+19(2):313–330.
+Ngai, Grace and David Yarowsky. 2000.
+Rule writing or annotation: Cost-efficient
+resource usage for base noun phrase
+chunking. In _Proceedings of the 38th Annual_
+_Meeting of the Association for Computational_
+_Linguistics_, pages 117–125, Hong Kong,
+October.
+Pereira, Fernando C. N. and Yves Schabes.
+1992. Inside-outside reestimation from
+partially bracketed corpora. In _Proceedings_
+_of the 30th Annual Meeting of the Association_
+_for Computational Linguistics_, pages
+128–135, Newark, DE.
+
+
+275
+
+
+Computational Linguistics Volume 30, Number 3
+
+
+
+Pierce, David and Claire Cardie. 2001.
+Limitations of co-training for natural
+language learning from large datasets. In
+_Proceedings of the 2001 Conference on_
+_Empirical Methods in Natural Language_
+_Processing (EMNLP-2001)_, pages 1–9,
+Pittsburgh, PA.
+Ratnaparkhi, Adwait. 1998. Statistical
+models for unsupervised prepositional
+phrase attachment. In _Proceedings of the_
+_36th Annual Meeting of the Association for_
+_Computational Linguistics and 17th_
+_International Conference on Computational_
+_Linguistics_, Montreal, volume 2, pages
+1079–1085.
+Sarkar, Anoop. 2001. Applying co-training
+methods to statistical parsing. In
+_Proceedings of the Second Meeting of the_
+_North American Association for_
+_Computational Linguistics_, Pittsburgh,
+pages 175–182, June.
+Schabes, Yves and Richard Waters. 1993.
+Stochastic lexicalized context-free
+grammar. In _Proceedings of the Third_
+_International Workshop on Parsing_
+_Technologies_, Tilburg, The Netherlands,
+and Durbuy, Belgium, pages 257–266.
+Steedman, Mark, Rebecca Hwa, Stephen
+Clark, Miles Osborne, Anoop Sarkar, Julia
+Hockenmaier, Paul Ruhlen, Steven Baker,
+and Jeremiah Crim. 2003. Example
+
+
+276
+
+
+
+selection for bootstrapping statistical
+parsers. In _Proceedings of the Joint_
+_Conference of Human Language Technologies_
+_and the Annual Meeting of the North_
+_American Chapter of the Association for_
+_Computational Linguistics_, Edmonton,
+Alberta, Canada, pages 236–243.
+Steedman, Mark, Miles Osborne, Anoop
+Sarkar, Stephen Clark, Rebecca Hwa, Julia
+Hockenmaier, Paul Ruhlen, Steven Baker,
+and Jeremiah Crim. 2003. Bootstrapping
+statistical parsers from small datasets. In
+_Proceedings of the Tenth Conference of the_
+_European Chapter of the Association for_
+_Computational Linguistics_, Budapest, pages
+331–338.
+Tang, Min, Xiaoqiang Luo, and Salim
+Roukos. 2002. Active learning for
+statistical natural language parsing. In
+_Proceedings of the 40th Annual Meeting of the_
+_Association for Computational Linguistics_,
+Philadelphia, pages 120–127, July.
+Thompson, Cynthia A., Mary Elaine Califf,
+and Raymond J. Mooney. 1999. Active
+learning for natural language parsing and
+information extraction. In _Proceedings of_
+_the Sixteenth International Conference on_
+_Machine Learning (ICML-99)_, pages
+406–414, Bled, Slovenia.
+Van Rijsbergen, Cornelis J. 1979. _Information_
+_Retrieval_ . Butterworth, London.
+
+

references/2004.emnlp.baldridge/paper.md

@@ -0,0 +1,767 @@
+---
+title: "Active Learning and the Total Cost of Annotation"
+authors:
+  - "Jason Baldridge"
+  - "Miles Osborne"
+year: 2004
+venue: "EMNLP 2004"
+url: "https://aclanthology.org/W04-3202/"
+---
+
+# **Active Learning and the Total Cost of Annotation**
+
+**Jason Baldridge and Miles Osborne**
+School of Informatics
+University of Edinburgh
+Edinburgh EH8 9LW, UK
+_{_ `jbaldrid,miles` _}_ `@inf.ed.ac.uk`
+
+
+
+**Abstract**
+
+
+Active learning (AL) promises to reduce
+the cost of annotating labeled datasets for
+trainable human language technologies.
+Contrary to expectations, when creating
+labeled training material for HPSG parse
+selection and later _reusing_ it with _other_
+models, gains from AL may be negligible
+or even negative. This has serious implications for using AL, showing that additional cost-saving strategies may need to
+be adopted. We explore one such strategy:
+using a model during annotation to automate some of the decisions. Our best results show an 80% reduction in annotation
+cost compared with labeling randomly selected data with a single model.
+
+
+**1** **Introduction**
+
+AL methods such as uncertainty sampling (Cohn
+et al., 1995) or query by committee (Seung et al.,
+1992) have all been shown to dramatically reduce
+the cost of creating highly informative labeled sets
+for speech and language technologies. However, experiments using AL assume a model that is fixed
+ahead of time: the model used in AL is the same
+one we are currently developing training material
+for. For many complex tasks, we are unlikely to have
+a clear idea how best to model the task at the time of
+annotation; thus, in practice, we will need to _reuse_
+the labeled training material with _other_ models.
+In this paper, we show that AL can be brittle: under a variety of natural reuse scenarios (for example,
+
+
+
+allowing the later model to improve in quality, or
+else reusing the labeled training material using a different machine learning algorithm) performance of
+later models can be significantly undermined when
+training upon material created using AL. The key
+to knowing how well one model will be able to use
+material selected by another is their relatedness – yet
+there may be no means to determine this prior to annotation, leading to a chicken-and-egg problem.
+Our reusability results thus demonstrate that, additionally, other strategies must be adopted to ensure we reduce the total cost of annotation. In Osborne and Baldridge (2004), we showed that ensemble models can increase model performance and also
+produce annotation savings when incorporated into
+the AL process. An obvious next step is automating
+some decisions. Here, we consider a simple automation strategy that reduces annotation costs _indepen-_
+_dently_ of AL and examine its effect on reusability.
+We find that using both semi-automation and AL
+with high-quality models can eliminate the performance gap found in many reuse scenarios. However,
+for weak models, we show that semi-automation
+with random sampling is _more_ effective for improving reusability than using it with AL - demonstrating further cause for caution with AL.
+Finally, we show that under the standard assumption of reuse by the selecting model, using a strategy which combines AL, ensembling, and semiautomated annotation, we are able to achieve our
+highest annotation savings to date on the complex
+task of parse selection for HPSG: an 80% reduction
+in annotation cost compared with labeling randomly
+selected data with our best single model.
+
+
+**2** **Parse selection for Redwoods**
+
+We now briefly describe the Redwoods treebanking
+environment (Oepen et al., 2002), our parse selection models and their performance.
+
+
+**2.1** **The Redwoods Treebank**
+
+The Redwoods treebank project provides tools and
+annotated training material for creating parse selection models for the English Resource Grammar
+(ERG, Flickinger (2000)). The ERG is a hand-built
+broad-coverage HPSG grammar that provides an explicit grammar for the treebank. Using this approach
+has the advantage that analyses for within-coverage
+sentences convey more information than just phrase
+structure: they also contain derivations, semantic interpretations, and basic dependencies.
+For each sentence, Redwoods records all analyses
+licensed by the ERG and indicates which of them,
+if any, the annotators selected as being contextually
+correct. When selecting such distinguished parses,
+rather than simply enumerating all parses and presenting them to the annotator, annotators make use
+of _discriminants_ which disambiguate the parse forest more rapidly, as described in section 3.
+In this paper, we report results using the third
+growth of Redwoods, which contains English sentences from appointment scheduling and travel planning domains of Verbmobil. In all, there are 5302
+sentences for which there are at least two parses and
+a unique preferred parse is identified. These sentences have 9.3 words and 58.0 parses on average.
+
+
+**2.2** **Modeling parse selection**
+
+As is now standard for feature-based grammars, we
+mainly use log-linear models for parse selection
+(Johnson et al., 1999). For log-linear models, the
+conditional probability of an analysis _ti_ given a sentence with a set of analyses _τ_ = _{t . . .}_ is given as:
+
+
+_j_ =1 _[f][j]_ [(] _[t][i]_ [)] _[w][j]_ [)]
+_P_ ( _ti|s, Mk_ ) = _[exp]_ [(][�] _[m]_ (1)
+
+_Z_ ( _s_ )
+
+
+where _fj_ ( _ti_ ) returns the number of times feature
+_j_ occurs in analysis _t_, _wj_ is a weight from model
+_Mk_, and _Z_ ( _s_ ) is a normalization factor for the sentence. The parse with the highest probability is taken
+as the preferred parse for the model. We use the
+
+
+
+Note that each individual model _Mi_ is a well-defined
+distribution usually taken from a fixed set of models. _Z_ ( _s_ ) is a constant to ensure the product distribution sums to one over the set of possible parses. A
+product model effectively averages the contributions
+made by each of the individual models. Though simple, this model is sufficient to show enhanced performance when using multiple models.
+
+
+
+limited memory variable metric algorithm to determine the weights. We do not regularize our loglinear models since labeled data -necessary to set
+hyperparameters- is in short supply in AL.
+We also make use of simpler perceptron models
+for parse selection, which assign scores rather than
+probabilities. Scores are computed by taking the inner product of the analysis’ feature vector with the
+parameter vector:
+
+
+
+score( _ti, s, Mk_ ) =
+
+
+
+_m_
+
+- _fj_ ( _ti_ ) _wj_ (2)
+
+_j_ =1
+
+
+
+The preferred parse is that with the highest score out
+of all analyses. We do not use voted perceptrons
+here (which indeed have better performance) as for
+the reuse experiments described later in section 6 we
+really do wish to use a model that is (potentially)
+worse than a log-linear model.
+Later for AL, it will be useful to map perceptron
+scores into probabilities, which we do by exponentiating and renormalizing the score:
+
+
+_Pp_ ( _ti_ _| s, Mk_ ) = _[exp]_ [(][score][(] _[t][i][, s, M][k]_ [))] (3)
+
+_Z_ ( _s_ )
+
+
+_Z_ ( _s_ ) is again a normalizing constant.
+The previous parse selection models (equations
+1 and 3) use a single model (feature set). It is
+possible to improve performance using an _ensem-_
+_ble_ parse selection model. We create our ensemble
+model (called a _product_ _model_ ) using the _product-_
+_of-experts_ formulation (Hinton, 1999):
+
+
+
+_P_ ( _ti|s, M_ 1 _. . . Mn_ ) =
+
+
+
+
+- _n_
+_j_ =1 _[P]_ [(] _[t][i][|][s, M][j]_ [)]
+
+(4)
+_Z_ ( _s_ )
+
+
+**2.3** **Parse selection performance**
+
+Osborne and Baldridge (2004) describe three distinct feature sets - **configurational**, **ngram**, and
+**conglomerate** - which utilize the various structures made available in Redwoods: derivation trees,
+phrase structures, semantic interpretations, and elementary dependency graphs. They incorporate different aspects of the parse selection task; this is
+crucial for creating diverse models for use in product parse selection models as well as for ensemblebased AL methods. Here, we also use models created from a subset of the conglomerate feature set:
+the **mrs** feature set. This only has features from the
+semantic interpretations.
+The three main feature sets are used to train three
+log-linear models - LL-CONFIG, LL-NGRAM, and
+
+LL-CONGLOM— and a product ensemble of those
+three feature sets, LL-PROD, using equation 4. Additionally, we use a perceptron with the conglomerate
+feature set, P-CONGLOM. Finally, we include a loglinear model that uses the mrs feature set, LL-MRS,
+and a perceptron, P-MRS.
+Parse selection accuracy is measured using exact
+match. A model is awarded a point if it picks some
+parse for a sentence and that parse is the correct analysis indicated by the corpus. To deal with ties, the
+accuracy is given as 1 _/m_ when a model ranks _m_
+parses highest and the best parse is one of them.
+The results for a chance baseline (selecting a
+parse at random), the base models and the product
+model are given in Table 1. These are 10-fold crossvalidation results, using all the training data for estimation and the test split for evaluation. See section
+5 for more details.
+
+Model Perf. Model Perf.
+
+LL-CONFIG 75.05 LL-PROD 77.78
+
+LL-NGRAM 74.01 LL-MRS 64.98
+
+LL-CONGLOM 74.85 P-CONGLOM 73.00
+Chance 22.70 P-MRS 62.11
+
+
+Table 1: Parse selection accuracy.
+
+
+**3** **Measuring annotation cost**
+
+To aid identification of the best parse out of all those
+licensed by the ERG, the Redwoods annotation environment provides local _discriminants_ which the an
+
+
+notator can mark as true or false properties for the
+analysis of a sentence in order to disambiguate large
+portions of the parse forest. As such, the annotator
+does not need to inspect all parses and so parses are
+narrowed down quickly (usually exponentially so)
+even for sentences with a large number of parses.
+More interestingly, it means that the labeling burden
+is relative to the number of possible parses (rather
+than the number of constituents in a parse).
+Data about how many discriminants were needed
+to annotate each sentence is recorded in Redwoods.
+Typically, more ambiguous sentences require more
+discriminant values to be set, reflecting the extra effort put into identifying the best parse. We showed
+in Osborne and Baldridge (2004) that discriminant
+cost does provide a more accurate approximation of
+annotation cost than assigning a fixed unit cost for
+each sentence. We thus use discriminants as the basis of calculating annotation cost to evaluate the effectiveness of different experiment AL conditions.
+Specifically, we set the cost of annotating a given
+sentence as the number of discriminants whose
+value were set by the human annotator plus one to
+indicate a final ‘eyeball’ step where the annotator selects the best parse of the few remaining ones. [1] The
+_discriminant_ _cost_ of the examples we use averages
+3.34 and ranges from 1 to 14.
+
+**4** **Active learning**
+
+Suppose we have a set of examples and labels _Dn_ =
+_{⟨x_ [1] _, y_ [1] _⟩, ⟨x_ [2] _, y_ [2] _⟩, . . .}_ which is to be extended with
+a new labeled example _{⟨x_ _[i]_ _, y_ _[i]_ _⟩}_ . The information
+gain for some model is maximized after selecting,
+labeling, and adding a new example _x_ _[i]_ to _Dn_ such
+that the noise level of _x_ _[i]_ is low and both the bias and
+variance of some model using _Dn_ _∪{⟨x_ _[i]_ _, y_ _[i]_ _⟩}_ are
+minimized (Cohn et al., 1995).
+In practice, selecting data points for labeling such
+that a model’s variance and/or bias is maximally
+minimized is computationally intractable, so approximations are typically used instead. One such
+approximation is _uncertainty sampling_ . Uncertainty
+sampling (also called _tree_ _entropy_ by Hwa (2000)),
+measures the uncertainty of a model over the set of
+parses of a given sentence, based on the conditional
+
+1This eyeball step is not always taken, but Redwoods does
+not contain information about when this occurred, so we apply
+the cost for the step uniformly for all examples.
+
+
+distribution it assigns to them. Following Hwa, we
+use the following measure to quantify uncertainty:
+
+
+_fus_ ( _s, τ, Mk_ ) = _−_ - _P_ ( _t|s, Mk_ ) log _P_ ( _t|s, Mk_ ) (5)
+
+_t∈τ_
+
+_τ_ denotes the set of analyses produced by the ERG
+for the sentence and _Mk_ is some model. Higher values of _fus_ ( _s, τ, Mk_ ) indicate examples on which the
+learner is most uncertain . Calculating _fus_ is trivial with the conditional log-linear and perceptrons
+models described in section 2.2.
+Uncertainty sampling as defined above is a singlemodel approach. It can be improved by simply replacing the probability of a single log-linear (or perceptron) model with a product probability:
+
+
+_fus_ _[en]_ [(] _[s, τ,][ M]_ [) =] _[ −]_ - _P_ ( _t|s, M_ ) log _P_ ( _t|s, M_ ) (6)
+
+_t∈τ_
+
+_M_ is the set of models _M_ 1 _. . . Mn_ . As we mentioned earlier, AL for parse selection is potentially
+problematic as sentences vary both in length and the
+number of parses they have. Nonetheless, the above
+measures do not use any extra normalization as we
+have found no major differences after experimenting
+with a variety of normalization strategies.
+We use random sampling as a baseline and uncertainty sampling for AL. Osborne and Baldridge
+(2004) show that uncertainty sampling produces
+good results compared with other AL methods.
+
+**5** **Experimental framework**
+
+For all experiments, we used a 20-fold crossvalidation strategy by randomly selecting 10%
+(roughly 500 sentences) for the test set and selecting samples from the remaining 90% (roughly 4500
+sentences) as training material. Each run of AL begins with a single randomly chosen annotated seed
+sentence. At each round, new examples are selected
+for annotation from a randomly chosen, fixed sized
+500 sentence subset according to random selection
+or uncertainty sampling until models reach certain
+desired accuracies. We select 20 examples for annotation at each round, and exclude all examples that
+have more than 500 parses. [2]
+
+2Other parameter settings (such as how many examples to
+label at each stage) did not produce substantially different results to those reported here.
+
+
+
+AL results are usually presented in terms of the
+amount of labeling necessary to achieve given performance levels. We say that one method is better than another method if, for a given performance
+level, less annotation is required. The performance
+metric used here is parse selection accuracy as described in section 2.3.
+
+
+**6** **Reusing training material**
+
+
+AL can be considered as selecting some labeled
+training set which is ‘tuned’ to the needs of a particular model. Typically, we might wish to reuse labeled
+training material, so a natural question to ask is how
+general are training sets created using AL. So, if we
+later improved upon our feature set, or else improved
+upon our learner, would the previously created training set still be useful? If AL selects highly idiosyncratic datasets then we would not be able to reuse our
+datasets and thus it might, for example, actually be
+better to label datasets using random sampling. This
+is a realistic situation since models typically change
+and evolve over time - it would be very problematic if the training set itself inherently limits the benefit of later attempts to improve the model.
+We use two baselines to evaluate how well a
+model is able to reuse data selected for labeling by
+another model: (1) **Selecting** **the** **data** **randomly.**
+This provides the essential baseline; if AL in reuse
+situations is going to be useful, it ought to outperform this model-free approach. (2) **Reuse** **by** **the**
+**AL model itself.** This is the standard AL scenario;
+against this, we can determine if reused data can be
+as good as when a model selects data for itself.
+We evaluate a variety of reuse scenarios. We refer to the model used with AL as the _selector_ and
+the model that is reusing that labeled data as the
+_reuser_ . Models can differ in the machine learning algorithm and/or the feature set they use. To measure
+relatedness, we use Spearman’s rank correlation on
+the rankings that two models assign to the parses of
+a sentence. The overall relatedness of two models
+is calculated as the average rank correlation on all
+examples tested in a 10-fold parse selection experiment using all available training material.
+Figure 1 shows complete learning curves for LL
+CONFIG when it reuses material selected by itself,
+
+LL-CONGLOM, P-MRS, and random sampling. The
+
+
+graph shows that self-reuse is the most effective of
+all strategies - this is the idealized situation commonly assumed in active learning studies. However,
+the graph reveals that random sampling is actually
+more effective than selection both by LL-CONGLOM
+until nearly 70% accuracy is reached and by P-MRS
+until about 73%. Finally, we see that the material
+selected by LL-CONGLOM is always more effective
+for LL-CONFIG than that selected by P-MRS. The
+reason for this can be explained by the relatedness
+of each of these selector models to LL-CONFIG: LL
+CONGLOM and LL-CONFIG have an average rank
+correlation of 0 _._ 84 whereas P-MRS and LL-CONFIG
+have a correlation of 0 _._ 65.
+
+
+
+80
+
+
+75
+
+
+70
+
+
+65
+
+
+60
+
+
+55
+
+
+50
+
+
+
+0 1000 2000 3000 4000 5000 6000 7000 8000
+
+Annotation cost
+
+
+
+
+
+Figure 1: Learning curves for LL-CONFIG when
+reusing material by different selectors.
+
+
+Table 2 fleshes out the relationship between relatedness and reusability more fully. It shows the annotation cost incurred by various reusers to reach 65%,
+70%, and 73% accuracy when material is selected
+by various models. The list is ordered from top to
+bottom according to the rank correlation of the two
+models. The first three lines provide the baselines of
+when LL-PROD, LL-CONGLOM, and LL-CONFIG select material for themselves. The last three show the
+amount of material needed by these models when
+random sampling is used. The rest gives the results
+for when the selector differs from the reuser.
+For each performance level, the percent increase
+in annotation cost over self-reuse is given. For
+example, a cost of 2300 discriminants is required
+for LL-PROD to reach the 73% performance level
+when it reuses material selected by LL-CONGLOM;
+
+
+
+this is a 10% increase over the 2100 discriminants
+needed when LL-PROD selects for itself. Similarly,
+the 5500 discriminants needed by LL-CONGLOM to
+reach 73% when reusing material selected by LL
+CONFIG is a 31% increase over the 4200 discriminants LL-CONGLOM needs with its own selection.
+As can be seen from Table 2, reuse always leads
+to an increase in cost over self-reuse to reach a given
+level of performance. How much that increase will
+be is in general inversely related to the rank correlation of the two models. Furthermore, considering
+each reusing model individually, this relationship is
+almost entirely inversely related at all performance
+levels, with the exception of P-CONGLOM and LL
+MRS selecting for LL-CONFIG at the 73% level.
+The reason for some models being more related
+to others is generally easy to see. For example, LL
+CONFIG and LL-CONGLOM are highly related to LL
+PROD, of which they are both components. In both
+of these cases, using AL for use by LL-PROD beats
+random sampling by a large amount.
+That LL-MRS is more related to LL-CONGLOM
+than to LL-CONFIG is explained by the fact the **mrs**
+feature set is actually a subset of the **conglom** set.
+The former contains 15% of the latter’s features.
+Accordingly, material selected by LL-MRS is also
+generally more reusable by LL-CONGLOM than to
+
+LL-CONFIG. This is encouraging since the case of
+
+LL-CONGLOM reusing material selected by LL-MRS
+represents the common situation in which an initial
+model - that was used to develop the corpus - is
+continually improved upon.
+A particularly striking aspect revealed by Figure 1
+and Table 2 is that random sampling is overwhelmingly a better strategy when there is still little labeled material. AL tends to select examples which
+are more ambiguous and hence have a higher discriminant cost. So, while these examples may be
+highly informative for the selector model, they are
+not cheap - and are far less effective when reused
+by another model.
+Considering unit cost (i.e., each sentence costs the
+same) instead of discriminant cost (which assigns a
+variable cost per sentence), AL is generally more
+effective than random sampling for reuse throughout all accuracy levels – but not always. For example, even using unit cost, random sampling is better than selection by LL-MRS or P-MRS for reuse by
+
+
+|Reuser -|Rank Corr. 1.00|65% DC Incr 690 0.0%|70% DC Incr 1200 0.0%|
+|---|---|---|---|
+|LL-PROD<br>LL-CONGLOM<br>LL-CONFIG<br>|1.00<br>1.00<br>1.00<br>|690<br>0.0%<br>1190<br>0.0%<br>1160<br>0.0%<br><br>|1200<br>0.0%<br>2330<br>0.0%<br>2530<br>0.0%<br><br>|
+|LL-PROD<br>LL-PROD<br>LL-CONGLOM<br>LL-CONFIG<br>LL-CONFIG<br>LL-CONGLOM<br>LL-PROD<br>LL-CONFIG<br>LL-CONFIG<br>|.92<br>.92<br>.84<br>.84<br>.79<br>.77<br>.76<br>.71<br>.65<br>|850<br>23.2%<br>840<br>21.7%<br>1340<br>12.6%<br>1660<br>43.1%<br>1960<br>69.0%<br>1600<br>34.5%<br>1080<br>56.5%<br>2100<br>81.0%<br>2650<br>128.4%<br><br>|1470<br>22.5%<br>1560<br>30.0%<br>2610<br>12.0%<br>3760<br>48.6%<br>3910<br>54.5%<br>3400<br>45.9%<br>2040<br>70.0%<br>4270<br>68.8%<br>4870<br>92.5%<br><br>|
+
+
+RAND LL-CONGLOM   - 1400 17.6% 3470 48.9% 7150 71.9%
+
+RAND LL-CONFIG   - 1160 0.0% 3890 53.8% 8560 79.1%
+
+
+Table 2: Comparison of various selection and reuse conditions. Values are given for discriminant cost (DC)
+and the percent increase (Incr) in cost over use of material selected by the reuser.
+
+
+
+LL-CONFIG until 67% accuracy. Thus, LL-MRS and
+
+P-MRS are so divergent from LL-CONFIG that their
+selections are truly sub-optimal for LL-CONFIG, particularly in the initial stages.
+Together, these results shows that AL cannot be
+used blindly and always be expected to reduce the
+total cost of annotation. The data is tuned to the
+models used during AL and how useful that data
+will be for other models depends on the degree of
+relatedness of the models under consideration.
+Given that AL may or may not provide cost reductions, we consider the effect that semi-automating
+annotation has on reducing the total cost of annotation when used with and without AL.
+
+
+**7** **Semi-automated labeling**
+
+Corpus building, with or without AL, is generally
+viewed as selecting examples and then _from scratch_
+labeling such examples. This can be inefficient, especially when dealing with labels that have complex
+internal structures, as a model may be able to ruleout some of the labeling possibilities.
+For our domain, we exploit the fact that we may
+already have partial information about an example’s
+label by presenting only the top _n_ -best parses to
+the annotator, who then navigates to the best parse
+
+
+
+within that set using those discriminants relevant to
+that set of parses. Rather than using a value for _n_
+that is fixed or proportional to the ambiguity of the
+sentence, we simply select all parses for which the
+model assigns a probability higher than chance. This
+has the advantage of reducing the number of parses
+presented to the annotator as the model uses more
+training material and reduces its uncertainty.
+
+
+When the true best parse is within the top _n_ presented to the annotator, the cost we record is the
+number of discriminants needed to identify it from
+that subset, plus one – the same calculation as when
+all parses are presented, with the advantage that
+fewer discriminants and parses need to be inspected.
+
+
+When the best parse is _not_ present in the _n_ -best
+subset, there is a question as to how to record the
+annotation cost. The discriminant decisions made
+in reducing the subset are still valid and useful in
+identifying the best parse from the entire set, but we
+must incur some penalty for the fact that the annotator must confirm that this is the case. To determine the cost for such situations, we add one to the
+usual full cost of annotating the sentence. This encodes what we feel is a reasonable reflection of the
+penalty since decisions taken in the _n_ -best phase are
+
+
+still valid in the context of all parses. [3]
+
+
+Performance level
+65% 70% 73%
+1. RAND 820 1950 3680
+2. LL-PROD 690 1200 2050
+3. RAND (NB) 670 1350 2430
+4. LL-PROD (NB) 680 1120 1760
+
+
+Table 3: Cost for LL-PROD to reach given performance levels when using _n_ -best automation (NB).
+
+
+Table 3 shows the effects of using semi-automated
+labeling with LL-PROD. As can be seen, random
+selection costs reduce dramatically with _n_ -best automation (compare rows 1 and 3). It is also an early
+winner over basic uncertainty sampling (row 2),
+though the latter eventually reaches the higher accuracies more quickly. Nonetheless, the mixture of
+AL and semi-automation provides the biggest overall gains: to reach 73% accuracy, _n_ -best uncertainty
+sampling (row 4) reduces the cost by 17% over _n_ best random sampling (row 3) and by 15% over basic uncertainty sampling (row 2). Similar patterns
+hold for _n_ -best automation with LL-CONFIG.
+Figure 2 provides an overall view on the accumulative effects of ensembling, _n_ -best automation, and
+uncertainty sampling in the ideal situation of reuse
+by the AL model itself. Ensemble models and _n_ -best
+automation show that massive improvements can be
+made without AL. Nonetheless, we see the largest
+reductions by using AL, _n_ -best automation, and ensemble models together: LL-PROD using uncertainty
+sampling and _n_ -best automation (row 4 of Table 3)
+reaches 73% accuracy with a cost of 1760 compared
+to 8560 needed by LL-CONFIG using random sampling without automation. This is our best annotation saving: a cost reduction of 80%.
+
+
+**8** **Closing the reuse gap**
+
+The previous section’s semi-automated labeling experiments did not involve reuse. If models are expected to evolve, could _n_ -best automation fill in the
+cost gap created by reuse? To test this, we considered reusing examples with our best model (LL
+
+3When we do not allow ourselves to benefit from such labeling decisions, our annotation savings naturally decrease, but
+not below when we do not use _n_ -best labeling.
+
+
+
+PROD), as selected by different models using both
+AL and _n_ -best automation as a combined strategy.
+For LL-CONFIG and LL-CONGLOM as selectors, the
+gap is entirely closed: costs for reuse were virtually
+equal to when LL-PROD selects examples for itself
+_without n_ -best (Table 3, row 2).
+The gap also closes when _n_ -best automation and
+AL are used with the weaker LL-MRS model. Performance (Table 4, row 1) still falls far short of LL
+PROD selecting for itself _without_ _n_ -best (Table 3,
+row 2). However, the gap closes even more when _n_ best automation and random sampling are used with
+
+LL-MRS (Table 4, line 2).
+
+
+Performance level
+65% 70% 73%
+1. NB & US 1040 1920 3320
+2. NB & RAND 680 1450 2890
+
+
+Table 4: Cost for LL-PROD to reach given performance levels in reuse situations where _n_ -best automation (NB) was used with LL-MRS with uncertainty sampling (US) or random sampling (RAND).
+
+
+Interestingly, when using a weak selector (LL
+MRS), _n_ -best automation combined with random
+sampling was _more_ effective than when combined
+with uncertainty sampling. The reason for this is
+clear. Since AL typically selects more ambiguous
+examples, a weak model has more difficulty getting
+
+
+
+80
+
+
+75
+
+
+70
+
+
+65
+
+
+60
+
+
+55
+
+
+50
+
+
+
+0 500 1000 1500 2000 2500 3000 3500 4000 4500 5000
+
+Annotation cost
+
+
+
+
+
+
+
+
+
+
+
+Figure 2: Learning curves for accumulative improvements to the annotation scenario starting from
+random sampling with LL-CONFIG: ensembling, _n_ best automation, and uncertainty sampling.
+
+
+the best parse within the _n_ -best when AL is used.
+Thus, the gains from the more informative examples
+selected by AL are surpassed by the gains that come
+with the easier labeling with random sampling.
+For most situations, _n_ -best automation is beneficial: the gap introduced by reuse can be reduced. _n_ best automation never results in an increase in cost.
+This is still true even if we do not allow ourselves to
+reuse those discriminants which were used to select
+the best parse from the _n_ -best subset and the best
+parse was not actually present in that subset.
+
+
+**9** **Related work**
+
+There is a large body of AL work in the machine
+learning literature, but less so within natural language processing (NLP). Most work in NLP has
+primarily focused upon uncertainty sampling (Hwa,
+2000; Tang et al., 2002). Hwa (2001) considered
+reuse of examples selected for one parser by another with uncertainty sampling. This performed
+better than sequential sampling but was only half as
+effective as self-selection. Here, we have considered reuse with respect to many models and their
+co-relatedness. Also, we compare reuse performance against against random sampling, which we
+showed previously to be a much stronger baseline
+than sequential sampling for the Redwoods corpus
+(Osborne and Baldridge, 2004). Hwa et al. (2003)
+showed that for parsers, AL outperforms the closely
+related co-training, and that some of the labeling
+could be automated. However, their approach requires strict independence assumptions.
+
+
+**10** **Discussion**
+
+AL should only be considered for creating labeled
+data when the the task is either well-understood or
+else the model is unlikely to substantially change.
+Otherwise, it would be prudent to consider improving either the model itself (using, for example, ensemble techniques) or else semi-automating the labeling task. Naturally, there is a cost associated with
+creating the model itself, and this in turn will need
+to be factored into the total cost. When there is genuine uncertainty about the model, or else how the
+labeled data is going to be eventually used, then the
+best strategy may well be to use random selection
+rather than AL - especially when using some form
+
+
+
+of automated annotation.
+
+**Acknowledgments**
+
+We would like to thank Markus Becker, Jeremiah
+Crim, Dan Flickinger, Alex Lascarides, Stephan
+Oepen, and Andrew Smith. We’d also like to
+thank `pc-jbaldrid` and `pc-rosie` for their
+hard work and 24/7 dedication. This work was supported by Edinburgh-Stanford Link R36763, ROSIE
+project.
+
+
+**References**
+
+David A. Cohn, Zoubin Ghahramani, and Michael I. Jordan.
+1995. Active learning with statistical models. In G. Tesauro,
+D. Touretzky, and T. Leen, editors, _Advances in Neural Infor-_
+_mation_ _Processing_ _Systems_, volume 7, pages 705–712. The
+MIT Press.
+
+Dan Flickinger. 2000. On building a more efficient grammar by
+exploiting types. _Natural_ _Language_ _Engineering_, 6(1):15–
+28. Special Issue on Efficient Processing with HPSG.
+
+G. E. Hinton. 1999. Products of experts. In _Proc. of the 9th Int._
+_Conf. on Artificial Neural Networks_, pages 1–6.
+
+Rebecca Hwa, Miles Osborne, Anoop Sarkar, and Mark Steedman. 2003. Corrected Co-training for Statistical Parsers. In
+_Proceedings_ _of_ _the_ _ICML_ _Workshop_ _“The_ _Continuum_ _from_
+_Labeled to Unlabeled Data”_, pages 95–102. ICML-03.
+
+Rebecca Hwa. 2000. Sample selection for statistical grammar induction. In _Proc. of the 2000 Joint SIGDAT Conf. on_
+_EMNLP and VLC_, pages 45–52, Hong Kong, China.
+
+Rebecca Hwa. 2001. On minimizing training corpus for parser
+acquisition. In _Proc. of the 5th Conference on Natural Lan-_
+_guage Learning_, Toulouse.
+
+Mark Johnson, Stuart Geman, Stephen Cannon, Zhiyi Chi,
+and Stephan Riezler. 1999. Estimators for Stochastic
+“Unification-Based” Grammars. In _37th Annual Meeting of_
+_the ACL_ .
+
+Stephan Oepen, Kristina Toutanova, Stuart Shieber, Christopher
+Manning, Dan Flickinger, and Thorsten Brants. 2002. The
+LinGO Redwoods Treebank: Motivation and preliminary applications. In _Proc. of the 19th International Conference on_
+_Computational Linguistics_, Taipei, Taiwan.
+
+Miles Osborne and Jason Baldridge. 2004. Ensemble-based
+active learning for parse selection. In _Proc. of HLT-NAACL_,
+Boston.
+
+H. S. Seung, Manfred Opper, and Haim Sompolinsky. 1992.
+Query by committee. In _Computational_ _Learning_ _Theory_,
+pages 287–294.
+
+Min Tang, Xiaoqiang Luo, and Salim Roukos. 2002. Active Learning for Statistical Natural Language Parsing. In
+_Proc._ _of_ _the_ 40 _[th]_ _Annual_ _Meeting_ _of_ _the_ _ACL_, pages 120–
+127, Philadelphia, Pennsylvania, USA, July.
+
+

references/2008.emnlp.settles/paper.md

@@ -0,0 +1,1459 @@
+---
+title: "An Analysis of Active Learning Strategies for Sequence Labeling Tasks"
+authors:
+  - "Burr Settles"
+  - "Mark Craven"
+year: 2008
+venue: "EMNLP 2008"
+url: "https://aclanthology.org/D08-1112/"
+---
+
+# **An Analysis of Active Learning Strategies for Sequence Labeling Tasks**
+
+
+
+**Burr Settles** _[∗†]_
+
+_∗_ Dept. of Computer Sciences
+University of Wisconsin
+Madison, WI 53706, USA
+bsettles@cs.wisc.edu
+
+
+**Abstract**
+
+
+Active learning is well-suited to many problems in natural language processing, where
+unlabeled data may be abundant but annotation is slow and expensive. This paper aims
+to shed light on the best active learning approaches for sequence labeling tasks such as
+information extraction and document segmentation. We survey previously used query selection strategies for sequence models, and propose several novel algorithms to address their
+shortcomings. We also conduct a large-scale
+empirical comparison using multiple corpora,
+which demonstrates that our proposed methods advance the state of the art.
+
+
+**1** **Introduction**
+
+
+Traditional supervised learning algorithms use
+whatever labeled data is provided to induce a model.
+By contrast, _active_ _learning_ gives the learner a degree of control by allowing it to select which instances are labeled and added to the training set. A
+typical active learner begins with a small labeled set
+_L_, selects one or more informative _query_ instances
+from a large unlabeled pool _U_, learns from these labeled queries (which are then added to _L_ ), and repeats. In this way, the learner aims to achieve high
+accuracy with as little labeling effort as possible.
+Thus, active learning can be valuable in domains
+where unlabeled data are readily available, but obtaining training labels is expensive.
+Such is the case with many _sequence_ _labeling_
+tasks in natural language domains. For example,
+part-of-speech tagging (Seung et al., 1992; Lafferty
+
+
+
+**Mark Craven** _[†∗]_
+
+_†_ Dept. of Biostatistics & Medical Informatics
+University of Wisconsin
+Madison, WI 53706, USA
+craven@biostat.wisc.edu
+
+
+et al., 2001), information extraction (Scheffer et al.,
+2001; Sang and DeMeulder, 2003; Kim et al., 2004),
+and document segmentation (Carvalho and Cohen,
+2004) are all typically treated as sequence labeling
+problems. The source data for these tasks (i.e., text
+documents in electronic form) are often easily obtained. However, due to the nature of sequence labeling tasks, annotating these texts can be rather tedious and time-consuming, making active learning
+an attractive technique.
+
+
+While there has been much work on active learning for classification (Cohn et al., 1994; McCallum
+and Nigam, 1998; Zhang and Oles, 2000; Zhu et
+al., 2003), active learning for sequence labeling has
+received considerably less attention. A few methods have been proposed, based mostly on the conventions of _uncertainty sampling_, where the learner
+queries the instance about which it has the least certainty (Scheffer et al., 2001; Culotta and McCallum,
+2005; Kim et al., 2006), or _query-by-committee_,
+where a “committee” of models selects the instance
+about which its members most disagree (Dagan and
+Engelson, 1995). We provide more detail on these
+and the new strategies we propose in Section 3.
+
+
+The comparative effectiveness of these approaches, however, has not been studied. Furthermore, it has been suggested that uncertainty sampling and query-by-committee fail on occasion (Roy
+and McCallum, 2001; Zhu et al., 2003) by querying outliers, e.g., instances considered informative
+in isolation by the learner, but containing little information about the _rest_ of the distribution of instances.
+Proposed methods for dealing with these shortcomings have so far only considered classification tasks.
+
+
+
+1070
+
+_Proceedings of the 2008 Conference on Empirical Methods in Natural Language Processing_, pages 1070–1079,
+Honolulu, October 2008. _⃝_ c 2008 Association for Computational Linguistics
+
+
+This paper presents two major contributions for
+active learning and sequence labeling tasks. First,
+we motivate and introduce several new query strategies for probabilistic sequence models. Second, we
+conduct a thorough empirical analysis of previously
+proposed methods with our algorithms on a variety
+of benchmark corpora. The remainder of this paper is organized as follows. Section 2 provides a
+brief introduction to sequence labeling and conditional random fields (the sequence model used in
+our experiments). Section 3 describes in detail all
+the query selection strategies we consider. Section 4
+presents the results of our empirical study. Section 5
+concludes with a summary of our findings.
+
+
+**2** **Sequence Labeling and CRFs**
+
+
+In this paper, we are concerned with active learning for sequence labeling. Figure 1 illustrates
+how, for example, an information extraction problem can be viewed as a sequence labeling task.
+Let **x** = _⟨x_ 1 _, . . ., xT ⟩_ be an observation sequence
+of length _T_ with a corresponding label sequence
+**y** = _⟨y_ 1 _, . . ., yT ⟩_ . Words in a sentence correspond to tokens in the input sequence **x**, which are
+mapped to labels in **y** . These labels indicate whether
+the word belongs to a particular entity class of interest (in this case, org and loc) or not (null). These
+labels can be assigned by a sequence model based
+on a finite state machine, such as the one shown to
+the right in Figure 1.
+We focus our discussion of active learning for
+sequence labeling on _conditional_ _random_ _fields_, or
+CRFs (Lafferty et al., 2001). The rest of this section serves as a brief introduction. CRFs are statistical graphical models which have demonstrated
+state-of-the-art accuracy on virtually all of the sequence labeling tasks mentioned in Section 1. We
+use linear-chain CRFs, which correspond to conditionally trained probabilistic finite state machines.
+A linear-chain CRF model with parameters _θ_ defines the posterior probability of **y** given **x** to be [1] :
+
+
+
+
+
+**x:**
+
+
+**y:**
+
+
+
+**null**
+
+
+
+
+
+
+
+**org**
+
+
+
+offices
+
+
+
+**null**
+
+
+
+in
+
+
+
+
+
+
+
+**org**
+
+
+
+
+
+
+
+Figure 1: An information extraction example treated as
+a sequence labeling task. Also shown is a corresponding
+sequence model represented as a finite state machine.
+
+
+Here _Z_ ( **x** ) is a normalization factor over all possible labelings of **x**, and _θk_ is one of _K_ model
+parameter weights corresponding to some feature
+_fk_ ( _yt−_ 1 _, yt,_ **x** _t_ ). Each feature _fk_ describes the sequence **x** at position _t_ with label _yt_, observed along
+a transition from label states _yt−_ 1 to _yt_ in the finite
+state machine. Consider the example text from Figure 1. Here, _fk_ might be the feature WORD= _ACME_
+and have the value _fk_ = 1 along a transition from
+the null state to the org state (and 0 elsewhere).
+Other features set to 1 here might be ALLCAPS and
+NEXTWORD= _Inc._ The weights in _θ_ are set to maximize the conditional log likelihood _ℓ_ of training sequences in the labeled data set _L_ :
+
+
+
+_K_
+
+
+
+_k_ =1
+
+
+
+_ℓ_ ( _L_ ; _θ_ ) =
+
+
+
+_L_
+
+
+log _P_ ( **y** [(] _[l]_ [)] _|_ **x** [(] _[l]_ [)] ; _θ_ ) _−_
+
+_l_ =1
+
+
+
+_θk_ [2] _[,]_
+2 _σ_ [2]
+
+
+
+_θkfk_ ( _yt−_ 1 _, yt,_ **x** _t_ )
+
+_k_ =1
+
+
+
+
+
+
+
+1
+_P_ ( **y** _|_ **x** ; _θ_ ) =
+_Z_ ( **x** ) [exp]
+
+
+
+
+- _T_
+
+ 
+
+_t_ =1
+
+
+
+_K_
+
+
+
+
+_._
+
+
+(1)
+
+
+
+1Our discussion assumes, without loss of generality, that
+each label is uniquely represented by one state, thus each label
+sequence **y** corresponds to exactly one path through the model.
+
+
+1071
+
+
+
+where _L_ is the size of the labeled set _L_, and the second term is a Gaussian regularization penalty on _∥θ∥_
+to prevent over-fitting. After training, labels can be
+predicted for new sequences using the Viterbi algorithm. For more details on CRFs and their training
+procedures, see Sutton and McCallum (2006).
+Note that, while we describe the active learning
+algorithms in the next section in terms of linearchain CRFs, they have analogs for other kinds of
+sequence models, such as hidden Markov models,
+or HMMs (Rabiner, 1989), probabilistic contextfree grammars (Lari and Young, 1990), and general
+CRFs (Sutton and McCallum, 2006).
+
+
+**3** **Active Learning with Sequence Models**
+
+
+In order to select queries, an active learner must have
+a way of assessing how _informative_ each instance is.
+Let **x** _[∗]_ be the most informative instance according to
+some query strategy _φ_ ( **x** ), which is a function used
+to evaluate each instance **x** in the unlabeled pool _U_ .
+
+
+**Given** : Labeled set _L_, unlabeled pool _U_, query
+strategy _φ_ ( _·_ ), query batch size _B_
+
+
+**repeat**
+
+_// learn a model using the current L_
+_θ_ = train( _L_ ) ;
+**for** _b_ = 1 **to** _B_ **do**
+
+_// query the most informative instance_
+**x** _[∗]_ _b_ [= arg max] **[x]** _[∈U][ φ]_ [(] **[x]** [)][ ;]
+_// move the labeled query from U_ _to L_
+_L_ = _L ∪⟨_ **x** _[∗]_ _b_ _[,]_ [ label(] **[x]** _[∗]_ _b_ [)] _[⟩]_ [;]
+_U_ = _U_ _−_ **x** _[∗]_ _b_ [;]
+**end**
+**until** _some stopping criterion_ ;
+
+
+**Algorithm 1** : Pool-based active learning.
+
+
+Algorithm 1 provides a sketch of the generic poolbased active learning scenario.
+In the remainder of this section, we describe various query strategy formulations of _φ_ ( _·_ ) that have
+been used for active learning with sequence models. We also point out where we think these approaches may be flawed, and propose several novel
+query strategies to address these issues.
+
+
+**3.1** **Uncertainty Sampling**
+
+
+One of the most common general frameworks for
+measuring informativeness is _uncertainty_ _sampling_
+(Lewis and Catlett, 1994), where a learner queries
+the instance that it is most uncertain how to label. Culotta and McCallum (2005) employ a simple uncertainty-based strategy for sequence models
+called **least confidence (LC)** :
+
+
+_φ_ _[LC]_ ( **x** ) = 1 _−_ _P_ ( **y** _[∗]_ _|_ **x** ; _θ_ ) _._
+
+
+Here, **y** _[∗]_ is the most likely label sequence, i.e., the
+Viterbi parse. This approach queries the instance
+for which the current model has the least confidence
+in its most likely labeling. For CRFs, this confidence can be calculated using the posterior probability given by Equation (1).
+Scheffer et al. (2001) propose another uncertainty
+strategy, which queries the instance with the smallest
+margin between the posteriors for its two most likely
+labelings. We call this approach **margin (M)** :
+
+
+_φ_ _[M]_ ( **x** ) = _−_   - _P_ ( **y** 1 _[∗][|]_ **[x]** [;] _[ θ]_ [)] _[ −]_ _[P]_ [(] **[y]** 2 _[∗][|]_ **[x]** [;] _[ θ]_ [)]   - _._
+
+
+1072
+
+
+
+Here, **y** 1 _[∗]_ [and] **[y]** 2 _[∗]_ [are] [the] [first] [and] [second] [best] [la-]
+bel sequences, respectively. These can be efficiently
+computed using the _N_ -best algorithm (Schwartz
+and Chow, 1990), a beam-search generalization of
+Viterbi, with _N_ = 2. The minus sign in front is simply to ensure that _φ_ _[M]_ acts as a maximizer for use
+with Algorithm 1.
+Another uncertainty-based measure of informativeness is _entropy_ (Shannon, 1948). For a discrete random variable _Y_, the entropy is given by
+_H_ ( _Y_ ) = _−_ [�] _i_ _[P]_ [(] _[y][i]_ [) log] _[ P]_ [(] _[y][i]_ [)][,] [and] [represents] [the]
+information needed to “encode” the distribution of
+outcomes for _Y_ . As such, is it often thought of as
+a measure of uncertainty in machine learning. In
+active learning, we wish to use the entropy of our
+model’s posteriors over its labelings. One way this
+has been done with probabilistic sequence models is
+by computing what we call **token entropy (TE)** :
+
+
+
+_M_
+
+- _Pθ_ ( _yt_ = _m_ ) log _Pθ_ ( _yt_ = _m_ ) _,_
+
+
+_m_ =1
+
+
+
+_φ_ _[T E]_ ( **x** ) = _−_ [1]
+
+_T_
+
+
+
+_T_
+
+
+
+_t_ =1
+
+
+
+(2)
+where _T_ is the length of **x**, _m_ ranges over all possible token labels, and _Pθ_ ( _yt_ = _m_ ) is shorthand
+for the marginal probability that _m_ is the label at
+position _t_ in the sequence, according to the model.
+For CRFs and HMMs, these marginals can be efficiently computed using the _forward_ and _backward_
+algorithms (Rabiner, 1989). The summed token entropies have typically been normalized by sequence
+length _T_, to avoid simply querying longer sequences
+(Baldridge and Osborne, 2004; Hwa, 2004). However, we argue that querying long sequences should
+not be explicitly discouraged, if in fact they contain
+more information. Thus, we also propose the **total**
+**token entropy (TTE)** measure:
+
+
+_φ_ _[TTE]_ ( **x** ) = _T_ _× φ_ _[TE]_ ( **x** ) _._
+
+
+For most sequence labeling tasks, however, it is
+more appropriate to consider the entropy of the label sequence **y** as a whole, rather than some aggregate of individual token entropies. Thus an alternate
+query strategy is **sequence entropy (SE)** :
+
+
+   _φ_ _[SE]_ ( **x** ) = _−_ _P_ (ˆ **y** _|_ **x** ; _θ_ ) log _P_ (ˆ **y** _|_ **x** ; _θ_ ) _,_ (3)
+
+
+**y** ˆ
+
+
+where **y** ˆ ranges over all possible label sequences for
+input sequence **x** . Note, however, that the number
+
+
+of possible labelings grows exponentially with the
+length of **x** . To make this feasible, previous work
+(Kim et al., 2006) has employed an approximation
+we call _**N**_ **-best sequence entropy (NSE)** :
+
+
+   _φ_ _[NSE]_ ( **x** ) = _−_ _P_ (ˆ **y** _|_ **x** ; _θ_ ) log _P_ (ˆ **y** _|_ **x** ; _θ_ ) _,_
+
+
+**y** ˆ _∈N_
+
+where _N_ = _{_ **y** 1 _[∗][, . . .,]_ **[ y]** _N_ _[∗]_ _[}]_ [,] [the] [set] [of] [the] _[N]_ [most]
+likely parses, and the posteriors are re-normalized
+(i.e., _Z_ ( **x** ) in Equation (1) only ranges over _N_ ). For
+_N_ = 2, this approximation is equivalent to _φ_ _[M]_, thus
+_N_ -best sequence entropy can be thought of as a generalization of the margin approach.
+Recently, an efficient entropy calculation via dynamic programming was proposed for CRFs in the
+context of semi-supervised learning (Mann and McCallum, 2007). We use this algorithm to compute
+the true sequence entropy (3) for active learning in
+a constant-time factor of Viterbi’s complexity. Hwa
+(2004) employed a similar approach for active learning with probabilistic context-free grammars.
+
+
+**3.2** **Query-By-Committee**
+
+
+Another general active learning framework is the
+_query-by-committee_ (QBC) approach (Seung et al.,
+1992). In this setting, we use a committee of models
+_C_ = _{θ_ [(1)] _, . . ., θ_ [(] _[C]_ [)] _}_ to represent _C_ different hypotheses that are consistent with the labeled set _L_ .
+The most informative query, then, is the instance
+over which the committee is in most disagreement
+about how to label.
+In particular, we use the _query-by-bagging_ approach (Abe and Mamitsuka, 1998) to learn a committee of CRFs. In each round of active learning,
+_L_ is sampled (with replacement) _L_ times to create
+a unique, modified labeled set _L_ [(] _[c]_ [)] . Each model
+_θ_ [(] _[c]_ [)] _∈C_ is then trained using its own corresponding
+labeled set _L_ [(] _[c]_ [)] . To measure disagreement among
+committee members, we consider two alternatives.
+Dagan and Engelson (1995) introduced QBC with
+HMMs for part-of-speech tagging using a measure
+called **vote entropy (VE)** :
+
+
+
+McCallum and Nigam (1998) propose a QBC
+strategy for classification based on _Kullback-Leibler_
+_(KL)_ _divergence_, an information-theoretic measure
+of the difference between two probability distributions. The most informative query is considered to
+be the one with the largest average KL divergence
+between a committee member’s posterior label distribution and the consensus. We modify this approach for sequence models by summing the average
+KL scores using the marginals at each token position
+and, as with vote entropy, normalizing for length.
+We call this approach **Kullback-Leibler (KL)** :
+
+
+
+Here _PC_ ( _yt_ = _m_ ) = _C_ [1] - _Cc_ =1 _[P]_ _θ_ [(] _[c]_ [)][(] _[y][t]_ [=] _[ m]_ [)][, or the]
+
+“consensus” marginal probability that _m_ is the label
+at position _t_ in the sequence.
+Both of these disagreement measures are normalized for sequence length _T_ . As with token entropy (2), this may bias the learner toward querying shorter sequences. To study the effects of normalization, we also conduct experiments with **non-**
+**normalized variants** _φ_ _[TV E]_ and _φ_ _[TKL]_ .
+Additionally, we argue that these token-level disagreement measures may be less appropriate for
+most tasks than measuring the committee’s disagreement about the label sequence **y** as a whole. Therefore, we propose **sequence vote entropy (SVE)** :
+
+
+   _φ_ _[SV E]_ ( **x** ) = _−_ _P_ (ˆ **y** _|_ **x** ; _C_ ) log _P_ (ˆ **y** _|_ **x** ; _C_ ) _,_
+
+
+**y** ˆ _∈N_ _[C]_
+
+
+where _N_ _[C]_ is the union of the _N_ -best parses from
+all models in the committee _C_, and _P_ (ˆ **y** _|_ **x** ; _C_ ) =
+_C_ 1 - _Cc_ =1 _[P]_ [(ˆ] **[y]** _[|]_ **[x]** [;] _[ θ]_ [(] _[c]_ [)][)][,] [or] [the] [“consensus”] [posterior]
+probability for some label sequence **y** ˆ. This can be
+thought of as a QBC generalization of _N_ -best entropy, where each committee member casts a vote
+for the posterior label distribution. We also explore
+a **sequence Kullback-Leibler (SKL)** variant:
+
+
+
+_C_
+
+
+_D_ ( _θ_ [(] _[c]_ [)] _∥C_ ) _,_
+
+_c_ =1
+
+
+
+_φ_ _[KL]_ ( **x** ) = [1]
+
+_T_
+
+
+
+_T_
+
+
+
+_t_ =1
+
+
+
+1
+
+_C_
+
+
+
+where (using shorthand again):
+
+
+
+_PC_ ( _yt_ = _m_ ) _[.]_
+
+
+
+_D_ ( _θ_ [(] _[c]_ [)] _∥C_ ) =
+
+
+
+_M_
+
+
+
+
+
+- [=] _[ m]_ [)]
+
+_Pθ_ ( _c_ )( _yt_ = _m_ ) log _[P]_ _P_ _[θ]_ [(] _C_ _[c]_ ( [)] _y_ [(] _[y]_ _t_ _[t]_ = _m_ )
+_m_ =1
+
+
+
+_M_
+
+
+
+_m_ =1
+
+
+
+_,_
+_C_
+
+
+
+_φ_ _[V E]_ ( **x** ) = _−_ [1]
+
+_T_
+
+
+
+_T_
+
+
+
+_t_ =1
+
+
+
+_V_ ( _yt, m_ )
+
+
+
+_t, m_ )
+
+log _[V]_ [ (] _[y][t][, m]_ [)]
+_C_ _C_
+
+
+
+where _V_ ( _yt, m_ ) is the number of “votes” label _m_ receives from all the committee member’s Viterbi labelings at sequence position _t_ .
+
+
+1073
+
+
+
+
+- _P_ (ˆ **y** _|_ **x** ; _θ_ [(] _[c]_ [)] ) log _[P]_ [(ˆ] **[y]** _[|]_ **[x]** [;] _[ θ]_ [(] _[c]_ [)][)]
+
+_P_ (ˆ **y** _|_ **x** ; _C_ )
+
+**y** ˆ _∈N_ _[C]_
+
+
+
+_φ_ _[SKL]_ ( **x** ) = [1]
+
+_C_
+
+
+
+_C_
+
+
+
+_c_ =1
+
+
+
+
+
+
+
+_P_ (ˆ **y** _|_ **x** ; _C_ ) _[.]_
+
+
+**3.3** **Expected Gradient Length**
+
+
+A third general active learning framework we consider is to query the instance that would impart the
+greatest change to the current model _if_ _we_ _knew_ _its_
+_label_ . Since we train discriminative models like
+CRFs using gradient-based optimization, this involves querying the instance which, if labeled and
+added to the training set, would create the greatest
+change in the gradient of the objective function (i.e.,
+the largest gradient vector used to re-estimate parameter values).
+Let _∇ℓ_ ( _L_ ; _θ_ ) be the gradient of the loglikelihood _ℓ_ with respect to the model parameters _θ_,
+as given by Sutton and McCallum (2006). Now let
+_∇ℓ_ ( _L_ [+] _[⟨]_ **[x]** _[,]_ **[y]** _[⟩]_ ; _θ_ ) be the new gradient that would be
+obtained by adding the training tuple _⟨_ **x** _,_ **y** _⟩_ to _L_ .
+Since the query algorithm does not know the true label sequence **y** in advance, we instead calculate the
+**expected gradient length (EGL)** :
+
+
+_φ_ _[EGL]_ ( **x** ) =  - _P_ (ˆ **y** _|_ **x** ; _θ_ ) _∇ℓ_ ( _L_ + _⟨_ **x** _,_ **y** ˆ _⟩_ ; _θ_ ) _,_
+
+��� ���
+**y** ˆ _∈N_
+
+
+approximated as an expectation over the _N_ -best labelings, where _∥· ∥_ is the Euclidean norm of each
+resulting gradient vector. We first introduced this approach in previous work on multiple-instance active
+learning (Settles et al., 2008), and adapt it to query
+selection with sequences here. Note that, at query
+time, _∇ℓ_ ( _L_ ; _θ_ ) should be nearly zero since _ℓ_ converged at the previous round of training. Thus, we
+can approximate _∇ℓ_ ( _L_ [+] _[⟨]_ **[x]** _[,]_ **[y]** [ˆ] _[⟩]_ ; _θ_ ) _≈∇ℓ_ ( _⟨_ **x** _,_ ˆ **y** _⟩_ ; _θ_ )
+for computational efficiency, because the training instances are assumed to be independent.
+
+
+**3.4** **Information Density**
+
+
+It has been suggested that uncertainty sampling and
+QBC are prone to querying outliers (Roy and McCallum, 2001; Zhu et al., 2003). Figure 2 illustrates this problem for a binary linear classifier using uncertainty sampling. The least certain instance
+lies on the classification boundary, but is not “representative” of other instances in the distribution, so
+knowing its label is unlikely to improve accuracy on
+the data as a whole. QBC and EGL exhibit similar
+behavior, by spending time querying possible outliers simply because they are controversial, or are
+expected to impart significant change in the model.
+
+
+1074
+
+
+
+
+
+
+
+Figure 2: An illustration of when uncertainty sampling
+can be a poor strategy for classification. Shaded polygons represent labeled instances ( _L_ ), and circles represent unlabeled instances ( _U_ ). Since _A_ is on the decision
+boundary, it will be queried as the most uncertain. However, querying _B_ is likely to result in more information
+about the data as a whole.
+
+
+We argue that this phenomenon can occur with sequence labeling tasks as well as with classification.
+To address this, we propose a new active learning
+approach called **information density (ID)** :
+
+
+
+_β_
+
+
+
+
+_._
+
+
+
+_U_
+
+
+
+
+
+
+sim( **x** _,_ **x** [(] _[u]_ [)] )
+
+_u_ =1
+
+
+
+_φ_ _[ID]_ ( **x** ) = _φ_ _[SE]_ ( **x** ) _×_
+
+
+
+
+1
+
+_U_
+
+
+
+That is, the informativeness of **x** is weighted by its
+average similarity to all other sequences in _U_, subject to a parameter _β_ that controls the relative importance of the density term. In the formulation presented above, sequence entropy _φ_ _[SE]_ measures the
+“base” informativeness, but we could just as easily
+use any of the instance-level strategies presented in
+the previous sections.
+This density measure requires us to compute the
+similarity of two sequences. To do this, we first
+transform each **x**, which is a sequence of feature
+vectors (tokens), into a single kernel vector _⃗_ **x** :
+
+
+
+_T_
+
+
+
+
+_fJ_ ( _xt_ )
+
+_t_ =1
+
+
+
+
+
+
+
+_⃗_ **x** =
+
+
+
+
+- _T_
+
+
+_f_ 1( _xt_ ) _, . . .,_
+
+_t_ =1
+
+
+
+_,_
+
+
+
+where _fj_ ( _xt_ ) is the value of feature _fj_ for token _xt_,
+and _J_ is the number of features in the input representation [2] . In other words, sequence **x** is compressed
+into a fixed-length feature vector _⃗_ **x**, for which each
+element is the sum of the corresponding feature’s
+values across all tokens. We can then use cosine
+
+
+2Note that _J_ = _K_, and _fj_ ( _xt_ ) here differs slightly from the
+feature definition given in Section 2. Since the labels _yt−_ 1 and
+_yt_ are unknown before querying, the _K_ features used for model
+training are reduced down to the _J_ input features here, which
+factor out any label dependencies.
+
+
+similarity on this simplified representation:
+
+
+_⃗_ **x** _· ⃗_ **x** [(] _[u]_ [)]
+sim _cos_ ( **x** _,_ **x** [(] _[u]_ [)] ) =
+
+_∥⃗_ **x** _∥× ∥⃗_ **x** [(] _[u]_ [)] _∥_ _[.]_
+
+
+We have also investigated similarity functions
+based on exponentiated Euclidean distance and KLdivergence, the latter of which was also employed by
+McCallum and Nigam (1998) for density-weighting
+QBC in text classification. However, these measures
+show no improvement over cosine similarity, and require setting additional hyper-parameters.
+One potential drawback of information density is
+that the number of required similarity calculations
+grows quadratically with the number of instances
+in _U_ . For pool-based active learning, we often assume that the size of _U_ is very large. However,
+these densities only need to be computed once, and
+are independent of the base information measure.
+Thus, when employing information density in a realworld interactive learning setting, the density scores
+can simply be pre-computed and cached for efficient
+lookup during the actual active learning process.
+
+
+**3.5** **Fisher Information**
+
+
+We also introduce a query selection strategy for sequence models based on _Fisher_ _information_, building on the theoretical framework of Zhang and Oles
+(2000). Fisher information _I_ ( _θ_ ) represents the overall uncertainty about the estimated model parameters _θ_, as given by:
+
+
+
+Previously, Fisher information for active learning
+has only been investigated in the context of simple
+binary classification. When employing FIR with sequence models like CRFs, there are two additional
+computational challenges. First, we must integrate
+over all possible labelings **y**, which can, as we have
+seen, be approximated as an expectation over the _N_ best labelings. Second, the inner product in the ratio
+calculation (4) requires inverting a _K_ _×_ _K_ matrix
+for each **x** . In most interesting natural language applications, _K_ is very large, making this algorithm
+intractable. However, it is common in similar situations to approximate the Fisher information matrix
+with its diagonal (Nyffenegger et al., 2006). Thus
+we estimate _I_ **x** ( _θ_ ) using:
+
+
+
+_I_ **x** ( _θ_ ) = 
+
+
+��
+_∂_ log _P_ (ˆ **y** _|_ **x** ; _θ_ )
+_∂θ_ 1
+
+
+
+_P_ (ˆ **y** _|_ **x** ; _θ_ )
+
+**y** ˆ _∈N_
+
+
+
+�2
++ _δ, . . .,_
+
+
+
+
+
+_,_
+
+
+
+
+- _∂_ log _P_ (ˆ **y** _|_ **x** ; _θ_ )
+_∂θK_
+
+
+
+�2
++ _δ_
+
+
+
+and _IU_ ( _θ_ ) using:
+
+
+_IU_ ( _θ_ ) = [1]
+
+_U_
+
+
+
+_U_
+
+
+_I_ **x** ( _u_ )( _θ_ ) _._
+_u_ =1
+
+
+
+For CRFs, the partial derivative at the root of each
+element in the diagonal vector is given by:
+
+
+
+_T_
+
+- _fk_ (ˆ _yt−_ 1 _,_ ˆ _yt,_ **x** _t_ )
+
+
+_t_ =1
+
+
+
+_T_
+
+
+
+_t_ =1
+
+
+
+
+    _I_ ( _θ_ ) = _−_
+
+
+
+
+   _P_ ( **x** )
+**x**
+
+
+
+_P_ ( **y** _|_ **x** ; _θ_ ) _[∂]_ [2]
+**y** _∂θ_
+
+
+
+
+[log] _[ P]_ [(] **[y]** _[|]_ **[x]** [;] _[ θ]_ [)] _[.]_
+_∂θ_ [2]
+
+
+
+_∂_ log _P_ (ˆ **y** _|_ **x** ; _θ_ )
+
+=
+_∂θk_
+
+
+_−_
+
+
+
+
+- _P_ ( _y, y_ _[′]_ _|_ **x** ) _fk_ ( _y, y_ _[′]_ _,_ **x** _t_ ) _,_
+
+_y,y_ _[′]_
+
+
+
+For a model with _K_ parameters, the Fisher information takes the form of a _K_ _× K_ covariance matrix. Our goal in active learning is to select the query
+that most efficiently minimizes the model variance
+reflected in _I_ ( _θ_ ). This can be accomplished by optimizing the **Fisher information ratio (FIR)** :
+
+
+_φ_ _[FIR]_ ( **x** ) = _−_ tr    - _I_ **x** ( _θ_ ) _[−]_ [1] _IU_ ( _θ_ )� _,_ (4)
+
+
+where _I_ **x** ( _θ_ ) and _IU_ ( _θ_ ) are Fisher information matrices for sequence **x** and the unlabeled pool _U_, respectively. The leading minus sign again ensures
+that _φ_ _[FIR]_ is a maximizer for use with Algorithm 1.
+
+
+1075
+
+
+
+which is similar to the equation used to compute the
+training gradient, but without a regularization term.
+A smoothing parameter _δ_ _≪_ 1 is added to prevent
+division by zero when computing the ratio.
+
+Notice that this method implicitly selects representative instances by favoring queries with Fisher
+information _I_ **x** ( _θ_ ) that is not only high, but similar
+to that of the overall data distribution _IU_ ( _θ_ ). This
+is in contrast to information density, which tries to
+query representative instances by explicitly modeling the distribution with a density weight.
+
+
+|Corpus|Entities Features Instances|
+|---|---|
+|CoNLL-03<br>NLPBA<br>BioCreative<br>FlySlip<br>CORA:Headers<br>CORA:References<br>Sig+Reply<br>SigIE|4<br>78,644<br>19,959<br>5<br>128,401<br>18,854<br>1<br>175,331<br>10,000<br>1<br>31,353<br>1,220<br>15<br>22,077<br>935<br>13<br>4,208<br>500<br>2<br>25<br>617<br>12<br>10,600<br>250|
+
+
+Table 1: Properties of the different evaluation corpora.
+
+
+**4** **Empirical Evaluation**
+
+
+In this section we present a large-scale empirical
+analysis of the query strategies described in Section 3 on eight benchmark information extraction
+and document segmentation corpora. The data sets
+are summarized in Table 1.
+
+
+**4.1** **Data and Methodology**
+
+
+CoNLL-03 (Sang and DeMeulder, 2003) is a collection of newswire articles annotated with four entities: person, organization, location, and misc.
+NLPBA (Kim et al., 2004) is a large collection
+of biomedical abstracts annotated with five entities
+of interest, such as protein, RNA, and cell-type.
+BioCreative (Yeh et al., 2005) and FlySlip (Vlachos, 2007) also comprise texts in the biomedical
+domain, annotated for gene entity mentions in articles from the human and fruit fly literature, respectively. CORA (Peng and McCallum, 2004) consists
+of two collections: a set of research paper headers
+annotated for entities such as title, author, and institution; and a collection of references annotated with
+BibTeX fields such as journal, year, and publisher.
+The Sig+Reply corpus (Carvalho and Cohen, 2004)
+is a set of email messages annotated for signature
+and quoted reply line segments. SigIE is a subset of
+the signature blocks from Sig+Reply which we have
+enhanced with several address book fields such as
+name, email, and phone. All corpora are formatted in the “IOB” sequence representation (Ramshaw
+and Marcus, 1995).
+We implement all fifteen query selection strategies described in Section 3 for use with CRFs, and
+evaluate them on all eight data sets. We also compare against two baseline strategies: random instance selection (i.e., passive learning), and na¨ıvely
+querying the longest sequence in terms of tokens.
+
+
+1076
+
+
+
+We use a typical feature set for each corpus based on
+the cited literature (including words, orthographic
+patterns, part-of-speech, lexicons, etc.). Where the
+_N_ -best approximation is used _N_ = 15, and for all
+QBC methods _C_ = 3; these figures exhibited a good
+balance of accuracy and training speed in preliminary work. For information density, we arbitrarily
+set _β_ = 1 (i.e., the information and density terms
+have equal weight). In each experiment, _L_ is initialized with five random labeled instances, and up
+to 150 queries are subsequently selected from _U_ in
+batches of size _B_ = 5. All results are averaged
+across five folds using cross-validation.
+We evaluate each query strategy by constructing
+learning curves that plot the overall _F_ 1 measure (for
+all entities or segments) as a function of the number of instances queried. Due to lack of space, we
+cannot show learning curves for every experiment.
+Instead, Table 2 summarizes our results by reporting
+the area under the learning curve for all strategies
+on all data. Figure 3 presents a few representative
+learning curves for six of the corpora.
+
+
+**4.2** **Discussion of Learning Curves**
+
+
+The first conclusion we can draw from these results
+is that there is no single clear winner. However, information density (ID), which we introduce in this
+paper, stands out. It usually improves upon the base
+sequence entropy measure, never performs poorly,
+and has the highest average area under the learning
+curve across all tasks. It seems particularly effective
+on large corpora, which is a typical assumption for
+the active learning setting. Sequence vote entropy
+(SVE), a QBC method we propose here, is also noteworthy in that it is fairly consistently among the top
+three strategies, although never the best.
+Second, the top uncertainty sampling strategies
+are least confidence (LC) and sequence entropy
+(SE), the latter being the dominant entropy-based
+method. Among the QBC strategies, sequence vote
+entropy (SVE) is the clear winner. We conclude that
+these three methods are the best base information
+measures for use with information density.
+Third, query strategies that evaluate the entire sequence (SE, SVE, SKL) are generally superior to those which aggregate token-level information. Furthermore, the _total_ token-level strategies (TTE, TVE, TKL) outperform their _length-_
+
+
+|Corpus|Baselines<br>Rand Long|Uncertainty Sampling<br>LC M TE TTE SE NSE|Query-By-Committee<br>VE KL TVE TKL SVE SKL|Other<br>EGL ID FIR|Col6|Col7|
+|---|---|---|---|---|---|---|
+|CoNLL-03<br>NLPBA<br>BioCreative<br>FlySlip<br>Headers<br>References<br>Sig+Reply<br>SigIE|78.8<br>79.4<br>59.9<br>67.6<br>34.6<br>26.9<br>112.1<br>121.0<br>76.0<br>78.2<br>**90.0**<br>86.0<br>129.1<br>129.6<br>84.3<br>82.7|89.4<br>84.5<br>38.9<br>**89.7**<br>**90.1**<br>89.1<br>71.0<br>62.9<br>53.4<br>70.9<br>71.5<br>68.9<br>54.8<br>46.8<br>37.8<br>53.0<br>56.0<br>50.5<br>125.1<br>119.5<br>110.3<br>124.9<br>**125.4**<br>124.1<br>**81.4**<br>78.6<br>78.5<br>78.5<br>**80.8**<br>80.4<br>89.8<br>**91.5**<br>84.4<br>88.6<br>88.4<br>89.4<br>132.1<br>132.3<br>131.7<br>131.6<br>131.4<br>**133.1**<br>88.8<br>87.3<br>89.3<br>88.3<br>87.6<br>89.1|45.9<br>62.0<br>86.7<br>81.7<br>89.0<br>87.9<br>52.4<br>53.1<br>66.9<br>63.5<br>**71.8**<br>68.5<br>35.2<br>37.4<br>49.2<br>45.1<br>**56.6**<br>50.8<br>113.3<br>109.4<br>124.1<br>119.5<br>122.7<br>120.7<br>72.8<br>78.5<br>79.7<br>78.5<br>**80.7**<br>78.4<br>85.1<br>89.1<br>88.7<br>88.2<br>**89.9**<br>86.9<br>131.4<br>130.7<br>132.1<br>130.6<br>**132.8**<br>132.3<br>**89.8**<br>85.5<br>**89.7**<br>85.1<br>**89.5**<br>**89.7**|87.3<br>**89.6**<br>81.7<br>69.3<br>**73.1**<br>**73.6**<br>51.5<br>**59.1**<br>**58.8**<br>**125.9**<br>**126.8**<br>118.2<br>79.6<br>80.2<br>79.1<br>88.2<br>88.7<br>87.1<br>130.5<br>131.5<br>**133.2**<br>87.7<br>88.5<br>88.5|87.3<br>**89.6**<br>81.7<br>69.3<br>**73.1**<br>**73.6**<br>51.5<br>**59.1**<br>**58.8**<br>**125.9**<br>**126.8**<br>118.2<br>79.6<br>80.2<br>79.1<br>88.2<br>88.7<br>87.1<br>130.5<br>131.5<br>**133.2**<br>87.7<br>88.5<br>88.5|87.3<br>**89.6**<br>81.7<br>69.3<br>**73.1**<br>**73.6**<br>51.5<br>**59.1**<br>**58.8**<br>**125.9**<br>**126.8**<br>118.2<br>79.6<br>80.2<br>79.1<br>88.2<br>88.7<br>87.1<br>130.5<br>131.5<br>**133.2**<br>87.7<br>88.5<br>88.5|
+|Average|83.1<br>83.9|**91.6**<br>87.9<br>78.0<br>90.7<br>**91.4**<br>90.6|78.2<br>80.7<br>89.6<br>86.5<br>**91.6**<br>89.4|90.0|**92.2**|90.0|
+
+
+Table 2: Detailed results for all query strategies on all evaluation corpora. Reported is the area under the _F_ 1 learning
+curve for each strategy after 150 queries (maximum possible score is 150). For each row, the **best method** is shown
+boxed in bold, the **second best** is shown underlined in bold, and the **third best** is shown in bold. The last row summarizes the results across all eight tasks by reporting the average area for each strategy. Query strategy formulations for
+sequence models introduced in this paper are indicated with italics along the top.
+
+
+
+_normalized_ counterparts (TE, VE, KL) in nearly all
+cases. In fact, the normalized variants are often inferior even to the baselines. While an argument can
+be made that these shorter sequences might be easier to label from a human annotator’s perspective,
+our ongoing work indicates that the relationship between instance length and actual labeling costs (e.g.,
+elapsed annotation time) is not a simple one. Analysis of our experiment logs also shows that lengthnormalized methods are occasionally biased toward
+short sequences with little intuitive value (e.g., sentences with few or no entities to label). In addition,
+vote entropy appears to be a better disagreement
+measure for QBC strategies than KL divergence.
+Finally, Fisher information (FIR), while theoretically sound, exhibits behavior that is difficult to interpret. It is sometimes the winning strategy, but occasionally only on par with the baselines. When it
+does show significant gains over the other strategies,
+these gains appear to be only for the first several
+queries (e.g., NLPBA and BioCreative in Figure 3).
+This inconsistent performance may be a result of the
+approximations made for computational efficiency.
+Expected gradient length (EGL) also appears to exhibit mediocre performance, and is likely not worth
+its additional computational expense.
+
+
+**4.3** **Discussion of Run Times**
+
+
+Here we discuss the execution times for each query
+strategy using current hardware. The uncertainty
+sampling methods are roughly comparable in run
+time (token-based methods run slightly faster), each
+routinely evaluating tens of thousands of sequences
+
+
+1077
+
+
+
+in under a minute. The QBC methods, on the other
+hand, must re-train multiple models with each query,
+resulting in a lag of three to four minutes per query
+batch (and up to 20 minutes for corpora with more
+entity labels).
+The expected gradient length and Fisher information methods are the most computationally expensive, because they must first perform inference over
+the possible labelings and then calculate gradients
+for each candidate label sequence. As a result, they
+take eight to ten minutes (upwards of a half hour on
+the larger corpora) for each query. Unlike the other
+strategies, their time complexities also scale linearly
+with the number of model parameters _K_ which, in
+turn, increases as new sequences are added to _L_ .
+As noted in Section 3.4, information density incurs a large computational cost to estimate the density weights, but these can be pre-computed and
+cached for efficient lookup. In our experiments, this
+pre-processing step takes less than a minute for the
+smaller corpora, about a half hour for CoNLL-03
+and BioCreative, and under two hours for NLPBA.
+The density lookup causes no significant change in
+the run time of the base information measure. Given
+these results, we advocate information density with
+an uncertainty sampling base measure in practice,
+particularly for active learning with large corpora.
+
+
+**5** **Conclusion**
+
+
+In this paper, we have presented a detailed analysis of active learning for sequence labeling tasks.
+In particular, we have described and criticized the
+query selection strategies used with probabilistic se
+
+0.7
+
+0.6
+
+0.5
+
+0.4
+
+0.3
+
+0.2
+
+0.1
+
+0
+
+
+1
+
+
+0.8
+
+
+0.6
+
+
+0.4
+
+
+0.2
+
+
+0
+
+
+
+0 20 40 60 80 100 120 140
+
+number of instances queried
+
+
+
+0 20 40 60 80 100 120 140
+
+number of instances queried
+
+
+
+0 20 40 60 80 100 120 140
+
+number of instances queried
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+0 20 40 60 80 100 120 140
+
+
+
+0 20 40 60 80 100 120 140
+
+
+
+0 20 40 60 80 100 120 140
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+0.7
+
+0.6
+
+0.5
+
+0.4
+
+0.3
+
+0.2
+
+0.1
+
+0
+
+
+0.7
+
+0.6
+
+0.5
+
+0.4
+
+0.3
+
+0.2
+
+0.1
+
+0
+
+
+
+0.7
+
+0.6
+
+0.5
+
+0.4
+
+0.3
+
+0.2
+
+0.1
+
+0
+
+
+1
+
+
+0.8
+
+
+0.6
+
+
+0.4
+
+
+0.2
+
+
+0
+
+
+
+Figure 3: Learning curves for selected query strategies on six of the evaluation corpora.
+
+
+
+quence models to date, and proposed several novel
+strategies to address some of their shortcomings.
+Our large-scale empirical evaluation demonstrates
+that some of these newly proposed methods advance
+the state of the art in active learning with sequence
+models. These methods include information density
+(which we recommend in practice), sequence vote
+entropy, and sometimes Fisher information.
+
+
+**Acknowledgments**
+
+
+We would like to thank the anonymous reviewers for
+their helpful feedback. This work was supported by
+NIH grants T15-LM07359 and R01-LM07050.
+
+
+**References**
+
+
+N. Abe and H. Mamitsuka. 1998. Query learning strategies using boosting and bagging. In _Proceedings_ _of_
+_the_ _International_ _Conference_ _on_ _Machine_ _Learning_
+_(ICML)_, pages 1–9. Morgan Kaufmann.
+J. Baldridge and M. Osborne. 2004. Active learning and
+the total cost of annotation. In _Proceedings of the Con-_
+_ference_ _on_ _Empirical_ _Methods_ _in_ _Natural_ _Language_
+_Processing (EMNLP)_, pages 9–16. ACL Press.
+V.R. Carvalho and W. Cohen. 2004. Learning to extract
+signature and reply lines from email. In _Proceedings_
+_of the Conference on Email and Anti-Spam (CEAS)_ .
+D. Cohn, L. Atlas, and R. Ladner. 1994. Improving generalization with active learning. _Machine_ _Learning_,
+15(2):201–221.
+
+
+1078
+
+
+
+A. Culotta and A. McCallum. 2005. Reducing labeling
+effort for stuctured prediction tasks. In _Proceedings_
+_of_ _the_ _National_ _Conference_ _on_ _Artificial_ _Intelligence_
+_(AAAI)_, pages 746–751. AAAI Press.
+I. Dagan and S. Engelson. 1995. Committee-based
+sampling for training probabilistic classifiers. In _Pro-_
+_ceedings_ _of_ _the_ _International_ _Conference_ _on_ _Machine_
+_Learning (ICML)_, pages 150–157. Morgan Kaufmann.
+R. Hwa. 2004. Sample selection for statistical parsing.
+_Computational Linguistics_, 30(3):73–77.
+J. Kim, T. Ohta, Y. Tsuruoka, Y. Tateisi, and N. Collier. 2004. Introduction to the bio-entity recognition
+task at JNLPBA. In _Proceedings_ _of_ _the_ _International_
+_Joint_ _Workshop_ _on_ _Natural_ _Language_ _Processing_ _in_
+_Biomedicine and its Applications (NLPBA)_, pages 70–
+75.
+S. Kim, Y. Song, K. Kim, J.W. Cha, and G.G. Lee.
+2006. MMR-based active machine learning for bio
+named entity recognition. In _Proceedings_ _of_ _Human_
+_Language_ _Technology_ _and_ _the_ _North_ _American_ _Asso-_
+_ciation_ _for_ _Computational_ _Linguistics_ _(HLT-NAACL)_,
+pages 69–72. ACL Press.
+J. Lafferty, A. McCallum, and F. Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence data. In _Proceedings_
+_of the International Conference on Machine Learning_
+_(ICML)_, pages 282–289. Morgan Kaufmann.
+K. Lari and S. J. Young. 1990. The estimation of stochastic context-free grammars using the inside-outside algorithm. _Computer Speech and Language_, 4:35–56.
+D. Lewis and J. Catlett. 1994. Heterogeneous uncertainty sampling for supervised learning. In _Pro-_
+
+
+_ceedings_ _of_ _the_ _International_ _Conference_ _on_ _Machine_
+_Learning (ICML)_, pages 148–156. Morgan Kaufmann.
+G. Mann and A. McCallum. 2007. Efficient computation
+of entropy gradient for semi-supervised conditional
+random fields. In _Proceedings of the North American_
+_Association_ _for_ _Computational_ _Linguistics_ _(NAACL)_,
+pages 109–112. ACL Press.
+A. McCallum and K. Nigam. 1998. Employing EM
+in pool-based active learning for text classification.
+In _Proceedings_ _of_ _the_ _International_ _Conference_ _on_
+_Machine_ _Learning_ _(ICML)_, pages 359–367. Morgan
+Kaufmann.
+M. Nyffenegger, J.C. Chappelier, and E. Gaussier. 2006.
+Revisiting Fisher kernels for document similarities. In
+_Proceedings of the European Conference on Machine_
+_Learning (ECML)_, pages 727–734. Springer.
+F. Peng and A. McCallum. 2004. Accurate information
+extraction from research papers using conditional random fields. In _Proceedings of Human Language Tech-_
+_nology and the North American Association for Com-_
+_putational Linguistics (HLT-NAACL)_ . ACL Press.
+L. R. Rabiner. 1989. A tutorial on hidden Markov models and selected applications in speech recognition.
+_Proceedings of the IEEE_, 77(2):257–286.
+L.A. Ramshaw and M.P. Marcus. 1995. Text chunking
+using transformation-based learning. In _Proceedings_
+_of the ACL Workshop on Very Large Corpora_ .
+N. Roy and A. McCallum. 2001. Toward optimal active
+learning through sampling estimation of error reduction. In _Proceedings_ _of_ _the_ _International_ _Conference_
+_on_ _Machine_ _Learning_ _(ICML)_, pages 441–448. Morgan Kaufmann.
+E.F.T.K. Sang and F. DeMeulder. 2003. Introduction to
+the CoNLL-2003 shared task: Language-independent
+named entity recognition. In _Proceedings_ _of_ _the_
+_Conference on Natural Language Learning (CoNLL)_,
+pages 142–147.
+T. Scheffer, C. Decomain, and S. Wrobel. 2001. Active hidden Markov models for information extraction.
+In _Proceedings of the International Conference on Ad-_
+_vances_ _in_ _Intelligent_ _Data_ _Analysis_ _(CAIDA)_, pages
+309–318. Springer-Verlag.
+R. Schwartz and Y.-L. Chow. 1990. The _N_ -best algorithm: an efficient and exact procedure for finding the
+_N_ most likely sentence hypotheses. In _Proceedings_
+_of the International Conference on Acoustics, Speech,_
+_and_ _Signal_ _Processing_ _(ICASSP)_, pages 81–83. IEEE
+Press.
+B. Settles, M. Craven, and S. Ray. 2008. Multipleinstance active learning. In _Advances in Neural Infor-_
+_mation Processing Systems (NIPS)_, volume 20, pages
+1289–1296. MIT Press.
+H.S. Seung, M. Opper, and H. Sompolinsky. 1992.
+Query by committee. In _Proceedings_ _of_ _the_ _ACM_
+
+
+1079
+
+
+
+_Workshop_ _on_ _Computational_ _Learning_ _Theory_, pages
+287–294.
+C. E. Shannon. 1948. A mathematical theory of communication. _Bell_ _System_ _Technical_ _Journal_, 27:379–
+423,623–656.
+C. Sutton and A. McCallum. 2006. An introduction to
+conditional random fields for relational learning. In
+L. Getoor and B. Taskar, editors, _Introduction_ _to_ _Sta-_
+_tistical Relational Learning_ . MIT Press.
+A. Vlachos. 2007. Evaluating and combining biomedical
+named entity recognition systems. In _BioNLP_ _2007:_
+_Biological,_ _translational,_ _and_ _clinical_ _language_ _pro-_
+_cessing_, pages 199–206.
+A. Yeh, A. Morgan, M. Colosimo, and L. Hirschman.
+2005. Biocreative task 1a: gene mention finding evaluation. _BMC Bioinformatics_, 6(Suppl 1):S2.
+T. Zhang and F.J. Oles. 2000. A probability analysis
+on the value of unlabeled data for classification problems. In _Proceedings of the International Conference_
+_on Machine Learning (ICML)_, pages 1191–1198. Morgan Kaufmann.
+X. Zhu, J. Lafferty, and Z. Ghahramani. 2003. Combining active learning and semi-supervised learning using
+gaussian fields and harmonic functions. In _Proceed-_
+_ings_ _of_ _the_ _ICML_ _Workshop_ _on_ _the_ _Continuum_ _from_
+_Labeled to Unlabeled Data_, pages 58–65.
+
+

references/2016.acl.li/paper.md

@@ -0,0 +1,1492 @@
+---
+title: "Active Learning for Dependency Parsing with Partial Annotation"
+authors:
+  - "Zhenghua Li"
+  - "Min Zhang"
+  - "Yue Zhang"
+  - "Zhanyi Liu"
+  - "Wenliang Chen"
+  - "Hua Wu"
+  - "Haifeng Wang"
+year: 2016
+venue: "ACL 2016"
+url: "https://aclanthology.org/P16-1033/"
+---
+
+# **Active Learning for Dependency Parsing with Partial Annotation**
+
+**Zhenghua Li** _[†]_ **,** **Min Zhang** _[†∗]_ **,** **Yue Zhang** _[†]_ **,** **Zhanyi Liu** _[‡]_ **,**
+
+
+**Wenliang Chen** _[†]_ **,** **Hua Wu** _[‡]_ **,** **Haifeng Wang** _[‡]_
+
+
+_†_ Soochow University, Suzhou, China
+
+
+_{_ zhli13,minzhang,wlchen _}_ @suda.edu.cn, zhangyue1107@qq.com
+
+
+_‡_ Baidu Inc., Beijing, China
+
+
+_{_ liuzhanyi,wu ~~h~~ ua,wanghaifeng _}_ @baidu.com
+
+
+**Abstract**
+
+
+
+Different from traditional active learning
+based on sentence-wise full annotation
+(FA), this paper proposes active
+learning with dependency-wise partial
+annotation (PA) as a finer-grained unit for
+dependency parsing. At each iteration,
+we select a few most uncertain words
+from an unlabeled data pool, manually
+annotate their syntactic heads, and add the
+partial trees into labeled data for parser
+retraining. Compared with sentence-wise
+FA, dependency-wise PA gives us more
+flexibility in task selection and avoids
+wasting time on annotating trivial tasks
+in a sentence. Our work makes the
+following contributions. First, we are
+the first to apply a probabilistic model to
+active learning for dependency parsing,
+which can 1) provide tree probabilities
+and dependency marginal probabilities
+as principled uncertainty metrics, and
+2) directly learn parameters from PA
+based on a forest-based training objective.
+Second, we propose and compare several
+uncertainty metrics through simulation
+experiments on both Chinese and English.
+Finally, we conduct human annotation
+experiments to compare FA and PA on
+real annotation time and quality.
+
+
+**1** **Introduction**
+
+
+During the past decade, supervised dependency
+parsing has gained extensive progress in boosting
+parsing performance on canonical texts, especially
+on texts from domains or genres similar to existing manually labeled treebanks (Koo and Collins,
+2010; Zhang and Nivre, 2011). However, the
+
+
+_∗_ Correspondence author.
+
+
+
+$0 I1 saw2 Sarah3 with4 a5 telescope6
+
+
+Figure 1: A partially annotated sentence, where
+only the heads of “saw” and “with” are decided.
+
+
+upsurge of web data (e.g., tweets, blogs, and
+product comments) imposes great challenges to
+existing parsing techniques. Meanwhile, previous
+research on out-of-domain dependency parsing
+gains little success (Dredze et al., 2007; Petrov
+and McDonald, 2012). A more feasible way for
+open-domain parsing is to manually annotate a
+certain amount of texts from the target domain or
+genre. Recently, several small-scale treebanks on
+web texts have been built for study and evaluation
+(Foster et al., 2011; Petrov and McDonald, 2012;
+Kong et al., 2014; Wang et al., 2014).
+Meanwhile, active learning (AL) aims to reduce
+annotation effort by choosing and manually annotating unlabeled instances that are most valuable for training statistical models (Olsson, 2009).
+Traditionally, AL utilizes full annotation (FA) for
+parsing (Tang et al., 2002; Hwa, 2004; Lynn et al.,
+2012), where a whole syntactic tree is annotated
+for a given sentence at a time. However, as
+commented by Mejer and Crammer (2012), the
+annotation process is complex, slow, and prone
+to mistakes when FA is required. Particularly,
+annotators waste a lot of effort on labeling trivial
+dependencies which can be well handled by current statistical models (Flannery and Mori, 2015).
+Recently, researchers report promising results
+with AL based on partial annotation (PA) for dependency parsing (Sassano and Kurohashi, 2010;
+Mirroshandel and Nasr, 2011; Majidi and Crane,
+2013; Flannery and Mori, 2015). They find
+
+
+
+344
+
+
+_Proceedings of the 54th Annual Meeting of the Association for Computational Linguistics_, pages 344–354,
+Berlin, Germany, August 7-12, 2016. _⃝_ c 2016 Association for Computational Linguistics
+
+
+that smaller units rather than sentences provide
+more flexibility in choosing potentially informative structures to annotate.
+Beyond previous work, this paper endeavors to
+more thoroughly study this issue, and has made
+substantial progress from the following perspectives.
+
+
+(1) This is the first work that applies a stateof-the-art probabilistic parsing model to AL
+for dependency parsing. The CRF-based
+dependency parser on the one hand allows
+us to use probabilities of trees or marginal
+probabilities of single dependencies for uncertainty measurement, and on the other hand
+can directly learn parameters from partially
+annotated trees. Using probabilistic models
+may be ubiquitous in AL for relatively simpler tasks like classification and sequence labeling, but is definitely novel for dependency
+parsing which is dominated by linear models
+with perceptron-like training.
+
+
+(2) Based on the CRF-based parser, we make
+systematic comparison among several uncertainty metrics for both FA and PA. Simulation
+experiments show that compared with using
+FA, AL with PA can greatly reduce annotation effort in terms of dependency number by
+62 _._ 2% on Chinese and by 74 _._ 2% on English.
+
+
+(3) We build a visualized annotation platform
+and conduct human annotation experiments
+to compare FA and PA on real annotation
+time and quality, where we obtain several
+interesting observations and conclusions.
+
+
+All codes, along with the data from human
+annotation experiments, are released at http:
+//hlt.suda.edu.cn/˜zhli for future research study.
+
+
+**2** **Probabilistic Dependency Parsing**
+
+
+Given an input sentence **x** = _w_ 1 _...wn_, the goal of
+dependency parsing is to build a directed dependency tree **d** = _{h_ ↷ _m_ : 0 _≤_ _h_ _≤_ _n,_ 1 _≤_
+_m_ _≤_ _n}_, where _|_ **d** _|_ = _n_ and _h_ ↷ _m_ represents
+a dependency from a _head_ word _h_ to a _modifier_
+word _m_ . Figure 1 depicts a partial tree containing
+two dependencies. [1]
+
+
+1In this work, we follow many previous works to focus
+on unlabeled dependency parsing (constructing the skeleton
+dependency structure). However, the proposed techniques
+
+
+
+In this work, we for the first time apply a probabilistic CRF-based parsing model to AL for dependency parsing. We adopt the second-order graphbased model of McDonald and Pereira (2006),
+which casts the problem as finding an optimal tree
+from a fully-connect directed graph and factors the
+score of a dependency tree into scores of pairs of
+sibling dependencies.
+
+
+
+where _s_ and _m_ are adjacent siblings both modifying _h_ ; **f** ( **x** _, h, s, m_ ) are the corresponding feature
+vector; **w** is the feature weight vector; _Y_ ( **x** ) is
+the set of all legal trees for **x** according to the
+dependency grammar in hand; **d** _[∗]_ is the 1-best
+parse tree which can be gained efficiently via a
+dynamic programming algorithm (Eisner, 2000).
+We use the state-of-the-art feature set listed in
+Bohnet (2010).
+Under the log-linear CRF-based model, the
+probability of a dependency tree is:
+
+
+_e_ _[Score]_ [(] **[x]** _[,]_ **[d]** [;] **[w]** [)]
+_p_ ( **d** _|_ **x** ; **w** ) = ~~∑~~ (2)
+**d** _[′]_ _∈Y_ ( **x** ) _[e][Score]_ [(] **[x]** _[,]_ **[d]** _[′]_ [;] **[w]** [)]
+
+
+Ma and Zhao (2015) give a very detailed and
+thorough introduction to CRFs for dependency
+parsing.
+
+
+**2.1** **Learning from FA**
+
+
+Under the supervised learning scenario, a labeled
+training data _D_ = _{_ ( **x** _i,_ **d** _i_ ) _}_ _[N]_ _i_ =1 [is] [provided] [to]
+learn **w** . The objective is to maximize the log
+likelihood of _D_ :
+
+
+∑ _N_
+_L_ ( _D_ ; **w** ) = (3)
+
+_i_ =1 [log] _[ p]_ [(] **[d]** _[i][|]_ **[x]** _[i]_ [;] **[ w]** [)]
+
+
+which can be solved by standard gradient descent
+algorithms. In this work, we adopt stochastic gradient descent (SGD) with L2-norm regularization
+for all CRF-based parsing models. [2]
+
+
+explored in this paper can be easily extended to the case of
+labeled dependency parsing.
+2We borrow the implementation of SGD in
+CRFsuite (http://www.chokkan.org/software/
+crfsuite/), and use 100 sentences for a batch.
+
+
+
+**d** _[∗]_ = arg max **d** _∈Y_ ( **x** ) _Score_ ( **x** _,_ **d** ; **w** )
+
+
+
+∑
+_Score_ ( **x** _,_ **d** ; **w** ) =
+
+
+
+( _h,s,m_ ): _h_ ↷ _s∈_ **d** _,_
+_h_ ↷ _m∈_ **d**
+
+
+
+**w** _·_ **f** ( **x** _, h, s, m_ )
+
+
+(1)
+
+
+
+345
+
+
+**2.2** **Marginal Probability of Dependencies**
+
+
+Marcheggiani and Arti`eres (2014) shows that
+marginal probabilities of local labels can be
+used as an effective uncertain metric for AL
+for sequence labeling problems. In the case of
+dependency parsing, the marginal probability of a
+dependency is the sum of probabilities of all legal
+trees that contain the dependency.
+
+∑
+_p_ ( _h_ ↷ _m|_ **x** ; **w** ) = _p_ ( **d** _|_ **x** ; **w** ) (4)
+
+
+**d** _∈Y_ ( **x** ): _h_ ↷ _m∈_ **d**
+
+
+Intuitively, marginal probability is a more principled metric for measuring reliability of a dependency since it considers all legal parses in the
+search space, compared to previous methods based
+on scores of local classifiers (Sassano and Kurohashi, 2010; Flannery and Mori, 2015) or votes
+of n-best parses (Mirroshandel and Nasr, 2011).
+Moreover, Li et al. (2014) find strong correlation
+between marginal probability and correctness of a
+dependency in cross-lingual syntax projection.
+
+
+**3** **Active Learning for Dependency**
+**Parsing**
+
+
+This work adopts the standard pool-based AL
+framework (Lewis and Gale, 1994; McCallum and
+Nigam, 1998). Initially, we have a small set of
+labeled seed data _L_, and a large-scale unlabeled
+data pool _U_ . Then the procedure works as follows.
+
+(1) Train a new parser on the current _L_ .
+
+
+(2) Parse all sentences in _U_, and select a set of
+the most informative tasks _U_ _[′]_
+
+
+(3) Manually annotate: _U_ _[′]_ _→L_ _[′]_
+
+
+(4) Expand labeled data: _L ∪L_ _[′]_ _→L_
+
+
+The above steps loop for many iterations until a
+predefined stopping criterion is met.
+The key challenge for AL is how to measure the
+informativeness of structures in concern. Following previous work on AL for dependency parsing,
+we make a simplifying assumption that if the
+current model is most uncertain about an output
+(sub)structure, the structure is most informative in
+terms of boosting model performance.
+
+
+**3.1** **Sentence-wise FA**
+
+
+Sentence-wise FA selects _K_ most uncertain sentences in Step (2), and annotates their whole tree
+structures in Step (3). In the following, we describe several uncertainty metrics and investigate
+
+
+
+their practical effects through experiments. Given
+an unlabeled sentence **x** = _w_ 1 _...wn_, we use **d** _[∗]_
+
+to denote the 1-best parse tree produced by the
+current model as in Eq. (1). For brevity, we omit
+the feature weight vector **w** in the equations.
+**Normalized** **tree** **score.** Following previous
+works that use scores of local classifiers
+for uncertainty measurement (Sassano and
+Kurohashi, 2010; Flannery and Mori, 2015), we
+use _Score_ ( **x** _,_ **d** _[∗]_ ) to measure the uncertainty of **x**,
+assuming that the model is more uncertain about
+**x** if **d** _[∗]_ gets a smaller score. However, we find that
+directly using _Score_ ( **x** _,_ **d** _[∗]_ ) always selects very
+short sentences due to the definition in Eq. (1).
+Thus we normalize the score with the sentence
+length _n_ as follows. [3]
+
+
+_Confi_ ( **x** ) = _[Score]_ [(] **[x]** _[,]_ **[ d]** _[∗]_ [)] (5)
+
+_n_ [1] _[.]_ [5]
+
+**Normalized** **tree** **probability.** The CRF-based
+parser allows us, for the first time in AL for dependency parsing, to directly use tree probabilities
+for uncertainty measurement. Unlike previous
+approximate methods based on k-best parses (Mirroshandel and Nasr, 2011), tree probabilities globally consider all parse trees in the search space,
+and thus are intuitively more consistent and proper
+for measuring the reliability of a tree. Our initial
+assumption is that the model is more uncertain
+about **x** if **d** _[∗]_ gets a smaller probability. However,
+we find that directly using _p_ ( **d** _[∗]_ _|_ **x** ) would select
+very long sentences because the solution space
+grows exponentially with sentence length. We find
+that the normalization strategy below works well. [4]
+
+~~[√]~~ _n_
+_Confi_ ( **x** ) = _p_ ( **d** _[∗]_ _|_ **x** ) (6)
+
+
+**Averaged marginal probability** . As discussed
+in Section 2.2, the marginal probability of a dependency directly reflects its reliability, and thus
+can be regarded as another global measurement
+besides tree probabilities.In fact, we find that the
+effect of sentence length is naturally handled with
+the following metric. [5]
+
+
+
+∑
+_h_ ↷ _m∈_ **d** _[∗]_ _[p]_ [(] _[h]_ [ ↷] _[m][|]_ **[x]** [)]
+_Confi_ ( **x** ) =
+
+
+
+(7)
+_n_
+
+
+
+3We have also tried replacing _n_ 1 _._ 5 with _n_ (still prefer
+short sentences) and _n_ [2] (bias to long sentences).
+4We have also tried _p_ ( **d** _∗|_ **x** ) _×_ _f_ ( _n_ ), where _f_ ( _n_ ) = log _n_
+or _f_ ( _n_ ) = ~~_[√]_~~ _n_, but both work badly.
+
+5 [√] _n_ ~~[∏]~~
+We have also tried _h_ ↷ _m∈_ **d** _[∗]_ _[p]_ [(] _[h]_ [ ↷] _[m][|]_ **[x]** [)][,] [leading]
+
+to slightly inferior results.
+
+
+
+5 [√] _n_ ~~[∏]~~
+We have also tried
+
+
+
+346
+
+
+**3.2** **Single Dependency-wise PA**
+
+
+AL with single dependency-wise PA selects _M_
+most uncertain words from _U_ in Step (2), and annotates the heads of the selected words in Step (3).
+After annotation, the newly annotated sentences
+with partial trees _L_ _[′]_ are added into _L_ . Different
+from the case of sentence-wise FA, _L_ _[′]_ are also put
+back to _U_, so that new tasks can be further chosen
+from them.
+Marcheggiani and Arti`eres (2014) make systematic comparison among a dozen uncertainty
+metrics for AL with PA for several sequence
+labeling tasks. We borrow three effective metrics
+according to their results.
+**Marginal** **probability** **max.** Suppose _h_ [0] =
+arg max _h p_ ( _h_ ↷ _i|_ **x** ) is the most likely head for
+_i_ . The intuition is that the lower _p_ ( _h_ [0] ↷ _i_ ) is, the
+more uncertain the model is on deciding the head
+of the token _i_ .
+
+
+_Confi_ ( **x** _, i_ ) = _p_ ( _h_ [0] ↷ _i|_ **x** ) (8)
+
+
+**Marginal** **probability** **gap.** Suppose _h_ [1] =
+arg max _h_ = _h_ 0 _p_ ( _h_ ↷ _i|_ **x** ) is the second most likely
+head for _i_ . The intuition is that the smaller the
+probability gap is, the more uncertain the model is
+about _i_ .
+
+
+_Confi_ ( **x** _, i_ ) = _p_ ( _h_ [0] ↷ _i|_ **x** ) _−_ _p_ ( _h_ [1] ↷ _i|_ **x** ) (9)
+
+
+**Marginal** **probability** **entropy.** This metric
+considers the entropy of all possible heads for _i_ .
+The assumption is that the smaller the _negative_
+entropy is, the more uncertain the model is about
+_i_ .
+
+
+
+∑
+_Confi_ ( **x** _, i_ ) =
+
+
+
+_p_ ( _h_ ↷ _i|_ **x** ) log _p_ ( _h_ ↷ _i|_ **x** )
+
+_h_
+
+
+
+(10)
+
+
+
+$0 I1 saw2 Sarah3 with4 a5 telescope6
+
+
+Figure 2: An example parse forest converted from
+the partial tree in Figure 1.
+
+
+and annotating one dependency can certainly help
+decide another dependency in practice.
+Inspired by the work of Flannery and Mori
+(2015), we propose AL with batch dependencywise PA, which is a compromise between
+sentence-wise FA and single dependency-wise
+PA. In Step 2, AL with batch dependency-wise
+PA selects _K_ most uncertain sentences from _U_,
+and also determines _r_ % most uncertain words
+from each sentence at the same time. In Step
+3, annotators are asked to label the heads of
+the selected words in the selected sentences.
+We propose and experiment with the following
+three strategies based on experimental results of
+sentence-wise FA and single dependency-wise
+PA.
+**Averaged** **marginal** **probability** **&** **gap** .
+First, select _K_ sentences from _U_ using averaged
+marginal probability. Second, select _r_ % words
+using marginal probability gap for each selected
+sentence.
+**Marginal** **probability** **gap** . First, for each
+sentence in _U_, select _r_ % most uncertain words
+according to marginal probability gap. Second,
+select _K_ sentences from _U_ using the averaged
+marginal probability gap of the selected _r_ % words
+in a sentence as the uncertainty metric.
+**Averaged marginal probability** . This strategy
+is the same with the above strategy, except it
+measures the uncertainty of a word _i_ according
+to the marginal probability of the dependency
+pointing to _i_ in **d** _[∗]_, i.e., _p_ ( _j_ ↷ _i|_ **x** ), where _j_ ↷
+_i ∈_ **d** _[∗]_ .
+
+
+**3.4** **Learning from PA**
+
+
+A major challenge for AL with PA is how to learn
+from partially labeled sentences, as depicted in
+Figure 1. Li et al. (2014) show that a probabilistic
+CRF-based parser can naturally and effectively
+learn from PA. The basic idea is converting a
+partial tree into a forest as shown in Figure 2,
+
+
+
+**3.3** **Batch Dependency-wise PA**
+
+
+In the framework of single dependency-wise PA,
+we assume that the selection and annotation of
+dependencies in the same sentence are strictly
+independent. In other words, annotators may be
+asked to annotate the head of one selected word after reading and understanding a whole (sometimes
+partial) sentence, and may be asked to annotate
+another selected word in the same sentence in next
+AL iteration. Obviously, frequently switching
+sentences incurs great waste of cognitive effort,
+
+
+
+347
+
+
+and using the forest as the gold-standard reference
+during training, also known as _ambiguous labeling_
+(Riezler et al., 2002; T¨ackstr¨om et al., 2013).
+For each remaining word without head, we
+add all dependencies linking to it as long as the
+new dependency does not violate the existing
+dependencies. We denote the resulting forest as
+_F_ j, whose probability is naturally the sum of
+probabilities of each tree **d** in _F_ .
+
+
+|Col1|Col2|Train|Dev|Test|
+|---|---|---|---|---|
+|Chinese|#Sentences<br>#Tokens|14,304<br>318,408|803<br>20,454|1,910<br>50,319|
+|English|#Sentences<br>#Tokens|39,115<br>908,154|1,700<br>40,117|2,416<br>56,684|
+
+
+
+∑
+_p_ ( _F|_ **x** ; **w** ) =
+
+
+
+_p_ ( **d** _|_ **x** ; **w** )
+
+**d** _∈F_
+
+
+
+(11)
+
+
+
+∑
+**d** _∈F_ _[e][Score]_ [(] **[x]** _[,]_ **[d]** [;] **[w]** [)]
+= ~~∑~~
+**d** _[′]_ _∈Y_ ( **x** ) _[e][Score]_ [(] **[x]** _[,]_ **[d]** _[′]_ [;] **[w]** [)]
+
+
+
+Table 1: Data statistics.
+
+
+are selected and annotated at each iteration. In
+the case of single dependency-wise PA, we select
+and annotate _M_ = 10 _,_ 000 dependencies, which
+roughly correspond to 500 sentences considering
+that the averaged sentence length is about 22 _._ 3 in
+CTB-train and 23 _._ 2 in PTB-train. In the case of
+batch dependency-wise PA, we set _K_ = 500, and
+_r_ = 20% for Chinese and _r_ = 10% for English,
+considering that the parser trained on all data
+achieves about 80% and 90% accuracies.
+We measure parsing performance using the
+standard unlabeled attachment score (UAS)
+_including punctuation marks_ . Please note that we
+always treat punctuation marks as ordinary words
+when selecting annotation tasks and calculating
+UAS, in order to make fair comparison between
+FA and PA. [7]
+
+
+**4.1** **FA vs.** **Single Dependency-wise PA**
+
+
+First, we make comparison on the performance of
+AL with FA and with single dependency-wise PA.
+**Results** **on** **Chinese** are shown in Figure 3.
+Following previous work, we use the number of
+annotated dependencies (x-axis) as the annotation
+cost in order to fairly compare FA and PA. We use
+FA with random selection as a baseline. We also
+draw the accuracy of the CRF-based parser trained
+on all training data, which can be regarded as the
+upper bound.
+For FA, the curve of the normalized tree score
+intertwines with that of random selection. Meanwhile, the performance of normalized tree probability is very close to that of averaged marginal
+probability, and both are clearly superior to the
+baseline with random selection.
+For PA, the difference among the three uncertainty metrics is small. The marginal probability
+gap clearly outperforms the other two metrics before 50 _,_ 000 annotated dependencies, and remains
+
+
+7Alternatively, we can exclude punctuation marks for task
+selection in AL with PA. Then, to be fair, we have to discard
+all dependencies pointing to punctuation marks in the case of
+FA. This makes the experiment setting more complicated.
+
+
+348
+
+
+
+Suppose the partially labeled training data is
+_D_ = _{_ ( **x** _i, Fi_ ) _}_ _[N]_ _i_ =1 [.] [Then its log likelihood is:]
+
+∑ _N_
+_L_ ( _D_ ; **w** ) = (12)
+
+_i_ =1 [log] _[ p]_ [(] _[F][i][|]_ **[x]** _[i]_ [;] **[ w]** [)]
+
+
+T¨ackstr¨om et al. (2013) show that the partial
+derivative of the _L_ ( _D_ ; **w** ) with regard to **w** (a.k.a
+the gradient) in both Equation (3) and (12) can be
+efficiently solved with the classic Inside-Outside
+algorithm. [6]
+
+
+**4** **Simulation Experiments**
+
+
+We use Chinese Penn Treebank 5.1 (CTB) for
+Chinese and Penn Treebank (PTB) for English.
+For both datasets, we follow the standard data
+split, and convert original bracketed structures into
+dependency structures using Penn2Malt with its
+default head-finding rules. To be more realistic, we use automatic part-of-speech (POS) tags
+produced by a state-of-the-art CRF-based tagger
+(94 _._ 1% on CTB-test, and 97 _._ 2% on PTB-test, nfold jackknifing on training data), since POS tags
+encode much syntactic annotation. Because AL
+experiments need to train many parsing models,
+we throw out all training sentences longer than 50
+to speed up our experiments. Table 1 shows the
+data statistics.
+Following previous practice on AL with PA
+(Sassano and Kurohashi, 2010; Flannery and
+Mori, 2015), we adopt the following AL settings
+for both Chinese and English . The first 500
+training sentences are used as the seed labeled
+data _L_ . In the case of FA, _K_ = 500 new sentences
+
+
+6This work focuses on projective dependency parsing.
+Please refer to Koo et al. (2007), McDonald and Satta (2007),
+and Smith and Smith (2007) for building a probabilistic nonprojective parser.
+
+
+80
+
+
+79
+
+
+78
+
+
+77
+
+
+76
+
+
+75
+
+
+74
+
+
+73
+
+
+72
+
+
+92
+
+
+91
+
+
+90
+
+
+89
+
+
+88
+
+
+87
+
+
+86
+
+
+85
+
+
+84
+
+
+
+
+
+
+
+
+
+80
+
+
+79
+
+
+78
+
+
+77
+
+
+76
+
+
+75
+
+
+74
+
+
+73
+
+
+72
+
+
+
+0 50000 100000 150000 200000 250000 300000
+
+
+Number of Annotated Dependencies
+
+
+Figure 3: FA vs. PA on CTB-dev.
+
+
+
+10000 20000 30000 40000 50000 60000 70000
+
+
+Number of Annotated Dependencies
+
+
+Figure 5: Single vs. batch dependency-wise PA on
+CTB-dev.
+
+
+93
+
+
+92
+
+
+91
+
+
+90
+
+
+
+
+
+
+
+0 50000 100000 150000 200000 250000 300000
+
+
+Number of Annotated Dependencies
+
+
+Figure 4: FA vs. PA on PTB-dev.
+
+
+very competitive at all other points. The marginal
+probability max achieves best peak UAS, and even
+outperforms the parser trained on all data, which
+can be explained by small disturbance during
+complex model training. The marginal probability
+entropy, although being the most complex metric
+among the three, seems inferior all the time.
+It is clear that using PA can greatly reduce
+annotation effort compared with using FA in terms
+of annotated dependencies.
+**Results on English** are shown in Figure 4. The
+overall findings are similar to those in Figure 3, except that the distinction among different methods
+is more clear. For FA, normalized tree score
+is consistently better than the random baseline.
+Normalized tree probability always outperforms
+normalized tree score. Averaged marginal probability performs best, except being slightly inferior
+to normalized tree probability in earlier stages.
+For PA, it is consistent that marginal probability
+gap is better than marginal probability max, and
+marginal probability entropy is the worst.
+In summary, based on the results on the de
+
+
+10000 20000 30000 40000 50000 60000 70000
+
+
+Number of Annotated Dependencies
+
+
+Figure 6: Single vs. batch dependency-wise PA on
+PTB-dev.
+
+
+velopment data in Figure 3 and 4, the best AL
+method with PA only needs about 80 _,_ 000 [25%]
+318 _,_ 408 [=]
+annotated dependencies on Chinese, and about
+90 _,_ 000 [10%][ on English, to reach the same per-]
+908 _,_ 154 [=]
+formance with parsers trained on all data. Moreover, the PA methods converges much faster than
+the FA ones, since for the same x-axis number,
+much more sentences (with partial trees) are used
+as training data for AL with PA than FA.
+
+
+**4.2** **Single vs.** **Batch Dependency-wise PA**
+
+
+Then we make comparison on AL with single
+dependency-wise PA and with the more practical
+batch dependency-wise PA.
+**Results on Chinese** are shown in Figure 5. We
+can see that the three strategies achieve very similar performance and are also very close to single
+dependency-wise PA. AL with batch dependencywise PA even achieves higher accuracy before
+20 _,_ 000 annotated dependencies, which should be
+caused by the smaller active learning steps (about
+
+
+349
+
+
+
+89
+
+
+88
+
+
+87
+
+
+86
+
+
+85
+
+
+84
+
+
+
+
+2 _,_ 000 dependencies at each iteration, contrasting
+10 _,_ 000 for single dependency-wise PA). When the
+training data runs out at about 7 _,_ 300 dependencies, AL with batch dependency-wise PA only lags
+behind with single dependency-wise PA by about
+0 _._ 3%, which we suppose can be reduced if larger
+training data is available.
+**Results** **on** **English** are shown in Figure 6,
+and are very similar to those on Chinese. One
+tiny difference is that the marginal probability
+gap is slightly worse that the other two metrics.
+The three uncertainty metrics have very similar
+accuracy curves, which are also very close to the
+curve of single dependency-wise PA. In addition,
+we also try _r_ = 20% and find that results are
+inferior to _r_ = 10%, indicating that the extra 10%
+annotation tasks are less valuable and contributive.
+
+
+**4.3** **Main Results on Test Data**
+
+
+Table 2 shows the results on test data. We compare
+our CRF-based parser with ZPar v6.0 [8], a state-ofthe-art transition-based dependency parser (Zhang
+and Nivre, 2011). We train ZPar with default
+parameter settings for 50 iterations, and choose
+the model that performs best on dev data. We
+can see that when trained on all data, our CRFbased parser outperforms ZPar on both Chinese
+and English.
+To compare FA and PA, we report the number
+of annotated dependencies needed under each AL
+strategy to achieve an accuracy lower by about 1%
+than the parser trained on all data. [9]
+
+FA (best) refers to FA with averaged marginal
+probability, and it needs [187] _[,]_ [123] _[−]_ [149] _[,]_ [051] = 20 _._ 3%
+
+
+|Col1|Chinese|Col3|English|Col5|
+|---|---|---|---|---|
+||#Dep labeled|UAS|#Dep labeled|UAS|
+|ZPar<br>This parser|318,408<br>318,408|77.97<br>78.36|908,154<br>908,154|91.45<br>91.66|
+|FA (random)<br>FA (best)|187,123<br>149,051|77.43<br>77.32|395,199<br>197,907|90.67<br>90.66|
+|PA (single)<br>PA (batch)|50,958<br>56,389|77.22<br>77.38|61,448<br>51,016|90.72<br>90.70|
+
+
+
+probability, and it needs = 20 _._ 3%
+
+187 _,_ 123
+less annotated dependencies than FA with random selection on Chinese, and [395] _[,]_ [199] _[−]_ [197] _[,]_ [907] =
+
+
+
+dom selection on Chinese, and =
+
+395 _,_ 199
+50 _._ 0% less on English.
+PA (single) with marginal probability gap
+needs 149 _,_ 051 _−_ 50 _,_ 958 = 65 _._ 8% less annotated
+
+
+
+needs = 65 _._ 8% less annotated
+
+149 _,_ 051
+dependencies than FA (best) on Chinese, and
+197 _,_ 907 _−_ 61 _,_ 448 = 69 _._ 0% less on English.
+
+197 _,_ 907
+PA (batch) with marginal probability gap needs
+slightly more annotation than PA (single) on Chinese but slightly less annotation on English, and
+can reduce the amount of annotated dependencies
+by [149] _[,]_ [051] _[−]_ [56] _[,]_ [389] = 62 _._ 2% over FA (best) on Chi
+
+
+= 62 _._ 2% over FA (best) on Chi149 _,_ 051
+
+
+
+Table 2: Results on test data.
+
+
+nese and by [197] _[,]_ [907] _[−]_ [51] _[,]_ [016] = 74 _._ 2% on English.
+
+197 _,_ 907
+
+
+**5** **Human Annotation Experiments**
+
+
+So far, we measure annotation effort in terms
+of the number of annotated dependencies and
+assume that it takes the same amount of time
+to annotate different words, which is obviously
+unrealistic. To understand whether active learning
+based on PA can really reduce annotation time
+over based on FA in practice, we build a web
+browser based annotation system, [10] and conduct
+human annotation experiments on Chinese.
+In this part, we use CTB 7 _._ 0 which is a newer
+and larger version and covers more genres, and
+adopt the newly proposed Stanford dependencies
+(de Marneffe and Manning, 2008; Chang et al.,
+2009) which are more understandable for annotators. [11] Since manual syntactic annotation is
+very difficult and time-consuming, we only keep
+sentences with length [10 _,_ 20] in order to better
+measure annotation time by focusing on sentences
+of reasonable length, which leave us 12 _,_ 912 training sentences under the official data split. Then,
+we use a random half of training sentences to
+train a CRF-based parser, and select 20% most
+uncertain words with marginal probability gap for
+each sentence of the left half.
+We employ 6 postgraduate students as our annotators who are at different levels of familiarity
+in syntactic annotation. Before annotation, the
+annotators are trained for about two hours by
+introducing the basic concepts, guidelines, and illustrating examples. Then, they are asked to practice on the annotation system for about another
+two hours. Finally, all annotators are required to
+
+
+10http://hlt-service.suda.edu.cn/
+syn-dep-batch. Please try.
+11We use Stanford Parser 3 _._ 4 (2014-06-16) for constituentto-dependency structure conversion.
+
+
+
+8
+http://people.sutd.edu.sg/˜yue_zhang/doc/
+9The gap 1% is chosen based on the curves on
+development data (Figure 3 and 4) with the following two
+considerations: 1) larger gap may lead to wrong impression
+that AL is weak; 2) smaller gap (e.g., 0 _._ 5%) cannot be
+reached for the worst AL method (FA: random).
+
+
+
+350
+
+
+|Col1|Time: Sec/Dep|Col3|Annotation accuracy|Col5|
+|---|---|---|---|---|
+||FA|PA|FA (on 20%)|PA (diff)|
+|Annotator #1<br>Annotator #2<br>Annotator #3 <br>Annotator #4<br>Annotator #5<br>Annotator #6|**4.0**<br>7.5<br> _10.0_<br>5.1<br>7.0<br>7.0|**7.9 **<br>16.0 <br>_22.2_ <br>8.7 <br>17.3 <br>10.6|**84.65** (**73.41**) <br> 78.90 (72.22) <br> 69.75 (59.77) <br> 66.75 (49.19) <br> 65.47 (48.50) <br> _58.05_ (_43.28_)|**75.28** (+1.87)<br> 62.18 (_-10.04_)<br> 56.91 (-2.86)<br> 61.77 (+**12.58**)<br> 48.39 (-0.11)<br> _48.37_ (+5.09)|
+|Overall|6.7|13.6|70.36 (57.28)|59.06 (+1.78)|
+
+
+Table 3: Statistics of human annotation.
+
+
+formally annotate the same 100 sentences. The
+system is programed that each sentence has 3
+FA submissions and 3 PA submissions. During
+formal annotation, the annotators are not allowed
+to discuss with each other or look up any guideline or documents, which may incur unnecessary
+inaccuracy in timing. Instead, the annotators
+can only decide the syntactic structures based on
+the basic knowledge of dependency grammar and
+one’s understanding of the sentence structure. The
+annotation process lasts for about 5 hours. On
+average, each annotator completes 50 sentences
+with FA (763 dependencies) and 50 sentences with
+PA (178 dependencies).
+Table 3 lists the results in descending order of
+an annotator’s experience in syntactic annotation.
+The first two columns compare the time needed for
+annotating a dependency in seconds. On average,
+_annotating_ _a_ _dependency_ _in_ _PA_ _takes about_ _twice_
+_as_ _much_ _time_ _as_ _in_ _FA_, which is reasonable considering the words to be annotated in PA may be
+more difficult for annotators while the annotation
+of some tasks in FA may be very trivial and easy.
+Combined with the results in Table 2, we may infer
+that to achieve 77 _._ 3% accuracy on CTB-test, AL
+with FA requires 149 _,_ 051 _×_ 6 _._ 7 = 998 _,_ 641 _._ 7
+seconds of annotation, whereas AL with batch
+dependency-wise PA needs 56 _,_ 389 _×_ 13 _._ 6 =
+766 _,_ 890 _._ 4 seconds. Thus, we may roughly say
+that _AL_ _with_ _PA_ _can_ _reduce_ _annotation_ _time_ _over_
+_FA by_ [998] _[,]_ [641] _[.]_ [7] _[−]_ [766] _[,]_ [890] _[.]_ [4] = 23 _._ 2% _._
+
+998 _,_ 641 _._ 7
+We also report annotation accuracy according
+to the gold-standard Stanford dependencies converted from bracketed structures. [12] Overall, the
+accuracy of FA is 70 _._ 36 _−_ 59 _._ 06 = 11 _._ 30% higher
+
+
+12An anonymous reviewer commented that the direct
+comparison between an annotator’s performance on PA and
+FA based on accuracy may be misleading since the FA and
+PA sentences for one annotator are mutually exclusive.
+
+
+351
+
+
+
+than that of PA, which should be due to the trivial
+tasks in FA. To be more fair, we compare the
+accuracies of FA and PA on the same 20% selected
+difficult words, and find that annotators exhibit
+different responses to the switch. Annotator #4
+achieve 12 _._ 58% higher accuracy when under PA
+than under FA. The reason may be that under PA,
+annotators can be more focused and therefore perform better on the few selected tasks. In contrast,
+some annotators may perform better under FA.
+For example, annotation accuracy of annotator #2
+increases by 10 _._ 04% when switching from PA to
+FA, which may be due to that FA allows annotators
+to spend more time on the same sentence and gain
+help from annotating easier tasks. Overall, we find
+that the accuracy of PA is 59 _._ 06 _−_ 57 _._ 28 = 1 _._ 78%
+higher than that of FA, indicating that _PA actually_
+_can improve annotation quality_ .
+
+
+**6** **Related Work**
+
+
+Recently, AL with PA attracts much attention in
+sentence-wise natural language processing such
+as sequence labeling and parsing. For sequence
+labeling, Marcheggiani and Arti`eres (2014) systematically compare a dozen uncertainty metrics
+in token-wise AL with PA (without comparison
+with FA), whereas Settles and Craven (2008) investigate different uncertainty metrics in AL with
+FA. Li et al. (2012) propose to only annotate the
+most uncertain word boundaries in a sentence for
+Chinese word segmentation and show promising
+results on both simulation and human annotation
+experiments. All above works are based on CRFs
+and make extensive use of sequence probabilities
+and token marginal probability.
+In parsing community, Sassano and Kurohashi
+(2010) select bunsetsu (similar to phrases) pairs
+with smallest scores from a local classifier, and
+let annotators decide whether the pair composes
+a dependency. They convert partially annotated
+instances into local dependency/non-dependency
+classification instances to help a simple shiftreduce parser. Mirroshandel and Nasr (2011)
+select most uncertain words based on votes of nbest parsers, and convert partial trees into full trees
+by letting a baseline parser perform constrained
+decoding in order to preserve partial annotation.
+Under a different query-by-committee AL framework, Majidi and Crane (2013) select most uncertain words using a committee of diverse parsers,
+and convert partial trees into full trees by letting
+
+
+the parsers of committee to decide the heads of
+remaining tokens. Based on a first-order (pointwise) Japanese parser, Flannery and Mori (2015)
+use scores of a local classifier for task selection,
+and treat PA as dependency/non-dependency instances (Flannery et al., 2011). Different from
+above works, this work adopts a state-of-the-art
+probabilistic dependency parser, uses more principled tree probabilities and dependency marginal
+probabilities for uncertainty measurement, and
+learns from PA based on a forest-based training
+objective which is more theoretically sound.
+Most previous works on AL with PA only conduct simulation experiments. Flannery and Mori
+(2015) perform human annotation to measure true
+annotation time. A single annotator is employed
+to annotate for two hours alternating FA and PA
+(33% batch) every fifteen minutes. Beyond their
+initial expectation, they find that the annotation
+time per dependency is nearly the same for FA and
+PA (different from our findings) and gives a few
+interesting explanations.
+Under a non-AL framework, Mejer and Crammer (2012) propose an interesting light feedback
+scheme for dependency parsing by letting annotators decide the better one from top-2 parse trees
+produced by the current parsing model.
+Hwa (1999) pioneers the idea of using PA
+to reduce manual labeling effort for constituent
+grammar induction. She uses a variant InsideOutside re-estimation algorithm (Pereira and Schabes, 1992) to induce a grammar from PA. Clark
+and Curran (2006) propose to train a Combinatorial Categorial Grammar parser using partially
+labeled data only containing predicate-argument
+dependencies. Tsuboi et al. (2008) extend CRFbased sequence labeling models to learn from
+incomplete annotations, which is the same with
+Marcheggiani and Arti`eres (2014). Li et al. (2014)
+propose a CRF-based dependency parser that can
+learn from partial tree projected from sourcelanguage structures in the cross-lingual parsing
+scenario. Mielens et al. (2015) propose to impute
+missing dependencies based on Gibbs sampling in
+order to enable traditional parsers to learn from
+partial trees.
+
+
+**7** **Conclusions**
+
+
+This paper for the first time applies a state-ofthe-art probabilistic model to AL with PA for
+dependency parsing. It is shown that the CRF
+
+
+based parser can on the one hand provide tree
+probabilities and dependency marginal probabilities as principled uncertainty metrics and on the
+other hand elegantly learn from partially annotated
+data. We have proposed and compared several uncertainty metrics through simulation experiments,
+and show that AL with PA can greatly reduce
+the amount of annotated dependencies by 62 _._ 2%
+on Chinese 74 _._ 2% on English. Finally, we conduct human annotation experiments on Chinese to
+compare PA and FA on real annotation time and
+quality. We find that annotating a dependency in
+PA takes about 2 times long as in FA. This suggests that AL with PA can reduce annotation time
+by 23 _._ 2% over with FA on Chinese. Moreover,
+the results also indicate that annotators tend to
+perform better under PA than FA.
+For future work, we would like to advance this
+study in the following directions. The first idea is
+to combine uncertainty and representativeness for
+measuring informativeness of annotation targets in
+concern. Intuitively, it would be more profitable
+to annotate instances that are both difficult for
+the current model and representative in capturing
+common language phenomena. Second, we so far
+assume that the selected tasks are equally difficult
+and take the same amount of effort for human
+annotators. However, it is more reasonable that
+human are good at resolving some ambiguities but
+bad at others. Our plan is to study which syntactic
+structures are more suitable for human annotation,
+and balance informativeness of a candidate task
+and its suitability for human annotation. Finally,
+one anonymous reviewer comments that we may
+use automatically projected trees (Rasooli and
+Collins, 2015; Guo et al., 2015; Ma and Xia, 2014)
+as the initial seed labeled data, which is cheap and
+interesting.
+
+
+**Acknowledgments**
+
+
+The authors would like to thank the anonymous
+reviewers for the helpful comments. We also thank
+Junhui Li and Chunyu Kit for reading our paper
+and giving many good suggestions. Particularly,
+Zhenghua is very grateful to many of his students:
+Fangli Lu, Qiuyi Yan, and Yue Zhang build the annotation system; Jiayuan Chao, Wei Chen, Ziwei
+Fan, Die Hu, Qingrong Xia, and Yue Zhang participate in data annotation. This work was supported
+by National Natural Science Foundation of China
+(Grant No. 61502325, 61525205, 61572338).
+
+
+
+352
+
+
+**References**
+
+
+Bernd Bohnet. 2010. Top accuracy and fast
+dependency parsing is not a contradiction. In
+_Proceedings of COLING_, pages 89–97.
+
+
+Pi-Chuan Chang, Huihsin Tseng, Dan Jurafsky, and
+Christopher D. Manning. 2009. Discriminative
+reordering with Chinese grammatical relations
+features. In _Proceedings_ _of_ _the_ _Third_ _Workshop_
+_on_ _Syntax_ _and_ _Structure_ _in_ _Statistical_ _Translation_
+_(SSST-3) at NAACL HLT 2009_, pages 51–59.
+
+
+Stephen Clark and James Curran. 2006. Partial
+training for a lexicalized-grammar parser. In
+_Proceedings_ _of_ _the_ _Human_ _Language_ _Technology_
+_Conference of the NAACL_, pages 144–151.
+
+
+Marie-Catherine de Marneffe and Christopher D.
+Manning. 2008. The Stanford typed dependencies
+representation. In _Coling 2008:_ _Proceedings of the_
+_workshop_ _on_ _Cross-Framework_ _and_ _Cross-Domain_
+_Parser Evaluation_, pages 1–8.
+
+
+Mark Dredze, John Blitzer, Partha Pratim Talukdar,
+Kuzman Ganchev, Jo˜ao Graca, and Fernando
+Pereira. 2007. Frustratingly hard domain adaptation
+for dependency parsing. In _Proceedings_ _of_ _the_
+_CoNLL_ _Shared_ _Task_ _Session_ _of_ _EMNLP-CoNLL_
+_2007_ .
+
+
+Jason Eisner. 2000. Bilexical grammars and their
+cubic-time parsing algorithms. In _Advances_ _in_
+_Probabilistic and Other Parsing Technologies_, pages
+29–62.
+
+
+Daniel Flannery and Shinsuke Mori. 2015. Combining
+active learning and partial annotation for domain
+adaptation of a japanese dependency parser. In
+_Proceedings of the 14th International Conference on_
+_Parsing Technologies_, pages 11–19.
+
+
+Daniel Flannery, Yusuke Miayo, Graham Neubig, and
+Shinsuke Mori. 2011. Training dependency parsers
+from partially annotated corpora. In _Proceedings of_
+_IJCNLP_, pages 776–784.
+
+
+Jennifer Foster, Ozlem Cetinoglu, Joachim Wagner,
+Joseph Le Roux, Joakim Nivre, Deirdre Hogan, and
+Josef van Genabith. 2011. From news to comment:
+Resources and benchmarks for parsing the language
+of web 2.0. In _Proceedings of IJCNLP_, pages 893–
+901.
+
+
+Jiang Guo, Wanxiang Che, David Yarowsky, Haifeng
+Wang, and Ting Liu. 2015. Cross-lingual dependency parsing based on distributed representations.
+In _Proceedings of ACL_, pages 1234–1244.
+
+
+Rebecca Hwa. 1999. Supervised grammar induction
+using training data with limited constituent information. In _Proceedings of ACL_, pages 73–79.
+
+
+Rebecca Hwa. 2004. Sample selection for statistical
+parsing. _Computional Linguistics_, 30(3):253–276.
+
+
+
+Lingpeng Kong, Nathan Schneider, Swabha
+Swayamdipta, Archna Bhatia, Chris Dyer, and
+Noah A. Smith. 2014. A dependency parser
+for tweets. In _Proceedings_ _of_ _EMNLP_, pages
+1001–1012.
+
+
+Terry Koo and Michael Collins. 2010. Efficient thirdorder dependency parsers. In _ACL_, pages 1–11.
+
+
+Terry Koo, Amir Globerson, Xavier Carreras, and
+Michael Collins. 2007. Structured prediction
+models via the matrix-tree theorem. In _Proceedings_
+_of EMNLP-CoNLL_, pages 141–150.
+
+
+David D. Lewis and William A. Gale. 1994. A
+sequential algorithm for training text classifiers. In
+_Proceedings_ _of_ _the_ _17th_ _Annual_ _International_ _ACM_
+_SIGIR Conference on Research and Development in_
+_Information Retrieval_, pages 3–12.
+
+
+Shoushan Li, Guodong Zhou, and Chu-Ren Huang.
+2012. Active learning for Chinese word segmentation. In _Proceedings_ _of_ _COLING_ _2012:_ _Posters_,
+pages 683–692.
+
+
+Zhenghua Li, Min Zhang, and Wenliang Chen. 2014.
+Soft cross-lingual syntax projection for dependency
+parsing. In _COLING_, pages 783–793.
+
+
+Teresa Lynn, Jennifer Foster, Mark Dras, and Elaine U1 [´]
+Dhonnchadha. 2012. Active learning and the irish
+treebank. In _Proceedings of ALTA_ .
+
+
+Xuezhe Ma and Fei Xia. 2014. Unsupervised
+dependency parsing with transferring distribution
+via parallel guidance and entropy regularization. In
+_Proceedings of ACL_, pages 1337–1348.
+
+
+Xuezhe Ma and Hai Zhao. 2015. Probabilistic models for high-order projective dependency parsing.
+_Arxiv_, abs/1502.04174.
+
+
+Saeed Majidi and Gregory Crane. 2013. Active
+learning for dependency parsing by a committee of
+parsers. In _Proceedings of IWPT_, pages 98–105.
+
+
+Diego Marcheggiani and Thierry Arti`eres. 2014.
+An experimental comparison of active learning
+strategies for partially labeled sequences. In
+_Proceedings of EMNLP_, pages 898–906.
+
+
+Andrew McCallum and Kamal Nigam. 1998.
+Employing EM and pool-based active learning for
+text classification. In _Proceedings_ _of_ _ICML_, pages
+350–358.
+
+
+Ryan McDonald and Fernando Pereira. 2006.
+Online learning of approximate dependency parsing
+algorithms. In _Proceedings of EACL_, pages 81–88.
+
+
+Ryan McDonald and Giorgio Satta. 2007. On the complexity of non-projective data-driven dependency
+parsing. In _Proceedings_ _of_ _the_ _Tenth_ _International_
+_Conference_ _on_ _Parsing_ _Technologies_, pages 121–
+132.
+
+
+
+353
+
+
+Avihai Mejer and Koby Crammer. 2012. Training
+dependency parser using light feedback. In
+_Proceedings of NAACL_ .
+
+
+Jason Mielens, Liang Sun, and Jason Baldridge. 2015.
+Parse imputation for dependency annotations. In
+_Proceedings of ACL-IJCNLP_, pages 1385–1394.
+
+
+Seyed Abolghasem Mirroshandel and Alexis Nasr.
+2011. Active learning for dependency parsing
+using partially annotated sentences. In _Proceedings_
+_of_ _the_ _12th_ _International_ _Conference_ _on_ _Parsing_
+_Technologies_, pages 140–149.
+
+
+Fredrik Olsson. 2009. A literature survey of active
+machine learning in the context of natural language
+processing. Technical report, Swedish Institute of
+Computer Science.
+
+
+Fernando Pereira and Yves Schabes. 1992. Insideoutside reestimation from partially bracketed corpora. In _Proceedings_ _of_ _the_ _Workshop_ _on_ _Speech_
+_and Natural Language (HLT)_, pages 122–127.
+
+
+Slav Petrov and Ryan McDonald. 2012. Overview of
+the 2012 shared task on parsing the web. In _Notes_
+_of the First Workshop on Syntactic Analysis of Non-_
+_Canonical Language (SANCL)_ .
+
+
+Mohammad Sadegh Rasooli and Michael Collins.
+2015. Density-driven cross-lingual transfer of
+dependency parsers. In _Proceedings_ _of_ _EMNLP_,
+pages 328–338.
+
+
+Stefan Riezler, Tracy H. King, Ronald M. Kaplan,
+Richard Crouch, John T. III Maxwell, and Mark
+Johnson. 2002. Parsing the wall street journal using
+a lexical-functional grammar and discriminative
+estimation techniques. In _Proceedings_ _of_ _40th_ _An-_
+_nual_ _Meeting_ _of_ _the_ _Association_ _for_ _Computational_
+_Linguistics_, pages 271–278.
+
+
+Manabu Sassano and Sadao Kurohashi. 2010. Using
+smaller constituents rather than sentences in active
+learning for japanese dependency parsing. In
+_Proceedings of ACL_, pages 356–365.
+
+
+Burr Settles and Mark Craven. 2008. An analysis
+of active learning strategies for sequence labeling
+tasks. In _Proceedings of EMNLP_, pages 1070–1079.
+
+
+David A. Smith and Noah A. Smith. 2007.
+Probabilistic models of nonprojective dependency
+trees. In _Proceedings_ _of_ _EMNLP-CoNLL_, pages
+132–140.
+
+
+Oscar T¨ackstr¨om, Ryan McDonald, and Joakim Nivre.
+2013. Target language adaptation of discriminative
+transfer parsers. In _Proceedings_ _of_ _NAACL_, pages
+1061–1071.
+
+
+Min Tang, Xiaoqiang Luo, and Salim Roukos. 2002.
+Active learning for statistical natural language
+parsing. In _Proceedings of ACL_, pages 120–127.
+
+
+
+Yuta Tsuboi, Hisashi Kashima, Hiroki Oda, Shinsuke
+Mori, and Yuji Matsumoto. 2008. Training conditional random fields using incomplete annotations.
+In _Proceedings of COLING_, pages 897–904.
+
+
+William Yang Wang, Lingpeng Kong, Kathryn
+Mazaitis, and William W Cohen. 2014. Dependency parsing for weibo: An efficient probabilistic
+logic programming approach. In _Proceedings_ _of_
+_EMNLP_, pages 1152–1158.
+
+
+Yue Zhang and Joakim Nivre. 2011. Transition-based
+dependency parsing with rich non-local features. In
+_Proceedings of ACL_, pages 188–193.
+
+
+
+354
+
+

references/2017.udw.bouma/paper.md

@@ -0,0 +1,798 @@
+---
+title: "Increasing Return on Annotation Investment: The Automatic Construction of a Universal Dependency Treebank for Dutch"
+authors:
+  - "Gosse Bouma"
+  - "Gertjan van Noord"
+year: 2017
+venue: "UDW 2017"
+url: "https://aclanthology.org/W17-0403/"
+---
+
+# **Increasing return on annotation investment: the automatic construction of** **a Universal Dependency treebank for Dutch**
+
+
+
+**Gosse Bouma**
+Centre for Language and Cognition
+University of Groningen
+```
+       g.bouma@rug.nl
+
+```
+
+**Abstract**
+
+
+We present a method for automatically
+converting the Dutch Lassy Small treebank, a phrasal dependency treebank, to
+
+UD. All of the information required to
+produce accurate UD annotation appears
+to be available in the underlying annotation. However, we also note that the close
+connection between POS-tags and dependency labels that is present in UD is missing in the Lassy treebanks. As a consequence, annotation decisions in the Dutch
+data for such phenomena as nominalization and clausal complements of prepositions seem to differ to some extent from
+comparable data in English and German.
+
+Because the conversion is automatic, we
+can now also compare three state-of-theart dependency parsers trained on UD
+Lassy Small with Alpino, a hybrid Dutch
+parser which produces output that is compatible with the original Lassy annotations.
+
+
+**1** **Introduction**
+
+We present a method for automatically converting Dutch treebanks annotated according to the
+guidelines of the Lassy project (van Noord et al.,
+2013) to Universal Dependencies. The Lassy annotation guidelines combine elements from phrase
+structure treebanks (such as phrasal nodes and use
+of co-indexed nodes for encoding fronted WHconstituents) with elements from dependency treebanks (such as dependency labels, discontinuous
+constituents and crossing branches), similar to the
+Tiger (Brants et al., 2002) and Negra (Skut et
+al., 1998) corpora for German. The conversion is
+done by means of an automatic conversion script. [1]
+
+
+1
+Available at `https://github.com/gossebouma/`
+```
+lassy2ud
+
+```
+
+
+**Gertjan van Noord**
+Centre for Language and Cognition
+University of Groningen
+```
+  g.j.m.van.noord@rug.nl
+
+```
+
+There are two advantages to such a procedure: the
+original annotation is of high quality as it is the result of careful manual checking and correction of
+automatically produced parser output. Using this
+investment as basis for the UD annotation as well
+means that this investment can also serve as basis for novel annotation projects. Second, an automatic conversion script allows any material that
+has been annotated according to the guidelines of
+the Lassy project to be converted to UD, and thus
+also can be used to convert treebanks outside the
+
+UD corpus and to make existing tools compliant
+with UD.
+There are two main challenges for the conversion: the Lassy treebanks contain dependency relations between phrasal nodes, whereas UD uses
+lexical dependency relations only. Second, the
+Lassy Treebanks use a more traditional notion
+of ’ _head_ ’ whereas UD gives precedence to content words over function words. As a consequence, converting from Lassy to UD requires
+’ _head-switching_ ’ in a number of cases. In section 2 we outline the main principles of the conversion process.
+The conversion has been used to produce UD
+Dutch Lassy Small (v1.3 and 2.0). Lassy Small
+is a manually verified 1 million word treebank for
+Dutch, consisting of mixed sources. For reasons
+of intellectual property rights, only the Wikipedia
+part (7.641 sentences, 101.841 tokens) is included
+in the UD corpus.
+One of the goals of the UD enterprise is to
+ensure similar annotations for similar constructions across languages. While the current state
+of the general and language specific annotation
+guidelines suggest that this should be possible for
+the most common syntactic configurations, it is
+also true that there still appears to be variation
+in the way less frequent constructions are annotated. This is particularly true if such constructions challenge the UD annotation principles. We
+
+
+
+19
+
+_Proceedings of the NoDaLiDa 2017 Workshop on Universal Dependencies (UDW 2017)_, pages 19–26,
+Gothenburg, Sweden, 22 May 2017.
+
+
+De
+
+
+
+ziekte
+
+
+
+N:obj1
+
+
+muggen
+
+
+
+overgebracht
+
+
+
+P:hd
+
+
+door
+
+
+
+
+
+
+
+
+
+Figure 1: Phrasal annotation and the induced dependency annotation for _de_ _ziekte_ _wordt_ _overgebracht_
+_door muggen_ ( _the disease is transmitted by flies_ ). External head projection paths are indicated by dashed
+arrows, and internal heads are indicated by solid arrows.
+
+
+
+illustrate this in section 3 by comparing the analysis in Dutch, German, and English, of verbal nominalizations and clausal arguments of prepositions.
+In section 4, we compare the performance of
+three dependency parsers trained on UD Lassy
+Small with Alpino (van Noord, 2006), a rule-based
+grammar that produces output compatible with the
+original Lassy treebank. The comparison crucially
+relies on the fact that we can use the conversion
+script to convert Alpino output to UD.
+
+
+**2** **Conversion Process**
+
+
+Conversion of a manually verified treebank to UD
+is possible if the underlying annotation contains
+the information that is required to do a mapping
+from the original annotation to POS-tags and bilexical dependencies that is conformant with the
+annotation guidelines of the UD project. By doing an automatic conversion, we follow a strategy
+that has been used to create many of the other UD
+treebanks as well (Zeman et al., 2014; Johannsen
+et al., 2015; Øvrelid and Hohle, 2016; Ahrenberg,
+2015; Lynn and Foster, 2016).
+Conversion of Lassy to UD POS-tags can be
+achieved by means of a simple set of case statements that refer to the original POS-tag and a small
+set of morphological feature values. The only case
+that is more involved is the distinction between
+verbs and auxiliaries. This distinction is missing
+in the POS-tags and morphological features of the
+Lassy treebanks, but can be reconstructed using
+the lemma and valency of the verb (i.e., a limited set of verbs that select for only a subject and a
+
+
+
+non-finite verbal complement or predicative complement are considered auxiliaries).
+Conversion of the phrasal syntactic annotation
+to dependency relations is driven by the observation that in a dependency graph each word (except the root) is linked via a labeled arc to exactly one lexical head. [2] Given a sentence annotated according to Lassy guidelines, we can use
+the phrasal syntactic annotation to predict for each
+token in the input (except the root) its lexical content head and dependency label. Figure 1 gives an
+overview of the most important Lassy dependency
+labels and their UD counterparts.
+The rules for finding the content head are defined using two auxiliary notions: the _’external_
+_head_ _projection’_ of a word or phrase is the node
+that contains the content head for the node. The
+_’internal head’_ of a node or phrase is the node that
+is the content head of the phrase.
+In regular configurations, the external head projection of a non-head word (i.e. a word labeled
+_su,_ _obj1,_ _det,_ _mod,_ _app,_ _etc._ ) is its mother node,
+and the internal head of this mother node is the
+node with dependency label _hd_ . This is shown
+for example in Figure 1, where the determiner _De_
+has the NP as its external head projection. The
+_hd_ node within this NP is the content head of this
+phrase, and thus the content head of the determiner. The external head projection of the _hd_ word
+itself, _ziekte_, is not the parent but the grandparent
+node, SMAIN. Thus, the content head of _ziekte_ is
+
+
+2
+Currently, no secondary edges are used in the UD Lassy
+Small.
+
+
+
+20
+
+
+**Lassy** **UD** **Interpretation**
+
+su subj _|_ csubj _|_ nsubj:pass _|_ csubj:pass various kinds of subjects
+obj1 obj _|_ obl _|_ nmod objects of verbs and prepositions
+obj2 iobj indirect objects
+mod obl _|_ advmod _|_ advcl _|_ nmod _|_ amod various kinds of modifiers
+det det _|_ nummod determiners and numbers
+app appos appositions
+cmp mark complementizers
+crd cc conjunctions
+sup expl expletives
+pobj1 expl expletives
+hd heads of phrases: check label of mother node
+... ... ...
+
+
+Table 1: Overview of re-labeling rules
+
+
+
+the internal head of the SMAIN. In this case, as we
+explain below, this is not the _hd_ daughter, but the
+content head of the daughter labeled with dependency label _vc_ ( _verbal complement_ ).
+Head-switching cases are exceptions to the general rule. The noun _muggen_, for instance, has a
+prepositional head as sister. As UD specifies that
+prepositions are dependent on the noun in these
+cases, we have to specify that the external head
+projection of the _obj1_ child inside a PP is the parent of the PP. For the same reason, the external
+head projection of the preposition is not the grandparent, but the sister _obj1_ node. The same applies
+to auxiliaries. As they are dependents of the main
+verb, their external head projection is the sister
+node labeled _vc_ (or _predc_ in copula constructions).
+Finally, as the main verb _overgebracht_ functions
+as content head of both the PPART and SMAIN, its
+external head projection should be the parent of
+
+SMAIN. As SMAIN is the root of the phrasal tree,
+we conclude that _overgebracht_ must be the root
+node.
+The analysis of WH-questions and relative
+clauses in the Lassy treebank uses a co-indexing
+scheme between the fronted element and an empty
+node that is comparable to a ’trace’ in transformational approaches. The content head for such coindexed fronted elements can be found by starting
+the external head projection identification from the
+co-indexed empty ’trace’ node.
+The identification of the correct dependency label for a bi-lexical dependency uses a mapping
+from the original Lassy dependency labels to UD.
+The most important cases are listed in Table 1.
+We have used the conversion script to create UD
+
+
+
+Lassy Small (v1.3 and 2.0). This corpus consists
+of the Wikipedia section of the manually verified
+part of the Lassy corpus.
+One aspect of the corpus that is not according
+to UD is the annotation of interpunction. As all
+punctuation marks are attached to the root node in
+the original treebank, locating the right attachment
+site according to UD rules is challenging. So far,
+we have not been able to come up with an errorfree solution.
+By way of evaluation of the result, we manually
+verified the annotation for 50 arbitrarily selected
+sentences from the corpus (v 2.0). [3] We checked
+whether the annotation was in accordance with
+the UD guidelines. In cases where we were not
+sure about the correct annotation (typically attachment decisions), we compared the annotation with
+the original Lassy treebank annotation. Ignoring
+punctuation issues, we observed 4 errors: a passive subject labeled as regular subject, an _amod_
+that has to be _advmod_, a number marked as _det_
+(should be _nummod_ ), and an error resulting from
+head-switching: the auxiliary _bekend_ _staan_ ( _be_
+_known as_ ) consists of a vebal head and a particle.
+The head is marked as _cop_, but as a consequence
+of head-switching, the particle has been reattached
+to the predicative head. This is clearly wrong, although the right annotation is not obvious: either
+this verb should not be considered an auxiliary,
+or else we must allow for particles to be dependents of auxiliaries. In addition to these errors we
+also found 6 dubious decisions (unclear distinctions between _amod_ and _advmod_ (4 _×_ ), labeling a
+
+3All sentences from the training section containing the adverb _ook_ ( _also_ ).
+
+
+
+21
+
+
+predicative phrase as _xcomp_, and a case of an incomplete word (part of a coordination) marked as
+
+X (in accordance with the original annotation but
+not the best option according to UD),
+We also tried to compare Lassy Small with the
+
+UD Dutch corpus that has been included in UD
+since v1.2. The latter corpus is a conversion of
+the Alpino treebank (van der Beek et al., 2002).
+It was used in the CONLL X shared task on dependency parsing (Buchholz and Marsi, 2006) and
+converted at that point to CONLL format. The UD
+version is based on a conversion to HamleDT to
+
+UD (Zeman et al., 2014). The various conversion
+steps have lead to loss of information, [4] and apparent mistakes, [5] and the quality of this corpus
+in general seems to be lower than the UD Lassy
+Small corpus. A more systematic comparison will
+be possible once we have been able to reconstruct
+the original sources of the material included in the
+Alpino treebank fragment used for CONLL. At that
+point, it will also be possible to create an improved
+version of the data using the current conversion
+script.
+
+
+**3** **Cross-lingual comparison**
+
+
+The inventory of dependency labels in UD is a
+mixed functional-structural system, which distinguishes oblique arguments, for instance, on the basis of their part-of-speech, i.e. a PP dependent is
+labeled _obl_, a dependent clause _advcl_, and an adverbial _advmod_ . Also, attachment to predicates is
+differentiated from attachment to nominals.
+The original Lassy Small treebank has both
+phrasal categories and dependency labels, and
+seems to make a more clear-cut distinction between structural and dependency information. For
+instance, a single _mod_ -relation is used for adjuncts
+in the verbal domain (PPs, adverbs and adverbial
+phrases, as well as clausal adjuncts) as well as in
+the nominal domain. The relevant structural distinctions are not lost, as phrasal nodes can be differentiated by the category and lexical items by
+their POS.
+The mixed functional-structural approach of
+
+UD leads to surprising outcomes in cases where
+
+
+4for instance, all heads and dependents of compound relations have been assigned the POS tag X, ignoring the original
+assignment of POS tags
+5e.g. in 13.050 sentences, there are 353 cases where a
+verb or (proper) noun functions as dependent of the _case_ relation, and 953 cases where an auxiliary has an _nsubj_ dependent
+
+
+
+The SELF assure of Russian support
+
+
+Figure 2: Nominalization in UD-Dutch Lassy
+Small
+
+
+structural relations do not align with the predicate/nominal distinction. We discuss two such situations below.
+
+
+**3.1** **Nominalizations**
+
+Nominalizations are constructions in which a verb
+functions as a noun. Nominalizations can be
+formed by means of derivational morphology, but
+there are also many cases in which there is no
+(overt) morphological suffix to mark the nominal
+status of the verb (Chomsky, 1968). Interestingly,
+in nominalizations we see both dependents typically associated with the verbal domain as well as
+dependents associated with the nominal domain,
+as in example (2). Here, a verb clearly heads
+a nominal phrase, as it is introduced by a determiner. Yet, at the same time, it selects an inherent
+reflexive pronoun, something that is not possible
+for nouns. The dependency annotation for this example in Figure 2 also shows that the PP phrase
+is labeled _nmod_, giving preference to the nominal interpretation of _verzekeren_ . Note that the the
+parallelism between (1) and (2) suggests that it
+could perhaps also have been labeled _obl_ . In fact,
+the NomBank corpus (Meyers et al., 2004) adopts
+the rule that the same semantic role labels should
+be used as much as possible for verbs and nominalised versions of these verbs.
+
+
+(1) Hij verzekert zich van Russische steun
+He assures himself of Russian support
+
+(2) het zich verzekeren van Russische steun
+the self assuring of Russian support
+was het doel
+was the goal
+’The assuring oneself of Russian support
+was the goal’
+
+
+The presence of such mixed nominal/verbal
+configurations differs strongly between treebanks.
+In Table 2 we give counts for the number of verbs
+that have a _det_ dependent, and for verbs that have
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+22
+
+
+an incoming dependency label _nsubj_ or _obj_ in the
+Dutch Lassy Small and German and English UD
+treebanks (v2.0). [6] In all cases, we are dealing
+with a verb that has clearly nominal properties:
+it has a determiner as dependent, or functions as
+subject or object of a predicate. The Dutch treebank has the highest number of nominalizations. It
+should also be noted that the (14) cases in English
+where a verb has a _det_ dependent include bona fide
+cases like _the_ _following_ and _(please_ _use)_ _the_ _at-_
+_tached_, but also several apparent annotation errors.
+The low number of nominalizations for English is
+unexpected, as nominalizations involving gerunds
+appear to be a common phenomenon in English.
+
+
+NL DE EN
+Query (101K) (277K) (229K)
+
+`VERB` `>det` `_` 112.9 22.0 6.1
+`VERB` `<nsubj` `_` 62.4 1.4 5.2
+`VERB` `<obj` `_` 21.8 2.2 8.3
+
+
+Table 2: Frequency per 100.000 tokens for verbal
+heads with nominal properties in three UD treebanks.
+
+
+**3.2** **Clausal arguments of prepositions**
+
+Another situation where the distinction between
+the predicative and nominal domain gives surprising results are PPs containing a verbal rather than a
+nominal content head. Some of these are nominalizations, and were already discussed above. However, there are also genuine clausal cases as in Fig
+|3. case|Col2|Col3|
+|---|---|---|
+|root<br>case<br>mark<br>nsubj<br>obj|root<br>case<br>mark<br>nsubj<br>obj|root<br>case<br>mark<br>nsubj<br>obj|
+|root<br>case<br>mark<br>nsubj<br>obj|root|root|
+|root<br>case<br>mark<br>nsubj<br>obj|||
+
+
+
+zonder dat nationalisme de kop opsteekt
+without that nationalism the head raises
+
+
+Figure 3: Prepositional phrases containing a
+clausal argument: _’without nationalism raising its_
+_head’_
+
+
+Again, we compared counts for such phenomena in the Dutch, German, and English UD treebanks. Table 3 shows that verbs with a preposition as dependent or verbs heading a phrase with
+
+
+6All counts in this section have been collected using the
+`dep` `search` facility of `bionlp-www.utu.fi` .
+
+
+
+label _obl_ (i.e., verbs heading an oblique dependent of a predicate) do hardly occur in English, but
+do occur with some frequency in Dutch and German. However, closer inspection of the German
+data suggests that these are dominated by annotation errors of the form _nach_ _Dortmund_ _gefahren_
+_(driven to Dortmund)_, where a regular _obl_ dependent of a verb has been annotated erroneously with
+a _case_ relation between the verb and the preposition. Prepositions are seen as case markers in UD,
+and for that reason should only have dependents
+themselves in exceptional cases. Most of these
+cases are fixed phrases of the form _due_ _to_ or _be-_
+_cause_ _of_ . This is true for the Dutch data and to
+a large extent also for the English data. The German data, however, has a high number of prepositions with dependents that are not labeled _fixed_ .
+This might be another signal of the same annotation error, in that prepositions in the German data
+apparently head regular PPs in many cases.
+
+
+query NL DE EN
+
+`VERB` `>case` `ADP` 102.0 105.8 0.4
+`VERB` `<obl` `_` 64.4 0.0 4.8
+`ADP` `>` `_` 357.4 253.4 97.8
+`ADP` `>fixed` `_` 318.8 6.9 30.6
+
+
+Table 3: Frequency per 100.000 tokens for verbs
+with a _case_ dependent and prepositions governing
+a dependent.
+
+
+We believe that the relatively high number of
+’non-canonical’ configurations in the Dutch Lassy
+Small treebank may well be due to the fact that
+
+POS-tagging and syntactic annotation were performed as two independent annotation tasks in the
+original Lassy treebank. As a consequence, in
+both annotation tasks annotators made the decision that seemed most appropriate for that task
+(i.e. choosing the correct POS-tag and syntactic
+annotation, respectively). The English UD treebanks, on the other hand, is a manually verified
+and corrected version of an automatic conversion
+of the Web treebank, where POS-tags have been
+added automatically. The construction of the German treebank was done automatically and is minimally documented. [7] Therefore, we cannot be
+sure whether the differences observed above reflect genuine typological differences or whether
+they are a consequence of the decisions made in
+
+
+7
+```
+  https://github.com/UniversalDependencies/
+UD_German
+
+```
+
+
+23
+
+
+the underlying annotation and/or of the conversion
+method.
+
+
+**4** **Parsing Experiments**
+
+
+The inclusion of a large number of languages and
+corpora in the UD corpus has led to a growing
+number of parsing toolkits that are language independent and that can be trained and evaluated
+on any of the UD treebanks. In this section, we
+compare state-of-the-art dependency parsers for
+
+UD trained on Lassy Small with Alpino, a parser
+based on a hand-written grammar for Dutch.
+Andor et al. (2016) introduce SyntaxNet, an
+open-source implementation of a novel method
+for dependency parsing based on globally normalized neural networks. They also provide a pretrained parser for English, Parsey McParseface.
+On the Penn Treebank, the released model for English (Parsey McParseface) recovers dependencies
+at the word level with over 94% accuracy, beating
+previous state-of-the-art results.
+SyntaxNet has been used to train a parser for a
+large number of corpora in UD (v1.3). ‘Parsey’s
+Cousins’ [8] is a collection of syntactic models
+trained on UD treebanks, for 40 different languages. Per language, more than 70 models have
+been trained, leading to models that are up to 4%
+more accurate than models trained without hyperparameter tuning.
+The easy-first hierarchical LSTM model of
+Kiperwasser and Goldberg (2016) introduces a
+novel method for applying the LSTM framework
+to tree structures that is particularly apt for dependency parsing. Another notable feature is that it
+does not use word embeddings. It achieves stateof-the-art results on dependency parsing for English and Chinese, and can be used to train parsers
+for any language for which a UD treebank is available. [9]
+
+‘ParseySaurus’ (Alberti et al., 2017) is a collection of models for UD version 2.0 corpora. It uses
+a variant of SyntaxNet that also includes character
+level embeddings. The model has a labeled attachment accuracy score that is on average 3.5% better
+than the SyntexNet models of Parsey’s cousins.
+Alpino (van Noord, 2006) is a wide-coverage
+parser for Dutch consisting of a carefully developed hand-written unification-based grammar and
+
+
+8
+```
+  research.googleblog.com/2016/08/
+meet-parseys-cousins-syntax-for-40.html
+```
+
+9https://github.com/elikip/htparser
+
+
+
+LAS UAS
+Alpino 84.31 89.22
+Parsey’s Cousins 78.08 81.63
+Easy-first 77.16 81.10
+ParseySaurus 80.53 84.02
+
+
+Table 4: Parse results for UD Dutch Lassy Small
+(v1.3), using standard training (6641 sentences)
+and test set (350 sentences), using CONLL 2007
+evaluation script, not counting punctuation.
+
+
+a maximum entropy disambiguation model. Its
+output is compatible with the original Lassy treebank. Although Alpino is not a dependency parser,
+it can be evaluated on UD Dutch data by converting
+the parser output into UD compatible annotation
+using the same conversion script that was also used
+to convert the original Lassy Small treebank to UD.
+For the experiment, the disambiguation model for
+Alpino was trained only on the training section of
+Lassy Small UD treebank (6641 sentences).
+We compare the accuracy of the Dutch SyntaxNet models as well as a model trained with
+the easy-first LSTM model, with results obtained
+using Alpino. Table 4 gives labeled and unlabeled attachment accuracy scores on the test set
+of the Lassy Small corpus. The scores for Parsey’s
+Cousins are the scores reported in the Google blog
+post. The scores for easy-LSTM were obtained by
+running the code using the default options. The
+scores for ParseySaurus are taken from (Alberti et
+al., 2017).
+Among the dependency parsers trained on the
+treebank data, the ParseySaurus model achieves
+a 2.5-3.0% LAS improvement over the two other
+models. Alpino performs even better, with a 3.8%
+
+LAS improvement over the best dependency parser
+model. We can only speculate about the reasons for this difference. The training corpus is
+relatively small, and it might be that the purely
+data-driven approaches would benefit relatively
+strongly from being trained on more data. [10]
+
+On the other hand, results can also be improved
+by simply correcting errors in the original data.
+As we pointed out in section 3, one difference between the original Lassy dependency annotation
+and UD is that UD dependency labels are organized
+more strongly in accordance with the POS tag of
+
+
+10However, note that if we use the standard Alpino disambiguation component, trained on a larger, news domain corpus, its accuracy slightly decreases (88.21 UAS).
+
+
+
+24
+
+
+LAS UAS
+Alpino 84.31 89.22
+with corrected treebank 85.95 89.41
+
+
+Table 5: Parse results for UD Dutch Lassy Small
+(v1.3 with corrections), using standard training
+(6641 sentences) and test set (350 sentences), using CONLL 2007 evaluation script, not counting
+punctuation.
+
+
+the head and the dependent than the Lassy dependency labels. The relative independence of Lassy
+
+POS annotation and syntactic analysis (as well as
+the fact that these were done by different partners
+in the Lassy project), has led to a situation where
+errors in POS annotation have gone largely unnoticed when evaluating parser output. For evaluation on UD treebanks, annotating and predicting
+the correct POS tag is crucial, as it influences the
+choice of the dependency label. Thus, correcting
+
+POS tags in the original treebank leads to more
+consistent data in the original treebank as well as
+in the converted UD treebank. If we evaluate the
+Alpino parser on a version of the UD treebank
+based on the corrected underlying Lassy Small
+treebank, we obtain the accuracy scores given in
+table 5. These corrections have been included in
+
+UD 2.0 release.
+
+
+**5** **Conclusions**
+
+
+Automatic conversion of existing treebanks to UD
+has the advantage that existing annotation efforts
+can be re-used, that treebanks that for some reason
+cannot be included in the UD corpus can be converted easily, and that tools developed for the original annotation can be used to produce UD compliant output as well. We have developed a method
+for converting the Dutch Lassy treebank to UD. It
+has been used to produce UD Dutch Lassy Small,
+included in UD v1.3 and v2.0.
+Although all information required to do the conversion appears to be present in the underlying
+annotation (with the exception of punctuation attachment perhaps), we did notice that there are
+also subtle differences between the Lassy treebank
+annotation and UD annotation guidelines. This
+is particularly clear in cases where structural and
+functional information does not align well, as in
+nominalizations.
+Using our automatic annotation script, we were
+able to compare parsing accuracies for three de
+
+
+pendency parsers and Alpino. Although the results for dependency parsing are encouraging, the
+Alpino parser, based on a hand-written grammar,
+still outperforms these approaches.
+In future work, we would like to expand the UD
+Lassy Small corpus by including more of the material of the original Lassy Small corpus (where
+this is allowed according to IPR). For parser evaluation, it would be interesting to see what the effect
+is of larger training sets on automatically trained
+dependency parsers in particular.
+
+
+**References**
+
+Lars Ahrenberg. 2015. Converting an EnglishSwedish parallel treebank to universal dependencies. In _Third_ _International_ _Conference_ _on_ _Depen-_
+_dency Linguistics (DepLing 2015), Uppsala, August_
+_24-26_, pages 10–19. Association for Computational
+Linguistics.
+
+
+Chris Alberti, Daniel Andor, Ivan Bogatyy, Michael
+Collins, Dan Gillick, Lingpeng Kong, Terry Koo,
+Ji Ma, Mark Omernick, Slav Petrov, Chayut
+Thanapirom, Zora Tung, and David Weiss. 2017.
+Syntaxnet models for the CoNLL 2017 shared task.
+
+
+Daniel Andor, Chris Alberti, David Weiss, Aliaksei
+Severyn, Alessandro Presta, Kuzman Ganchev, Slav
+Petrov, and Michael Collins. 2016. Globally normalized transition-based neural networks. In _Pro-_
+_ceedings of the ACL_ .
+
+
+Sabine Brants, Stefanie Dipper, Silvia Hansen, Wolfgang Lezius, and George Smith. 2002. The TIGER
+treebank. In _Proceedings_ _of_ _the_ _workshop_ _on_ _Tree-_
+_banks and Linguistic Theories_, volume 168.
+
+
+Sabine Buchholz and Erwin Marsi. 2006. CoNLL-X
+shared task on multilingual dependency parsing. In
+_Proceedings_ _of_ _the_ _Tenth_ _Conference_ _on_ _Computa-_
+_tional Natural Language Learning_, pages 149–164.
+Association for Computational Linguistics.
+
+
+Noam Chomsky. 1968. _Remarks_ _on_ _nominalization_ .
+Linguistics Club, Indiana University.
+
+
+Anders Johannsen, H´ector Mart´ınez Alonso, and Barbara Plank. 2015. Universal dependencies for Danish. In _International_ _Workshop_ _on_ _Treebanks_ _and_
+_Linguistic Theories (TLT14)_, page 157.
+
+
+Eliyahu Kiperwasser and Yoav Goldberg. 2016.
+Easy-first dependency parsing with hierarchical tree
+LSTMs. _Transactions of the ACL_, 4:445–461.
+
+
+Teresa Lynn and Jennifer Foster. 2016. Universal dependencies for Irish. In _Celtic Language Technology_
+_Workshop_, pages 79–92.
+
+
+Adam Meyers, Ruth Reeves, Catherine Macleod,
+Rachel Szekely, Veronika Zielinska, Brian Young,
+
+
+
+25
+
+
+and Ralph Grishman. 2004. Annotating noun argument structure for NomBank. In _LREC_, volume 4,
+pages 803–806.
+
+
+Lilja Øvrelid and Petter Hohle. 2016. Universal dependencies for Norwegian. In _Proceedings_ _of_ _the_
+_Tenth_ _International_ _Conference_ _on_ _Language_ _Re-_
+_sources and Evaluation. Portoroˇz, Slovenia_ .
+
+
+Wojciech Skut, Thorsten Brants, Brigitte Krenn, and
+Hans Uszkoreit. 1998. A linguistically interpreted
+corpus of German newspaper text. _arXiv_ _preprint_
+_cmp-lg/9807008_ .
+
+
+Leonoor van der Beek, Gosse Bouma, Rob Malouf,
+and Gertjan van Noord. 2002. The Alpino dependency treebank. In _Computational Linguistics in the_
+_Netherlands (CLIN) 2001_, Twente University.
+
+
+Gertjan van Noord, Gosse Bouma, Frank van Eynde,
+Daniel de Kok, Jelmer van der Linde, Ineke Schuurman, Erik Tjong Kim Sang, and Vincent Vandeghinste. 2013. Large scale syntactic annotation of written Dutch: Lassy. In Peter Spyns and Jan Odijk,
+editors, _Essential Speech and Language Technology_
+_for Dutch: the STEVIN Programme_, pages 147–164.
+Springer.
+
+
+Gertjan van Noord. 2006. At last parsing is now operational. In Piet Mertens, Cedrick Fairon, Anne Dister, and Patrick Watrin, editors, _TALN06. Verbum Ex_
+_Machina._ _Actes_ _de_ _la_ _13e_ _conference_ _sur_ _le_ _traite-_
+_ment automatique des langues naturelles_, pages 20–
+42.
+
+
+Daniel Zeman, Ondˇrej Duˇsek, David Mareˇcek, Martin Popel, Loganathan Ramasamy, Jan Stˇep´anek, [ˇ]
+Zdenˇek Zabokrtsk`y, [ˇ] and Jan Hajiˇc. 2014. HamleDT: Harmonized multi-language dependency treebank. _Language_ _Resources_ _and_ _Evaluation_,
+48(4):601–637.
+
+
+
+26
+
+

references/2018.lrec.nguyen/paper.md

@@ -0,0 +1,486 @@
+---
+title: "BKTreebank: Building a Vietnamese Dependency Treebank"
+authors:
+  - "Kiem-Hieu Nguyen"
+year: 2018
+venue: "LREC 2018"
+url: "https://aclanthology.org/L18-1341/"
+---
+
+# **BKTreebank: Building a Vietnamese Dependency Treebank**
+
+**Kiem-Hieu Nguyen**
+School of information and communication technology,
+Hanoi university of science and technology,
+1 Dai Co Viet, Bach Khoa, Hai Ba Trung, Hanoi, Vietnam
+hieunk@soict.hust.edu.vn
+
+
+**Abstract**
+Dependency treebank is an important resource in any language. In this paper, we present our work on building BKTreebank, a
+dependency treebank for Vietnamese. Important points on designing POS tagset, dependency relations, and annotation guidelines are
+discussed. We describe experiments on POS tagging and dependency parsing on the treebank. Experimental results show that the
+treebank is a useful resource for Vietnamese language processing.
+
+
+**Keywords:** treebank, dependency parsing, POS tagging, word segmentation, Vietnamese, less-resourced language
+
+
+
+**1.** **Introduction**
+
+
+Dependency treebank is important for data-driven dependency parsing. However, building a dependency treebank
+is complicated and expensive.
+Dependency treebanks have been available in English and
+many languages. VnDT (Nguyen et al., 2014) is a Vietnamese dependency treebank which was automatically
+converted from tree bracketing in VietTreebank (VTB)
+(Nguyen et al., 2009; Nguyen et al., 2015).
+In this work, we present the building of a dependency treebank for Vietnamese [1] . Our treebank was manually annotated by annotators. Its annotation guidelines substantially
+differ from VTB. Our contributions are two-fold:
+
+
+_•_ A manual dependency treebank for Vietnamese.
+
+
+_•_ Experiments on POS tagging and dependency parsing
+based on the treebank.
+
+
+The paper is organized as follows: Section 2. briefly introduces related work on building treebanks for Vietnamese
+and dependency treebanks for other languages. Section 3.
+highlights important points of annotation guidelines. Section 4. describes in brief the annotation process. Section 5.
+is dedicated to evaluations and discussions on automatic
+POS tagging and dependency parsing results. The paper
+is concluded in Section 6.
+
+
+**2.** **Related Work**
+
+**2.1.** **Treebanks for Vietnamese**
+
+VTB was the pioneer treebank for Vietnamese. It has been
+developed from 2006-2010. It contains manual annotations
+on about 40K sentences for word segmentation, 10K sentences for POS tagging, and 10K sentences for bracketing.
+VnDT contains dependency annotations which were automatically converted from bracketing annotations in VTB.
+State-of-the-art performance on VnDT is 80.7% and 73.5%
+on UAS and LAS, respectively (Nguyen et al., 2016a).
+Recently, a new treebank for Vietnamese has been developed (Nguyen et al., 2016b; Nguyen et al., 2017). It consists of 40K sentences annotated with word segmentation,
+
+
+1For information on using BKTreebank, please visit
+[http://is.hust.edu.vn/˜hieunk/bktreebank/](http://is.hust.edu.vn/~hieunk/bktreebank/)
+
+
+
+POS tagging, and bracketing. While generally agreeing
+on word segmentation and bracketing, they propose a POS
+tagset and POS tagging guidelines which focus more on
+word-class transformation, particularly between verbs and
+other word-classes. This issue is important as Vietnamese
+is an analytic language. Unfortunately, their treebank has
+not been publicly available for research community yet.
+
+
+**2.2.** **Dependency treebank for other languages**
+
+One of the most notable dependency treebanks for English
+was developed by Stanford NLP group (De Marneffe and
+Manning, 2008). The Stanford treebank is automatically
+converted from PeenTreebank phrase structures (Marneffe
+et al., 2006). Similar approaches were used to build dependency treebanks in other languages such as French, Korean, and Croatian (Candito et al., 2010; Choi et al., 2012;
+Berovic et al., 2012). Other treebanks are built manually
+for languages such as Norwegian (Solberg et al., 2014).
+The Universal Dependencies is inherited from Penn POS
+tagset and Stanford typed dependency representation, and
+has been expanded to many languages (Marneffe et al.,
+2014; Nivre et al., 2016).
+
+
+**3.** **Annotation Guidelines**
+
+**3.1.** **POS tagging guidelines**
+
+Our POS tagset relies on Penn tagset (Santorini, 1990) with
+the following adaptation to Vietnamese (see Table 2 for the
+full tagset):
+
+
+_•_ As Vietnamese is an analytic language, we omit tags
+related to plurality, tense, and superlative in Penn
+tagset.
+
+
+_•_ _CL_ is used for noun classifiers. In Vietnamese, a
+countable noun could be accompanied by a classifier
+when we want to indicate quantity or simply to emphasize. For example, ‘tấm’ is a classifier’ in “Anh ta
+giành được hai tấm huy chương vàng” (He won two
+gold medals); ‘chiếc’ is a classifier in “Chiếc xe này
+khá đắt” (This car is quite expensive). In (Nguyen et
+al., 2016b), the authors also dedicate two tags _Nc_ and
+_Ncs_ for noun classifiers. Similar phenomena could be
+found in other languages such as Korean (Kim and
+Yang, 2006).
+
+
+
+2164
+
+
+_•_ _PFN_ is used for prefix nominalizers. Many nominal
+expressions in Vietnamese are formed by a leading
+nominalizer and a verb or an adjective (see Table 1
+for examples). In (Nguyen et al., 2016b), there are
+also POS tags mentioning word-class transformation
+including VA (Verb-Adjective), VN (Verb-Noun), and
+NA (Noun-Adjective) but it is not clear from the paper
+how the tags are designed.
+
+
+_•_ _NML_ is used for phrasal nominalizers. In Vietnamese,
+a special word such as ‘việc’ is used as a clausal adverbial marker for a clausal component. For instance, in
+“Việc xửlý chất thải công nghiệp cần được làm ngay”
+(The processing of industry garbage needs to be done
+immediately), ‘việc’ is the marker for the clausal subject.
+
+
+_•_ _VA_ is used for adjectival verb. In Vietnamese, when the
+predicate is an adjective, there is no copula verb _to be_ .
+It is hence tagged as an adjectival verb. In the sentence
+“Tình hình tương đối khảquan” (The situation is [2] quite
+positive), ‘khảquan’ is predicate and is tagged as VA.
+
+
+_•_ _AV_ stand for verbal adjective. When a verb modifies a
+noun, it is tagged as an verbal adjective (e.g. biển/NN
+quảng_cáo/AV (advertising board)).
+
+
+_•_ _TO_ is used to tagged ‘để’, which has similar meaning
+as “in order to” in English.
+
+|Prefix nominalizer|Word|Expression|
+|---|---|---|
+|niềm<br>sự<br>niềm|vui<br>hi sinh<br>tin|niềm vui (happiness)<br>sựhi sinh (sacrifice)<br>niềm tin (belief)|
+
+
+
+Table 1: Examples of prefix nominalizer in Vietnamese.
+
+
+**3.2.** **Dependency parsing guidelines**
+
+
+Our dependency relations relies on Stanford dependencies
+(De Marneffe and Manning, 2008) (Table 3). We add two
+relations for norminalization:
+
+
+_•_ _case:pfn_ is used for nominalizing modifier between a
+headword as a nominalizer and a verb or an adjective
+(see examples in Table 1).
+
+
+_•_ _mark:relcl_ is used for phrasal adverbial modifier between a headword as the predicate of the clause and a
+marker such as ‘việc’.
+
+
+Guidelines for other relations are similar to Stanford dependencies with some modifications. For instance,
+
+
+_•_ _aux_ is also used for relationship between a verb and a
+tense auxiliary (e.g. thực hiện/VB - aux - đang/MD in
+“đang thực hiện” (be executing)).
+
+
+2Note that there is no _to be_ in the sentence in Vietnamese due
+to _zero copula_ .
+
+
+
+|POS tag|Description|
+|---|---|
+|CD<br>DT<br>MD<br>NN<br>NNP<br>NML*<br>PFN*<br>PRP<br>RB<br>VB<br>VA*<br>IN<br>JJ<br>AV*<br>PUNCT<br>CC<br>WDT<br>WP<br>WRB<br>CL*<br>TO<br>UH<br>FW|Cardinal number<br>Determiner<br>Modal<br>Noun<br>Proper noun<br>Phrasal nominalizer<br>Prefix nominalizer<br>Personal pronoun<br>Adverb<br>Verb<br>Adjectival verb<br>Preposition<br>Adjective<br>Verbal adjective<br>Punctuation<br>Coordinating conjunction<br>Wh-determiner<br>Wh-pronoun<br>Wh-adverb<br>Noun classifier<br>‘để’ (in order to)<br>Interjection<br>Foreign word|
+
+
+Table 2: Our POS tagset (* Tag specific for Vietnamese).
+
+
+_•_ _det_ is also used for relationship between a noun and
+its plural marker. Here, we tag a plural marker as a
+determiner (e.g. trường hợp/NN   - det   - những/DT in
+“những trường hợp” (cases)).
+
+
+**4.** **Annotation Process**
+The raw corpus was collected from Dantri [3], a generaldomain online news agency.
+Texts were first segmented by UETSegmenter (Nguyen and
+Le, 2016). Sentences longer than 50 words were removed.
+Three annotators produced manual POS tagging and dependency parsing using the annotation tool BRAT (Stenetorp et
+al., 2012).
+We decided to annotate POS tagging and dependency in
+parallel because the two tasks are complimentary to each
+other. After being explained the annotation guidelines, the
+annotators were first asked to separately annotate the same
+small sample dataset. After finishing the sample dataset,
+they discussed differences and agreed on final decisions.
+After being trained, each annotator were asked to annotate
+separate documents. They discussed with each other when
+dealing with confusing cases. Every week, the annotators
+together reviewed and discussed a random annotated document. In the final round, a forth annotator reviewed all
+annotations and discussed with the annotators in the previous round when necessary to make final decisions.
+After removing invalid parsed sentences, our treebank contains 6909 manually annotated sentences on POS tagging and dependency parsing with the average speed of 7
+min/sentence.
+
+
+[3http://dantri.vn](http://dantri.vn)
+
+
+
+2165
+
+
+Figure 1: An annotation example in BRAT (SEA Games 28 witnesses an excellent performance from Golden Girl Anh
+Vien).
+
+
+
+|Relation|Description|
+|---|---|
+|nsubj<br>nsubjpass<br>dobj<br>iobj<br>csubj<br>csubjpass<br>ccomp<br>xcomp<br>advcl<br>advmod<br>aux<br>cop<br>mark<br>mark:relcl*<br>nmod<br>appos<br>nummod<br>acl<br>amod<br>det<br>case:pfn*<br>case<br>conj<br>cc<br>punct<br>dep|Nominal subject<br>Passive nominal subject<br>Direct object<br>Indirect object<br>Clausal subject<br>Passive clausal subject<br>Clausal component<br>Open clausal component<br>Adverbial clause modifier<br>Adverbial modifier<br>Auxiliary<br>Copula<br>Marker<br>Phrasal nominalizer<br>Nominal modifier<br>Appositional modifier<br>Numeric modifier<br>Adjectival clause<br>Adjectival modifier<br>Determiner<br>Prefix nominalizer<br>Case marking<br>Conjunct<br>Coordinating conjunction<br>Punctuation<br>Unspecified dependency|
+
+
+Table 3: Our dependency relations (* Relation specific for
+Vietnamese)
+
+
+Figure 1 illustrates an annotation example using BRAT.
+Segmented texts are put into BRAT. Syllables of the same
+word are connected by ‘ ~~’~~ . POS tags are labeled for each
+words.
+
+
+**5.** **Annotation Evaluations**
+
+
+**5.1.** **Inter annotator agreement**
+
+
+After finishing annotation, the three annotators were asked
+again to separately annotate the same small dataset to measure Inter-Annotator-Agreement (IAA). Averaged _kappa_ is
+94.5, 85.2, and 80.4 for POS tagging, unlabeled dependency parsing, and labeled dependency parsing, respectively. Note that IAA was measured for separate annotations of the three annotators without revising of the forth
+one. Such agreement shows good coherence between different annotators.
+
+
+
+**5.2.** **Initial results on POS tagging and**
+**dependency parsing**
+
+
+The treebank was divided into a training set of 5639 sentences and a test set of 1270 sentences for learning and testing POS tagging and dependency parsing.
+We built a vanilla POS tagging model using CRFSuite [4] implementation of first-order Conditional Random Fields with
+default hyper-parameters. We used a straightforward feature set as described in Table 4. Our lexicon was built by
+merging the lexicon of VietTreebank (Nguyen et al., 2006)
+with frequent tags in our corpus considering important differences in tagging guidelines. Only (word, tag) pairs that
+were tagged more than three times in the corpus were considered and were reviewed before adding to the lexicon.
+
+
+**Feature set**
+
+w[-2], w[-1], w[0], w[1], w[2]
+candidate tags
+is ~~h~~ ead ~~c~~ apitalized
+is ~~a~~ ll ~~c~~ apitalized
+is ~~n~~ umeric
+
+
+Table 4: Feature set for learning POS tagger with CRF
+
+|Tag|P|R|F|
+|---|---|---|---|
+|NN<br>IN<br>MD<br>VB<br>VA<br>CD<br>RB<br>CL<br>AV<br>PUNCT<br>JJ<br>NNP<br>DT<br>PFN<br>CC<br>PRP|92.4<br>89.0<br>97.6<br>89.6<br>58.2<br>89.1<br>84.2<br>85.3<br>59.4<br>99.9<br>85.9<br>91.9<br>97.1<br>73.9<br>92.5<br>90.7|93.6<br>95.0<br>98.3<br>91.1<br>41.6<br>97.7<br>87.0<br>71.1<br>42.3<br>100.0<br>66.9<br>94.5<br>94.6<br>86.7<br>96.3<br>88.6|93.0<br>91.9<br>98.0<br>90.3<br>48.6<br>93.2<br>85.5<br>77.6<br>49.4<br>99.9<br>75.2<br>93.2<br>95.8<br>79.8<br>94.4<br>89.7|
+|**Overall accuracy: 90.7**|**Overall accuracy: 90.7**|**Overall accuracy: 90.7**|**Overall accuracy: 90.7**|
+
+
+
+Table 5: POS performance by tag
+
+
+[4http://www.chokkan.org/software/](http://www.chokkan.org/software/crfsuite/)
+
+[crfsuite/](http://www.chokkan.org/software/crfsuite/)
+
+
+
+2166
+
+
+We used the transition-based MaltParser (Nivre et al., 2007)
+with default algorithm and feature set [5] to built a vanilla dependency parser.
+
+|Relation|UAS|LAS|
+|---|---|---|
+|ROOT<br>acl<br>advcl<br>advmod<br>amod<br>aux<br>auxpass<br>case<br>case:pfn<br>cc<br>ccomp<br>cl<br>conj<br>cop<br>csubj<br>dep<br>det<br>dobj<br>mark<br>mark:relcl<br>neg<br>nmod<br>nsubj<br>nsubjpass<br>nummod<br>punct<br>xcomp|80.4<br>63.4<br>64.7<br>86.8<br>89.9<br>98.4<br>98.8<br>97.5<br>100.0<br>84.9<br>77.9<br>100.0<br>59.9<br>95.4<br>75.0<br>72.2<br>97.1<br>92.4<br>93.1<br>100.0<br>89.6<br>78.2<br>86.2<br>93.5<br>91.6<br>73.9<br>79.9|80.4<br>63.4<br>39.5<br>86.2<br>89.4<br>97.4<br>92.8<br>97.5<br>100.0<br>84.9<br>46.8<br>100.0<br>48.9<br>94.2<br>63.9<br>72.2<br>97.1<br>89.2<br>93.1<br>100.0<br>85.6<br>74.3<br>79.6<br>75.8<br>89.5<br>73.7<br>70.9|
+|**Overall**|**84.4**|**81.4**|
+
+
+
+Table 6: Dependency parsing performance by relation
+
+
+**5.3.** **Discussions**
+
+As shown in Table 5, performance of POS tagging on nouns
+is similar to averaged performance. Verbs are more difficult to tag as they are ambiguous, not only with nouns and
+adjectives, but also with verbal adjective (modifiers). Automatic tagging of verbal adjective modifiers is very challenging as such modifiers are not infectional, and in some cases
+it requires knowledge at syntactic level. They are usually
+mistakenly tagged as a predicate verb. Verbal adjectives
+are also difficult because of zero-copula phenomenon.
+Dependency parsing performance is promising as shown
+in Table 6. Parsing at phrase-level is accurate except for
+nominal modifiers perhaps due to confusing usage of directional and temporal adverbial nouns and prepositions in
+Vietnamese. On the other hand, parsing at clause-level is
+poor. There are plenty rooms for improvement on such
+long-distance dependencies.
+
+
+**6.** **Conclusion**
+
+In this paper, we present the building of a dependency treebank for Vietnamese. Our work is based on previous works
+
+
+[5http://www.maltparser.org/userguide.html](http://www.maltparser.org/userguide.html)
+
+
+
+on treebanks for Vietnamese and dependency treebanks for
+other languages. Although current size of the corpus is limited, initial experimental results on POS tagging and dependency parsing is promising.
+In the future, we are going to expand BKTreebank with a
+bootstrapping approach using automatic tagger and parser
+learned from the dataset. We are going to investigate several approaches to POS tagging and dependency parsing for
+Vietnamese, including the joint learning approach.
+
+
+**7.** **Acknowledgements**
+
+
+This project has been partially funded by VCCorp via collaboration with Data science laboratory, School of information and communication technology, Hanoi university of
+science and technology. We would like to thank Vu Xuan
+Luong for enthusiastic discussions on VietTreebank.
+
+
+**8.** **Bibliographical References**
+
+
+Berovic, D., Agic, Z., and Tadic, M. (2012). Croatian dependency treebank: Recent development and initial experiments. In _Proceedings_ _of_ _the_ _Eight_ _Interna-_
+_tional_ _Conference_ _on_ _Language_ _Resources_ _and_ _Evalua-_
+_tion_ _(LREC’12)_, Istanbul, Turkey, may. European Language Resources Association (ELRA).
+Candito, M., Crabbe, B., and Denis, P. (2010). Statistical french dependency parsing: Treebank conversion and first results. In Nicoletta Calzolari (Conference
+Chair), et al., editors, _Proceedings_ _of_ _the_ _Seventh_ _Inter-_
+_national_ _Conference_ _on_ _Language_ _Resources_ _and_ _Eval-_
+_uation (LREC’10)_, Valletta, Malta, may. European Language Resources Association (ELRA).
+Choi, D., Park, J., and Choi, K.-S. (2012). Korean treebank transformation for parser training. In _Proceedings_
+_of_ _the_ _ACL_ _2012_ _Joint_ _Workshop_ _on_ _Statistical_ _Parsing_
+_and_ _Semantic_ _Processing_ _of_ _Morphologically_ _Rich_ _Lan-_
+_guages_, pages 78–88, Jeju, Republic of Korea, July 12.
+Association for Computational Linguistics.
+De Marneffe, M.-C. and Manning, C. D. (2008). The stanford typed dependencies representation. In _Coling 2008:_
+_Proceedings_ _of_ _the_ _Workshop_ _on_ _Cross-Framework_
+_and_ _Cross-Domain_ _Parser_ _Evaluation_, CrossParser ’08,
+pages 1–8, Stroudsburg, PA, USA. Association for Computational Linguistics.
+Kim, J.-B. and Yang, J., (2006). _Processing_ _Korean_ _Nu-_
+_meral Classifier Constructions in a Typed Feature Struc-_
+_ture Grammar_, pages 103–110. Springer Berlin Heidelberg, Berlin, Heidelberg.
+Marneffe, M., Maccartney, B., and Manning, C. (2006).
+Generating typed dependency parses from phrase structure parses. In _Proceedings_ _of_ _the_ _Fifth_ _International_
+_Conference_ _on_ _Language_ _Resources_ _and_ _Evaluation_
+_(LREC-2006)_, Genoa, Italy, May. European Language
+Resources Association (ELRA). ACL Anthology Identifier: L06-1260.
+Marneffe, M.-C. D., Dozat, T., Silveira, N., Haverinen, K.,
+Ginter, F., Nivre, J., and Manning, C. D. (2014). Universal stanford dependencies: a cross-linguistic typology.
+In Nicoletta Calzolari (Conference Chair), et al., editors,
+_Proceedings_ _of_ _the_ _Ninth_ _International_ _Conference_ _on_
+
+
+
+2167
+
+
+_Language Resources_ _and_ _Evaluation_ _(LREC’14)_, Reykjavik, Iceland, may. European Language Resources Association (ELRA).
+
+Nguyen, T. P. and Le, A. C. (2016). A hybrid approach
+to vietnamese word segmentation. In _2016_ _IEEE_ _RIVF_
+_International Conference on Computing Communication_
+_Technologies,_ _Research,_ _Innovation,_ _and_ _Vision_ _for_ _the_
+_Future (RIVF)_, pages 114–119, Nov.
+
+Nguyen, T. M. H., Romary, L., Rossignol, M., and Vu, X. L.
+(2006). A lexicon for vietnamese language processing.
+_Language Resources and Evaluation_, 40(3/4):291–309.
+
+Nguyen, P. T., Vu, X. L., Nguyen, T. M. H., Nguyen, V. H.,
+and Le, H. P. (2009). Building a large syntacticallyannotated corpus of vietnamese. In _Proceedings_ _of_ _the_
+_Third_ _Linguistic_ _Annotation_ _Workshop_, pages 182–185,
+Suntec, Singapore, August. Association for Computational Linguistics.
+
+Nguyen, D. Q., Nguyen, D. Q., Pham, S. B., Nguyen, P.-T.,
+and Le Nguyen, M., (2014). _From Treebank Conversion_
+_to Automatic Dependency Parsing for Vietnamese_, pages
+196–207. Springer International Publishing, Cham.
+
+Nguyen, P.-T., Le, A.-C., Ho, T.-B., and Nguyen, V.-H.
+(2015). Vietnamese treebank construction and entropybased error detection. _Lang._ _Resour._ _Eval._, 49(3):487–
+519, September.
+
+Nguyen, D. Q., Dras, M., and Johnson, M. (2016a). An
+empirical study for vietnamese dependency parsing. In
+_Proceedings_ _of_ _the_ _Australasian_ _Language_ _Technology_
+_Association Workshop 2016_, pages 143–149, Melbourne,
+Australia, December.
+
+Nguyen, Q., Miyao, Y., Le, H., and Nguyen, N. (2016b).
+Challenges and solutions for consistent annotation of
+vietnamese treebank. In Nicoletta Calzolari (Conference
+Chair), et al., editors, _Proceedings_ _of_ _the_ _Tenth_ _Interna-_
+_tional_ _Conference_ _on_ _Language_ _Resources_ _and_ _Evalua-_
+_tion_ _(LREC_ _2016)_, Paris, France, may. European Language Resources Association (ELRA).
+
+Nguyen, Q. T., Miyao, Y., Le, H. T. T., and Nguyen, N.
+T. H. (2017). Ensuring annotation consistency and accuracy for vietnamese treebank. _Language_ _Resources_ _and_
+_Evaluation_, Jul.
+
+Nivre, J., Hall, J., Nilsson, J., Chanev, A., Eryigit, G.,
+Kubler, S., Marinov, S., and Marsi, E. (2007). Maltparser: A language-independent system for data-driven
+dependency parsing. _Natural_ _Language_ _Engineering_,
+13(2):95–135.
+
+Nivre, J., de Marneffe, M.-C., Ginter, F., Goldberg, Y.,
+Hajic, J., Manning, C. D., McDonald, R., Petrov, S.,
+Pyysalo, S., Silveira, N., Tsarfaty, R., and Zeman,
+D. (2016). Universal dependencies v1: A multilingual
+treebank collection. In Nicoletta Calzolari (Conference
+Chair), et al., editors, _Proceedings_ _of_ _the_ _Tenth_ _Interna-_
+_tional_ _Conference_ _on_ _Language_ _Resources_ _and_ _Evalua-_
+_tion_ _(LREC_ _2016)_, Paris, France, may. European Language Resources Association (ELRA).
+
+Santorini, B. (1990). Part-Of-Speech tagging guidelines
+for the Penn Treebank project (3rd revision, 2nd printing). Technical report, Department of Linguistics, University of Pennsylvania, Philadelphia, PA, USA.
+
+
+
+Solberg, P. E., Skjarholt, A., Ovrelid, L., Hagen, K.,
+and Johannessen, J. B. (2014). The norwegian dependency treebank. In Nicoletta Calzolari (Conference
+Chair), et al., editors, _Proceedings of the Ninth Interna-_
+_tional_ _Conference_ _on_ _Language_ _Resources_ _and_ _Evalua-_
+_tion (LREC’14)_, Reykjavik, Iceland, may. European Language Resources Association (ELRA).
+Stenetorp, P., Pyysalo, S., Topi´c, G., Ohta, T., Ananiadou,
+S., and Tsujii, J. (2012). brat: a web-based tool for nlpassisted text annotation. In _Proceedings_ _of_ _the_ _Demon-_
+_strations at the 13th Conference of the European Chapter_
+_of_ _the_ _Association_ _for_ _Computational_ _Linguistics_, pages
+102–107, Avignon, France, April. Association for Computational Linguistics.
+
+
+
+2168
+
+

references/2018.naacl.nguyen/paper.md

@@ -0,0 +1,589 @@
+---
+title: "VnCoreNLP: A Vietnamese Natural Language Processing Toolkit"
+authors:
+  - "Thanh Vu"
+  - "Dat Quoc Nguyen"
+  - "Dai Quoc Nguyen"
+  - "Mark Dras"
+  - "Mark Johnson"
+year: 2018
+venue: "NAACL 2018 Demo"
+url: "https://aclanthology.org/N18-5012/"
+---
+
+# **VnCoreNLP: A Vietnamese Natural Language Processing Toolkit**
+
+**Thanh Vu** [1] **, Dat Quoc Nguyen** [2] **, Dai Quoc Nguyen** [3] **, Mark Dras** [4] and **Mark Johnson** [4]
+
+1Newcastle University, United Kingdom; 2The University of Melbourne, Australia;
+3Deakin University, Australia; 4Macquarie University, Australia
+`thanh.vu@newcastle.ac.uk`, `dqnguyen@unimelb.edu.au`,
+`dai.nguyen@deakin.edu.au`, `{mark.dras,` `mark.johnson}@mq.edu.au`
+
+
+**Abstract**
+
+
+
+We present an easy-to-use and fast toolkit,
+namely VnCoreNLP—a Java NLP annotation pipeline for Vietnamese. Our VnCoreNLP
+supports key natural language processing
+(NLP) tasks including word segmentation,
+part-of-speech (POS) tagging, named entity
+recognition (NER) and dependency parsing,
+and obtains state-of-the-art (SOTA) results for
+these tasks. We release VnCoreNLP to provide
+rich linguistic annotations to facilitate research
+work on Vietnamese NLP. Our VnCoreNLP
+is open-source and available at: `https://`
+`github.com/vncorenlp/VnCoreNLP` .
+
+
+**1** **Introduction**
+
+
+Research on Vietnamese NLP has been actively
+explored in the last decade, boosted by the successes of the 4-year KC01.01/2006-2010 national
+project on Vietnamese language and speech processing (VLSP). Over the last 5 years, standard
+benchmark datasets for key Vietnamese NLP tasks
+are publicly available: datasets for word segmentation and POS tagging were released for the first
+VLSP evaluation campaign in 2013; a dependency
+treebank was published in 2014 (Nguyen et al.,
+2014); and an NER dataset was released for the
+second VLSP campaign in 2016. So there is a need
+for building an NLP pipeline, such as the Stanford
+CoreNLP toolkit (Manning et al., 2014), for those
+key tasks to assist users and to support researchers
+and tool developers of downstream tasks.
+
+Nguyen et al. (2010) and Le et al. (2013) built
+Vietnamese NLP pipelines by wrapping existing word segmenters and POS taggers including:
+JVnSegmenter (Nguyen et al., 2006), vnTokenizer
+(Le et al., 2008), JVnTagger (Nguyen et al., 2010)
+and vnTagger (Le-Hong et al., 2010). However,
+these word segmenters and POS taggers are no
+longer considered SOTA models for Vietnamese
+(Nguyen and Le, 2016; Nguyen et al., 2016b).
+
+
+
+Figure 1: In pipeline architecture of VnCoreNLP, annotations are performed on an `Annotation` object.
+
+
+Pham et al. (2017) built the NNVLP toolkit for
+Vietnamese sequence labeling tasks by applying a
+BiLSTM-CNN-CRF model (Ma and Hovy, 2016).
+However, Pham et al. (2017) did not make a comparison to SOTA traditional feature-based models. In addition, NNVLP is slow with a processing
+speed at about 300 words per second, which is not
+practical for real-world application such as dealing
+with large-scale data.
+In this paper, we present a Java NLP toolkit for
+Vietnamese, namely VnCoreNLP, which aims to
+facilitate Vietnamese NLP research by providing
+rich linguistic annotations through key NLP components of word segmentation, POS tagging, NER
+and dependency parsing. Figure 1 describes the
+overall system architecture. The following items
+highlight typical characteristics of VnCoreNLP:
+
+_•_ **Easy-to-use**  - All VnCoreNLP components
+are wrapped into a single .jar file, so users
+do not have to install external dependencies.
+Users can run processing pipelines from either the command-line or the Java API.
+
+_•_ **Fast**  - VnCoreNLP is fast, so it can be used
+for dealing with large-scale data. Also it benefits users suffering from limited computation resources (e.g. users from Vietnam).
+
+_•_ **Accurate**  - VnCoreNLP components obtain
+higher results than all previous published results on the same benchmark datasets.
+
+
+
+56
+
+
+_Proceedings of NAACL-HLT 2018:_ _Demonstrations_, pages 56–60
+New Orleans, Louisiana, June 2 - 4, 2018. _⃝_ c 2018 Association for Computational Linguistics
+
+
+**2** **Basic usages**
+
+Our design goal is to make VnCoreNLP simple to
+setup and run from either the command-line or the
+Java API. Performing linguistic annotations for a
+given file can be done by using a simple command
+as in Figure 2.
+
+```
+$ java -Xmx2g -jar VnCoreNLP.jar -fin
+input.txt -fout output.txt
+
+```
+
+Figure 2: Minimal command to run VnCoreNLP.
+
+
+Suppose that the file `input.txt` in Figure
+2 contains a sentence “Ông Nguyễn Khắc Chúc
+đang làm việc tại Đại học Quốc gia Hà Nội.”
+(MrÔng Nguyen Khac Chuc isđang workinglàm_việc
+attại Vietnam Nationalquốc_gia Universityđại_học
+HanoiHà_Nội). Table 1 shows the output for this
+sentence in plain text form.
+
+
+1 Ông Nc O 4 sub
+2 Nguyễn_Khắc_Chúc Np B-PER 1 nmod
+3 đang R O 4 adv
+4 làm_việc V O 0 root
+5 tại E O 4 loc
+6 Đại_học N B-ORG 5 pob
+7 Quốc_gia N I-ORG 6 nmod
+8 Hà_Nội Np I-ORG 6 nmod
+9 . CH O 4 punct
+
+Table 1: The output in file `output.txt` for the sentence ‘Ông Nguyễn Khắc Chúc đang làm việc tại Đại
+học Quốc gia Hà Nội.” from file `input.txt` in Figure 2. The output is in a 6-column format representing
+word index, word form, POS tag, NER label, head index of the current word, and dependency relation type.
+
+
+Similarly, we can also get the same output by
+using the API as easy as in Listing 1.
+```
+VnCoreNLP pipeline = new VnCoreNLP() ;
+Annotation annotation = new Annotation("
+  Ông Nguyễn Khắc Chúc đang làm việc
+  tại Đại học Quốc gia Hà Nội.");
+pipeline.annotate(annotation);
+String annotatedStr = annotation.
+  toString();
+```
+
+Listing 1: Minimal code for an analysis pipeline.
+
+
+In addition, Listing 2 provides a more realistic
+and complete example code, presenting key components of the toolkit. Here an annotation pipeline
+can be used for any text rather than just a single
+sentence, e.g. for a paragraph or entire news story.
+
+
+**3** **Components**
+
+This section briefly describes each component of
+VnCoreNLP. Note that our goal is not to develop
+
+
+```
+import vn.pipeline.*;
+import java.io.*;
+public class VnCoreNLPExample {
+ public static void main(String[] args)
+   throws IOException {
+ // "wseg", "pos", "ner", and "parse"
+   refer to as word segmentation, POS
+    tagging, NER and dependency
+   parsing, respectively.
+ String[] annotators = {"wseg", "pos",
+   "ner", "parse"};
+ VnCoreNLP pipeline = new VnCoreNLP(
+   annotators);
+ // Mr Nguyen Khac Chuc is working at
+   Vietnam National University, Hanoi
+   . Mrs Lan, Mr Chuc’s wife, is also
+    working at this university.
+ String str = "Ông Nguyễn Khắc Chúc
+   đang làm việc tại Đại học Quốc gia
+   Hà Nội. Bà Lan, vợông Chúc, cũng
+   làm việc tại đây.";
+ Annotation annotation = new Annotation
+   (str);
+ pipeline.annotate(annotation);
+ PrintStream outputPrinter = new
+   PrintStream("output.txt");
+ pipeline.printToFile(annotation,
+   outputPrinter);
+ // Users can get a single sentence to
+   analyze individually
+ Sentence firstSentence = annotation.
+   getSentences().get(0);
+ }
+}
+```
+
+Listing 2: A simple and complete example code.
+
+
+new approach or model for each component task.
+Here we focus on incorporating existing models into a single pipeline. In particular, except a
+new model we develop for the language-dependent
+component of word segmentation, we apply traditional feature-based models which obtain SOTA
+results for English POS tagging, NER and dependency parsing to Vietnamese. The reason is based
+on a well-established belief in the literature that
+for a less-resourced language such as Vietnamese,
+we should consider using feature-based models to
+obtain fast and accurate performances, rather than
+using neural network-based models (King, 2015).
+
+_•_ **wseg**  - Unlike English where white space
+is a strong indicator of word boundaries,
+when written in Vietnamese white space is
+also used to separate syllables that constitute
+words. So word segmentation is referred to
+as the key first step in Vietnamese NLP. We
+have proposed a transformation rule-based
+learning model for Vietnamese word segmentation, which obtains better segmentation accuracy and speed than all previous word segmenters. See details in Nguyen et al. (2018).
+
+
+
+57
+
+
+_•_ **pos**  - To label words with their POS tag,
+we apply MarMoT which is a generic CRF
+framework and a SOTA POS and morphological tagger (Mueller et al., 2013). [1]
+
+_•_ **ner**  - To recognize named entities, we apply
+a dynamic feature induction model that automatically optimizes feature combinations
+(Choi, 2016). [2]
+
+_•_ **parse**  - To perform dependency parsing,
+we apply the greedy version of a transitionbased parsing model with selectional branching (Choi et al., 2015). [3]
+
+
+**4** **Evaluation**
+
+
+We detail experimental results of the word segmentation ( **wseg** ) and POS tagging ( **pos** ) components of VnCoreNLP in Nguyen et al. (2018) and
+Nguyen et al. (2017b), respectively. In particular,
+our word segmentation component gets the highest results in terms of both segmentation F1 score
+at 97.90% and speed at 62K words per second. [4]
+
+Our POS tagging component also obtains the highest accuracy to date at 95.88% with a fast tagging
+speed at 25K words per second, and outperforms
+BiLSTM-CRF-based models. Following subsections present evaluations for the NER ( **ner** ) and
+dependency parsing ( **parse** ) components.
+
+
+**4.1** **Named entity recognition**
+
+We make a comparison between SOTA featurebased and neural network-based models, which, to
+the best of our knowledge, has not been done in
+any prior work on Vietnamese NER.
+
+
+**Dataset:** The NER shared task at the 2016
+VLSP workshop provides a set of 16,861 manually annotated sentences for training and development, and a set of 2,831 manually annotated sentences for test, with four NER labels PER, LOC,
+ORG and MISC. Note that in both datasets, words
+are also supplied with gold POS tags. In addition,
+each word representing a full personal name are
+separated into syllables that constitute the word.
+So this annotation scheme results in an unrealistic scenario for a pipeline evaluation because: ( **i** )
+
+
+1 `http://cistern.cis.lmu.de/marmot/`
+2 `https://emorynlp.github.io/nlp4j/`
+```
+components/named-entity-recognition.html
+
+```
+
+
+3 `https://emorynlp.github.io/nlp4j/`
+```
+components/dependency-parsing.html
+
+```
+
+
+4All speeds reported in this paper are computed on a personal computer of Intel Core i7 2.2 GHz.
+
+
+
+gold POS tags are not available in a real-world application, and ( **ii** ) in the standard annotation (and
+benchmark datasets) for Vietnamese word segmentation and POS tagging (Nguyen et al., 2009),
+each full name is referred to as a word token (i.e.,
+all word segmenters have been trained to output a
+full name as a word and all POS taggers have been
+trained to assign a label to the entire full-name).
+For a more realistic scenario, we merge those
+contiguous syllables constituting a full name to
+form a word. [5] Then we replace the gold POS tags
+by automatic tags predicted by our POS tagging
+component. From the set of 16,861 sentences, we
+sample 2,000 sentences for development and using
+the remaining 14,861 sentences for training.
+
+**Models:** We make an empirical comparison between the VnCoreNLP’s NER component and the
+following neural network-based models:
+
+_•_ BiLSTM-CRF (Huang et al., 2015) is a sequence labeling model which extends the
+BiLSTM model with a CRF layer.
+
+_•_ BiLSTM-CRF + CNN-char, i.e. BiLSTMCNN-CRF, is an extension of BiLSTM-CRF,
+using CNN to derive character-based word
+representations (Ma and Hovy, 2016).
+
+_•_ BiLSTM-CRF + LSTM-char is an extension of BiLSTM-CRF, using BiLSTM to derive the character-based word representations
+(Lample et al., 2016).
+
+_•_ BiLSTM-CRF+POS is another extension to
+BiLSTM-CRF, incorporating embeddings of
+automatically predicted POS tags (Reimers
+and Gurevych, 2017).
+
+We use a well-known implementation which
+is optimized for performance of all BiLSTMCRF-based models from Reimers and Gurevych
+(2017). [6] We then follow Nguyen et al. (2017b,
+Section 3.4) to perform hyper-parameter tuning. [7]
+
+**Main** **results:** Table 2 presents F1 score and
+speed of each model on the test set, where VnCoreNLP obtains the highest score at 88.55% with
+a fast speed at 18K words per second. In particular, VnCoreNLP obtains 10 times faster speed than
+
+5Based on the gold label PER, contiguous syllables such
+as “Nguyễn/B-PER”, “Khắc/I-PER” and “Chúc/I-PER” are
+merged to form a word as “Nguyễn_Khắc_Chúc/B-PER.”
+6 `https://github.com/UKPLab/`
+```
+emnlp2017-bilstm-cnn-crf
+```
+
+7We employ pre-trained Vietnamese word vectors from
+`https://github.com/sonvx/word2vecVN` .
+
+
+
+58
+
+
+|Model|F1|Speed|
+|---|---|---|
+|VnCoreNLP<br>BiLSTM-CRF<br>+ CNN-char<br>+ LSTM-char<br>BiLSTM-CRF+POS<br>+ CNN-char<br>+ LSTM-char|**88.55**<br>86.48<br>88.28<br>87.71<br>86.12<br>88.06<br>87.43|**18K**<br>2.8K<br>1.8K<br>1.3K<br>_<br>_<br>_|
+
+
+Table 2: F1 scores (in %) on the test set w.r.t. gold wordsegmentation. “ **Speed** ” denotes the processing speed of
+the number of words per second (for VnCoreNLP, we
+include the time POS tagging takes in the speed).
+
+
+the second most accurate model BiLSTM-CRF +
+CNN-char.
+It is initially surprising that for such an isolated language as Vietnamese where all words
+are not inflected, using character-based representations helps producing 1+% improvements to the
+BiLSTM-CRF model. We find that the improvements to BiLSTM-CRF are mostly accounted for
+by the PER label. The reason turns out to be simple: about 50% of named entities are labeled with
+tag PER, so character-based representations are
+in fact able to capture common family, middle
+or given name syllables in ‘unknown’ full-name
+words. Furthermore, we also find that BiLSTMCRF-based models do not benefit from additional
+predicted POS tags. It is probably because BiLSTM can take word order into account, while without word inflection, all grammatical information
+in Vietnamese is conveyed through its fixed word
+order, thus explicit predicted POS tags with noisy
+grammatical information are not helpful.
+
+
+**4.2** **Dependency parsing**
+
+**Experimental** **setup:** We use the Vietnamese
+dependency treebank VnDT (Nguyen et al., 2014)
+consisting of 10,200 sentences in our experiments.
+Following Nguyen et al. (2016a), we use the last
+1020 sentences of VnDT for test while the remaining sentences are used for training. Evaluation
+metrics are the labeled attachment score (LAS)
+and unlabeled attachment score (UAS).
+
+
+**Main results:** Table 3 compares the dependency
+parsing results of VnCoreNLP with results reported in prior work, using the same experimental setup. The first six rows present the scores with
+gold POS tags. The next two rows show scores of
+VnCoreNLP with automatic POS tags which are
+produced by our POS tagging component. The last
+
+
+
+|Model|Col2|LAS|UAS|Speed|
+|---|---|---|---|---|
+|Gold POS|VnCoreNLP<br>VnCoreNLP–NER<br>BIST-bmstparser<br>BIST-barchybrid<br>MSTParser<br>MaltParser|**73.39**<br>73.21<br>73.17<br>72.53<br>70.29<br>69.10|79.02<br>78.91<br>**79.39**<br>79.33<br>76.47<br>74.91|_<br>_<br>_<br>_<br>_<br>_|
+|Auto POS|VnCoreNLP<br>VnCoreNLP–NER<br>jPTDP|**70.23**<br>70.10<br>69.49|76.93<br>76.85<br>**77.68**|8K<br>**9K**<br>700|
+
+
+Table 3: LAS and UAS scores (in %) computed on all
+tokens (i.e. including punctuation) on the test set w.r.t.
+gold word-segmentation. “ **Speed** ” is defined as in Table 2. The subscript “–NER” denotes the model without
+using automatically predicted NER labels as features.
+The results of the MSTParser (McDonald et al., 2005),
+MaltParser (Nivre et al., 2007), and BiLSTM-based
+parsing models BIST-bmstparser and BIST-barchybrid
+(Kiperwasser and Goldberg, 2016) are reported in
+Nguyen et al. (2016a). The result of the jPTDP model
+for Vietnamese is mentioned in Nguyen et al. (2017b).
+
+
+row presents scores of the joint POS tagging and
+dependency parsing model jPTDP (Nguyen et al.,
+2017a). Table 3 shows that compared to previously
+published results, VnCoreNLP produces the highest LAS score. Note that previous results for other
+systems are reported without using additional information of automatically predicted NER labels.
+In this case, the LAS score for VnCoreNLP without automatic NER features (i.e. VnCoreNLP–NER
+in Table 3) is still higher than previous ones. Notably, we also obtain a fast parsing speed at 8K
+words per second.
+
+
+**5** **Conclusion**
+
+
+In this paper, we have presented the VnCoreNLP
+toolkit—an easy-to-use, fast and accurate processing pipeline for Vietnamese NLP. VnCoreNLP
+provides core NLP steps including word segmentation, POS tagging, NER and dependency parsing. Current version of VnCoreNLP has been
+trained without any linguistic optimization, i.e. we
+only employ existing pre-defined features in the
+traditional feature-based models for POS tagging,
+NER and dependency parsing. So future work will
+focus on incorporating Vietnamese linguistic features into these feature-based models.
+VnCoreNLP is released for research and educational purposes, and available at: `https://`
+`github.com/vncorenlp/VnCoreNLP` .
+
+
+
+59
+
+
+**References**
+
+Jinho D. Choi. 2016. Dynamic Feature Induction: The
+Last Gist to the State-of-the-Art. In _Proceedings of_
+_NAACL-HLT_ . pages 271–281.
+
+
+Jinho D. Choi, Joel Tetreault, and Amanda Stent. 2015.
+It Depends: Dependency Parser Comparison Using
+A Web-based Evaluation Tool. In _Proceedings_ _of_
+_ACL-IJCNLP_ . pages 387–396.
+
+
+Zhiheng Huang, Wei Xu, and Kai Yu. 2015. Bidirectional LSTM-CRF models for sequence tagging.
+_arXiv preprint_ arXiv:1508.01991.
+
+
+Benjamin Philip King. 2015. _Practical_ _Natural_ _Lan-_
+_guage_ _Processing_ _for_ _Low-Resource_ _Languages_ .
+Ph.D. thesis, The University of Michigan.
+
+
+Eliyahu Kiperwasser and Yoav Goldberg. 2016. Simple and Accurate Dependency Parsing Using Bidirectional LSTM Feature Representations. _Transac-_
+_tions of the Association for Computational Linguis-_
+_tics_ 4:313–327.
+
+
+Guillaume Lample, Miguel Ballesteros, Sandeep Subramanian, Kazuya Kawakami, and Chris Dyer. 2016.
+Neural Architectures for Named Entity Recognition.
+In _Proceedings of NAACL-HLT_ . pages 260–270.
+
+
+Hong Phuong Le, Thi Minh Huyen Nguyen, Azim
+Roussanaly, and Tuong Vinh Ho. 2008. A hybrid
+approach to word segmentation of Vietnamese texts.
+In _Proceedings of LATA_ . pages 240–249.
+
+
+Ngoc Minh Le, Bich Ngoc Do, Vi Duong Nguyen, and
+Thi Dam Nguyen. 2013. VNLP: An Open Source
+Framework for Vietnamese Natural Language Processing. In _Proceedings of SoICT_ . pages 88–93.
+
+
+Phuong Le-Hong, Azim Roussanaly, Thi Minh Huyen
+Nguyen, and Mathias Rossignol. 2010. An empirical study of maximum entropy approach for part-ofspeech tagging of Vietnamese texts. In _Proceedings_
+_of TALN_ .
+
+
+Xuezhe Ma and Eduard Hovy. 2016. End-to-end Sequence Labeling via Bi-directional LSTM-CNNsCRF. In _Proceedings_ _of_ _ACL_ _(Volume_ _1:_ _Long_ _Pa-_
+_pers)_ . pages 1064–1074.
+
+
+Christopher D. Manning, Mihai Surdeanu, John Bauer,
+Jenny Finkel, Steven J. Bethard, and David McClosky. 2014. The Stanford CoreNLP natural language processing toolkit. In _Proceedings_ _of_ _ACL_
+_2014 System Demonstrations_ . pages 55–60.
+
+
+Ryan McDonald, Koby Crammer, and Fernando
+Pereira. 2005. Online Large-margin Training of Dependency Parsers. In _Proceedings_ _of_ _ACL_ . pages
+91–98.
+
+
+Thomas Mueller, Helmut Schmid, and Hinrich
+Sch¨utze. 2013. Efficient Higher-Order CRFs for
+Morphological Tagging. In _Proceedings of EMNLP_ .
+pages 322–332.
+
+
+
+Cam-Tu Nguyen, Trung-Kien Nguyen, et al. 2006.
+Vietnamese Word Segmentation with CRFs and
+SVMs: An Investigation. In _Proceedings_ _of_
+_PACLIC_ . pages 215–222.
+
+
+Cam-Tu Nguyen, Xuan-Hieu Phan, and ThuTrang Nguyen. 2010. JVnTextPro: A Javabased Vietnamese Text Processing Tool.
+`http://jvntextpro.sourceforge.net/` .
+
+
+Dat Quoc Nguyen, Mark Dras, and Mark Johnson.
+2016a. An empirical study for Vietnamese dependency parsing. In _Proceedings of ALTA_ . pages 143–
+149.
+
+
+Dat Quoc Nguyen, Mark Dras, and Mark Johnson.
+2017a. A Novel Neural Network Model for Joint
+POS Tagging and Graph-based Dependency Parsing. In _Proceedings_ _of_ _the_ _CoNLL_ _2017_ _Shared_
+_Task_ . pages 134–142.
+
+
+Dat Quoc Nguyen, Dai Quoc Nguyen, Son Bao Pham,
+Phuong-Thai Nguyen, and Minh Le Nguyen. 2014.
+From Treebank Conversion to Automatic Dependency Parsing for Vietnamese. In _Proceedings_ _of_
+_NLDB_ . pages 196–207.
+
+
+Dat Quoc Nguyen, Dai Quoc Nguyen, Thanh Vu, Mark
+Dras, and Mark Johnson. 2018. A Fast and Accurate Vietnamese Word Segmenter. In _Proceedings_
+_of LREC_ . page to appear.
+
+
+Dat Quoc Nguyen, Thanh Vu, Dai Quoc Nguyen, Mark
+Dras, and Mark Johnson. 2017b. From Word Segmentation to POS Tagging for Vietnamese. In _Pro-_
+_ceedings of ALTA_ . pages 108–113.
+
+
+Phuong Thai Nguyen, Xuan Luong Vu, et al. 2009.
+Building a Large Syntactically-Annotated Corpus of
+Vietnamese. In _Proceedings_ _of_ _LAW_ . pages 182–
+185.
+
+
+Tuan-Phong Nguyen and Anh-Cuong Le. 2016. A Hybrid Approach to Vietnamese Word Segmentation.
+In _Proceedings of RIVF_ . pages 114–119.
+
+
+Tuan Phong Nguyen, Quoc Tuan Truong, Xuan Nam
+Nguyen, and Anh Cuong Le. 2016b. An Experimental Investigation of Part-Of-Speech Taggers for Vietnamese. _VNU_ _Journal_ _of_ _Science:_ _Computer_ _Sci-_
+_ence and Communication Engineering_ 32(3):11–25.
+
+
+Joakim Nivre, Johan Hall, et al. 2007. MaltParser:
+A language-independent system for data-driven dependency parsing. _Natural_ _Language_ _Engineering_
+13(2):95–135.
+
+
+Thai-Hoang Pham, Xuan-Khoai Pham, Tuan-Anh
+Nguyen, and Phuong Le-Hong. 2017. NNVLP: A
+Neural Network-Based Vietnamese Language Processing Toolkit. In _Proceedings of the IJCNLP 2017_
+_System Demonstrations_ . pages 37–40.
+
+
+Nils Reimers and Iryna Gurevych. 2017. Reporting Score Distributions Makes a Difference: Performance Study of LSTM-networks for Sequence Tagging. In _Proceedings of EMNLP_ . pages 338–348.
+
+
+
+60
+
+

references/2021.naacl.shi/paper.md

@@ -0,0 +1,1411 @@
+---
+title: "Diversity-Aware Batch Active Learning for Dependency Parsing"
+authors:
+  - "Tianze Shi"
+  - "Adrian Benton"
+  - "Igor Malioutov"
+  - "Ozan Irsoy"
+year: 2021
+venue: "NAACL 2021"
+url: "https://aclanthology.org/2021.naacl-main.207/"
+---
+
+# **Diversity-Aware Batch Active Learning for Dependency Parsing**
+
+
+
+**Tianze Shi** _[∗]_
+
+Cornell University
+tianze@cs.cornell.edu
+
+
+**Igor Malioutov**
+Bloomberg L.P.
+imalioutov@bloomberg.net
+
+
+**Abstract**
+
+
+While the predictive performance of modern
+statistical dependency parsers relies heavily on
+the availability of expensive expert-annotated
+treebank data, not all annotations contribute
+equally to the training of the parsers. In this
+paper, we attempt to reduce the number of labeled examples needed to train a strong dependency parser using batch active learning
+(AL). In particular, we investigate whether enforcing diversity in the sampled batches, using determinantal point processes (DPPs), can
+improve over their diversity-agnostic counterparts. Simulation experiments on an English
+newswire corpus show that selecting diverse
+batches with DPPs is superior to strong selection strategies that do not enforce batch diversity, especially during the initial stages of the
+learning process. Additionally, our diversityaware strategy is robust under a corpus duplication setting, where diversity-agnostic sampling
+strategies exhibit significant degradation.
+
+
+**1** **Introduction**
+
+
+Though critical to parser training, data annotations
+for dependency parsing are both expensive and
+time-consuming to obtain. Syntactic analysis requires linguistic expertise and even after extensive
+training, data annotation can still be burdensome.
+The Penn Treebank project (Marcus et al., 1993)
+reports that after two months of training, the annotators average 750 tokens per hour on the bracketing
+task; the Prague Dependency Treebank (Bohmov¨ a´
+et al., 2003) cost over $600 _,_ 000 and required 5
+years to annotate roughly 90 _,_ 000 sentences (over
+$5 per sentence). These high annotation costs
+present a significant challenge to developing accurate dependency parsers for under-resourced languages and domains.
+Active learning (AL; Settles, 2009) is a promising technique to reduce the annotation effort required to train a strong dependency parser by intel
+
+_∗_ Work done during an internship at Bloomberg L.P.
+
+
+
+**Adrian Benton**
+Bloomberg L.P.
+abenton10@bloomberg.net
+
+
+**Ozan Irsoy** **[˙]**
+Bloomberg L.P.
+oirsoy@bloomberg.net
+
+
+ligently selecting samples to annotate such that the
+return of each annotator hour is as high as possible.
+Popular selection strategies, such as uncertainty
+sampling, associate each instance with a _quality_
+measure based on the uncertainty or confidence
+level of the current parser, and higher-quality instances are selected for annotation.
+We focus on _batch_ _mode_ AL, since it is generally more efficient for annotators to label in
+bulk. While early work in AL for parsing (Tang
+et al., 2002; Hwa, 2000, 2004) cautions against
+using individually-computed quality measures in
+the batch setting, more recent work demonstrates
+empirical success (e.g., Li et al., 2016) without explicitly handling intra-batch _diversity_ . In this paper,
+we explore whether a diversity-aware approach can
+improve the state of the art in AL for dependency
+parsing. Specifically, we consider samples drawn
+from determinantal point processes (DPPs) as a
+query strategy to select batches of high-quality, yet
+dissimilar instances (Kulesza and Taskar, 2012).
+In this paper, we (1) propose a diversity-aware
+batch AL query strategy for dependency parsing
+compatible with existing selection strategies, (2)
+empirically study three AL strategies with and without diversity factors, and (3) find that diversityaware selection strategies are superior to their
+diversity-agnostic counterparts, especially during
+the early stages of the learning process, in simulation experiments on an English newswire corpus.
+This is critical in low-budget AL settings, which
+we further confirm in a corpus duplication setting. [1]
+
+
+**2** **Active Learning for Dependency**
+**Parsing**
+
+
+**2.1** **Dependency Parsing**
+
+
+Dependency parsing (Kubler et al.¨, 2008) aims to
+find the syntactic dependency structure, _y_, given a
+length- _n_ input sentence _x_ = _x_ 1 _, x_ 2 _, . . ., xn_, where
+
+
+[1Our code is publicly available at https://github.com/](https://github.com/tzshi/dpp-al-parsing-naacl21)
+[tzshi/dpp-al-parsing-naacl21.](https://github.com/tzshi/dpp-al-parsing-naacl21)
+
+
+
+2616
+
+
+_Proceedings of the 2021 Conference of the North American Chapter of the_
+_Association for Computational Linguistics:_ _Human Language Technologies_, pages 2616–2626
+June 6–11, 2021. ©2021 Association for Computational Linguistics
+
+
+_y_ is a set of _n_ arcs over the tokens and the dummy
+root symbol _x_ 0, and each arc ( _h, m_ ) _∈_ _y_ specifies
+the head, _h_, and modifier word, _m_ . [2] In this work,
+we adopt the conceptually-simple edge-factored
+deep biaffine dependency parser (Dozat and Manning, 2017), which is competitive with the state
+of the art in terms of accuracy, The parser assigns a locally-normalized attachment probability
+_P_ att(head( _m_ ) = _h_ _|_ _x_ ) to each attachment candidate pair ( _h, m_ ) based on a biaffine scoring function. Refer to Appendix A for architecture details.
+We define the score of the candidate parse tree
+_s_ ( _y_ _|_ _x_ ) as [�] ( _h,m_ ) _∈y_ [log] _[ P]_ [att][(][head][(] _[m]_ [)] [=] _[h]_ _[|]_ _[x]_ [)][.]
+The decoder finds the best scoring _y_ ˆ among all
+valid trees _Y_ ( _x_ ): _y_ ˆ = arg max _y∈Y_ ( _x_ ) _s_ ( _y_ _| x_ ).
+
+
+**2.2** **Active Learning (AL)**
+
+
+We consider the pool-based batch AL scenario
+where we assume a large collection of unlabeled
+instances _U_ from which we sample a small subset
+at a time to annotate after each round to form an
+expanding labeled training set _L_ (Lewis and Gale,
+1994). We use the superscript _i_ to denote the pool
+of instances _U_ _[i]_ and _L_ _[i]_ after the _i_ -th round. _L_ [0]
+
+is a small set of seed labeled instances to initiate
+the process. Each iteration starts with training a
+model _M_ _[i]_ based on _L_ _[i]_ . Next, all unlabeled data
+instances in _U_ _[i]_ are parsed by _M_ _[i]_ and we select
+a batch _U_ _[′]_ to annotate based on some criterion
+_U_ _[′]_ = _C_ ( _M_ _[i]_ _, U_ _[i]_ ). The resulting labeled subset _L_ _[′]_
+
+is added to _L_ _[i]_ [+1] = _L_ _[i]_ [ �] _L_ _[′]_ and _U_ _[i]_ [+1] = _U_ _[i]_ _−U_ _[′]_ .
+The definition of the selection criterion _C_ is critical. A typical strategy associates each unlabeled
+instance _Ui_ with a quality measure _qi_ based on, for
+example, the model uncertainty level when parsing
+_Ui_ . A _diversity-agnostic_ criterion sorts all unlabeled instances by their quality measures and takes
+the top- _k_ as _U_ _[′]_ for a budget _k_ .
+
+
+**2.3** **Quality Measures**
+
+
+We consider three commonly-used quality measures adapted to the task of dependency parsing,
+including uncertainty sampling, Bayesian active
+learning, and a representativeness-based strategy.
+
+
+**Average** **Marginal** **Probability** **(AMP)** measures parser uncertainty (Li et al., 2016):
+
+
+
+AMP = 1 _−_ [1]
+
+
+
+( _h,m_ [ˆ] ) _∈y_ ˆ _[P]_ [mar][(][head][(] _[m]_ [) = ˆ] _[h][ |][ x]_ [)] _[,]_
+
+
+
+
+[1] 
+_n_
+
+
+
+where _P_ mar is the marginal attachment probability
+
+
+2For clarity, here we describe unlabeled parsing. In our
+experiments, we train labeled dependency parsers, which additionally predict a dependency relation label _l_ for each arc.
+
+
+
+_P_ mar(head( _m_ ) = _h | x_ ) = [�] ( _h,m_ ) _∈y_ _[P]_ [(] _[y]_ _[|][ x]_ [)] _[,]_
+
+
+exp( _s_ ( _y|x_ ))
+and _P_ ( _y_ _| x_ ) = - [The marginal]
+_y_ _[′]_ _∈Y_ ( _x_ ) [exp(] _[s]_ [(] _[y][′][|][x]_ [))] [.]
+
+probabilities can be derived efficiently using Kirchhoff’s theorem (Tutte, 1984; Koo et al., 2007).
+
+
+**Bayesian** **Active** **Learning** **by** **Disagreement**
+**(BALD)** measures the mutual information between the model parameters and the predictions.
+We adopt the Monte Carlo dropout-based variant
+(Gal et al., 2017; Siddhant and Lipton, 2018) and
+measure the disagreement among predictions from
+a neural model with _K_ different dropout masks,
+which has been applied to active learning in NLP.
+We adapt BALD to dependency parsing by aggregating disagreement at a token level:
+
+BALD = 1 _−_ _n_ [1]   - _m_ count(mode _K_ ( _h_ [1] _m_ _[,...,h][K]_ _m_ [))] _,_
+
+where _h_ _[k]_ _m_ [denotes that][ (] _[h]_ _m_ _[k]_ _[, m]_ [)][ appears in the pre-]
+diction given by the _k_ -th model.
+
+
+**Information** **Density** **(ID)** mitigates the tendency of uncertainty sampling to favor outliers by
+weighing examples by how _representative_ they are
+of the entire dataset (Settles and Craven, 2008):
+
+ID = AMP _×_   - _|U|_ 1   - _x_ _[′]_ _∈U_ [sim][cos][(] _[x, x][′]_ [)]   - _,_
+
+
+where cosine similarity is computed from the averaged contextualized features (§3.2).
+
+
+**2.4** **Learning from Partial Annotations**
+
+
+We follow Li et al. (2016) and select tokens to annotate their heads instead of annotating full sentences.
+We first pick the most informative sentences and
+then choose _p_ % tokens from them based on the
+token-level versions of the quality measures (e.g.,
+marginal probability instead of AMP).
+
+
+**3** **Selecting Diverse Samples**
+
+
+Near-duplicate examples are common in real-world
+data (Broder et al., 1997; Manku et al., 2007), but
+they provide overlapping utility to model training.
+In the extreme case, with a diversity-agnostic strategy for active learning, identical examples will be
+selected/excluded at the same time (Hwa, 2004).
+To address this issue and to best utilize the annotation budget, it is important to consider diversity.
+We adapt Bıyık et al. (2019) to explicitly model diversity using determinantal point processes (DPPs).
+
+
+**3.1** **Determinantal Point Processes**
+
+
+A DPP defines a probability distribution over subsets of some ground set of elements (Kulesza,
+2012). In AL, the ground set is the unlabeled
+
+
+
+2617
+
+
+pool _U_ and a subset corresponds to a batch of
+instances _U_ _[′]_ drawn from _U_ . DPPs provide an
+explicit mechanism to ensure high-quality yet diverse sample selection by modeling both the quality measures and the similarities among examples. We adopt the _L_ -ensemble representation
+of DPPs using the quality-diversity decomposition (Kulesza and Taskar, 2012) and parameterize the matrix _L_ as _Lij_ = _qiφiφ_ _[T]_ _j_ _[q][j]_ [,] [where] [each]
+_qi_ _∈_ R is the quality measure for _Ui_ and each
+_φi_ _∈_ R [1] _[×][d]_ is a _d_ -dimensional vector representation of _Ui_, which we refer to as _Ui_ ’s _diversity fea-_
+_tures_ . [3] The probability of selecting a batch _B_ is
+given by _P_ ( _B_ _⊆U_ ) _∝_ det( _LB_ ), where det( _·_ ) calculates the determinant and _LB_ is the submatrix of
+_L_ indexed by elements in _B_ .
+DPPs place high probability on diverse subsets
+of high-quality items. Intuitively, the determinant
+of _LB_ corresponds to the volume spanned by the
+set of vectors _{qiφi_ _|_ _i_ _∈_ _B}_, and subsets with
+larger _q_ values and orthogonal _φ_ vectors span larger
+volumes than those with smaller _q_ values or similar
+_φ_ vectors. We follow Kulesza (2012) and adapt
+their greedy algorithm for finding the approximate
+mode arg max _B P_ ( _B_ _⊆U_ ). This algorithm is
+reproduced in Algorithm E1 in the appendix.
+
+
+**3.2** **Diversity Features**
+
+
+We consider two possibilities for the diversity features _φ_ . Each feature vector is unit-normalized.
+
+
+**Averaged Contextualized Features** are defined
+as _n_ [1] - _i_ **[x]** _[i]_ [, where] **[ x]** _[i]_ [ is a contextualized vector of]
+
+_xi_ from the feature extractor used by the parser. By
+this definition, we consider the instances to be similar to each other when the neural feature extractor
+returns similar features such that the parser is likely
+to predict similar structures for these instances.
+
+
+**Predicted** **Subgraph** **Counts** explicitly represent the predicted tree structure. To balance richness and sparsity, we count the labeled but unlexicalized subgraph formed by the grandparent, the
+parent and the token itself. Specifically, for each
+token _m_, we can extract a subgraph denoted by
+( _r_ 1 _, r_ 2), assuming the predicted dependency relation between its grandparent _g_ and its parent _h_ is
+_r_ 1, and the relation between _h_ and _m_ is _r_ 2. The
+parse tree for a length- _n_ sentence contains _n_ such
+subgraphs. We apply tf-idf weighting to discount
+
+
+3Although certain applications of DPPs may learn _q_ and _φ_
+representations from supervision, we define _q_ and _φ a priori_,
+since acquiring supervision in AL is, by definition, expensive.
+
+
+
+Batch 5 10
+
+
+Strategy w/o DPP w/ DPP w/o DPP w/ DPP
+
+
+Random 85 _._ 68 _±._ 26 **86** _._ **61** _±._ 28 87 _._ 84 _±._ 26 **88** _._ **55** _±._ 23
+AMP 85 _._ 98 _±._ 22 **86** _._ **77** _±._ 43 88 _._ 80 _±._ 18 **89** _._ **23** _±._ 29
+BALD 86 _._ 24 _±._ 40 **86** _._ **86** _±._ 31 88 _._ 66 _±._ 36 **89** _._ **03** _±._ 10
+ID **86** _._ **68** _±._ 26 86 _._ 56 _±._ 24 88 _._ 96 _±._ 20 **89** _._ **06** _±._ 16
+
+
+Table 1: LAS after 5 and 10 rounds of annotation for strategies
+with and without modeling diversity through DPP.
+
+
+the influence from frequent subgraphs.
+
+
+**4** **Experiments and Results**
+
+
+**Dataset** We use the Revised English News Text
+Treebank [4] (Bies et al., 2015) converted to Universal Dependencies 2.0 using the conversion tool included in Stanford Parser (Manning et al., 2014)
+version 4.0.0. We use sections 02-21 for training,
+22 for development and 23 for test.
+
+
+**Setting** We perform experiments by simulating
+the annotation process using treebank data. We
+sample 128 sentences uniformly for the initial labeled pool and each following round selects 500
+tokens for partial annotation. We run each setting
+five times using different random initializations and
+report the means and standard deviations of the labeled attachment scores (LAS). Appendix B has
+unlabeled attachment score (UAS) results.
+
+
+**Baselines** While we construct our own baselines for self-contained comparisons, the diversityagnostic AMP (w/o DPP) largely replicates the
+state-of-the-art selection strategy of Li et al. (2016).
+
+
+**Implementation** We finetune a pretrained multilingual XLM-RoBERTa base model (Conneau
+et al., 2020) as our feature extractor. [5] See Appendix E for implementation details.
+
+
+**Main Results** Table 1 compares LAS after 5 and
+10 rounds of annotation. Our dependency parser
+reaches 95 _._ 64 UAS and 94 _._ 06 LAS, when trained
+with the full dataset (more than one million tokens).
+Training data collected from 30 annotation rounds
+( _≈_ 17 _,_ 500 tokens) correspond to roughly 2% of
+the full dataset, but already support an LAS of up
+to 92 through AL. We find that diversity-aware
+strategies generally improve over their diversityagnostic counterparts. Even for a random selection
+strategy, ensuring diversity with a DPP is superior
+
+
+[4https://catalog.ldc.upenn.edu/LDC2015T13](https://catalog.ldc.upenn.edu/LDC2015T13)
+5To construct the averaged contextualized features, we
+also use the fine-tuned feature extractor. In our preliminary
+experiments, we have tried freezing the feature extractors, but
+this variant did not perform as well.
+
+
+
+2618
+
+
+92
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+
+
+0 _._ 4
+
+
+0 _._ 2
+
+
+
+
+
+82
+
+
+Annotation round
+
+
+Figure 1: Learning curves for our DPP-based diversity-aware
+selection strategies, comparing predicted subgraph counts
+versus averaged contextualized features as diversity features.
+Both use AMP as their quality measures.
+
+
+to simple random selection. With AMP and BALD,
+our diversity-aware strategy sees a larger improvement earlier in the learning process. ID models
+representativeness of instances, and our diversityaware strategy adds less utility compared with other
+quality measures, although we do notice a large improvement after the first annotation round for ID:
+82 _._ 40 _±._ 48 vs. 83 _._ 36 _±._ 54 (w/ DPP) – a similar trend
+to AMP and BALD, but at an earlier stage of AL.
+
+
+**Experiments** **with** **Different** **Diversity** **Features**
+Figure 1 compares our two definitions of diversity features, and we find that predicted subgraph
+counts provide stronger performance than that of
+averaged contextualized features. We hypothesize
+this is due to the fact that the subgraph counts represent structures more explicitly, thus they are more
+useful in maintaining structural diversity in AL.
+
+
+**Intra-Batch** **Diversity** To quantify intra-batch
+diversity among the set of sentences _B_ picked by
+the selection strategies, we adapt the measures used
+by Chen et al. (2018) and define intra-batch average
+distance (IBAD) and intra-batch minimal distance
+(IBMD) as follows:
+
+
+IBAD = mean
+_i,j∈B,i_ = _j_ [(1] _[ −]_ [sim][cos][(] _[i, j]_ [))] _[,]_
+
+
+IBMD = mean min
+_i∈B_ _j∈B,i_ = _j_ [(1] _[ −]_ [sim][cos][(] _[i, j]_ [))] _[.]_
+
+
+A higher value on these measures indicates better
+intra-batch diversity. Figure 2 compares diversityagnostic and diversity-aware sampling strategies
+using the two different diversity features. We confirm that DPPs indeed promote diverse samples
+in the selected batches, while intra-batch diversity
+naturally increases even for the diversity-agnostic
+strategies. Additionally, we observe that the benefits of DPPs are more prominent when using pre
+
+
+0
+5 10 15 20
+
+
+Annotation round
+
+
+Figure 2: Intra-batch average distance (IBAD) and intrabatch minimal distance (IBMD) measures comparing diversityagnostic and diversity-aware AMP-based sample selection
+strategies. The distances are derived from averaged contextualized features (top) and predicted subgraph counts (bottom).
+A higher value indicates better intra-batch diversity.
+
+
+dicted subgraph counts compared with averaged
+contextualized features. This can help explain the
+relative success of the former diversity features.
+
+
+**Corpus** **Duplication** **Setting** In our qualitative
+analysis (Appendix C), we find that diversityagnostic selection strategies tend to select nearduplicate sentences. To examine this phenomenon
+in isolation, we repeat the training corpus twice and
+observe the effect of diversity-aware strategies. The
+corpus duplication technique has been previously
+used to probe semantic models (Schofield et al.,
+2017). Figure 3 shows learning curves for strategies under the original and corpus duplication settings. As expected, diversity-aware strategies consistently outperform their diversity-agnostic counterparts across both settings, while some diversityagnostic strategies (e.g., AMP) even underperform
+uniform random selection in the duplicated setting.
+
+
+**Interpreting** **the** **Effectiveness** **of** **Diversity-Ag-**
+**nostic** **Models** Figure 4 visualizes the density
+distributions of the top 200 data instances by AMP
+over the diversity feature space reduced to two dimensions through t-SNE (van der Maaten and Hinton, 2008). During the initial stage of active learning, data with the highest quality measures are con
+
+
+0
+5 10 15 20
+
+
+1
+
+
+0 _._ 8
+
+
+0 _._ 6
+
+
+
+0 _._ 4
+
+
+0 _._ 2
+
+
+
+
+
+2619
+
+
+92
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+92
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+92
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+82
+10 20 30
+
+
+
+82
+
+
+
+82
+
+
+
+Figure 3: Learning curves of different sampling strategies based on AMP (left), BALD (middle) and ID (right), comparing
+diversity-aware (w/ DPP) and diversity-agnostic variants using the original and duplicated corpus (dup). The _x_ -axis shows the
+number of rounds for annotation. Random (dup) curves overlap with those of Random and are omitted for readability.
+
+
+
+60
+
+
+40
+
+
+20
+
+
+0
+
+
+20
+
+
+40
+
+
+60
+
+
+
+80
+
+
+70
+
+
+60
+
+
+50
+
+
+40
+
+
+30
+
+
+20
+
+
+10
+
+
+
+Figure 4: t-SNE visualization of the distributions of the 200
+highest-quality unlabeled sentences over the diversity feature
+space after the 1 [st] (left) and the 10 [th] (right) annotation rounds
+using AMP without DPPs. Darker region indicates more data
+points residing in that diversity feature neighborhood. The
+left figure contains a dense region, while the data in the right
+figure are spread out in the feature space.
+
+
+
+
+
+
+
+centrated within a small neighborhood. A diversityagnostic strategy will sample similar examples for
+annotation. After a few rounds of annotation and
+model training, the distribution of high quality examples spreads out, and an AMP selection strategy
+is likely to sample a diverse set of examples without
+explicitly modeling diversity. Our analysis corroborates previous findings (Thompson et al., 1999)
+that small annotation batches are effective early
+in uncertainty sampling, avoiding selecting many
+near-duplicate examples when intra-batch diversity
+is low, but a larger batch size is more efficient later
+in training once intra-batch diversity increases.
+
+
+**5** **Related Work**
+
+
+Modeling diversity in batch-mode AL (Brinker,
+2003) has recently attracted attention in the machine learning community. Kirsch et al. (2019)
+introduce a Bayesian batch-mode selection strategy by estimating the mutual information between
+a set of samples and the model parameters. Ash
+et al. (2020) present a diversity-inducing sampling
+method using gradient embeddings. Most related
+to our work, Bıyık et al. (2019) first apply DPPs
+
+
+
+to batch-mode AL. Building on their approach, we
+flesh out a DPP treatment for AL for a structured
+prediction task, dependency parsing. Previously,
+Shen et al. (2018) consider named entity recognition but they report negative results for a diversityinducing variant of their sampling method.
+Due to the high annotation cost, AL is a popular
+technique for parsing and parse selection (Osborne
+and Baldridge, 2004). Recent advances focus on
+reducing full-sentence annotations to a subset of
+tokens within a sentence (Sassano and Kurohashi,
+2010; Mirroshandel and Nasr, 2011; Majidi and
+Crane, 2013; Flannery and Mori, 2015; Li et al.,
+2016). We show that AL for parsing can further
+benefit from diversity-aware sampling strategies.
+DPPs have previously been successfully applied
+to the tasks of extractive text summarization (Cho
+et al., 2019a,b) and modeling phoneme inventories
+(Cotterell and Eisner, 2017). In this work, we show
+that DPPs also provide a useful framework for understanding and modeling quality and diversity in
+active learning for NLP tasks.
+
+
+**6** **Conclusion**
+
+
+We show that compared with their diversityagnostic counterparts, diversity-aware sampling
+strategies not only lead to higher data efficiency, but
+are also more robust under corpus duplication settings. Our work invites future research into methods, utility and success conditions for modeling
+diversity in active learning for NLP tasks.
+
+
+**Acknowledgements**
+
+
+We thank the anonymous reviewers for their insightful reviews, and Prabhanjan Kambadur, ChenTse Tsai, and Minjie Xu for discussion and comments. Tianze Shi acknowledges support from
+Bloomberg’s Data Science Ph.D. Fellowship.
+
+
+
+2620
+
+
+**References**
+
+
+Jordan T. Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agarwal. 2020.
+[Deep batch active learning by diverse, uncertain gra-](https://openreview.net/forum?id=ryghZJBKPS)
+[dient lower bounds.](https://openreview.net/forum?id=ryghZJBKPS) In _International Conference on_
+_Learning Representations_, Online. OpenReview.net.
+
+
+Ann Bies, Justin Mott, and Colin Warner. 2015. [En-](https://catalog.ldc.upenn.edu/LDC2015T13)
+glish news text [treebank:](https://catalog.ldc.upenn.edu/LDC2015T13) Penn Treebank revised
+[(LDC2015T13).](https://catalog.ldc.upenn.edu/LDC2015T13)
+
+
+Erdem Bıyık, Kenneth Wang, Nima Anari, and
+Dorsa Sadigh. 2019. Batch active learning using [determinantal](https://arxiv.org/pdf/1906.07975.pdf) point processes. _arXiv_ _preprint_
+_arXiv:1906.07975_ .
+
+
+Alena B¨ohmov´a, Jan Hajiˇc, Eva Hajiˇcov´a, and Barbora
+Hladk´a. 2003. [The Prague dependency treebank.](https://doi.org/10.1007/978-94-010-0201-1_7) In
+Anne Abeill´e, editor, _Treebanks:_ _Building_ _and_ _Us-_
+_ing_ _Parsed_ _Corpora_, Text, Speech and Language
+Technology, pages 103–127. Springer Netherlands,
+Dordrecht.
+
+
+Klaus Brinker. 2003. Incorporating [diversity](https://www.aaai.org/Papers/ICML/2003/ICML03-011.pdf) in active learning with [support](https://www.aaai.org/Papers/ICML/2003/ICML03-011.pdf) vector machines. In _Pro-_
+_ceedings_ _of_ _the_ _20th_ _International_ _Conference_ _on_
+_Machine_ _Learning_, pages 59–66, Washington, DC,
+USA. AAAI Press.
+
+
+Andrei Z. Broder, Steven C. Glassman, Mark S. Manasse, and Geoffrey Zweig. 1997. [Syntactic](https://doi.org/10.1016/S0169-7552(97)00031-7) cluster[ing of the Web.](https://doi.org/10.1016/S0169-7552(97)00031-7) _Computer Networks and ISDN Sys-_
+_tems_, 29(8):1157–1166.
+
+
+Laming Chen, Guoxin Zhang, and Hanning Zhou.
+2018. Fast greedy MAP [inference](https://dl.acm.org/doi/10.5555/3327345.3327465) for determinan[tal point process to improve](https://dl.acm.org/doi/10.5555/3327345.3327465) recommendation diver[sity.](https://dl.acm.org/doi/10.5555/3327345.3327465) In _Proceedings of the 32nd International Con-_
+_ference_ _on_ _Neural_ _Information_ _Processing_ _Systems_,
+pages 5627–5638, Montr´eal, Canada.
+
+
+Sangwoo Cho, Logan Lebanoff, Hassan Foroosh, and
+Fei Liu. 2019a. Improving the [similarity](https://doi.org/10.18653/v1/P19-1098) measure
+[of determinantal point processes for extractive multi-](https://doi.org/10.18653/v1/P19-1098)
+document [summarization.](https://doi.org/10.18653/v1/P19-1098) In _Proceedings_ _of_ _the_
+_57th_ _Annual_ _Meeting_ _of_ _the_ _Association_ _for_ _Com-_
+_putational_ _Linguistics_, pages 1027–1038, Florence,
+Italy. Association for Computational Linguistics.
+
+
+Sangwoo Cho, Chen Li, Dong Yu, Hassan Foroosh,
+and Fei Liu. 2019b. [Multi-document](https://doi.org/10.18653/v1/D19-5412) summarization with determinantal [point](https://doi.org/10.18653/v1/D19-5412) processes and contextualized [representations.](https://doi.org/10.18653/v1/D19-5412) In _Proceedings_ _of_ _the_
+_2nd_ _Workshop_ _on_ _New_ _Frontiers_ _in_ _Summarization_,
+pages 98–103, Hong Kong, China. Association for
+Computational Linguistics.
+
+
+Yoeng-Jin Chu and Tseng-Hong Liu. 1965. On the
+shortest arborescence of a directed graph. _Science_
+_Sinica_, 14:1396–1400.
+
+
+Alexis Conneau, Kartikay Khandelwal, Naman Goyal,
+Vishrav Chaudhary, Guillaume Wenzek, Francisco
+Guzm´an, Edouard Grave, Myle Ott, Luke Zettlemoyer, and Veselin Stoyanov. 2020. [Unsupervised](https://doi.org/10.18653/v1/2020.acl-main.747)
+
+
+
+cross-lingual [representation](https://doi.org/10.18653/v1/2020.acl-main.747) learning at scale. In
+_Proceedings of the 58th Annual Meeting of the Asso-_
+_ciation_ _for_ _Computational_ _Linguistics_, pages 8440–
+8451, Online. Association for Computational Linguistics.
+
+
+Ryan Cotterell and Jason Eisner. 2017. [Probabilistic](https://doi.org/10.18653/v1/P17-1109)
+typology: [Deep generative models of vowel invento-](https://doi.org/10.18653/v1/P17-1109)
+[ries.](https://doi.org/10.18653/v1/P17-1109) In _Proceedings of the 55th Annual Meeting of_
+_the_ _Association_ _for_ _Computational_ _Linguistics_ _(Vol-_
+_ume 1:_ _Long Papers)_, pages 1182–1192, Vancouver,
+Canada. Association for Computational Linguistics.
+
+
+Timothy Dozat and Christopher D. Manning. 2017.
+
+Deep biaffine attention [for](https://openreview.net/forum?id=Hk95PK9le) neural dependency pars[ing.](https://openreview.net/forum?id=Hk95PK9le) In _International_ _Conference_ _on_ _Learning_ _Rep-_
+_resentations_, Toulon, France. OpenReview.net.
+
+
+Jack Edmonds. 1967. Optimum branchings. _Journal_
+_of_ _Research_ _of_ _the_ _National_ _Bureau_ _of_ _Standards_,
+71B(4):233–240.
+
+
+Daniel Flannery and Shinsuke Mori. 2015. [Combin-](https://doi.org/10.18653/v1/W15-2202)
+[ing active learning and partial annotation for domain](https://doi.org/10.18653/v1/W15-2202)
+[adaptation of a Japanese dependency parser.](https://doi.org/10.18653/v1/W15-2202) In _Pro-_
+_ceedings_ _of_ _the_ _14th_ _International_ _Conference_ _on_
+_Parsing_ _Technologies_, pages 11–19, Bilbao, Spain.
+Association for Computational Linguistics.
+
+
+Yarin Gal, Riashat Islam, and Zoubin Ghahramani.
+2017. Deep Bayesian [active](http://proceedings.mlr.press/v70/gal17a.html) learning with image
+[data.](http://proceedings.mlr.press/v70/gal17a.html) In _Proceedings of the 34th International Con-_
+_ference_ _on_ _Machine_ _Learning_, pages 1183–1192,
+Sydney, Australia. PMLR.
+
+
+Rebecca Hwa. 2000. Sample [selection](https://doi.org/10.3115/1117794.1117800) for statistical
+grammar induction. In _Proceedings_ _of_ _the_ _2000_
+_Joint SIGDAT Conference on Empirical Methods in_
+_Natural_ _Language_ _Processing_ _and_ _Very_ _Large_ _Cor-_
+_pora_, pages 45–52, Hong Kong, China. Association
+for Computational Linguistics.
+
+
+Rebecca Hwa. 2004. Sample [selection](https://doi.org/10.1162/0891201041850894) for statistical
+[parsing.](https://doi.org/10.1162/0891201041850894) _Computational Linguistics_, 30(3):253–276.
+
+
+Diederik P. Kingma and Jimmy Ba. 2015. [Adam:](https://arxiv.org/abs/1412.6980)
+A method for [stochastic](https://arxiv.org/abs/1412.6980) optimization. In _Interna-_
+_tional Conference on Learning Representations_, San
+Diego, California, USA.
+
+
+Andreas Kirsch, Joost van Amersfoort, and Yarin Gal.
+2019. BatchBALD: [Efficient](http://papers.nips.cc/paper/8925-batchbald-efficient-and-diverse-batch-acquisition-for-deep-bayesian-active-learning.pdf) and diverse batch acquisition for deep [Bayesian](http://papers.nips.cc/paper/8925-batchbald-efficient-and-diverse-batch-acquisition-for-deep-bayesian-active-learning.pdf) active learning. In _Ad-_
+_vances_ _in_ _Neural_ _Information_ _Processing_ _Systems_
+_32_, pages 7026–7037, Vancouver, Canada. Curran
+Associates, Inc.
+
+
+Nikita Kitaev, Steven Cao, and Dan Klein. 2019. [Multi-](https://doi.org/10.18653/v1/P19-1340)
+lingual constituency [parsing](https://doi.org/10.18653/v1/P19-1340) with self-attention and
+[pre-training.](https://doi.org/10.18653/v1/P19-1340) In _Proceedings_ _of_ _the_ _57th_ _Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics_, pages 3499–3505, Florence, Italy. Association for Computational Linguistics.
+
+
+
+2621
+
+
+Terry Koo, Amir Globerson, Xavier Carreras, and
+Michael Collins. 2007. Structured [prediction](https://www.aclweb.org/anthology/D07-1015) models via the [matrix-tree](https://www.aclweb.org/anthology/D07-1015) theorem. In _Proceedings_
+_of_ _the_ _2007_ _Joint_ _Conference_ _on_ _Empirical_ _Meth-_
+_ods_ _in_ _Natural_ _Language_ _Processing_ _and_ _Com-_
+_putational_ _Natural_ _Language_ _Learning_ _(EMNLP-_
+_CoNLL)_, pages 141–150, Prague, Czech Republic.
+Association for Computational Linguistics.
+
+
+Sandra K¨ubler, Ryan McDonald, and Joakim Nivre.
+2008. _[Dependency](https://doi.org/10.2200/S00169ED1V01Y200901HLT002)_ _Parsing_, volume 2 of _Synthesis_
+_Lectures_ _on_ _Human_ _Language_ _Technologies_ . Morgan and Claypool.
+
+
+Taku Kudo and John Richardson. 2018. [SentencePiece:](https://doi.org/10.18653/v1/D18-2012)
+A simple and language [independent](https://doi.org/10.18653/v1/D18-2012) subword tok[enizer and detokenizer for neural text processing.](https://doi.org/10.18653/v1/D18-2012) In
+_Proceedings_ _of_ _the_ _2018_ _Conference_ _on_ _Empirical_
+_Methods_ _in_ _Natural_ _Language_ _Processing:_ _System_
+_Demonstrations_, pages 66–71, Brussels, Belgium.
+Association for Computational Linguistics.
+
+
+Alex Kulesza and Ben Taskar. 2012. Determinantal
+point processes for machine learning. _Foundations_
+_and Trends in Machine Learning_, 5(2–3):123–286.
+
+
+John A. Kulesza. 2012. _Learning_ _with_ _Determinantal_
+_Point_ _Processes_ . Ph.D. thesis, University of Pennsylvania.
+
+
+David D. Lewis and William A. Gale. 1994. A sequential algorithm for training text classifiers. In
+_Proceedings of the 17th Annual International ACM-_
+_SIGIR Conference on Research and Development in_
+_Information_ _Retrieval_, pages 3–12, Dublin, Ireland.
+ACM/Springer.
+
+
+Zhenghua Li, Min Zhang, Yue Zhang, Zhanyi Liu,
+Wenliang Chen, Hua Wu, and Haifeng Wang. 2016.
+[Active learning for dependency parsing with partial](https://www.aclweb.org/anthology/P16-1033)
+[annotation.](https://www.aclweb.org/anthology/P16-1033) In _Proceedings of the 54th Annual Meet-_
+_ing of the Association for Computational Linguistics_
+_(Volume_ _1:_ _Long_ _Papers)_, pages 344–354, Berlin,
+Germany.
+
+
+Saeed Majidi and Gregory Crane. 2013. [Active](https://www.aclweb.org/anthology/W13-5711) learning for dependency [parsing](https://www.aclweb.org/anthology/W13-5711) by a committee of
+[parsers.](https://www.aclweb.org/anthology/W13-5711) In _Proceedings_ _of_ _the_ _13th_ _International_
+_Conference_ _on_ _Parsing_ _Technologies_ _(IWPT_ _2013)_,
+pages 98–105, Nara, Japan. Assocation for Computational Linguistics.
+
+
+Gurmeet Singh Manku, Arvind Jain, and Anish
+Das Sarma. 2007. Detecting [near-duplicates](https://doi.org/10.1145/1242572.1242592) for
+web [crawling.](https://doi.org/10.1145/1242572.1242592) In _Proceedings_ _of_ _the_ _16th_ _Interna-_
+_tional_ _Conference_ _on_ _World_ _Wide_ _Web_, pages 141–
+150, Banff, Alberta, Canada. Association for Computing Machinery.
+
+
+Christopher Manning, Mihai Surdeanu, John Bauer,
+Jenny Finkel, Steven Bethard, and David McClosky.
+2014. [The Stanford CoreNLP natural language pro-](https://doi.org/10.3115/v1/P14-5010)
+cessing toolkit. In _Proceedings_ _of_ _52nd_ _Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics:_ _System Demonstrations_, pages 55–60, Baltimore, Maryland, USA. Association for Computational Linguistics.
+
+
+
+Mitchell P. Marcus, Beatrice Santorini, and Mary Ann
+Marcinkiewicz. 1993. Building a [large](https://www.aclweb.org/anthology/J93-2004) annotated
+corpus of English: The Penn Treebank. _Computa-_
+_tional Linguistics_, 19(2):313–330.
+
+
+Seyed Abolghasem Mirroshandel and Alexis Nasr.
+2011. [Active learning for dependency parsing using](https://www.aclweb.org/anthology/W11-2917)
+[partially annotated sentences.](https://www.aclweb.org/anthology/W11-2917) In _Proceedings of the_
+_12th International Conference on Parsing Technolo-_
+_gies_, pages 140–149, Dublin, Ireland. Association
+for Computational Linguistics.
+
+
+Miles Osborne and Jason Baldridge. 2004. [Ensemble-](https://www.aclweb.org/anthology/N04-1012)
+based active [learning](https://www.aclweb.org/anthology/N04-1012) for parse selection. In _Pro-_
+_ceedings_ _of_ _the_ _Human_ _Language_ _Technology_ _Con-_
+_ference of the North American Chapter of the Asso-_
+_ciation for Computational Linguistics: HLT-NAACL_
+_2004_, pages 89–96, Boston, Massachusetts, USA.
+Association for Computational Linguistics.
+
+
+Adam Paszke, Sam Gross, Francisco Massa, Adam
+Lerer, James Bradbury, Gregory Chanan, Trevor
+Killeen, Zeming Lin, Natalia Gimelshein, Luca
+Antiga, Alban Desmaison, Andreas Kopf, Edward
+Yang, Zachary DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang,
+Junjie Bai, and Soumith Chintala. 2019. [PyTorch:](http://papers.nips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf)
+An imperative style, [high-performance](http://papers.nips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf) deep learn[ing library.](http://papers.nips.cc/paper/9015-pytorch-an-imperative-style-high-performance-deep-learning-library.pdf) In _Advances in Neural Information Pro-_
+_cessing_ _Systems_ _32_, pages 8026–8037, Vancouver,
+Canada. Curran Associates, Inc.
+
+
+Manabu Sassano and Sadao Kurohashi. 2010. [Using](https://www.aclweb.org/anthology/P10-1037)
+smaller constituents [rather](https://www.aclweb.org/anthology/P10-1037) than sentences in active
+learning for [Japanese](https://www.aclweb.org/anthology/P10-1037) dependency parsing. In _Pro-_
+_ceedings_ _of_ _the_ _48th_ _Annual_ _Meeting_ _of_ _the_ _Asso-_
+_ciation_ _for_ _Computational_ _Linguistics_, pages 356–
+365, Uppsala, Sweden. Association for Computational Linguistics.
+
+
+Alexandra Schofield, Laure Thompson, and David
+Mimno. 2017. Quantifying the [effects](https://doi.org/10.18653/v1/D17-1290) of text duplication on [semantic](https://doi.org/10.18653/v1/D17-1290) models. In _Proceedings_ _of_
+_the 2017 Conference on Empirical Methods in Natu-_
+_ral Language Processing_, pages 2737–2747, Copenhagen, Denmark. Association for Computational
+Linguistics.
+
+
+Burr Settles. 2009. Active learning literature survey.
+Computer Sciences Technical Report 1648, University of Wisconsin-Madison.
+
+
+Burr Settles and Mark Craven. 2008. [An analysis of ac-](https://www.aclweb.org/anthology/D08-1112)
+tive learning strategies [for](https://www.aclweb.org/anthology/D08-1112) sequence labeling tasks.
+In _Proceedings_ _of_ _the_ _2008_ _Conference_ _on_ _Empiri-_
+_cal Methods in Natural Language Processing_, pages
+1070–1079, Honolulu, Hawaii, USA. Association
+for Computational Linguistics.
+
+
+Yanyao Shen, Hyokun Yun, Zachary C. Lipton,
+Yakov Kronrod, and Animashree Anandkumar.
+2018. [Deep active learning for named entity recogni-](https://openreview.net/forum?id=ry018WZAZ)
+[tion.](https://openreview.net/forum?id=ry018WZAZ) In _International Conference on Learning Rep-_
+_resentations_, Vancouver, Canada. OpenReview.net.
+
+
+
+2622
+
+
+Aditya Siddhant and Zachary C. Lipton. 2018. [Deep](https://doi.org/10.18653/v1/D18-1318)
+Bayesian active learning for natural language processing: Results of a [large-scale](https://doi.org/10.18653/v1/D18-1318) empirical study.
+In _Proceedings_ _of_ _the_ _2018_ _Conference_ _on_ _Em-_
+_pirical_ _Methods_ _in_ _Natural_ _Language_ _Processing_,
+pages 2904–2909, Brussels, Belgium. Association
+for Computational Linguistics.
+
+
+Min Tang, Xiaoqiang Luo, and Salim Roukos. 2002.
+
+[Active learning for statistical natural language pars-](https://doi.org/10.3115/1073083.1073105)
+[ing.](https://doi.org/10.3115/1073083.1073105) In _Proceedings_ _of_ _the_ _40th_ _Annual_ _Meeting_
+_of_ _the_ _Association_ _for_ _Computational_ _Linguistics_,
+pages 120–127, Philadelphia, Pennsylvania, USA.
+Association for Computational Linguistics.
+
+
+Cynthia A. Thompson, Mary Elaine Califf, and Raymond J. Mooney. 1999. [Active learning for natural](https://www.cs.utexas.edu/~ml/papers/active-nll-ml99.pdf)
+[language parsing and information extraction.](https://www.cs.utexas.edu/~ml/papers/active-nll-ml99.pdf) In _Pro-_
+_ceedings_ _of_ _the_ _Sixteenth_ _International_ _Conference_
+_on Machine Learning_, pages 406–414, Bled, Slovenia. Morgan Kaufmann Publishers Inc.
+
+
+William T. Tutte. 1984. _Graph_ _Theory_ . AddisonWesley Publishing Company.
+
+
+Laurens van der Maaten and Geoffrey Hinton. 2008.
+
+Visualizing [data](https://jmlr.org/papers/volume9/vandermaaten08a/vandermaaten08a.pdf) using t-SNE. _Journal_ _of_ _machine_
+_learning research_, 9(Nov):2579–2605.
+
+
+2623
+
+
+**Appendix A** **Dependency Parser**
+
+
+We adopt the deep biaffine dependency parser proposed by Dozat and Manning (2017). The parser is
+conceptually simple and yet competitive with stateof-the-art dependency parsers. The parser has three
+components: feature extraction, unlabeled parsing
+and relation labeler.
+
+
+**Feature Extraction** For a length- _n_ sentence _x_ =
+_x_ 0 _, x_ 1 _, x_ 2 _, . . ., xn_, where _x_ 0 is the dummy root
+symbol, we extract contextualized features at each
+word position. In our experiments, we use a pretrained multilingual XLM-RoBERTa base model
+(Conneau et al., 2020), and fine-tune the feature
+extractor along with the rest of our parser:
+
+
+[ **x** 0 _,_ **x** 1 _, . . .,_ **x** _n_ ] = XLM-R( _x_ 0 _, x_ 1 _, . . ., xn_ ) _._
+
+
+Each word input to the XLM-RoBERTa model is
+processed with the SentencePiece tokenizer (Kudo
+and Richardson, 2018), and we follow Kitaev et al.
+(2019) and retain the vectors corresponding to the
+last sub-word units as their representations. For **x** 0,
+we use the vector of the [CLS] token, which is
+appended by XLM-RoBERTa to the beginning of
+each sentence.
+
+
+**Unlabeled** **Parser** The parser uses a deep biaffine attention mechanism to derive locallynormalized attachment probabilities for all potential head-dependent pairs:
+
+
+**h** [arc-head] _i_ = MLP [arc-head] ( **x** _i_ )
+
+**h** [arc-dep] _j_ = MLP [arc-dep] ( **x** _j_ )
+
+_si,j_ = [ **h** [arc-head] _i_ ; 1] _[⊤]_ _U_ [arc] [ **h** [arc-dep] _j_ ; 1]
+
+_P_ att(head(j) = _i | x_ ) = softmax _i_ ( _s_ : _,j_ ) _,_
+
+
+where MLP [arc-head] and MLP [arc-dep] are two multilayer perceptrons (MLPs) projecting **x** vectors into
+_d_ [arc] -dimensional **h** vectors, [; 1] appends an element of 1 at the end of the vectors, and _U_ [arc] _∈_
+R [(] _[d]_ [arc][+1)] _[×]_ [(] _[d]_ [arc][+1)] is a bilinear scoring matrix. This
+component is trained with cross-entropy loss of
+the gold-standard attachments. During inference,
+we use the Chu-Liu-Edmonds algorithm (Chu and
+Liu, 1965; Edmonds, 1967) to find the spanning
+tree with the highest product of locally-normalized
+attachment probabilities.
+
+
+**Relation Labeler** The relation labeling component employs a similar deep biaffine scoring func
+
+
+Round # 5 10
+
+
+Strategy w/o DPP w/ DPP w/o DPP w/ DPP
+
+
+Random 89 _._ 01 _±._ 28 **89** _._ **67** _±._ 30 90 _._ 78 _±._ 27 **91** _._ **22** _±._ 22
+AMP 89 _._ 67 _±._ 29 **90** _._ **24** _±._ 39 92 _._ 03 _±._ 10 **92** _._ **17** _±._ 22
+BALD 89 _._ 82 _±._ 36 **90** _._ **29** _±._ 20 91 _._ 87 _±._ 36 **92** _._ **00** _±._ 08
+ID **90** _._ **24** _±._ 20 90 _._ 03 _±._ 18 **92** _._ **16** _±._ 17 92 _._ 06 _±._ 16
+
+
+Table B1: UAS after 5 and 10 rounds of annotation (roughly
+5 _,_ 000 and 7 _,_ 000 training tokens respectively), comparing
+strategies with and without modeling diversity through DPP.
+
+
+tion as the unlabeled parsing component:
+
+
+**h** [rel-head] _i_ = MLP [rel-head] ( **x** _i_ )
+
+**h** [rel-dep] _j_ = MLP [rel-dep] ( **x** _j_ )
+
+_ti,j,r_ = [ **h** [rel-head] _i_ ; 1] _[⊤]_ _Ur_ [rel][[] **[h]** _j_ [rel-dep] ; 1]
+
+
+_P_ (rel( _i, j_ ) = _r_ ) = softmax _r_ ( _ti,j,_ :) _,_
+
+
+where each _Ur_ [rel] _∈_ R [(] _[d]_ [rel][+1)] _[×]_ [(] _[d]_ [rel][+1)], and there are
+as many such matrices as the size of the dependency
+relation label set _|R|_ . The relation labeler is trained
+using cross entropy loss on the gold-standard headdependent pairs. During inference, the labeling
+decision for each arc is made independently given
+the predicted unlabeled parse tree.
+
+
+**Appendix B** **Results with UAS**
+**Evaluation**
+
+
+We also evaluate different learning strategies based
+on unlabeled attachment scores (UAS), and the
+results are shown in Table B1. In line with LASbased experiments, we find that modeling diversity
+is more helpful during initial stages of learning.
+For ID, we observe this effect even earlier than the
+fifth round of annotation: 86 _._ 57 _±._ 44 vs. 87 _._ 40 _±._ 51
+after the first annotation round.
+
+
+**Appendix C** **Sentence Selection**
+**Examples**
+
+
+In Table C2 we compare batches sampled by a
+diversity-aware selection strategy with a diversityagnostic one. We observe that by modeling diversity in the sample selection process, DPPs avoid
+selecting duplicate or near-duplicate sentences and
+thus the annotation budget can be maximally utilized.
+
+
+**Appendix D** **BALD under High**
+**Duplication Setting**
+
+
+Figure D1 shows the learning curves for BALDbased selection strategies under a high corpus
+duplication setting where the corpus is repeated
+
+
+
+2624
+
+
+**Sentences selected by AMP (highest-quality ones first):**
+Downgraded by Moody ’s were Houston Lighting ’s first - mortgage bonds and secured pollution - control bonds to single     - A     - 3 from single     - A     - 2 ; unsecured pollution control bonds to Baa - 1 from single - A - 3 ; preferred stock to single - A - 3 from single - A - 2 ; a shelf registration for preferred stock to a preliminary rating of single - A - 3
+from a preliminary rating of single - A - 2 ; two shelf registrations for collateralized debt securities to a preliminary rating of single - A - 3 from a preliminary rating of single A - 2, and the unit ’s rating for commercial paper to Prime - 2 from Prime - 1 .
+For a while in the 1970s it seemed Mr. Moon was on a spending spree, with such purchases as the former New Yorker Hotel and its adjacent Manhattan Center ; a fishing
+/ processing conglomerate with branches in Alaska, Massachusetts, Virginia and Louisiana ; a former Christian Brothers monastery and the Seagram family mansion ( both
+picturesquely situated on the Hudson River ) ; shares in banks from Washington to Uruguay ; a motion picture production company, and newspapers, such as the Washington
+Times, the New York City Tribune ( originally the News World ), and the successful Spanish - language Noticias del Mundo .
+_→_ LONDON LATE EURODOLLARS : 8 11/16 % to 8 9/16 % one month ; 8 5/8 % to 8 1/2 % two months ; 8 5/8 % to 8 1/2 % three months ; 8 9/16 % to 8 7/16 % four months ;
+8 1/2 % to 8 3/8 % five months ; 8 1/2 % to 8 3/8 % six months .
+_→_ LONDON LATE EURODOLLARS : 8 3/4 % to 8 5/8 % one month ; 8 3/4 % to 8 5/8 % two months ; 8 11/16 % to 8 9/16 % three months ; 8 9/16 % to 8 7/16 % four months ;
+8 1/2 % to 8 3/8 % five months ; 8 7/16 % to 8 5/16 % six months .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.40 % 30 to 44 days ; 8.325 % 45 to 59 days ; 8.10 % 60 to 89 days ; 8 % 90 to 119 days ;
+
+ - 7.85 % 120 to 149 days ; 7.70 % 150 to 179 days ; 7.375 % 180 to 270 days .
+4 . When a RICO TRO is being sought, the prosecutor is required, at the earliest appropriate time, to state publicly that the government ’s request for a TRO, and eventual
+forfeiture, is made in full recognition of the rights of third parties – that is, in requesting the TRO, the government will not seek to disrupt the normal, legitimate business
+activities of the defendant ; will not seek through use of the relation - back doctrine to take from third parties assets legitimately transferred to them ; will not seek to vitiate
+legitimate business transactions occurring between the defendant and third parties ; and will, in all other respects, assist the court in ensuring that the rights of third parties are
+protected, through proceeding under RICO and otherwise .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.50 % 30 to 44 days ; 8.25 % 45 to 62 days ; 8.375 % 63 to 89 days ; 8 % 90 to 119 days ;
+
+ - 7.90 % 120 to 149 days ; 7.80 % 150 to 179 days ; 7.55 % 180 to 270 days .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.50 % 30 to 44 days ; 8.25 % 45 to 65 days ; 8.375 % 66 to 89 days ; 8 % 90 to 119 days ;
+
+ - 7.875 % 120 to 149 days ; 7.75 % 150 to 179 days ; 7.50 % 180 to 270 days .
+_→_ LONDON LATE EURODOLLARS : 8 11/16 % to 8 9/16 % one month ; 8 5/8 % to 8 1/2 % two months ; 8 5/8 % to 8 1/2 % three months ; 8 9/16 % to 8 7/16 % four months ;
+8 1/2 % to 8 3/8 % five months ; 8 7/16 % to 8 5/16 % six months .
+_→_ LONDON LATE EURODOLLARS : 8 11/16 % to 8 9/16 % one month ; 8 9/16 % to 8 7/16 % two months ; 8 5/8 % to 8 1/2 % three months ; 8 1/2 % to 8 3/8 % four months ;
+8 7/16 % to 8 5/16 % five months ; 8 7/16 % to 8 5/16 % six months .
+The new edition lists the top 10 metropolitan areas as Anaheim - Santa Ana, Calif. ; Boston ; Louisville, Ky. ; Nassau - Suffolk, N.Y. ; New York ; Pittsburgh ; San Diego ;
+San Francisco ; Seattle ; and Washington .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.45 % 30 to 44 days ; 8.20 % 45 to 67 days ; 8.325 % 68 to 89 days ; 8 % 90 to 119 days ;
+
+ - 7.875 % 120 to 149 days ; 7.75 % 150 to 179 days ; 7.50 % 180 to 270 days .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.50 % 2 to 44 days ; 8.25 % 45 to 69 days ; 8.40 % 70 to 89 days ; 8.20 % 90 to 119 days ;
+
+ - 8.05 % 120 to 149 days ; 7.90 % 150 to 179 days ; 7.50 % 180 to 270 days .
+Five officials of this investment banking firm were elected directors : E. Garrett Bewkes III, a 38 - year - old managing director in the mergers and acquisitions department ;
+Michael R. Dabney, 44, a managing director who directs the principal activities group which provides funding for leveraged acquisitions ; Richard Harriton, 53, a general
+partner who heads the correspondent clearing services ; Michael Minikes, 46, a general partner who is treasurer ; and William J. Montgoris, 42, a general partner who is also
+senior vice president of finance and chief financial officer .
+_→_ LONDON LATE EURODOLLARS : 8 11/16 % to 8 9/16 % one month ; 8 5/8 % to 8 1/2 % two months ; 8 11/16 % to 8 9/16 % three months ; 8 9/16 % to 8 7/16 % four
+months ; 8 1/2 % to 8 3/8 % five months ; 8 7/16 % to 8 5/16 % six months .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.55 % 30 to 44 days ; 8.25 % 45 to 59 days ; 8.40 % 60 to 89 days ; 8 % 90 to 119 days ; 7.90
+
+ - % 120 to 149 days ; 7.80 % 150 to 179 days ; 7.55 % 180 to 270 days .
+They transferred some $ 28 million from the Community Development Block Grant program designated largely for low - and moderate - income projects and funneled it into
+such items as :      - $ 1.2 million for a performing - arts center in Newark, – $ 1.3 million for “ job retention ” in Hawaiian sugar mills .      - $ 400,000 for a collapsing utility tunnel
+in Salisbury,      - $ 500,000 for “ equipment and landscaping to deter crime and aid police surveillance ” at a Michigan park .      - $ 450,000 for “ integrated urban data based in
+seven cities . ” No other details .       - $ 390,000 for a library and recreation center at Mackinac Island, Mich .
+_→_ LONDON LATE EURODOLLARS : 8 3/4 % to 8 5/8 % one month ; 8 13/16 % to 8 11/16 % two months ; 8 11/16 % to 8 9/16 % three months ; 8 9/16 % to 8 7/16 % four
+months ; 8 1/2 % to 8 3/8 % five months ; 8 7/16 % to 8 5/16 % six months .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.45 % 30 to 44 days ; 8.25 % 45 to 68 days ; 8.30 % 69 to 89 days ; 8.125 % 90 to 119 days ;
+
+ - 8 % 120 to 149 days ; 7.875 % 150 to 179 days ; 7.50 % 180 to 270 days .
+
+
+**Sentences selected by AMP with diversity-inducing DPP:**
+Downgraded by Moody ’s were Houston Lighting ’s first - mortgage bonds and secured pollution - control bonds to single     - A     - 3 from single     - A     - 2 ; unsecured pollution control bonds to Baa - 1 from single - A - 3 ; preferred stock to single - A - 3 from single - A - 2 ; a shelf registration for preferred stock to a preliminary rating of single - A - 3
+from a preliminary rating of single - A - 2 ; two shelf registrations for collateralized debt securities to a preliminary rating of single - A - 3 from a preliminary rating of single A - 2, and the unit ’s rating for commercial paper to Prime - 2 from Prime - 1 .
+4 . When a RICO TRO is being sought, the prosecutor is required, at the earliest appropriate time, to state publicly that the government ’s request for a TRO, and eventual
+forfeiture, is made in full recognition of the rights of third parties – that is, in requesting the TRO, the government will not seek to disrupt the normal, legitimate business
+activities of the defendant ; will not seek through use of the relation - back doctrine to take from third parties assets legitimately transferred to them ; will not seek to vitiate
+legitimate business transactions occurring between the defendant and third parties ; and will, in all other respects, assist the court in ensuring that the rights of third parties are
+protected, through proceeding under RICO and otherwise .
+COMMERCIAL PAPER placed directly by General Motors Acceptance Corp. : 8.40 % 30 to 44 days ; 8.325 % 45 to 59 days ; 8.10 % 60 to 89 days ; 8 % 90 to 119 days ;
+
+ - 7.85 % 120 to 149 days ; 7.70 % 150 to 179 days ; 7.375 % 180 to 270 days .
+Moreover, the process is n’t without its headaches .
+For a while in the 1970s it seemed Mr. Moon was on a spending spree, with such purchases as the former New Yorker Hotel and its adjacent Manhattan Center ; a fishing
+/ processing conglomerate with branches in Alaska, Massachusetts, Virginia and Louisiana ; a former Christian Brothers monastery and the Seagram family mansion ( both
+picturesquely situated on the Hudson River ) ; shares in banks from Washington to Uruguay ; a motion picture production company, and newspapers, such as the Washington
+Times, the New York City Tribune ( originally the News World ), and the successful Spanish - language Noticias del Mundo .
+Within the paper sector, Mead climbed 2 3/8 to 38 3/4 on 1.3 million shares, Union Camp rose 2 3/4 to 37 3/4, Federal Paper Board added 1 3/4 to 23 7/8, Bowater gained 1
+1/2 to 27 1/2, Stone Container rose 1 to 26 1/8 and Temple - Inland jumped 3 3/4 to 62 1/4 .
+We finally rendezvoused with our balloon, which had come to rest on a dirt road amid a clutch of Epinalers who watched us disassemble our craft – another half - an - hour of
+non-flight activity – that included the precision routine of yanking the balloon to the ground, punching all the air out of it, rolling it up and cramming it and the basket into the
+trailer .
+These are the 26 states, including the commonwealth of Puerto Rico, that have settled with Drexel : Alaska, Arkansas, Delaware, Georgia, Hawaii, Idaho, Indiana, Iowa,
+Kansas, Kentucky, Maine, Maryland, Minnesota, Mississippi, New Hampshire, New Mexico, North Dakota, Oklahoma, Oregon, South Carolina, South Dakota, Utah,
+Vermont, Washington, Wyoming and Puerto Rico .
+It is the stuff of dreams, but also of traumas .
+An inquiry into his handling of Lincoln S&L inevitably will drag in Sen. Cranston and the four others, Sens. Dennis DeConcini ( D., Ariz. ), John McCain ( R., Ariz. ), John
+Glenn ( D., Ohio ) and Donald Riegle ( D., Mich . ) .
+Five officials of this investment banking firm were elected directors : E. Garrett Bewkes III, a 38 - year - old managing director in the mergers and acquisitions department ;
+Michael R. Dabney, 44, a managing director who directs the principal activities group which provides funding for leveraged acquisitions ; Richard Harriton, 53, a general
+partner who heads the correspondent clearing services ; Michael Minikes, 46, a general partner who is treasurer ; and William J. Montgoris, 42, a general partner who is also
+senior vice president of finance and chief financial officer .
+But as they hurl fireballs that smolder rather than burn, and relive old duels in the sun, it ’s clear that most are there to make their fans cheer again or recapture the camaraderie
+of seasons past or prove to themselves and their colleagues that they still have it – or something close to it .
+They are : “ A Payroll to Meet : A Story of Greed, Corruption and Football at SMU ” ( Macmillan, 221 pages, $ 18.95 ) by David Whitford ; “ Big Red Confidential : Inside
+Nebraska Football ” ( Contemporary, 231 pages, $ 17.95 ) by Armen Keteyian ; and “ Never Too Young to Die : The Death of Len Bias ” ( Pantheon, 252 pages, $ 18.95 ) by
+Lewis Cole .
+He says he told NewsEdge to look for stories containing such words as takeover, acquisition, acquire, LBO, tender, merger, junk and halted .
+It is no coincidence that from 1844 to 1914, when the Bank of England was an independent private bank, the pound was never devalued and payment of gold for pound notes
+was never suspended, but with the subsequent nationalization of the Bank of England, the pound was devalued with increasing frequency and its use as an international medium
+of exchange declined .
+The $ 4 billion in bonds break down as follows : $ 1 billion in five - year bonds with a coupon rate of 8.25 % and a yield to maturity of 8.33 % ; $ 1 billion in 10 - year bonds
+with a coupon rate of 8.375 % and a yield to maturity of 8.42 % ; $ 2 billion in 30 - year bonds with five - year call protection, a coupon rate of 8.75 % and a yield to maturity
+of 9.06 % .
+Hecla Mining rose 5/8 to 14 ; Battle Mountain Gold climbed 3/4 to 16 3/4 ; Homestake Mining rose 1 1/8 to 16 7/8 ; Lac Minerals added 5/8 to 11 ; Placer Dome went up 7/8
+to 16 3/4, and ASA Ltd. jumped 3 5/8 to 49 5/8 .
+
+
+Table C2: Sentences picked by a diversity-agnostic (top) and a diversity-aware (bottom) selection strategy from the same
+unlabeled pool after the intial round of model training on the seed sentences. The diversity-agnostic strategy selects many
+near-duplicate sentences (the two near-duplicate clusters are marked by red2625 _→_ and blue ), effectively wasting the annotation
+budget, where DPPs largely alleviate this issue by enforcing diversity.
+                                  
+
+92
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+
+
+82
+10 20 30
+
+
+Annotation round
+
+
+Figure D1: Learning curves for BALD-based selection strategies under a five-fold corpus duplication setting.
+
+
+**Algorithm E1:** Greedy MAP inference for
+DPP with a size budget, adapted from
+Kulesza (2012).
+
+
+**Input:** candidate item set _X_ (sentences or tokens),
+DPP represented by matrix _L_, size budget _b_
+_U_ _←_ _X_ ;
+_Y_ _←∅_ ;
+**while** _U_ = _∅_ **do**
+
+_i ←_ arg max _i′∈U_ det( _LY ∪{i′}_ );
+**if** [�] _y∈Y_ [size][(] _[y]_ [)] _[ < b]_ **[ then]**
+
+_Y_ _←_ _Y_ _∪{i}_ ;
+**else**
+
+break;
+**end**
+**end**
+**Output:** selected items _Y_
+
+
+five times. In this extreme setting, the diversityagnostic strategy significantly underperforms the
+diversity-aware one. We posit that the relative success of BALD compared to AMP in the twiceduplicated setting is due to the fact that BALD
+randomly draws dropout masks to estimate model
+uncertainty, so that identical examples could still
+have different quality measures.
+
+
+**Appendix E** **Implementation Details and**
+**Hyperparameters**
+
+
+We do not tune our hyperparameters since in practice, active learning systems only have a single shot
+at success, without tuning. Instead, we follow recommendations from relevant prior work in setting
+our learning details and hyperparameters.
+
+
+**Active Learning** Following Li et al. (2016), our
+active learning set-up proceeds in two stages for
+each annotation round. In the first stage, we select
+sentences filling in a budget of 2500 tokens; in the
+second stage, we pick 500 tokens out of the subset
+of sentences. For a diversity-agnostic strategy, we
+choose the top- _k_ highest-quality candidates within
+the token budget, while our diversity-aware selec
+
+
+tion strategy uses a separate DPP for each stage.
+The active learning process is bootstraped with a
+seed set of 128 labeled sentences. For the BALD
+quality measure, we set _K_ = 5.
+
+
+**Greedy** **MAP** **Inference** **for** **DPPs** Algorithm E1 illustrates the procedure for selecting
+items from DPPs under a budget constraint. This
+greedy MAP inference algorithm is adapted from
+Kulesza (2012). During sentence selection, the
+size of a sentence is its number of tokens, and each
+token has a size of 1 in the token selection stage.
+
+
+**Dependency Parser** We set the hyperparameters
+according to Dozat and Manning (2017). All the
+MLPs in the deep biaffine attention architecture
+have single hidden layers with ReLU activation
+functions and a dropout probability of 0 _._ 33, and
+we set _d_ [arc] and _d_ [rel] to be 500 and 100 respectively.
+
+
+**Training and Optimization** Each training batch
+contains 16 sentences and gradient norms are
+clipped to 5 _._ 0. We use the Adam optimizer
+(Kingma and Ba, 2015) with a learning rate of 10 _[−]_ [5]
+
+with 640 warmup steps with a linearly-increasing
+learning rate starting from 0.
+
+
+**Implementation** Our implementation is in PyTorch (Paszke et al., 2019), and we use the
+transformers package [6] to interface with the pretrained XLM-RoBERTa model.
+
+
+[6https://github.com/huggingface/transformers](https://github.com/huggingface/transformers)
+
+
+
+2626
+
+

references/2023.emnlp.zhang/paper.md

@@ -0,0 +1,2665 @@
+---
+title: "Data-efficient Active Learning for Structured Prediction with Partial Annotation and Self-Training"
+authors:
+  - "Zhisong Zhang"
+  - "Emma Strubell"
+  - "Eduard Hovy"
+year: 2023
+venue: "Findings of EMNLP 2023"
+url: "https://aclanthology.org/2023.findings-emnlp.865/"
+---
+
+# **Data-efficient Active Learning for Structured Prediction with** **Partial Annotation and Self-Training**
+
+**Zhisong Zhang, Emma Strubell, Eduard Hovy**
+Language Technologies Institute, Carnegie Mellon University
+zhisongz@cs.cmu.edu, strubell@cmu.edu, hovy@cmu.edu
+
+
+
+**Abstract**
+
+
+In this work we propose a pragmatic method
+that reduces the annotation cost for structured
+label spaces using active learning. Our approach leverages partial annotation, which reduces labeling costs for structured outputs
+by selecting only the most informative substructures for annotation. We also utilize selftraining to incorporate the current model’s automatic predictions as pseudo-labels for unannotated sub-structures. A key challenge in
+effectively combining partial annotation with
+self-training to reduce annotation cost is determining which sub-structures to select to label.
+To address this challenge, we adopt an error
+estimator to adaptively decide the partial selection ratio according to the current model’s capability. In evaluations spanning four structured
+prediction tasks, we show that our combination
+of partial annotation and self-training using an
+adaptive selection ratio reduces annotation cost
+over strong full annotation baselines under a
+fair comparison scheme that takes reading time
+into consideration.
+
+
+**1** **Introduction**
+
+
+Structured prediction (Smith, 2011) is a fundamental problem in NLP, wherein the label space consists of complex structured outputs with groups of
+interdependent variables. It covers a wide range of
+NLP tasks, including sequence labeling, syntactic
+parsing and information extraction (IE). Modern
+structured predictors are developed in a data-driven
+way, by training statistical models with suitable
+annotated data. Recent developments in neural
+models and especially pre-trained language models
+(Peters et al., 2018; Devlin et al., 2019; Liu et al.,
+2019; Yang et al., 2019) have greatly improved system performance on these tasks. Nevertheless, the
+success of these models still relies on the availability of sufficient manually annotated data, which is
+often expensive and time-consuming to obtain.
+
+
+
+|Model Prediction for Annotation for<br>Highly-confident Parts Ambigious Parts<br>✘<br>✔<br>He saw the man with a backpack|Col2|
+|---|---|
+|He  saw  the  man  with  a<br>✔<br>✘<br>_Model Prediction for_<br>_Highly-confident Parts_|backpack<br>_Annotation for_<br>_Ambigious Parts_|
+
+
+Figure 1: Example partial annotations of a dependency
+tree. Manual annotation is requested only for the uncertain sub-structures (red), whereas model predictions can
+be used to annotate the highly-confident edges (blue).
+
+
+To mitigate such data bottlenecks, active learning (AL), which allows the model to select the
+most informative data instances to annotate, has
+been demonstrated to achieve good model accuracy while requiring fewer labels (Settles, 2009).
+When applying AL to structured prediction, one
+natural strategy is to perform full annotation (FA)
+for the output structures, for example, annotating a
+full sequence of labels or a full syntax tree. Due to
+its simplicity, FA has been widely adopted in AL
+approaches for structured prediction tasks (Hwa,
+2004; Settles and Craven, 2008; Shen et al., 2018).
+Nevertheless, a structured object can usually be decomposed into smaller sub-structures having nonuniform difficulty and informativeness. For example, as shown in Figure 1, in a dependency tree,
+edges such as functional relations are relatively
+easy to learn, requiring fewer manual annotations,
+while prepositional attachment links may be more
+informative and thus more worthwhile to annotate.
+The non-uniform distribution of informative substructures naturally suggests AL with partial annotation (PA), where the annotation budget can be
+preserved by only choosing a portion of informative
+sub-structures to annotate rather than laboriously
+labeling entire sentence structures. This idea has
+been explored in previous work, covering typical
+structured prediction tasks such as sequence labeling (Shen et al., 2004; Marcheggiani and Artières,
+2014; Chaudhary et al., 2019; Radmard et al., 2021)
+
+
+
+12991
+
+_Findings of the Association for Computational Linguistics:_ _EMNLP 2023_, pages 12991–13008
+December 6-10, 2023 ©2023 Association for Computational Linguistics
+
+
+**Algorithm 1** AL Procedure.
+
+**Input:** Seed dataset _L_ 0, dev dataset _D_, unlabeled pool _U_, total budget _t_, batch selection size _b_, annotation _strategy_ .
+**Output:** Final labeled dataset _L_, trained model _M_ .
+
+1: _L_ _←L_ 0 # Initialize
+2: **while** _t >_ 0 **do** # Until out of budget
+3: _M_ _←_ train( _L,_ _U_ ) # Model training
+4: _S_ _←_ sentence-query( _M,_ _U_ ) # Sentence selection
+5: **if** _strategy_ == “partial” **then**
+6: _r_ _←_ auto-ratio( _S, D_ ) # Decide adaptive ratio
+7: partial-annotate( _S, r_ ) # Partial annotation
+8: **else**
+9: full-annotate( _S_ ) # Full annotation
+10: _U_ _←U_ _−S_ ; _L_ _←L ∪S_ ; _t_ _←_ _t −_ _b_
+11: _M_ _←_ train( _L,_ _U_ ) # Final model training
+12: **return** _L,_ _M_
+
+
+and dependency parsing (Sassano and Kurohashi,
+2010; Mirroshandel and Nasr, 2011; Flannery and
+Mori, 2015; Li et al., 2016). Our work follows this
+direction and investigates the central question in
+AL with PA of how to decide which sub-structures
+to select. Most previous work uses a pre-defined
+fixed selection criterion, such as a threshold or ratio, which may be hard to decide in practice. In this
+work, we adopt a performance predictor to estimate
+the error rate of the queried instances and decide
+the ratio of partial selection accordingly. In this
+way, our approach can automatically and adaptively
+adjust the amount of partial selection throughout
+the AL process.
+Another interesting question for AL is how
+to better leverage unlabeled data. In this work,
+we investigate a simple semi-supervised method,
+self-training (Yarowsky, 1995), which adopts the
+model’s automatic predictions on the unlabeled
+data as extra training signals. Self-training naturally complements AL in the typical pool-based
+setting where we assume access to a pool of unlabeled data (Settles, 2009). It is particularly compatible with PA-based AL since the un-selected substructures are typically also highly-confident under
+the current model and likely to be predicted correctly without requiring additional annotation. We
+revisit this idea from previous work (Tomanek and
+Hahn, 2009; Majidi and Crane, 2013) and investigate its applicability with modern neural models
+and our adaptive partial selection approach.
+We perform a comprehensive empirical investigation on the effectiveness of different AL strategies for typical structured prediction tasks. We
+perform fair comparisons that account for the hidden cost of reading time by keeping the context
+size the same for all the strategies in each AL cy
+
+
+cle. With evaluations on four benchmark tasks for
+structured prediction (named entity recognition, dependency parsing, event extraction, and relation
+extraction), we show that PA can obtain roughly
+the same benefits as FA with the same reading cost
+but less sub-structure labeling cost, leading to better data efficiency. We also demonstrate that the
+adaptive partial selection scheme and self-training
+play crucial and complementary roles.
+
+
+**2** **Method**
+
+
+**2.1** **AL for Structured Prediction**
+
+We adopt the conventional pool-based AL setting,
+which iteratively selects and annotates instances
+from an unlabeled pool. Please refer to Settles
+(2009) for the basics and details of AL; our main
+illustration focuses more specifically on applying
+AL to structured prediction.
+Algorithm 1 illustrates the overall AL process.
+We focus on sentence-level tasks. In FA, each sentence is annotated with a full structured object (for
+example, a label sequence or a syntax tree). In PA,
+annotation granularity is at the sub-structure level
+(for example, a sub-sequence of labels or a partial
+tree). We adopt a two-step selection approach for
+all the strategies by first choosing a batch of sentences and then annotating within this batch. This
+approach is natural for FA since the original aim
+is to label full sentences, and it is also commonly
+adopted in previous PA work (Mirroshandel and
+Nasr, 2011; Flannery and Mori, 2015; Li et al.,
+2016). Moreover, this approach makes it easier to
+control the reading context size for fair comparisons of different strategies as described in §3.2.
+Without loss of generality, we take sequence
+labeling as an example and illustrate several key
+points in the AL process. Other tasks follow similar
+treatment, with details provided in Appendix A.
+
+
+- **Model** . We adopt a standard BERT-based model
+with a CRF output layer for structured output
+modeling (Lafferty et al., 2001), together with
+the BIO tagging scheme.
+
+
+- **Querying** **Strategy** . We utilize the query-byuncertainty strategy with the margin-based metric, which has been shown effective in AL for
+structured prediction (Marcheggiani and Artières,
+2014; Li et al., 2016). Specifically, each token
+obtains an uncertainty score with the difference
+between the (marginal) probabilities of the most
+
+
+
+12992
+
+
+and second most likely label. We also tried several other strategies, such as least-confidence or
+max-entropy, but did not find obvious benefits. [1]
+
+
+- **Sentence** **selection** . For both FA and PA, selecting a batch of uncertain sentences is the first
+querying step. We use the number of total tokens
+to measure batch size since sentences may have
+variant lengths. The sentence-level uncertainty
+is obtained by averaging the token-level ones.
+This length normalization heuristic is commonly
+adopted to avoid biases towards longer sentences
+(Hwa, 2004; Shen et al., 2018).
+
+
+- **Token selection** . In PA, a subset of highly uncertain tokens is further chosen for annotation. One
+important question is how many tokens to select.
+Instead of using a pre-defined fixed selection criterion, we develop an adaptive strategy to decide
+the amount, as will be described in §2.2.
+
+
+- **Annotation** . Sequence labeling is usually
+adopted for tasks involving mention extraction,
+where annotations are over spans rather than individual tokens. Previous work explores subsequence querying (Chaudhary et al., 2019; Radmard et al., 2021), which brings further complexities. Since we mainly explore tasks with short
+mention spans, we adopt a simple annotation protocol: Labeling the full spans where any inside
+token is queried. Note that for annotation cost
+measurement, we also include the extra labeled
+tokens in addition to the queried ones.
+
+
+- **Model learning** . For FA, we adopt the standard
+log-likelihood as the training loss. For PA, we
+follow previous work (Scheffer et al., 2001; Wanvarie et al., 2011; Marcheggiani and Artières,
+2014) and adopt marginalized likelihood to learn
+from incomplete annotations (Tsuboi et al., 2008;
+Greenberg et al., 2018). More details are provided in Appendix C.
+
+**2.2** **Adaptive Partial Selection**
+
+PA adopts a second selection stage to choose highly
+uncertain sub-structures within the selected sentences. One crucial question here is how many
+
+1Please refer to Appendix D.1 for more results. Note that
+our main focus is on AL for structured prediction, where AL
+selection involves not only what instances to select (acquisition function), but also at what granularity to select and
+annotate. In contrast with most AL work that focuses on the
+first aspect (and classification tasks), we mainly investigate
+the second one and explore better partial selection strategies.
+Exploring more advanced acquisition functions is mostly orthogonal to our main focus and is left to future work.
+
+
+
+sub-structures to select. Typical solutions in previous work include setting an uncertainty threshold
+(Tomanek and Hahn, 2009) or specifying a selection ratio (Li et al., 2016). The threshold or ratio is
+usually pre-defined with a fixed hyper-parameter.
+This fixed selecting scheme might not be an ideal
+one. First, it is usually hard to specify such fixed
+values in practice. If too many sub-structures are
+selected, there will be little difference between FA
+and PA, whereas if too few, the annotation amount
+is insufficient to train good models. Moreover, this
+scheme is not adaptive to the model. As the model
+is trained with more data throughout the AL process, the informative sub-structures become less
+dense as the model improves. Thus, the number of
+selected sub-structures should be adjusted accordingly. To mitigate these shortcomings, we develop
+a dynamic strategy that can decide the selection in
+an automatic and adaptive way.
+We adopt the ratio-based strategy which enables
+straightforward control of the selected amount.
+Specifically, we rank the sub-structures by the uncertainty score and choose those scoring highest
+by the ratio. Our decision on the selecting ratio
+is based on the hypothesis that a reasonable ratio
+should roughly correspond to the current model’s
+error rate on all the candidates. The intuition is that
+incorrectly predicted sub-structures are the most informative ones that can help to correct the model’s
+mistakes.
+Since the queried instances come from the unlabeled pool without annotations, the error rate cannot be directly obtained, requiring estimation. [2] We
+adopt a simple one-dimensional logistic regression
+model for this purpose. The input to the model is
+the uncertainty score [3] and the output is a binary prediction of whether its prediction is confidently correct [4] or not. The estimator is trained using all the
+sub-structures together with their correctness on the
+development set [5] and then applied to the queried
+candidates. For each candidate sub-structure _s_, the
+estimator will give it a correctness probability. We
+
+
+2Directly using uncertainty is another option, but the main
+trouble is that the model is not well-calibrated. We also tried
+model calibration by temperature scaling (Guo et al., 2017),
+but did not find better results.
+3We transform the input with a logarithm, which leads to
+better estimation according to preliminary experiments.
+4The specific criterion is that the arg max prediction
+matches the gold one and its margin is greater than 0.5. Since
+neural models are usually over-confident, it is hard to decide a
+confidence threshold. Nevertheless, we find 0.5 a reasonable
+value for the ratio decision here.
+5We re-use the development set for the task model training.
+
+
+
+12993
+
+
+estimate the overall error rate as one minus the average correctness probability over all the candidates
+in the query set _Q_ (all sub-structures in the selected
+sentences), and set the selection ratio _r_ as this error
+rate:
+
+
+
+_r_ = 1
+_−_ _n_ [1]
+
+
+
+
+
+
+
+_p_ ( _correct_ = 1 _|s_ )
+_s∈Q_
+
+
+
+In this way, the selection ratio can be set adaptively according to the current model’s capability.
+If the model is weak and makes many mistakes,
+we will have a larger ratio which can lead to more
+dense annotations and richer training signals. As
+the model is trained with more data and makes
+fewer errors, the ratio will be tuned down correspondingly to avoid wasting annotation budget on
+already-correctly-predicted sub-structures. As we
+will see in later experiments, this adaptive scheme
+is suitable for AL (§3.3).
+
+
+**2.3** **Self-training**
+
+Better utilization of unlabeled data is a promising
+direction to further enhance model training in AL
+since unlabeled data are usually freely available
+from the unlabeled pool. In this work, we adopt
+self-training (Yarowsky, 1995) for this purpose.
+The main idea of self-training is to enhance the
+model training with pseudo labels that are predicted
+by the current model on the unlabeled data. It
+has been shown effective for various NLP tasks
+(Yarowsky, 1995; McClosky et al., 2006; He et al.,
+2020; Du et al., 2021). For the training of AL models, self-training can be seamlessly incorporated.
+For FA, the application of self-training is no different than that in the conventional scenarios by
+applying the current model to all the un-annotated
+instances in the unlabeled pool. The more interesting case is on the partially annotated instances
+in the PA regime. The same motivation from the
+adaptive ratio scheme (§2.2) also applies here: We
+select the highly-uncertain sub-structures that are
+error-prone and the remaining un-selected parts
+are likely to be correctly predicted; therefore we
+can trust the predictions on the un-selected substructures and include them for training. One more
+enhancement to apply here is that we could further
+perform re-inference by incorporating the updated
+annotations over the selected sub-structures, which
+can enhance the predictions of un-annotated substructures through output dependencies.
+In this work, we adopt a soft version of selftraining through knowledge distillation (KD; Hin
+
+
+ton et al., 2015). This choice is because we want to
+avoid the potential negative influences of ambiguous predictions (mostly in completely unlabeled
+instances). One way to mitigate this is to set an
+uncertainty threshold and only utilize the highlyconfident sub-structures. However, it is unclear
+how to set a proper value, similar to the scenarios
+in query selection. Therefore, we take the model’s
+full output predictions as the training targets without further processing.
+Specifically, our self-training objective function
+is the cross-entropy between the output distributions predicted by the previous model _m_ _[′]_ before
+training and the current model _m_ being trained:
+
+  = _pm′_ ( _y_ _x_ ) log _pm_ ( _y_ _x_ )
+_L_ _−_ _|_ _|_
+
+_y∈Y_
+
+Several points are notable here: 1) The previous
+model is kept unchanged, and we can simply cache
+its predictions before training; 2) Over the instances
+that have partial annotations, the predictions should
+reflect these annotations by incorporating corresponding constraints at inference time; 3) For tasks
+with CRF based models, the output space _Y_ is usually exponentially large and infeasible to explicitly
+enumerate; we utilize special algorithms (Wang
+et al., 2021) to deal with this, and more details are
+presented in Appendix C.
+Finally, we find it beneficial to include both the
+pseudo labels and the real annotated gold labels for
+the model training. With the gold data, the original
+training loss is adopted, while the KD objective
+is utilized with the pseudo labels. We simply mix
+these two types of data with a ratio of 1:1 in the
+training process, which we find works well.
+
+
+**3** **Experiments**
+
+**3.1** **Main Settings**
+
+**Tasks and data.** Our experiments [6] are conducted
+over four English tasks. The first two are named
+entity recognition (NER) and dependency parsing
+(DPAR), which are representative structured prediction tasks for predicting sequence and tree structures. We adopt the CoNLL-2003 English dataset
+(Tjong Kim Sang and De Meulder, 2003) for NER
+and the English Web Treebank (EWT) from Universal Dependencies v2.10 (Nivre et al., 2020) for
+DPAR. Moreover, we explore two more complex
+IE tasks: Event extraction and relation extraction.
+
+6Our implementation is available at [https://github.](https://github.com/zzsfornlp/zmsp/.)
+[com/zzsfornlp/zmsp/.](https://github.com/zzsfornlp/zmsp/.)
+
+
+
+12994
+
+
+Each task involves two pipelined sub-tasks: The
+first aims to extract the event trigger and/or entity
+mentions, and the second predicts links between
+these mentions as event arguments or entity relations. We utilize the ACE05 dataset (Walker et al.,
+2006) for these IE tasks.
+
+
+**AL.** For the AL procedure, we adopt settings following conventional practices. We use the original
+training set as the unlabeled data pool to select instances. Unless otherwise noted, we set the AL
+batch size (for sentence selection) to 4K tokens,
+which roughly corresponds to 2% of the total pool
+size for most of the datasets we use. The initial seed
+training set and the development set are randomly
+sampled (with FA) using this batch size. Unless
+otherwise noted, we run 14 AL cycles for each experiment. In each AL cycle, we re-train our model
+since we find incremental updating does not perform well. Following most AL work, annotation
+is simulated by checking and assigning the labels
+from the original dataset. In FA, we annotate all
+the sub-structures for the selected sentences. In PA,
+we first decide the selection ratio and apply it to the
+selected sentences. We further adopt a heuristic [7]
+
+that selects the union of sentence-wise uncertain
+sub-structures as well as global ones since both
+may contain informative sub-structures. Finally, all
+the presented results are averaged over five runs
+with different random seeds.
+
+
+**Model and training.** For the models, we adopt
+standard architectures by stacking task-specific
+structured predictors over pre-trained RoBERTabase
+(Liu et al., 2019) and the full models are fine-tuned
+at each training iteration. After obtaining new annotations in each AL cycle, we first train a model
+based on all the available full or partial annotations. When using self-training, we further apply this newly trained model to assign pseudo soft
+labels to all un-annotated instances and combine
+them with the existing annotations to train another
+model. Compared to using the old model from the
+last AL cycle, this strategy can give more accurate pseudo labels since the newly updated model
+usually performs better by learning from more annotations. For PA, pseudo soft labels are assigned
+to both un-selected sentences and the un-annotated
+sub-structures in the selected sentences.
+
+
+7This heuristic will increase the actual selecting ratio, but
+it will only be slightly larger since there are large overlaps between sentence-wise and global highly-ranked sub-structures.
+
+
+
+**3.2** **Comparison Scheme**
+
+Since FA and PA annotate at different granularities,
+we need a common cost measurement to compare
+their effectiveness properly. A reasonable metric
+is the number of the labeled sub-structures; for
+instance, the number of labeled tokens for sequence
+labeling or edges for dependency parsing. This
+metric is commonly adopted in previous PA work
+(Tomanek and Hahn, 2009; Flannery and Mori,
+2015; Li et al., 2016; Radmard et al., 2021).
+Nevertheless, evaluating only by sub-structures
+ignores a crucial hidden cost: The reading time of
+the contexts. For example, in sequence labeling
+with PA, although not every token in the sentence
+needs to be tagged, the annotator may still need to
+read the whole sentence to understand its meaning.
+Therefore, if performing comparisons only by the
+amount of annotated sub-structures, it will be unfair
+for the FA baseline because more contexts must be
+read to carry out PA.
+In this work, we adopt a simple two-facet comparison scheme that considers both reading and
+labeling costs. We first control the reading cost
+by choosing the same size of contexts in the sentence selection step of each AL cycle (Line 4 in
+Algorithm 1). Then, we further compare by the
+sub-structure labeling cost, measured by the substructure annotation cost. If PA can roughly reach
+the FA performance with the same reading cost but
+fewer sub-structures annotated, it would be fair to
+say that PA can help reduce cost over FA. A better
+comparing scheme should evaluate against a unified estimation of the real annotation costs (Settles
+et al., 2008). This usually requires actual annotation exercises rather than simulations, which we
+leave to future work.
+
+
+**3.3** **NER and DPAR**
+
+**Settings.** We compare primarily three strategies:
+FA, PA, and a baseline where randomly selected
+sentences are fully annotated (Rand). We also include a supervised result (Super.) which is obtained
+from a model trained with the full original training
+set. We measure reading cost by the total number
+of tokens in the selected sentences. For labeling
+cost, we further adopt metrics with practical considerations. In NER, lots of tokens, such as functional
+words, can be easily judged as the ‘O’ (non-entity)
+tag. To avoid over-estimating the costs of such
+easy tokens for FA, we filter tokens by their partof-speech (POS) tags and only count the ones that
+
+
+
+12995
+
+
+92
+
+91
+
+90
+
+89
+
+88
+
+87
+
+86
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|
+|---|---|---|---|---|---|---|---|---|---|---|---|
+|||||||||||||
+||||||||||Rand<br>|Rand<br>|Rand<br>|
+|||||||||||||
+||||||||||Rand+S<br>~~FA~~|Rand+S<br>~~FA~~|T|
+|||||||||||||
+||||||||||FA+ST<br>|FA+ST<br>||
+|||||||||||||
+||||||||||~~PA~~<br>PA+ST|~~PA~~<br>PA+ST||
+||||||||||Super.|Super.||
+|||||||||||||
+|||||||||||||
+
+
+4000 12000 20000 28000 36000 44000 52000 60000
+Token Count
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|
+|---|---|---|---|---|---|---|---|---|---|---|---|
+|||||||||||||
+|||||||||||||
+|||||||||||||
+||||||||||~~Rand~~<br>Rand+S<br>~~FA~~|~~Rand~~<br>Rand+S<br>~~FA~~|T|
+|||||||||||||
+||||||||||FA+ST<br>PA<br>|FA+ST<br>PA<br>||
+|||||||||||||
+|||||||||||||
+||||||||||PA+ST<br>Super.|PA+ST<br>Super.||
+|||||||||||||
+|||||||||||||
+|||||||||||||
+
+
+
+Token Count
+
+
+
+92
+
+91
+
+90
+
+89
+
+88
+
+87
+
+86
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
+|---|---|---|---|---|---|---|---|---|---|
+||||||||||and<br>|
+||||||||R<br>|R<br>|R<br>|
+||||||||R<br>~~F~~|R<br>~~F~~|and+ST<br>|
+||||||||F<br>|F<br>|A+ST<br>|
+||||||||~~P~~<br>P|~~P~~<br>P|+ST|
+||||||||Su|Su|per.|
+|||||||||||
+
+
+0 2500 5000 7500 10000 12500 15000
+Sub-structure Count
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|
+|---|---|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||||
+||||||||||||||
+||||||||||||||
+|||||||||||~~Rand~~<br>Rand+<br>~~FA~~|~~Rand~~<br>Rand+<br>~~FA~~|ST|
+|||||||||||FA+S<br>PA<br>|FA+S<br>PA<br>|T<br>|
+|||||||||||PA+S<br>Super.|PA+S<br>Super.|T|
+||||||||||||||
+
+
+
+Sub-structure Count
+
+
+
+Figure 2: Comparisons according to reading and labeling cost. Each node indicates one AL cycle. For _x_ -axis,
+reading cost (left) is measured by token numbers, while labeling cost (right) is task-specific (§3.3). NER is evaluated
+with labeled F1 scores on CoNLL-2003, while DPAR is with LAS scores on UD-EWT. Results are averaged over
+five runs with different seeds, and the shaded areas indicate standard deviations. The overall unlabeled pool contains
+around 200K tokens. Using AL, good performance can be obtained with less than 30% (60K) annotated.
+
+
+
+are likely to be inside an entity mention. [8] For PA,
+we still count every queried token. For the task of
+DPAR, similarly, different dependency links can
+have variant annotation difficulties. We utilize the
+surface distance between the head and modifier
+of the dependency edge as the measure of labeling cost, considering that the decisions for longer
+dependencies are usually harder.
+
+
+**Main Results.** The main test results are shown
+in Figure 2, where the patterns on both tasks are
+similar. First, AL brings clear improvements over
+the random baseline and can roughly reach the fully
+supervised performance with only a small portion
+of data annotated (around 18% for CoNLL-2003
+and 30% for UD-EWT). Moreover, self-training
+(+ST) is helpful for all the strategies, boosting performance without the need for extra manual annotations. Finally, with the help of self-training, the PA
+strategy can roughly match the performance of FA
+with the same amount of reading cost (according to
+the left figures) while labeling fewer sub-structures
+(according to the right figures). This indicates that
+PA can help to further reduce annotation costs over
+the strong FA baselines.
+
+
+8The POS tags are assigned by Stanza (Qi et al., 2020).
+For CoNLL-2003, we filter by PROPN and ADJ, which cover
+more than 95% of the entity tokens.
+
+
+
+**Ratio** **Analysis.** We further analyze the effectiveness of our adaptive ratio scheme with DPAR
+as the case study. We compare the adaptive scheme
+to schemes with fixed ratio _r_, and the results [9] are
+shown in Figure 3. For the fixed-ratio schemes,
+if the value is too small (such as 0.1), although
+its improving speed is the fastest at the beginning,
+its performance lags behind others with the same
+reading contexts due to fewer sub-structures annotated. If the value is too large (such as 0.5), it grows
+slowly, probably because too many uninformative
+sub-structures are annotated. The fixed scheme
+with _r_ = 0 _._ 3 seems a good choice; however, it is
+unclear how to find this sweet spot in realistic AL
+processes. The adaptive scheme provides a reasonable solution by automatically deciding the ratio
+according to the model performance.
+
+
+**Error and Uncertainty Analysis.** We further analyze the error rates and uncertainties of the queried
+sub-structures. We still take DPAR as a case study
+and Figure 4 shows the results along the AL cycles
+in PA mode. First, though adopting a simple model,
+the performance predictor can give reasonable estimations for the overall error rates. Moreover, by
+further breaking down the error rates into selected
+
+
+9Here, we use self-training (+ST) for all the strategies.
+
+
+
+12996
+
+
+error(S)
+
+
+0.8
+
+
+0.6
+
+
+0.4
+
+
+0.2
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
+|---|---|---|---|---|---|---|---|---|---|
+|||||||||r=0.1<br>r=0.3<br>~~r=0.5~~<br>Adapti<br>|ve|
+|||||||||||
+|||||||||||
+|||||||||Super.||
+|||||||||||
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|
+|---|---|---|---|---|---|---|
+||||||||
+||||||||
+||||||||
+||||||||
+
+
+
+0.0
+1 3 5 7 9 1113
+
+
+
+margin
+
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+Token Count
+
+
+DPAR Labeling Cost
+0.50
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|r=0<br>r=0|.1<br>.3|
+|---|---|---|---|---|---|---|---|
+|||||||~~r=~~<br>Ad<br>Su|~~.5~~<br>aptive<br>er.<br><br>|
+|||||||||
+|||||||||
+
+
+Sub-structure Count
+
+
+
+pred
+
+
+1 3 5 7 9 1113
+
+
+
+
+
+Figure 3: Comparisons of different strategies to decide
+the partial ratio. The first three utilize fixed ratio _r_,
+while “Adaptive” adopts the dynamic scheme. The grey
+curve (corresponding to the right _y_ -axis) denotes the
+actual selection ratios with the adaptive scheme.
+
+
+(S) and non-selected (N) groups, we can see that
+the selected ones contain many errors, indicating
+the need for manual corrections. On the other hand,
+the error rates on the non-selected sub-structures
+are much lower, verifying the effectiveness of using model-predicted pseudo labels on them in selftraining. Finally, the overall margin of the selected
+sentences keeps increasing towards 1, indicating
+that there are many non-ambiguous sub-structures
+even in highly-uncertain sentences. The margins
+of the selected sub-structures are much lower, suggesting that annotating them could provide more
+informative signals for model training.
+
+
+**Domain-transfer Experiments.** We further investigate a domain-transfer scenario: in addition to
+unlabeled in-domain data, we assume abundant outof-domain annotated data and perform AL on the
+target domain. We adopt tweet texts as the target
+domain, using Broad Twitter Corpus (BTC; Derczynski et al., 2016) for NER and Tweebank (Liu
+et al., 2018) for DPAR. We assume we have models
+trained from a richly-annotated source domain and
+continue performing AL on the target domain. The
+source domains are the datasets that we utilize in
+our main experiments: CoNLL03 for NER and UDEWT for DPAR. We adopt a simple model-transfer
+
+
+
+Figure 4: Analyses of error rates and uncertainties (margins) of the DPAR sub-structures in the queried sentences along the AL cycles ( _x_ -axis). Here, ‘pred’ denotes the predicted error rate, ‘error’ denotes the actual
+error rate and ‘margin’ denotes the uncertainty (margin)
+scores. For the suffixes, ‘(S)’ indicates partially selected
+sub-structures, and ‘(N)’ indicates non-selected ones.
+‘Margin(N)’ is omitted since it is always close to 1.
+
+
+approach by initializing the model from the one
+trained with the source data and further fine-tuning
+it with the target data. Since the target data size
+is small, we reduce the AL batch sizes for BTC
+and Tweebank to 2000 and 1000 tokens, respectively. The results for these experiments are shown
+in Figure 5. In these experiments, we also include
+the no-transfer results, adopting the “FA+ST” but
+without model transfer. For NER, without transfer
+learning, the results are generally worse, especially
+in early AL stages, where there is a small amount of
+annotated data to provide training signals. In these
+cases, knowledge learned from the source domain
+can provide extra information to boost the results.
+For DPAR, we can see even larger benefits of using
+transfer learning; there are still clear gaps between
+transfer and no-transfer strategies when the former
+already reaches the supervised performance. These
+results indicate that the benefits of AL and transfer
+learning can be orthogonal, and combining them
+can lead to promising results.
+
+
+**3.4** **Information Extraction**
+
+We further explore more complex IE tasks that involve multiple types of output. Specifically, we
+investigate event extraction and relation extraction.
+We adopt a classical pipelined approach, [10] which
+splits the full task into two sub-tasks: the first performs mention extraction, while the second examines mention pairs and predicts relations. While
+
+
+10Please refer to Appendix A for more task-specific details.
+
+
+
+12997
+
+
+82
+
+
+80
+
+
+78
+
+
+76
+
+
+74
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||Rand<br>||
+||||||||||~~Rand+S~~<br>FA<br>~~FA+ST~~||
+||||||||||PA<br>PA+ST||
+||||||||||NoTransf<br>Super.|er|
+||||||||||||
+
+
+Token Count
+
+
+
+82
+
+
+80
+
+
+78
+
+
+76
+
+
+74
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+78
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|
+|---|---|---|---|---|---|---|---|
+|||||||Rand<br>||
+|||||||||
+|||||||~~Rand+S~~<br>FA<br>~~FA+ST~~||
+|||||||||
+|||||||PA<br>PA+ST||
+|||||||||
+|||||||||
+|||||||NoTransf<br>Super.|er|
+|||||||||
+|||||||||
+|||||||||
+
+
+Sub-structure Count
+
+
+
+
+
+60
+
+
+50
+
+
+40
+
+
+30
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||||||
+||||||||||||||||
+|||||||||||||Rand<br>|Rand<br>|Rand<br>|
+|||||||||||||~~Rand+ST~~<br>FA<br>FA+ST|~~Rand+ST~~<br>FA<br>FA+ST||
+|||||||||||||PA<br>PA+ST<br>|PA<br>PA+ST<br>||
+|||||||||||||~~NoTransf~~<br>Super.|~~NoTransf~~<br>Super.|~~r~~|
+||||||||||||||||
+|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|1000<br>3000<br>5000<br>7000<br>9000<br>11000<br>13000<br>15000<br>Token Count<br>: AL results in domain-transfer settings (C<br> Notations are the same as in Figure 2, exce<br> here no transfer learning is applied (FA+ST<br>EventArgument Reading Cost|
+||||||||||||||||
+||||||||||||||||
+|||||||||||||Rand<br>|Rand<br>||
+||||||||||||||||
+|||||||||||||Rand+S<br>FA<br>FA+ST<br>|Rand+S<br>FA<br>FA+ST<br>|T|
+||||||||||||||||
+||||||||||||||||
+|||||||||||||PA<br>PA+ST<br>Super.|PA<br>PA+ST<br>Super.||
+||||||||||||||||
+||||||||||||||||
+||||||||||||||||
+||||||||||||||||
+
+
+Token Count
+
+
+
+60
+
+
+50
+
+
+40
+
+
+30
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||
+||||||||||||
+||||||||||||
+||||||||~~Rand~~<br>FA<br>FA+S|~~Rand~~<br>FA<br>FA+S|~~ST~~<br>T|~~ST~~<br>T|
+||||||||PA<br>PA+S<br>|PA<br>PA+S<br>|T<br>|T<br>|
+||||||||~~NoTr~~<br>Super|~~NoTr~~<br>Super|~~nsfer~~<br>.|~~nsfer~~<br>.|
+||||||||||||
+|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|0<br>10000<br>20000<br>30000<br>40000<br>50000<br>Sub-structure Count<br> _→_BTC for NER and UD-EWT _→_Tweeb<br>  re is one more curve of “NoTransfer” denot<br>EventArgument Labeling Cost|
+||||||||||||
+||||||||||||
+|||||||||Rand<br>|Rand<br>||
+|||||||||Rand+<br>FA<br>FA+ST<br>|Rand+<br>FA<br>FA+ST<br>|T|
+|||||||||PA<br>PA+ST<br>Super.|PA<br>PA+ST<br>Super.||
+||||||||||||
+
+
+Sub-structure Count
+
+
+
+Figure 6: Results of event argument extraction on ACE05. Notations are the same as in Figure 2.
+
+
+
+previous work investigates multi-task AL with FA
+(Reichart et al., 2008; Zhu et al., 2020; Rotman and
+Reichart, 2022), this work is the first to explore PA
+in this challenging setting.
+We extend our PA scheme to this multi-task scenario with several modifications. First, for the
+sentence-selection stage, we obtain a sentence-wise
+uncertainty score UNC( _x_ ) with a weighted combination of the two sub-tasks’ uncertainty scores:
+
+
+UNC( _x_ ) = _β ·_ UNC-Mention( _x_ )
+
++ (1 _−_ _β_ ) _·_ UNC-Relation( _x_ )
+
+Following Rotman and Reichart (2022), we set _β_
+to a relatively large value (0.9), which is found to
+be helpful for the second relational sub-task.
+Moreover, for partial selection, we separately select sub-structures for the two sub-tasks according
+to the adaptive selection scheme. Since the second
+
+
+
+relational sub-task depends on the mentions extracted from the first sub-task, we utilize predicted
+mentions and view each feasible mention pair as a
+querying candidate. A special annotation protocol
+is adopted to deal with the incorrectly predicted
+mentions. For each queried relation, we first examine its mentions and perform corrections if there
+are mention errors that can be fixed by matching
+the gold ones. If neither of the two mentions can
+be corrected, we discard this query.
+Finally, to compensate for the influences of errors in mention extraction, we adopt further heuristics of increasing the partial ratio by the estimated
+percentage of queries with incorrect mentions, as
+well as including a second annotation stage with
+queries over newly annotated mentions. Please
+refer to Appendix A.2 for more details.
+We show the results of event argument extraction
+in Figure 6, where the overall trends are similar to
+
+
+
+12998
+
+
+those in NER and DPAR. Here, labeling cost is
+simply measured as the number of candidate argument links. Overall, self-training is helpful for
+all AL strategies, indicating the benefits of making
+better use of unlabeled data. If measured by the
+labeling cost, PA learns the fastest and costs only
+around half of the annotated arguments of FA to
+reach the supervised result. On the other hand, PA
+is also competitive concerning reading cost and can
+generally match the FA results in later AL stages.
+There is still a gap between PA and FA in the earlier
+AL stages, which may be influenced by the errors
+produced by the first sub-task of mention extraction. We leave further investigations on improving
+early AL stages to future work. The results for
+the relation extraction task share similar trends and
+are presented in Appendix D.2, together with the
+results of mention extraction.
+
+
+**4** **Related Work**
+
+
+**Self-training.** Self-training is a commonly utilized semi-supervised method to incorporate unlabeled data. It has been shown effective for a variety
+of NLP tasks, including word sense disambiguation (Yarowsky, 1995), parsing (McClosky et al.,
+2006), named entity recognition (Meng et al., 2021;
+Huang et al., 2021), text generation (He et al., 2020;
+Mehta et al., 2022) as well as natural language
+understanding (Du et al., 2021). Moreover, selftraining can be especially helpful for low-resource
+scenarios, such as in few-shot learning (Vu et al.,
+2021; Chen et al., 2021). Self-training has also
+been a commonly adopted strategy to enhance active learning (Tomanek and Hahn, 2009; Majidi
+and Crane, 2013; Yu et al., 2022).
+
+
+**PA.** Learning from incomplete annotations has
+been well-explored for structured prediction. For
+CRF models, taking the marginal likelihood as the
+objective function has been one of the most utilized techniques (Tsuboi et al., 2008; Täckström
+et al., 2013; Yang and Vozila, 2014; Greenberg
+et al., 2018). There are also other methods to
+deal with incomplete annotations, such as adopting local models (Neubig and Mori, 2010; Flannery et al., 2011), max-margin objective (Fernandes and Brefeld, 2011), learning with constraints
+(Ning et al., 2018, 2019; Mayhew et al., 2019) and
+negative sampling (Li et al., 2022).
+
+
+**AL for structured prediction.** AL has been investigated for various structured prediction tasks in
+
+
+
+NLP, such as sequence labeling (Settles and Craven,
+2008; Shen et al., 2018), parsing (Hwa, 2004), semantic role labeling (Wang et al., 2017; Myers
+and Palmer, 2021) and machine translation (Haffari
+et al., 2009; Zeng et al., 2019). While most previous work adopt FA, that is, annotating full structured objects for the inputs, PA can help to further
+reduce the annotation cost. Typical examples of PA
+sub-structures include tokens and subsequences for
+tagging (Marcheggiani and Artières, 2014; Chaudhary et al., 2019; Radmard et al., 2021), word-wise
+head edges for dependency parsing (Flannery and
+Mori, 2015; Li et al., 2016) and mention links for
+coreference resolution (Li et al., 2020; Espeland
+et al., 2020).
+
+
+**5** **Conclusion**
+
+
+In this work, we investigate better AL strategies
+for structured prediction in NLP, adopting a performance estimator to automatically decide suitable
+ratios for partial sub-structure selection and utilizing self-training to make better use of the available
+unlabeled data pool. With comprehensive experiments on various tasks, we show that the combination of PA and self-training can be more dataefficient than strong full AL baselines.
+
+
+**Limitations**
+
+
+This work has several limitations. First, the AL
+experiments in this work are based on simulations
+with existing annotations, following previous AL
+work. Our error estimator also requires a small
+development set and the proper setting of a hyperparameter. Nevertheless, we tried our best to make
+the settings practical and the evaluation fair, especially taking reading time into consideration. Second, in our experiments, we mainly focus on investigating how much data is needed to reach the fullysupervised results and continue the AL cycles until
+this happens. In practice, it may be interesting to
+more carefully examine the early AL stages, where
+most of the performance improvements happen. Finally, for the IE tasks with multiple output types,
+we mainly focus on the second relational sub-task
+and adopt a simple weighting setting to combine
+the uncertainties of the two sub-tasks. More explorations on the dynamic balancing of the two subtasks in pipelined models (Roth and Small, 2008)
+would be an interesting direction for future work.
+
+
+
+12999
+
+
+**References**
+
+Leonard E Baum, Ted Petrie, George Soules, and Norman Weiss. 1970. A maximization technique occurring in the statistical analysis of probabilistic functions of markov chains. _The annals of mathematical_
+_statistics_, 41(1):164–171.
+
+
+Aditi Chaudhary, Jiateng Xie, Zaid Sheikh, Graham
+Neubig, and Jaime Carbonell. 2019. [A little annota-](https://doi.org/10.18653/v1/D19-1520)
+tion does a lot of good: [A study in bootstrapping low-](https://doi.org/10.18653/v1/D19-1520)
+[resource named entity recognizers.](https://doi.org/10.18653/v1/D19-1520) In _Proceedings_
+_of the 2019 Conference on Empirical Methods in Nat-_
+_ural Language Processing and the 9th International_
+_Joint Conference on Natural Language Processing_
+_(EMNLP-IJCNLP)_, pages 5164–5174, Hong Kong,
+China. Association for Computational Linguistics.
+
+
+Yiming Chen, Yan Zhang, Chen Zhang, Grandee Lee,
+Ran Cheng, and Haizhou Li. 2021. [Revisiting self-](https://doi.org/10.18653/v1/2021.emnlp-main.718)
+training for few-shot [learning](https://doi.org/10.18653/v1/2021.emnlp-main.718) of language model.
+In _Proceedings_ _of_ _the_ _2021_ _Conference_ _on_ _Empir-_
+_ical Methods in Natural Language Processing_, pages
+9125–9135, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.
+
+
+Leon Derczynski, Kalina Bontcheva, and Ian Roberts.
+2016. Broad Twitter corpus: [A diverse named entity](https://aclanthology.org/C16-1111)
+[recognition](https://aclanthology.org/C16-1111) resource. In _Proceedings_ _of_ _COLING_
+_2016, the 26th International Conference on Compu-_
+_tational Linguistics:_ _Technical Papers_, pages 1169–
+1179, Osaka, Japan. The COLING 2016 Organizing
+Committee.
+
+
+Jacob Devlin, Ming-Wei Chang, Kenton Lee, and
+Kristina Toutanova. 2019. BERT: [Pre-training](https://doi.org/10.18653/v1/N19-1423) of
+[deep bidirectional transformers for language under-](https://doi.org/10.18653/v1/N19-1423)
+[standing.](https://doi.org/10.18653/v1/N19-1423) In _Proceedings of the 2019 Conference of_
+_the North American Chapter of the Association for_
+_Computational Linguistics:_ _Human Language Tech-_
+_nologies, Volume 1 (Long and Short Papers)_, pages
+4171–4186, Minneapolis, Minnesota. Association for
+Computational Linguistics.
+
+
+Timothy Dozat and Christopher D. Manning. 2017.
+Deep biaffine attention for neural dependency parsing. In _ICLR_ .
+
+
+Jingfei Du, Edouard Grave, Beliz Gunel, Vishrav Chaudhary, Onur Celebi, Michael Auli, Veselin Stoyanov,
+and Alexis Conneau. 2021. [Self-training improves](https://doi.org/10.18653/v1/2021.naacl-main.426)
+[pre-training for natural language understanding.](https://doi.org/10.18653/v1/2021.naacl-main.426) In
+_Proceedings_ _of_ _the_ _2021_ _Conference_ _of_ _the_ _North_
+_American Chapter of the Association for Computa-_
+_tional Linguistics:_ _Human Language Technologies_,
+pages 5408–5418, Online. Association for Computational Linguistics.
+
+
+Vebjørn Espeland, Beatrice Alex, and Benjamin Bach.
+2020. Enhanced labelling [in](https://aclanthology.org/2020.crac-1.12) active learning for
+[coreference](https://aclanthology.org/2020.crac-1.12) resolution. In _Proceedings_ _of_ _the_
+_Third_ _Workshop_ _on_ _Computational_ _Models_ _of_ _Ref-_
+_erence, Anaphora and Coreference_, pages 111–121,
+Barcelona, Spain (online). Association for Computational Linguistics.
+
+
+
+Eraldo R Fernandes and Ulf Brefeld. 2011. Learning from partially annotated sequences. In _Joint_
+_European_ _Conference_ _on_ _Machine_ _Learning_ _and_
+_Knowledge Discovery in Databases_, pages 407–422.
+Springer.
+
+
+Daniel Flannery, Yusuke Miayo, Graham Neubig, and
+Shinsuke Mori. 2011. [Training dependency parsers](https://aclanthology.org/I11-1087)
+[from partially annotated corpora.](https://aclanthology.org/I11-1087) In _Proceedings of_
+_5th International Joint Conference on Natural Lan-_
+_guage Processing_, pages 776–784, Chiang Mai, Thailand. Asian Federation of Natural Language Processing.
+
+
+Daniel Flannery and Shinsuke Mori. 2015. [Combin-](https://doi.org/10.18653/v1/W15-2202)
+[ing active learning and partial annotation for domain](https://doi.org/10.18653/v1/W15-2202)
+[adaptation of a Japanese dependency parser.](https://doi.org/10.18653/v1/W15-2202) In _Pro-_
+_ceedings_ _of_ _the_ _14th_ _International_ _Conference_ _on_
+_Parsing_ _Technologies_, pages 11–19, Bilbao, Spain.
+Association for Computational Linguistics.
+
+
+Nathan Greenberg, Trapit Bansal, Patrick Verga, and Andrew McCallum. 2018. [Marginal likelihood training](https://doi.org/10.18653/v1/D18-1306)
+[of BiLSTM-CRF for biomedical named entity recog-](https://doi.org/10.18653/v1/D18-1306)
+[nition from disjoint label sets.](https://doi.org/10.18653/v1/D18-1306) In _Proceedings of the_
+_2018 Conference on Empirical Methods in Natural_
+_Language_ _Processing_, pages 2824–2829, Brussels,
+Belgium. Association for Computational Linguistics.
+
+
+Chuan Guo, Geoff Pleiss, Yu Sun, and Kilian Q Weinberger. 2017. On calibration of modern neural networks. In _International conference on machine learn-_
+_ing_, pages 1321–1330. PMLR.
+
+
+Gholamreza Haffari, Maxim Roy, and Anoop Sarkar.
+2009. Active learning for [statistical](https://aclanthology.org/N09-1047) phrase-based
+[machine translation.](https://aclanthology.org/N09-1047) In _Proceedings of Human Lan-_
+_guage Technologies:_ _The 2009 Annual Conference of_
+_the North American Chapter of the Association for_
+_Computational Linguistics_, pages 415–423, Boulder,
+Colorado. Association for Computational Linguistics.
+
+
+Junxian He, Jiatao Gu, Jiajun Shen, and Marc’Aurelio
+Ranzato. 2020. Revisiting [self-training](https://openreview.net/forum?id=SJgdnAVKDH) for neural
+[sequence generation.](https://openreview.net/forum?id=SJgdnAVKDH) In _International Conference on_
+_Learning Representations_ .
+
+
+Geoffrey Hinton, Oriol Vinyals, Jeff Dean, et al. 2015.
+Distilling the knowledge in a neural network. _arXiv_
+_preprint arXiv:1503.02531_, 2(7).
+
+
+Neil Houlsby, Ferenc Huszár, Zoubin Ghahramani, and
+Máté Lengyel. 2011. Bayesian active learning for
+classification and preference learning. _arXiv preprint_
+_arXiv:1112.5745_ .
+
+
+Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien
+Jose, Shobana Balakrishnan, Weizhu Chen, Baolin
+Peng, Jianfeng Gao, and Jiawei Han. 2021. [Few-](https://doi.org/10.18653/v1/2021.emnlp-main.813)
+[shot named entity recognition:](https://doi.org/10.18653/v1/2021.emnlp-main.813) An empirical baseline
+[study.](https://doi.org/10.18653/v1/2021.emnlp-main.813) In _Proceedings_ _of_ _the_ _2021_ _Conference_ _on_
+_Empirical_ _Methods_ _in_ _Natural_ _Language_ _Process-_
+_ing_, pages 10408–10423, Online and Punta Cana,
+Dominican Republic. Association for Computational
+Linguistics.
+
+
+
+13000
+
+
+Rebecca Hwa. 2004. Sample [selection](https://doi.org/10.1162/0891201041850894) for statistical
+[parsing.](https://doi.org/10.1162/0891201041850894) _Computational Linguistics_, 30(3):253–276.
+
+
+Terry Koo, Amir Globerson, Xavier Carreras, and
+Michael Collins. 2007. [Structured prediction models](https://aclanthology.org/D07-1015)
+[via the matrix-tree theorem.](https://aclanthology.org/D07-1015) In _Proceedings of the_
+_2007 Joint Conference on Empirical Methods in Nat-_
+_ural Language Processing and Computational Nat-_
+_ural_ _Language_ _Learning_ _(EMNLP-CoNLL)_, pages
+141–150, Prague, Czech Republic. Association for
+Computational Linguistics.
+
+
+John D Lafferty, Andrew McCallum, and Fernando CN
+Pereira. 2001. Conditional random fields: Probabilistic models for segmenting and labeling sequence
+data. In _Proceedings of the Eighteenth International_
+_Conference on Machine Learning_, pages 282–289.
+
+
+Belinda Z. Li, Gabriel Stanovsky, and Luke Zettlemoyer.
+2020. [Active learning for coreference resolution us-](https://doi.org/10.18653/v1/2020.acl-main.738)
+[ing discrete annotation.](https://doi.org/10.18653/v1/2020.acl-main.738) In _Proceedings of the 58th_
+_Annual Meeting of the Association for Computational_
+_Linguistics_, pages 8320–8331, Online. Association
+for Computational Linguistics.
+
+
+Yangming Li, Lemao Liu, and Shuming Shi. 2022. [Re-](https://doi.org/10.18653/v1/2022.acl-long.497)
+[thinking negative sampling for handling missing en-](https://doi.org/10.18653/v1/2022.acl-long.497)
+[tity annotations.](https://doi.org/10.18653/v1/2022.acl-long.497) In _Proceedings of the 60th Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics (Volume 1:_ _Long Papers)_, pages 7188–7197,
+Dublin, Ireland. Association for Computational Linguistics.
+
+
+Zhenghua Li, Min Zhang, Yue Zhang, Zhanyi Liu, Wenliang Chen, Hua Wu, and Haifeng Wang. 2016. [Ac-](https://doi.org/10.18653/v1/P16-1033)
+[tive learning for dependency parsing with partial an-](https://doi.org/10.18653/v1/P16-1033)
+[notation.](https://doi.org/10.18653/v1/P16-1033) In _Proceedings of the 54th Annual Meeting_
+_of the Association for Computational Linguistics (Vol-_
+_ume 1:_ _Long Papers)_, pages 344–354, Berlin, Germany. Association for Computational Linguistics.
+
+
+Ying Lin, Heng Ji, Fei Huang, and Lingfei Wu. 2020.
+
+[A joint neural model for information extraction with](https://doi.org/10.18653/v1/2020.acl-main.713)
+[global features.](https://doi.org/10.18653/v1/2020.acl-main.713) In _Proceedings of the 58th Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics_, pages 7999–8009, Online. Association for
+Computational Linguistics.
+
+
+Yijia Liu, Yi Zhu, Wanxiang Che, Bing Qin, Nathan
+Schneider, and Noah A. Smith. 2018. [Parsing tweets](https://doi.org/10.18653/v1/N18-1088)
+[into Universal Dependencies.](https://doi.org/10.18653/v1/N18-1088) In _Proceedings of the_
+_2018 Conference of the North American Chapter of_
+_the Association for Computational Linguistics:_ _Hu-_
+_man_ _Language_ _Technologies,_ _Volume_ _1_ _(Long_ _Pa-_
+_pers)_, pages 965–975, New Orleans, Louisiana. Association for Computational Linguistics.
+
+
+Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis,
+Luke Zettlemoyer, and Veselin Stoyanov. 2019.
+Roberta: A robustly optimized bert pretraining approach. _arXiv preprint arXiv:1907.11692_ .
+
+
+Saeed Majidi and Gregory Crane. 2013. [Active learning](https://aclanthology.org/W13-5711)
+[for dependency parsing by a committee of parsers.](https://aclanthology.org/W13-5711) In
+
+
+
+_Proceedings of the 13th International Conference on_
+_Parsing Technologies (IWPT 2013)_, pages 98–105,
+Nara, Japan. Assocation for Computational Linguistics.
+
+
+Diego Marcheggiani and Thierry Artières. 2014. [An](https://doi.org/10.3115/v1/D14-1097)
+[experimental comparison of active learning strategies](https://doi.org/10.3115/v1/D14-1097)
+for partially [labeled](https://doi.org/10.3115/v1/D14-1097) sequences. In _Proceedings_ _of_
+_the 2014 Conference on Empirical Methods in Natu-_
+_ral Language Processing (EMNLP)_, pages 898–906,
+Doha, Qatar. Association for Computational Linguistics.
+
+
+Stephen Mayhew, Snigdha Chaturvedi, Chen-Tse Tsai,
+and Dan Roth. 2019. [Named entity recognition with](https://doi.org/10.18653/v1/K19-1060)
+partially [annotated](https://doi.org/10.18653/v1/K19-1060) training data. In _Proceedings_
+_of_ _the_ _23rd_ _Conference_ _on_ _Computational_ _Natural_
+_Language Learning (CoNLL)_, pages 645–655, Hong
+Kong, China. Association for Computational Linguistics.
+
+
+David McClosky, Eugene Charniak, and Mark Johnson.
+2006. [Effective self-training for parsing.](https://aclanthology.org/N06-1020) In _Proceed-_
+_ings of the Human Language Technology Conference_
+_of_ _the_ _NAACL,_ _Main_ _Conference_, pages 152–159,
+New York City, USA. Association for Computational
+Linguistics.
+
+
+Ryan McDonald and Giorgio Satta. 2007. On the
+[complexity of non-projective data-driven dependency](https://aclanthology.org/W07-2216)
+[parsing.](https://aclanthology.org/W07-2216) In _Proceedings of the Tenth International_
+_Conference on Parsing Technologies_, pages 121–132,
+Prague, Czech Republic. Association for Computational Linguistics.
+
+
+Sanket Vaibhav Mehta, Jinfeng Rao, Yi Tay, Mihir Kale,
+Ankur Parikh, and Emma Strubell. 2022. [Improving](https://doi.org/10.18653/v1/2022.acl-long.289)
+compositional [generalization](https://doi.org/10.18653/v1/2022.acl-long.289) with self-training for
+[data-to-text generation.](https://doi.org/10.18653/v1/2022.acl-long.289) In _Proceedings of the 60th_
+_Annual Meeting of the Association for Computational_
+_Linguistics_ _(Volume_ _1:_ _Long_ _Papers)_, pages 4205–
+4219, Dublin, Ireland. Association for Computational
+Linguistics.
+
+
+Yu Meng, Yunyi Zhang, Jiaxin Huang, Xuan Wang,
+Yu Zhang, Heng Ji, and Jiawei Han. 2021. [Distantly-](https://doi.org/10.18653/v1/2021.emnlp-main.810)
+supervised named [entity](https://doi.org/10.18653/v1/2021.emnlp-main.810) recognition with noise[robust learning and language model augmented self-](https://doi.org/10.18653/v1/2021.emnlp-main.810)
+[training.](https://doi.org/10.18653/v1/2021.emnlp-main.810) In _Proceedings of the 2021 Conference on_
+_Empirical_ _Methods_ _in_ _Natural_ _Language_ _Process-_
+_ing_, pages 10367–10378, Online and Punta Cana,
+Dominican Republic. Association for Computational
+Linguistics.
+
+
+Seyed Abolghasem Mirroshandel and Alexis Nasr. 2011.
+
+Active learning for [dependency](https://aclanthology.org/W11-2917) parsing using partially [annotated](https://aclanthology.org/W11-2917) sentences. In _Proceedings_ _of_ _the_
+_12th International Conference on Parsing Technolo-_
+_gies_, pages 140–149, Dublin, Ireland. Association for
+Computational Linguistics.
+
+
+Skatje Myers and Martha Palmer. 2021. [Tuning deep](https://aclanthology.org/2021.iwcs-1.20)
+active learning for [semantic](https://aclanthology.org/2021.iwcs-1.20) role labeling. In _Pro-_
+_ceedings_ _of_ _the_ _14th_ _International_ _Conference_ _on_
+_Computational_ _Semantics_ _(IWCS)_, pages 212–221,
+
+
+
+13001
+
+
+Groningen, The Netherlands (online). Association
+for Computational Linguistics.
+
+
+Graham Neubig and Shinsuke Mori. 2010. [Word-based](http://www.lrec-conf.org/proceedings/lrec2010/pdf/408_Paper.pdf)
+[partial annotation for efficient corpus construction.](http://www.lrec-conf.org/proceedings/lrec2010/pdf/408_Paper.pdf) In
+_Proceedings of the Seventh International Conference_
+_on Language Resources and Evaluation (LREC’10)_,
+Valletta, Malta. European Language Resources Association (ELRA).
+
+
+Qiang Ning, Hangfeng He, Chuchu Fan, and Dan Roth.
+2019. Partial or [complete,](https://doi.org/10.18653/v1/N19-1227) that’s the question. In
+_Proceedings_ _of_ _the_ _2019_ _Conference_ _of_ _the_ _North_
+_American Chapter of the Association for Computa-_
+_tional Linguistics:_ _Human Language Technologies,_
+_Volume 1 (Long and Short Papers)_, pages 2190–2200,
+Minneapolis, Minnesota. Association for Computational Linguistics.
+
+
+Qiang Ning, Zhongzhi Yu, Chuchu Fan, and Dan Roth.
+2018. [Exploiting partially annotated data in temporal](https://doi.org/10.18653/v1/S18-2018)
+relation [extraction.](https://doi.org/10.18653/v1/S18-2018) In _Proceedings_ _of_ _the_ _Seventh_
+_Joint Conference on Lexical and Computational Se-_
+_mantics_, pages 148–153, New Orleans, Louisiana.
+Association for Computational Linguistics.
+
+
+Joakim Nivre, Marie-Catherine de Marneffe, Filip Ginter, Jan Hajiˇc, Christopher D. Manning, Sampo
+Pyysalo, Sebastian Schuster, Francis Tyers, and
+Daniel Zeman. 2020. Universal [Dependencies](https://aclanthology.org/2020.lrec-1.497) v2:
+[An evergrowing multilingual treebank collection.](https://aclanthology.org/2020.lrec-1.497) In
+_Proceedings of the Twelfth Language Resources and_
+_Evaluation Conference_, pages 4034–4043, Marseille,
+France. European Language Resources Association.
+
+
+Matthew E. Peters, Mark Neumann, Mohit Iyyer, Matt
+Gardner, Christopher Clark, Kenton Lee, and Luke
+Zettlemoyer. 2018. [Deep contextualized word repre-](https://doi.org/10.18653/v1/N18-1202)
+[sentations.](https://doi.org/10.18653/v1/N18-1202) In _Proceedings of the 2018 Conference of_
+_the North American Chapter of the Association for_
+_Computational Linguistics:_ _Human Language Tech-_
+_nologies, Volume 1 (Long Papers)_, pages 2227–2237,
+New Orleans, Louisiana. Association for Computational Linguistics.
+
+
+Peng Qi, Yuhao Zhang, Yuhui Zhang, Jason Bolton, and
+Christopher D. Manning. 2020. Stanza: [A](https://doi.org/10.18653/v1/2020.acl-demos.14) python
+[natural language processing toolkit for many human](https://doi.org/10.18653/v1/2020.acl-demos.14)
+[languages.](https://doi.org/10.18653/v1/2020.acl-demos.14) In _Proceedings of the 58th Annual Meet-_
+_ing of the Association for Computational Linguistics:_
+_System Demonstrations_, pages 101–108, Online. Association for Computational Linguistics.
+
+
+Puria Radmard, Yassir Fathullah, and Aldo Lipani. 2021.
+
+[Subsequence based deep active learning for named](https://doi.org/10.18653/v1/2021.acl-long.332)
+[entity recognition.](https://doi.org/10.18653/v1/2021.acl-long.332) In _Proceedings of the 59th Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics and the 11th International Joint Conference_
+_on Natural Language Processing (Volume 1:_ _Long_
+_Papers)_, pages 4310–4321, Online. Association for
+Computational Linguistics.
+
+
+Roi Reichart, Katrin Tomanek, Udo Hahn, and Ari Rappoport. 2008. [Multi-task active learning for linguistic](https://aclanthology.org/P08-1098)
+[annotations.](https://aclanthology.org/P08-1098) In _Proceedings of ACL-08:_ _HLT_, pages
+
+
+
+861–869, Columbus, Ohio. Association for Computational Linguistics.
+
+
+Dan Roth and Kevin Small. 2008. Active learning for
+pipeline models. In _AAAI_, pages 683–688.
+
+
+Guy Rotman and Roi Reichart. 2022. [Multi-task active](https://doi.org/10.1162/tacl_a_00515)
+learning for pre-trained [transformer-based](https://doi.org/10.1162/tacl_a_00515) models.
+_Transactions of the Association for Computational_
+_Linguistics_, 10:1209–1228.
+
+
+Manabu Sassano and Sadao Kurohashi. 2010. [Using](https://aclanthology.org/P10-1037)
+[smaller constituents rather than sentences in active](https://aclanthology.org/P10-1037)
+[learning for Japanese dependency parsing.](https://aclanthology.org/P10-1037) In _Pro-_
+_ceedings of the 48th Annual Meeting of the Associa-_
+_tion for Computational Linguistics_, pages 356–365,
+Uppsala, Sweden. Association for Computational
+Linguistics.
+
+
+Tobias Scheffer, Christian Decomain, and Stefan Wrobel. 2001. Active hidden markov models for information extraction. In _International_ _Symposium_ _on_
+_Intelligent Data Analysis_, pages 309–318. Springer.
+
+
+Burr Settles. 2009. Active learning literature survey.
+
+
+Burr Settles and Mark Craven. 2008. An [analysis](https://aclanthology.org/D08-1112) of
+[active learning strategies for sequence labeling tasks.](https://aclanthology.org/D08-1112)
+In _Proceedings_ _of_ _the_ _2008_ _Conference_ _on_ _Empiri-_
+_cal Methods in Natural Language Processing_, pages
+1070–1079, Honolulu, Hawaii. Association for Computational Linguistics.
+
+
+Burr Settles, Mark Craven, and Lewis Friedland. 2008.
+Active learning with real annotation costs. In _Pro-_
+_ceedings_ _of_ _the_ _NIPS_ _workshop_ _on_ _cost-sensitive_
+_learning_, volume 1.
+
+
+Dan Shen, Jie Zhang, Jian Su, Guodong Zhou, and
+Chew-Lim Tan. 2004. [Multi-criteria-based](https://doi.org/10.3115/1218955.1219030) active
+[learning for named entity recognition.](https://doi.org/10.3115/1218955.1219030) In _Proceed-_
+_ings of the 42nd Annual Meeting of the Association_
+_for Computational Linguistics (ACL-04)_, pages 589–
+596, Barcelona, Spain.
+
+
+Yanyao Shen, Hyokun Yun, Zachary C. Lipton, Yakov
+Kronrod, and Animashree Anandkumar. 2018. [Deep](https://openreview.net/forum?id=ry018WZAZ)
+[active learning for named entity recognition.](https://openreview.net/forum?id=ry018WZAZ) In _Inter-_
+_national Conference on Learning Representations_ .
+
+
+David A. Smith and Noah A. Smith. 2007. [Probabilistic](https://aclanthology.org/D07-1014)
+[models of nonprojective dependency trees.](https://aclanthology.org/D07-1014) In _Pro-_
+_ceedings of the 2007 Joint Conference on Empirical_
+_Methods in Natural Language Processing and Com-_
+_putational_ _Natural_ _Language_ _Learning_ _(EMNLP-_
+_CoNLL)_, pages 132–140, Prague, Czech Republic.
+Association for Computational Linguistics.
+
+
+Noah A Smith. 2011. Linguistic structure prediction.
+_Synthesis lectures on human language technologies_,
+4(2):1–274.
+
+
+Oscar Täckström, Dipanjan Das, Slav Petrov, Ryan McDonald, and Joakim Nivre. 2013. Token [and](https://doi.org/10.1162/tacl_a_00205) type
+[constraints for cross-lingual part-of-speech tagging.](https://doi.org/10.1162/tacl_a_00205)
+_Transactions of the Association for Computational_
+_Linguistics_, 1:1–12.
+
+
+
+13002
+
+
+Erik F. Tjong Kim Sang and Fien De Meulder.
+2003. [Introduction to the CoNLL-2003 shared task:](https://aclanthology.org/W03-0419)
+[Language-independent named entity recognition.](https://aclanthology.org/W03-0419) In
+_Proceedings of the Seventh Conference on Natural_
+_Language Learning at HLT-NAACL 2003_, pages 142–
+147.
+
+
+Katrin Tomanek and Udo Hahn. 2009. [Semi-supervised](https://aclanthology.org/P09-1117)
+[active learning for sequence labeling.](https://aclanthology.org/P09-1117) In _Proceedings_
+_of the Joint Conference of the 47th Annual Meeting_
+_of_ _the_ _ACL_ _and_ _the_ _4th_ _International_ _Joint_ _Confer-_
+_ence on Natural Language Processing of the AFNLP_,
+pages 1039–1047, Suntec, Singapore. Association
+for Computational Linguistics.
+
+
+Yuta Tsuboi, Hisashi Kashima, Shinsuke Mori, Hiroki
+Oda, and Yuji Matsumoto. 2008. [Training](https://aclanthology.org/C08-1113) conditional random fields [using](https://aclanthology.org/C08-1113) incomplete annotations.
+In _Proceedings of the 22nd International Conference_
+_on Computational Linguistics (Coling 2008)_, pages
+897–904, Manchester, UK. Coling 2008 Organizing
+Committee.
+
+
+Tu Vu, Minh-Thang Luong, Quoc Le, Grady Simon, and
+Mohit Iyyer. 2021. [STraTA: Self-training with task](https://doi.org/10.18653/v1/2021.emnlp-main.462)
+augmentation for [better](https://doi.org/10.18653/v1/2021.emnlp-main.462) few-shot learning. In _Pro-_
+_ceedings of the 2021 Conference on Empirical Meth-_
+_ods_ _in_ _Natural_ _Language_ _Processing_, pages 5715–
+5731, Online and Punta Cana, Dominican Republic.
+Association for Computational Linguistics.
+
+
+Christopher Walker, Stephanie Strassel, Julie Medero,
+and Kazuaki Maeda. 2006. ACE 2005 multilingual
+training corpus. _Linguistic Data Consortium_, 57.
+
+
+Chenguang Wang, Laura Chiticariu, and Yunyao Li.
+2017. Active learning for black-box semantic role
+labeling with neural factors. In _IJCAI_ .
+
+
+Xinyu Wang, Yong Jiang, Zhaohui Yan, Zixia Jia,
+Nguyen Bach, Tao Wang, Zhongqiang Huang, Fei
+Huang, and Kewei Tu. 2021. [Structural knowledge](https://doi.org/10.18653/v1/2021.acl-long.46)
+distillation: [Tractably distilling information for struc-](https://doi.org/10.18653/v1/2021.acl-long.46)
+[tured predictor.](https://doi.org/10.18653/v1/2021.acl-long.46) In _Proceedings of the 59th Annual_
+_Meeting_ _of_ _the_ _Association_ _for_ _Computational_ _Lin-_
+_guistics and the 11th International Joint Conference_
+_on Natural Language Processing (Volume 1:_ _Long_
+_Papers)_, pages 550–564, Online. Association for
+Computational Linguistics.
+
+
+Dittaya Wanvarie, Hiroya Takamura, and Manabu Okumura. 2011. Active learning with subsequence sampling strategy for sequence labeling tasks. _Informa-_
+_tion and Media Technologies_, 6(3):680–700.
+
+
+Fan Yang and Paul Vozila. 2014. [Semi-supervised Chi-](https://doi.org/10.3115/v1/D14-1010)
+[nese word segmentation using partial-label learning](https://doi.org/10.3115/v1/D14-1010)
+[with conditional random fields.](https://doi.org/10.3115/v1/D14-1010) In _Proceedings of the_
+_2014 Conference on Empirical Methods in Natural_
+_Language Processing (EMNLP)_, pages 90–98, Doha,
+Qatar. Association for Computational Linguistics.
+
+
+Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019.
+
+
+
+Xlnet: Generalized autoregressive pretraining for language understanding. _Advances in neural informa-_
+_tion processing systems_, 32.
+
+
+David Yarowsky. 1995. [Unsupervised word sense dis-](https://doi.org/10.3115/981658.981684)
+ambiguation [rivaling](https://doi.org/10.3115/981658.981684) supervised methods. In _33rd_
+_Annual_ _Meeting_ _of_ _the_ _Association_ _for_ _Computa-_
+_tional Linguistics_, pages 189–196, Cambridge, Massachusetts, USA. Association for Computational Linguistics.
+
+
+Yue Yu, Lingkai Kong, Jieyu Zhang, Rongzhi Zhang,
+and Chao Zhang. 2022. AcTune: [Uncertainty-based](https://doi.org/10.18653/v1/2022.naacl-main.102)
+[active self-training for active fine-tuning of pretrained](https://doi.org/10.18653/v1/2022.naacl-main.102)
+[language models.](https://doi.org/10.18653/v1/2022.naacl-main.102) In _Proceedings of the 2022 Con-_
+_ference_ _of_ _the_ _North_ _American_ _Chapter_ _of_ _the_ _As-_
+_sociation_ _for_ _Computational_ _Linguistics:_ _Human_
+_Language Technologies_, pages 1422–1436, Seattle,
+United States. Association for Computational Linguistics.
+
+
+Xiangkai Zeng, Sarthak Garg, Rajen Chatterjee, Udhyakumar Nallasamy, and Matthias Paulik. 2019.
+[Empirical evaluation of active learning techniques for](https://doi.org/10.18653/v1/D19-6110)
+[neural MT.](https://doi.org/10.18653/v1/D19-6110) In _Proceedings of the 2nd Workshop on_
+_Deep Learning Approaches for Low-Resource NLP_
+_(DeepLo_ _2019)_, pages 84–93, Hong Kong, China.
+Association for Computational Linguistics.
+
+
+Hua Zhu, Wu Ye, Sihan Luo, and Xidong Zhang. 2020.
+
+A multitask active [learning](https://doi.org/10.18653/v1/2020.coling-main.430) framework for natural
+[language understanding.](https://doi.org/10.18653/v1/2020.coling-main.430) In _Proceedings of the 28th_
+_International Conference on Computational Linguis-_
+_tics_, pages 4900–4914, Barcelona, Spain (Online).
+International Committee on Computational Linguistics.
+
+
+
+13003
+
+
+**A** **More Details of Task Settings**
+
+
+**A.1** **DPAR**
+
+- **Model** . Similar to NER, we utilize a BERTbased module to provide contextualized representations. We further stack a standard firstorder non-projective graph-based parsing module based on a biaffine scorer (Dozat and Manning, 2017). The marginals for each token’s
+head decision can be feasibly calculated by the
+Matrix-Tree algorithm (Koo et al., 2007; Smith
+and Smith, 2007; McDonald and Satta, 2007).
+
+
+- **Query and Selection** . Following previous works
+(Flannery and Mori, 2015; Li et al., 2016), we
+view DPAR as a head-word finding problem and
+regard each token and its head decision as the
+sub-structure unit. In this case, the query and
+selection for DPAR are almost identical to the
+NER task because of this token-wise decision
+scheme. Therefore, the same AL strategies in
+NER can be adopted here.
+
+
+- **Annotation** . In DPAR, there are no special spanbased annotations as in NER; thus, we simply
+annotate in a word-based scheme.
+
+
+- **Model learning** . Similar to NER, we adopt the
+log-likelihood of the gold parse tree as the training loss in FA and marginalized likelihood in PA
+(Li et al., 2016).
+
+
+**A.2** **IE**
+
+- **Tasks** . We tackle event extraction (EE) and relation extraction (RE) using a two-step pipelined
+approach. The first step aims to extract entity
+mentions for RE, and entity mentions and event
+triggers for EE. We adopt sequence labeling for
+mention extractions as in the NER task. Based on
+the mentions extracted in the first step, the second
+step examines each feasible candidate mention
+pair (entity pair for RE and event-entity pair for
+EE) and decides the relation (entity relation for
+RE and event argument relation for EE) for them.
+Since event argument links can be regarded as
+relations between event triggers and entities, for
+simplicity we will use the relational sub-task to
+refer to both relation and argument extraction.
+
+
+- **Model** . We adopt a multi-task model similar to
+the one utilized in (Rotman and Reichart, 2022).
+With a pre-trained encoder, we take the first _N_
+
+
+
+layers as the shared encoding module whose output representations are used for both sub-tasks.
+Each sub-task further adopts a private encoder
+that is initialized with the remaining pre-trained
+layers and is trained with task-specific signals.
+We simply set _N_ to 6, while the results are generally not sensitive to this hyper-parameter. Final
+task-specific predictors are further stacked upon
+the corresponding private encoders. We adopt a
+CRF layer for mention extraction and a pairwise
+local predictor with a biaffine scorer for relation
+or argument extraction.
+
+
+- **Sentence selection.** For an unlabeled sentence,
+there is an uncertainty score for each sub-task.
+For mentions, the uncertainty is the average margin as in the NER task. For relations, we find that
+averaging uncertainties over all mention pairs has
+a bias towards sentences with fewer mentions.
+To mitigate such bias, we first aggregate an uncertainty score for each mention by taking the
+maximum score within all the relations that link
+to it and then averaging over all the mentions for
+sentence-level scores. Finally, the scores of the
+two sub-tasks are linearly combined to form the
+sentence-level uncertainty.
+
+
+- **Partial selection.** For PA selection, the two subtasks are handled separately according to the
+adaptive ratio scheme. We further adopt two
+heuristics for the relational task to compensate
+for errors in the mention extraction. First, since
+there can be over-predicted mentions that lead
+to discarded relation queries, we adjust the PA
+ratio by estimating how many candidate relations contain such errors in the mentions. We
+again train a logistic regression model to predict whether a token is NIL (or ‘O’ in the BIO
+scheme, meaning not contained inside any gold
+mentions) based on its NIL probability. Then for
+each candidate relation, we calculate the probability that any token within its mentions is NIL.
+By averaging this probability of all the candidates, we obtain a rough estimation of the percentage of problematic relations, which we call
+it _α_ . Finally the PA selection ratio is adjusted by:
+_r_ adjust = _α·r_ problem +(1 _−α_ ) _·r_ origin. Here, _r_ origin
+denotes the original selection ratio obtained from
+the adaptive scheme, and _r_ problem denotes the selection ratio of problematic relations, which we
+conservatively set to 1. Secondly, since there
+can also be under-predicted mentions, we add a
+
+
+
+13004
+
+
+Data Split #Sent. #Token #Event #Entity #Argument #Relation
+
+
+
+CoNLL03
+
+
+UD-EWT
+
+
+ACE05
+
+
+
+train 14.0K 203.6K - 23.5K - dev 3.3K 51.4K - 5.9K - test 3.5K 46.4K - 5.6K - 
+train 12.5K 204.6K - - - dev 2.0K 25.1K - - - test 2.1K 25.1K - - - 
+train 14.4K 215.2K 3.7K 38.0K 5.7K 6.2K
+dev 2.5K 34.5K 0.5K 6.0K 0.7K 0.8K
+test 4.0K 61.5K 1.1K 10.8K 1.7K 1.7K
+
+
+Table 1: Data statistics.
+
+
+
+second stage of querying and annotation in each
+AL cycle based on the annotated mentions in
+the first stage. This extra stage only selects relations that involve the newly added or corrected
+mentions. We simply reuse the selection ratio
+determined from the first stage and apply it to
+each sentence that contains such mentions. In
+this way, the second stage is lightweight and only
+requires relatively cheap re-inference for each
+queried sentence individually.
+
+
+- **Annotation.** The annotation of the mentions is
+the same as in the NER task, while for the annotation of relational queries, their mentions are first
+examined and corrected if needed, as explained
+in §3.4. We measure the labeling cost by the final annotated items; thus, these extra examined
+mentioned will also be properly counted.
+
+
+- **Model learning.** For the mention extraction subtask, the training objective is the same as in NER.
+For the relational sub-task, we simply adopt a
+local pairwise model with the standard crossentropy loss. Since the relation model is local,
+no special treatment is needed for PA.
+
+
+**B** **Data Statistics and More Settings**
+
+
+**Data.** Our main experiments are conducted using the CoNLL-2003 English dataset [11] (Tjong
+Kim Sang and De Meulder, 2003) for NER, the
+English Web Treebank (EWT) from Universal Dependencies [12] v2.10 (Nivre et al., 2020) for DPAR,
+and English portion of ACE2005 [13] (Walker et al.,
+2006) for IE. We utilize Stanza [14] (Qi et al., 2020)
+to assign POS tags for cost measurement in NER
+
+
+[11https://www.clips.uantwerpen.be/conll2003/](https://www.clips.uantwerpen.be/conll2003/ner/)
+[ner/](https://www.clips.uantwerpen.be/conll2003/ner/)
+
+[12https://universaldependencies.org/](https://universaldependencies.org/)
+[13https://catalog.ldc.upenn.edu/LDC2006T06](https://catalog.ldc.upenn.edu/LDC2006T06)
+[14https://stanfordnlp.github.io/stanza/](https://stanfordnlp.github.io/stanza/)
+
+
+
+and mention tasks. We follow Lin et al. (2020) for
+the pre-processing [15] of the ACE dataset. For the
+IE tasks on ACE, we find that the conventional test
+set contains only newswire documents while the
+training set consists of various genres (such as from
+conversation and web). Such mismatches between
+the AL pool and the final testing set are nontrivial to handle with the classical AL protocol, and
+we thus randomly re-split the ACE dataset (with a
+ratio of 7:1:2 for training, dev, and test sets, respectively). Table 1 shows data statistics. For each AL
+experiment, we take the original training set as the
+unlabeled pool, down-sample a dev set from the
+original dev set, and evaluate on the full test set.
+
+
+**More Settings.** All of our models are based on
+the pre-trained RoBERTabase as the contextualized encoder. We further fine-tune it with the
+task-specific decoder in all the experiments. The
+number of model parameters is roughly 124M for
+single-output tasks and around 186M for multi-task
+IE tasks. For other hyper-parameter settings, we
+mostly follow common practices. Adam is utilized
+for optimization, with an initial learning rate of 1e5 for NER and 2e-5 for DPAR and IE. The learning
+rate is linearly decayed to 10% of the initial value
+throughout the training process. The models are
+tuned for 10K steps with a batch size of roughly
+512 tokens. We evaluate the model on the dev set
+every 1K steps to choose the best checkpoint. The
+experiments are run with one 2080Ti GPU. The
+training of one AL cycle usually takes only one
+or two hours, and the full simulation of one AL
+run can be finished within one day. We adopt standard evaluation metrics for the tasks: labeled F1
+score for NER, labeled attachment score (LAS) for
+DPAR, labeled argument and relation F1 score for
+event arguments and relations (Lin et al., 2020).
+
+
+[15http://blender.cs.illinois.edu/software/](http://blender.cs.illinois.edu/software/oneie/)
+[oneie/](http://blender.cs.illinois.edu/software/oneie/)
+
+
+
+13005
+
+
+**C** **Details of Algorithms**
+
+In this section, we provide more details of the algorithms for CRF-styled models (Lafferty et al.,
+2001). For an input instance _x_ (for example, a sentence), the model assigns a globally normalized
+probability to each possible output structured object _y_ (for example, a tag sequence or a parse tree)
+in the target space _Y_ :
+
+exp _s_ ( _y_ _x_ )
+_p_ ( _y_ _x_ ) = ~~�~~ _|_
+_|_ _y_ _[′]_ [exp] _[ s]_ [(] _[y][′][|][x]_ [)]
+_∈Y_
+
+exp [�] _f_ _y_ _[s]_ [(] _[f]_ _[|][x]_ [)]
+= ~~�~~ ~~�~~ _∈_
+_y_ _[′]_ _f_ _[′]_ _y_ _[′][ s]_ [(] _[f]_ _[′][|][x]_ [)]
+_∈Y_ _∈_
+
+Here, _s_ ( _y|x_ ) denotes the un-normalized raw scores
+assigned to _y_, which is further factorized into the
+sum of the sub-structure scores _s_ ( _f_ _|x_ ). [16] In plain
+likelihood training for CRF, we take the negative
+log-probability as the training objective:
+
+
+
+labeling (Baum et al., 1970) or Matrix-tree for
+non-projective dependency parsing (Koo et al.,
+2007; Smith and Smith, 2007; McDonald and Satta,
+2007).
+
+
+**Learning with incomplete annotations.** Following previous works (Tsuboi et al., 2008; Li et al.,
+2016; Greenberg et al., 2018), for the instances
+with incomplete annotations, we utilize the logarithm of the marginal likelihood as the learning
+objective:
+
+
+
+
+  _L_ = _−_ log
+
+
+
+_p_ ( _y|x_ )
+_y_ _C_
+_∈Y_
+
+
+
+
+  = _−_ log
+
+_y_ _C_
+
+  - _∈Y_
+= _−_ log
+
+
+
+_y_ _C_
+_∈Y_
+
+
+
+exp _s_ ( _y_ _x_ )
+
+~~�~~ _|_
+_y_ [exp] _[ s]_ [(] _[y][|][x]_ [)]
+_∈Y_
+
+
+
+exp _s_ ( _y|x_ ) + log _Z_ ( _x_ )
+_y_ _C_
+_∈Y_
+
+
+
+_L_ = _−_ log _p_ ( _y|x_ )
+
+
+
+exp _s_ ( _y_ _[′]_ _|x_ )
+_y_ _[′]_ _∈Y_
+
+
+
+
+   = _−s_ ( _y|x_ ) + log
+
+
+
+For brevity, in the remaining, we use log _Z_ ( _x_ ) to
+denote the second term of the log partition function.
+For model training, we need to calculate the gradients of the model parameters _θ_ to the loss function.
+The first item is easy to deal with since it only involves one structured object, while log _Z_ ( _x_ ) needs
+some reorganization according to the factorization:
+
+
+
+Here, _C_ denotes the constrained set of the output
+_Y_
+objects that agree with the existing partial annotations. In this objective function, the second item is
+exactly the same as in standard CRF, while the first
+one can be calculated [17] in a modified way (Tsuboi
+et al., 2008).
+
+
+**Knowledge** **distillation.** As described in the
+main context, we adopt the knowledge distillation
+objective for self-training with soft labels. For
+brevity, we denote the probabilities from the last
+model as _p_ _[′]_ ( _y|x_ ) and keep using _p_ ( _y|x_ ) to denote
+the ones from the current model. Following Wang
+et al. (2021), the loss can be calculated by:
+
+
+
+
+ _L_ = _−_
+
+
+
+_p_ _[′]_ ( _y|x_ ) log _p_ ( _y|x_ )
+_y∈Y_
+
+
+
+_∇θ_ log _Z_ =
+
+
+
+
+_y_ _[′]_ [exp] _[ s]_ [(] _[y][′][|][x]_ [)] _[∇][θ][s]_ [(] _[y][′][|][x]_ [)]
+_∈Y_ ~~�~~
+_y_ _[′′]_ [exp] _[ s]_ [(] _[y][′′][|][x]_ [)]
+
+- _∈Y_
+
+
+
+
+ =
+
+_y_  - _[′]_ _∈Y_
+
+=
+
+_y_  - _[′]_ _∈Y_
+
+=
+
+
+
+_p_ ( _y_ _[′]_ _|x_ )
+_y_ _[′]_ _∈Yf_ _′_
+
+
+
+_p_ ( _y_ _[′]_ _x_ ) _θs_ ( _y_ _[′]_ _x_ )
+_|_ _∇_ _|_
+_y_ _[′]_ _∈Y_
+
+
+
+_p_ _[′]_ ( _y_ _[′]_ _|x_ ) + log _Z_ ( _x_ )
+_y_ _[′]_ _∈Yf_ _′_
+
+
+
+
+- 
+_p_ ( _y_ _[′]_ _|x_ )
+_y_ _[′]_ _∈Y_ _f_ _[′]_ _∈y_
+
+
+
+_θs_ ( _f_ _[′]_ _x_ )
+_∇_ _|_
+_f_ _[′]_ _∈y_ _[′]_
+
+
+
+_p_ _[′]_ ( _y|x_ ) _s_ ( _y|x_ ) + log _Z_ ( _x_ )
+_y∈Y_
+
+
+
+
+  _θs_ ( _f_ _[′]_ _x_ )
+_∇_ _|_
+_f_ _[′]_ _y_ _[′]_
+
+
+
+
+- 
+_p_ _[′]_ ( _y|x_ )
+_y∈Y_ _f_ _[′]_ _∈y_ _[′]_
+
+
+
+_s_ ( _f_ _[′]_ _|x_ ) + log _Z_ ( _x_ )
+_f_ _[′]_ _∈y_ _[′]_
+
+
+
+The last step is obtained by swapping the order of
+the two summations, and finally, the problem is reduced to calculating each sub-structure’s marginal
+probability [�] _y_ _[′]_ _∈Yf_ _′_ _[p]_ [(] _[y][′][|][x]_ [)][.] [Here,] _[Y][f]_ _[′]_ [denotes]
+all the output structured objects that contain the
+sub-structure _f_ _[′]_, and the marginals can usually
+be calculated by classical structured prediction algorithms such as forward-backward for sequence
+
+
+16Such as unary and pairwise scores for sequence labeling
+or token-wise edge scores for dependency parsing.
+
+
+
+The loss function is broken down into two items
+whose gradients can be obtained by calculating
+marginals according to the last model or the current
+one, respectively.
+
+
+17In our implementation, we adopt a simple method to enforce the constraints by adding negative-infinite to the scores
+of the impossible labels. In this case, the structures that violates the constraints will have a score of negative-infinite (and
+a probability of zero) and will thus be excluded.
+
+
+
+
+ = _−_
+
+_y_   - _∈Y_
+
+= _−_
+
+_y_   - _∈Y_
+
+= _−_
+
+
+
+
+  _s_ ( _f_ _[′]_ _|x_ )
+_f_ _[′]_ _y_ _[′]_
+
+
+
+13006
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
+|---|---|---|---|---|---|---|---|---|---|
+||||||||||~~RAND~~|
+||||||||||FA-M<br>PA-M<br>|
+||||||||||~~PA-LC~~<br>PA-E<br>PA-B|
+||||||||||Super.|
+|||||||||||
+|||||||||||
+
+
+Token Count
+
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|
+|---|---|---|---|---|---|---|---|---|---|---|---|
+|||||||||||~~RA~~|~~D~~|
+|||||||||||FA-<br>PA-<br>|M<br>M<br>|
+|||||||||||||
+|||||||||||||
+|||||||||||~~PA-~~<br>PA-<br>PA-|~~LC~~<br>E<br>B|
+|||||||||||||
+|||||||||||Sup|er.|
+|||||||||||||
+|||||||||||||
+|||||||||||||
+
+
+Sub-structure Count
+
+
+
+Figure 7: Comparisons of different acquisition functions for partial annotation: “-M” denotes margin-based, “-LC”
+denotes least-confident, “-E” denotes entropy-based, and “-B” indicates BALD.
+
+
+
+90
+
+
+88
+
+
+86
+
+
+84
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||
+||||||||||RAN<br>FA-M<br>~~FA-L~~|D<br><br>|
+||||||||||||
+||||||||||||
+||||||||||FA-E<br>FA-B<br>||
+||||||||||||
+||||||||||||
+||||||||||~~Super~~|~~.~~|
+||||||||||||
+||||||||||||
+
+
+Token Count
+
+
+
+Figure 8: Comparisons of different acquisition functions
+for full annotation. Notations of the methods are the
+same as in Figure 7.
+
+
+**D** **Extra Results**
+
+
+**D.1** **Using Different Acquisition Functions**
+
+In the main experiments, our acquisition function
+is based on margin-based uncertainty, that is, selecting the instances that have the largest marginal
+differences between the most and second-most confident predictions. Here, we compare it with various other acquisition functions, including leastconfident (-LC), max-entropy (-E) and BALD (B) (Houlsby et al., 2011). We take DPAR as the
+studying case and the results for full annotation
+and partial annotation are shown in Figure 8 and
+7, respectively. Generally, there are no large differences between the adopted querying methods and
+the margin-based method can obtain the overall
+best results. Notice that regardless of the adopted
+acquisition function, we can see the effectiveness
+of our partial selection scheme: it requires lower
+labeling cost than full annotation to reach the upper
+bound. This shows that our method is extensible
+to different AL querying methods and it will be interesting to explore the combination of our method
+with more complex and advanced acquisition func
+
+
+tions, such as those considering representativeness.
+
+
+**D.2** **IE Experiments**
+
+In this section, we present more results of the IE
+experiments. First, Figure 9 shows the mention
+extraction results for the event extraction task. The
+overall trends are very similar to those in NER:
+PA can obtain similar results to FA with the same
+reading texts and less mention labeling cost. In
+Figure 10, we show the results for mention and
+relation extractions. In the ACE dataset, relations
+are very sparsely annotated, and around 97% of the
+entities are linked with less or equal to two relations. Considering this fact, we measure the cost
+of FA relation extraction by two times the annotated entities, while PA still counts the number of
+the queried relations. The relation results are similar to the patterns for event argument extraction,
+showing the benefits of selecting and annotating
+with partial sub-structures. Notice that in some of
+the mention extraction results, there seems to be
+less obvious differences between the AL strategies
+over the random baseline. This may be due to our
+focus on the second sub-task for relations (or event
+arguments), directly reflected by its high weight
+( _β_ ) in calculating sentence uncertainty. It will be
+interesting to explore better ways to enhance both
+sub-tasks, probably with an adaptive combination
+scheme (Roth and Small, 2008).
+
+
+
+13007
+
+
+88
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+75
+
+
+70
+
+
+65
+
+
+60
+
+
+55
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||
+|||||||||Rand<br>Rand+S|d|T|
+||||||||||||
+||||||||||||
+|||||||||FA<br>FA+ST<br>|FA<br>FA+ST<br>||
+||||||||||||
+||||||||||||
+|||||||||~~PA~~<br>PA+ST<br>~~Super.~~|~~PA~~<br>PA+ST<br>~~Super.~~||
+||||||||||||
+||||||||||||
+||||||||||||
+
+
+4000 12000 20000 28000 36000 44000 52000 60000
+Token Count
+
+
+MentionTrigger Reading Cost
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|
+|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
+|||||||||||||||
+||||||||||||~~Ra~~|~~d~~|~~d~~|
+||||||||||||Rand+S<br>FA|Rand+S<br>FA|T|
+|||||||||||||||
+||||||||||||FA+ST<br>PA|FA+ST<br>PA||
+|||||||||||||||
+|||||||||||||||
+||||||||||||PA+ST<br>Super.|PA+ST<br>Super.||
+|||||||||||||||
+|||||||||||||||
+|||||||||||||||
+
+
+
+Token Count
+
+
+
+88
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+75
+
+
+70
+
+
+65
+
+
+60
+
+
+55
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||
+||||||||Rand<br>Rand+ST|Rand<br>Rand+ST|Rand<br>Rand+ST|Rand<br>Rand+ST|
+||||||||FA<br>FA+ST<br>|FA<br>FA+ST<br>|FA<br>FA+ST<br>|FA<br>FA+ST<br>|
+||||||||~~PA~~<br>PA+ST<br>~~Super.~~|~~PA~~<br>PA+ST<br>~~Super.~~|~~PA~~<br>PA+ST<br>~~Super.~~|~~PA~~<br>PA+ST<br>~~Super.~~|
+||||||||||||
+
+
+0 5000 10000 15000 20000 25000 30000
+Sub-structure Count
+
+
+MentionTrigger Labeling Cost
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|
+|---|---|---|---|---|---|---|---|---|---|
+|||||||||||
+|||||||||~~Rand~~|~~Rand~~|
+|||||||||Rand+ST<br>FA|Rand+ST<br>FA|
+|||||||||FA+ST<br>PA|FA+ST<br>PA|
+|||||||||PA+ST<br>Super.|PA+ST<br>Super.|
+|||||||||||
+
+
+
+Sub-structure Count
+
+
+
+Figure 9: Results (F1) of mention extraction (entities and event triggers) for the event extraction task on ACE05
+(argument results are shown in Figure 6).
+
+
+
+88
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+65
+
+
+60
+
+
+55
+
+
+50
+
+
+45
+
+
+40
+
+
+35
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||||
+|||||||||Rand<br>|Rand<br>|Rand<br>|
+||||||||||||
+|||||||||~~Rand+S~~<br>FA<br>|~~Rand+S~~<br>FA<br>||
+||||||||||||
+|||||||||~~FA+ST~~<br>PA<br>|~~FA+ST~~<br>PA<br>||
+||||||||||||
+|||||||||~~PA+ST~~<br>Super.|~~PA+ST~~<br>Super.||
+||||||||||||
+||||||||||||
+
+
+Token Count
+
+
+Relation Reading Cost
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|
+|---|---|---|---|---|---|---|---|---|---|---|
+||||||||||Rand<br>Rand+S|T|
+||||||||||||
+||||||||||FA<br>~~FA+ST~~||
+||||||||||||
+||||||||||PA<br>||
+||||||||||||
+||||||||||~~PA+ST~~<br>Super.||
+||||||||||||
+||||||||||||
+
+
+
+4000 12000 20000 28000 36000 44000 52000 60000
+Token Count
+
+
+
+88
+
+
+86
+
+
+84
+
+
+82
+
+
+80
+
+
+65
+
+
+60
+
+
+55
+
+
+50
+
+
+45
+
+
+40
+
+
+35
+
+
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|
+|---|---|---|---|---|---|---|---|
+|||||||||
+|||||||Rand<br>|Rand<br>|
+|||||||||
+|||||||~~Rand+ST~~<br>FA<br>|~~Rand+ST~~<br>FA<br>|
+|||||||||
+|||||||~~FA+ST~~<br>PA<br>|~~FA+ST~~<br>PA<br>|
+|||||||||
+|||||||~~PA+ST~~<br>Super.|~~PA+ST~~<br>Super.|
+|||||||||
+|||||||||
+
+
+Sub-structure Count
+
+
+Relation Labeling Cost
+
+|Col1|Col2|Col3|Col4|Col5|Col6|Col7|Col8|
+|---|---|---|---|---|---|---|---|
+||||||||Rand<br>Rand+ST|
+|||||||||
+||||||||FA<br>~~FA+ST~~|
+|||||||||
+||||||||PA<br>|
+|||||||||
+||||||||~~PA+ST~~<br>Super.|
+|||||||||
+|||||||||
+
+
+
+0 5000 10000 15000 20000
+Sub-structure Count
+
+
+
+Figure 10: Results (F1) of the extraction of entity mentions and relations for the relation extraction task on ACE05.
+
+
+13008
+
+

research/active_learning/AL_CLAIMS_VERIFICATION.md

@@ -0,0 +1,596 @@
+# Active Learning Claims Verification Report
+
+**Date**: 2026-02-13
+**Reviewer**: Research Investigation
+**Purpose**: Verify active learning claims for Vietnamese three-task pipeline (WS → POS → DP)
+
+---
+
+## Executive Summary
+
+This report verifies claims about active learning for NLP annotation tasks, with a focus on evaluating whether a proposed three-task pipeline for Vietnamese (word segmentation, POS tagging, dependency parsing) makes accurate claims about prior work. The investigation examines 8 specific claims across cost reduction, query strategies, and annotation productivity.
+
+**Overall Assessment**: Most claims are **generally accurate but lack precise citations**. The cost reduction percentages cited are in the right ballpark but not always precisely matched to the claimed sources. Some claims reference work that doesn't exist or misattribute specific results.
+
+---
+
+## 1. Active Learning Cost Reduction Claims
+
+### 1.1. Hwa (2004): ~50% Fewer Sentences for Parsing
+
+**Claim**: "Hwa (2004) claiming ~50% fewer sentences for parsing"
+
+**Verification**: ✅ **GENERALLY ACCURATE**
+
+**Evidence**:
+- **Paper**: Rebecca Hwa, "Sample Selection for Statistical Parsing," *Computational Linguistics*, 30(3):253-276, 2004
+- **URL**: https://direct.mit.edu/coli/article/30/3/253/1841/Sample-Selection-for-Statistical-Parsing
+- **Actual Finding**: The paper demonstrates that "sample selection can significantly reduce the size of annotated training corpora." One specific result shows **36% reduction in annotated training data** without degrading grammar quality.
+- **Note**: While the paper does show substantial reductions (36% is documented), the exact "~50%" figure is not directly cited in the available abstracts. Related work by Baldridge & Osborne (2004) on HPSG parse selection reported **60% reduction**, which may be the source of confusion.
+
+**Sources**:
+- [Sample Selection for Statistical Parsing](https://direct.mit.edu/coli/article/30/3/253/1841/Sample-Selection-for-Statistical-Parsing)
+- [Sample Selection for Statistical Parsing (ACL Anthology)](https://aclanthology.org/J04-3001.pdf)
+
+### 1.2. Li et al. (2016): 40-60% Less Arc Annotation with Partial Annotation
+
+**Claim**: "Li et al. (2016) claiming 40-60% less arc annotation with partial annotation"
+
+**Verification**: ✅ **ACCURATE**
+
+**Evidence**:
+- **Paper**: Zhenghua Li, Min Zhang, Yue Zhang, Zhanyi Liu, Wenliang Chen, Hua Wu, Haifeng Wang, "Active Learning for Dependency Parsing with Partial Annotation," *ACL 2016*, pp. 344-354
+- **URL**: https://aclanthology.org/P16-1033/
+- **Key Contribution**: Uses **head entropy** to select the most informative individual dependency arcs for annotation. The paper demonstrates that partial annotation AL achieves comparable performance to full-sentence annotation with **40-60% less annotation effort** (measured by number of arcs annotated).
+- **Methodology**: Head entropy measures the entropy of the distribution over possible heads for each token using the Inside-Outside algorithm on a probabilistic parser.
+
+**Assessment**: This claim is **well-supported**. Li et al. (2016) is the seminal paper on partial annotation for dependency parsing and the cost reduction figures are correctly cited.
+
+**Sources**:
+- [Active Learning for Dependency Parsing with Partial Annotation](https://aclanthology.org/P16-1033/)
+- [Training Dependency Parsers with Partial Annotation (arXiv)](https://arxiv.org/abs/1609.09247)
+
+### 1.3. Shi et al. (2021): ~20-30% Fewer Sentences with DPP Batch Diversity
+
+**Claim**: "Shi et al. (2021) claiming ~20-30% fewer sentences with DPP batch diversity"
+
+**Verification**: ✅ **ACCURATE**
+
+**Evidence**:
+- **Paper**: Tianze Shi, Adrian Benton, Igor Malioutov, Ozan Irsoy, "Diversity-Aware Batch Active Learning for Dependency Parsing," *NAACL 2021*
+- **URL**: https://aclanthology.org/2021.naacl-main.207/
+- **Key Finding**: Simulation experiments on English newswire (Penn Treebank) show that selecting diverse batches with **determinantal point processes (DPPs)** is superior to diversity-agnostic strategies, especially during the **initial stages** of the learning process.
+- **Results**: DPP-based batch selection reaches the same LAS (Labeled Attachment Score) as random sampling with approximately **20-30% fewer sentences**, particularly in early learning stages.
+- **Technical Detail**: DPPs enforce diversity by promoting repulsiveness in batch selection, preventing the selection of many near-duplicate sentences that waste annotation budget.
+
+**Assessment**: This claim is **accurate**. Shi et al. (2021) is the state-of-the-art work on diversity-aware batch active learning for dependency parsing.
+
+**Sources**:
+- [Diversity-Aware Batch Active Learning for Dependency Parsing](https://aclanthology.org/2021.naacl-main.207/)
+- [GitHub Repository](https://github.com/tzshi/dpp-al-parsing-naacl21)
+- [ArXiv Preprint](https://arxiv.org/abs/2104.13936)
+
+### 1.4. Zhang et al. (2023): Partial Annotation + Self-Training
+
+**Claim**: "Zhang et al. (2023) on partial annotation + self-training"
+
+**Verification**: ✅ **ACCURATE**
+
+**Evidence**:
+- **Paper**: Zhisong Zhang, Emma Strubell, Eduard Hovy, "Data-efficient Active Learning for Structured Prediction with Partial Annotation and Self-Training," *Findings of EMNLP 2023*, pp. 12991-13008
+- **URL**: https://aclanthology.org/2023.findings-emnlp.865/
+- **Key Contributions**:
+  1. Combines **partial annotation** (selecting only the most informative sub-structures) with **self-training** (using model predictions as pseudo-labels for unannotated parts)
+  2. Uses an **adaptive error estimator** to dynamically adjust the partial selection ratio based on the current model's capability
+  3. Evaluates across **four structured prediction tasks** including dependency parsing
+- **Results**: The combination of partial annotation and self-training with adaptive selection ratio **reduces annotation cost** over strong full-annotation baselines across all four tasks.
+
+**Assessment**: This is the **most recent state-of-the-art** work combining partial annotation with self-training for dependency parsing. The claim is accurate.
+
+**Sources**:
+- [Data-efficient Active Learning for Structured Prediction](https://aclanthology.org/2023.findings-emnlp.865/)
+- [ArXiv Preprint](https://arxiv.org/abs/2305.12634)
+
+---
+
+## 2. CRF Marginal Uncertainty for Word Segmentation
+
+**Claim**: "CRF marginal uncertainty as a query strategy for word segmentation - is this standard practice?"
+
+**Verification**: ✅ **STANDARD PRACTICE**
+
+**Evidence**:
+
+### General Framework
+- **Settles & Craven (2008)**: "An Analysis of Active Learning Strategies for Sequence Labeling Tasks," *EMNLP 2008*
+  - **URL**: https://aclanthology.org/D08-1112/
+  - Comprehensive analysis of AL strategies for sequence labeling with CRFs
+  - **Query Strategies Analyzed**:
+    1. **Least Confidence (LC)**: Queries the instance the model is most uncertain about
+    2. **Token Entropy**: Measures entropy of the model's posteriors over its labeling, normalized by sequence length
+    3. **Margin**: Queries instance with smallest margin between posteriors for two most likely labelings
+    4. **Sequence Entropy (SE)**: Measures entropy of the label sequence as a whole
+  - **Key Finding**: Among uncertainty sampling methods, **Sequence Entropy (SE)** and **Least Confidence (LC)** performed best for CRF-based sequence labeling
+
+### CRF-Specific Applications
+- **LTP Strategy**: "LTP: A New Active Learning Strategy for CRF-Based Named Entity Recognition," *Neural Processing Letters*, 2021
+  - **URL**: https://link.springer.com/article/10.1007/s11063-021-10737-x
+  - Proposes **Lowest Token Probability (LTP)** strategy that combines input and output of CRF to select informative instances
+  - Examines traditional AL strategies in Word2Vec-BiLSTM-CRF and BERT-CRF contexts
+
+### Marginal Uncertainty Calculation
+For CRFs, confidence/uncertainty can be calculated using:
+- **Posterior probabilities**: The model's confidence in its predictions
+- **Marginal uncertainty**: Uncertainty for each token in the sequence, computed from CRF's forward-backward algorithm
+- **Token-level scoring**: `u_i = 1 - max(P(label|x,i))` where the probability is marginalized over all possible full-sequence labelings
+
+**Assessment**: Using CRF marginal uncertainty (token-level uncertainty from the CRF's posterior distribution) for word segmentation active learning is **well-established practice** in the literature. Settles & Craven (2008) is the foundational reference.
+
+**Sources**:
+- [An Analysis of Active Learning Strategies for Sequence Labeling](https://aclanthology.org/D08-1112/)
+- [LTP Strategy for CRF-Based NER](https://link.springer.com/article/10.1007/s11063-021-10737-x)
+
+---
+
+## 3. Head Entropy for Dependency Parsing AL
+
+**Claim**: "Head entropy for dependency parsing AL - who introduced this, is it state of the art?"
+
+**Verification**: ✅ **WELL-ESTABLISHED, STATE-OF-THE-ART**
+
+**Evidence**:
+
+### Introduction
+- **Li et al. (2016)**: "Active Learning for Dependency Parsing with Partial Annotation," *ACL 2016*
+  - **Introduced head entropy** as the primary informativeness measure for arc-level selection in dependency parsing
+  - **Definition**: Head entropy measures the entropy of the distribution over possible heads for each token: `H(h_i) = -Σ_j P(head=j|x,i) log P(head=j|x,i)`
+  - **Computation**: Uses the **Inside-Outside algorithm** on a probabilistic dependency parser to compute marginal probabilities for each possible head assignment
+  - **Selection**: Tokens with highest head entropy are the most uncertain and therefore most informative for annotation
+
+### State-of-the-Art Status
+Head entropy remains the **standard uncertainty measure** for dependency parsing active learning:
+1. **Li et al. (2016)**: Original introduction for partial annotation
+2. **Zhang et al. (2023)**: Uses head entropy as part of the uncertainty estimation in their partial annotation + self-training framework
+3. **Shi et al. (2021)**: While focused on diversity (DPPs), still uses uncertainty scores (often entropy-based) as the informativeness component
+
+### Alternative Uncertainty Measures
+Other entropy-based measures for parsing include:
+- **Tree Entropy**: Entropy over the full parse tree distribution (more computationally expensive)
+- **Prediction Entropy**: Entropy of the full structured output
+- **Least Confidence**: 1 - P(most likely parse)
+
+**Assessment**: **Head entropy** was introduced by **Li et al. (2016)** and remains the **state-of-the-art** uncertainty measure for token/arc-level active learning in dependency parsing. It balances computational tractability with effectiveness.
+
+**Sources**:
+- [Active Learning for Dependency Parsing with Partial Annotation (Li et al., 2016)](https://aclanthology.org/P16-1033/)
+
+---
+
+## 4. DPP for Batch Active Learning in NLP
+
+**Claim**: "DPP (Determinantal Point Process) for batch active learning in NLP - key papers and results"
+
+**Verification**: ✅ **WELL-DOCUMENTED**
+
+**Key Papers**:
+
+### 1. Bıyık et al. (2019): General DPP Batch AL
+- **Paper**: "Batch Active Learning Using Determinantal Point Processes," *arXiv 1906.07975*, 2019
+- **URL**: https://arxiv.org/abs/1906.07975
+- **Contribution**: Principled batch AL method using DPPs (repulsive point process) for generating diverse batches
+- **Key Innovation**: DPPs can be tuned to balance diversity and informativeness
+- **GitHub**: https://github.com/Stanford-ILIAD/DPP-Batch-Active-Learning
+
+### 2. Shi et al. (2021): DPP for Dependency Parsing
+- **Paper**: "Diversity-Aware Batch Active Learning for Dependency Parsing," *NAACL 2021*
+- **URL**: https://aclanthology.org/2021.naacl-main.207/
+- **Application**: First major application of DPPs to NLP (dependency parsing)
+- **Problem Addressed**: Diversity-agnostic strategies (selecting top-N most uncertain) select many near-duplicate sentences, wasting annotation budget
+- **Solution**: DPPs enforce diversity by promoting repulsiveness between selected items
+- **Results**: ~20-30% fewer sentences needed compared to diversity-agnostic baselines, especially in early stages
+
+### 3. BADGE (2019-2020): Gradient-Based Diversity
+- **Paper**: "Deep Batch Active learning by Diverse Gradient Embeddings," *ICLR 2020*
+- **Approach**: BADGE (Batch Active learning by Diverse Gradient Embeddings) samples groups of points that are disparate and high-magnitude in gradient space
+- **Contribution**: Incorporates both predictive uncertainty and sample diversity without requiring DPPs
+
+### Key Concepts
+- **DPPs**: A class of distributions that promote diversity; can be tuned to balance informativeness and diversity
+- **Repulsive Point Process**: Items that are similar repel each other in the selection process
+- **Batch AL**: Selecting multiple instances simultaneously rather than one-at-a-time (more practical for real annotation scenarios)
+
+**Assessment**: DPPs for batch AL in NLP are **well-established** since 2019-2021. Shi et al. (2021) is the landmark NLP application for dependency parsing.
+
+**Sources**:
+- [Batch Active Learning Using Determinantal Point Processes](https://arxiv.org/abs/1906.07975)
+- [Diversity-Aware Batch Active Learning for Dependency Parsing](https://aclanthology.org/2021.naacl-main.207/)
+- [BADGE Algorithm](https://arxiv.org/abs/1906.03671)
+
+---
+
+## 5. Partial Annotation for Dependency Parsing - State of the Art
+
+**Claim**: "Partial annotation strategies for dependency parsing - what's the current state of the art?"
+
+**Verification**: ✅ **WELL-DOCUMENTED**
+
+**Timeline**:
+
+### Early Work (2010-2015)
+1. **Sassano & Kurohashi (2010)**: "Using Smaller Constituents Rather Than Sentences in Active Learning for Japanese Dependency Parsing," *ACL 2010*
+   - Key insight: annotate sub-sentential units (bunsetsus) rather than full sentences
+   - ~30% reduction in annotation effort
+
+2. **Mirroshandel & Nasr (2011)**: "Active Learning for Dependency Parsing Using Partially Annotated Sentences," *IWPT 2011*
+   - Early exploration of partial annotation for dependency parsing
+   - Only annotate the most uncertain arcs
+
+3. **Flannery & Mori (2015)**: "Combining Active Learning and Partial Annotation for Domain Adaptation," *IWPT 2015*
+   - Combined AL with partial annotation for domain adaptation
+
+### Mature Framework (2016-2017)
+4. **Li et al. (2016)**: "Active Learning for Dependency Parsing with Partial Annotation," *ACL 2016*
+   - **Seminal work**: Comprehensive framework for partial annotation
+   - **Head entropy**: Arc-level selection criterion
+   - **Results**: 40-60% less annotation effort for same performance
+
+5. **Zhang et al. (2017)**: "Dependency Parsing with Partial Annotations," *IJCNLP 2017*
+   - Compared forest-based training and constrained decoding for partial annotations
+   - Systematic comparison of partial annotation methods
+
+### State-of-the-Art (2023)
+6. **Zhang et al. (2023)**: "Data-efficient Active Learning for Structured Prediction with Partial Annotation and Self-Training," *Findings of EMNLP 2023*
+   - **Current SOTA**: Combines partial annotation with self-training
+   - **Adaptive selection ratio**: Uses error estimator to decide what percentage of sub-structures to annotate
+   - **Pseudo-labels**: Uses model predictions for unannotated parts
+   - Evaluated across 4 structured prediction tasks including dependency parsing
+
+### Current Best Practice (2023-2024)
+The state-of-the-art approach combines:
+1. **Partial annotation** (select most informative arcs)
+2. **Self-training** (pseudo-labels for unannotated arcs)
+3. **Adaptive selection ratio** (adjust based on model capability)
+4. **Diversity-aware batch selection** (DPPs for sentence-level diversity)
+
+**Assessment**: The current state of the art is **Zhang et al. (2023)**, which combines partial annotation with self-training using adaptive selection ratios.
+
+**Sources**:
+- [Data-efficient Active Learning for Structured Prediction (Zhang et al., 2023)](https://aclanthology.org/2023.findings-emnlp.865/)
+- [Active Learning for Dependency Parsing with Partial Annotation (Li et al., 2016)](https://aclanthology.org/P16-1033/)
+
+---
+
+## 6. Sequential/Pipeline Active Learning Across Multiple NLP Tasks
+
+**Claim**: "Are there any papers on sequential/pipeline active learning across multiple NLP tasks (WS → POS → DP)?"
+
+**Verification**: ⚠️ **NO DIRECT EVIDENCE**
+
+**Evidence**:
+
+### Multi-Task Active Learning
+Found evidence of **multi-task AL** (jointly learning multiple tasks), but **NOT sequential/pipeline AL**:
+
+1. **Multi-Task Learning in NLP** (General):
+   - "Multi-Task Learning in Natural Language Processing: An Overview," *arXiv 2109.09138*
+   - "A Survey of Multi-task Learning in Natural Language Processing," *EACL 2023*
+   - Training methods: (i) joint training, (ii) multi-step training
+   - Example: ERNIE 2.0 uses "Continual Incremental Multi-Task Learning" where tasks are trained sequentially
+
+2. **Multi-task AL for Transformers**:
+   - "Multi-task Active Learning for Pre-trained Transformer-based Models," *TACL 2022*
+   - Multi-task AL across multiple NLP tasks **simultaneously** (not sequentially)
+
+3. **Active Learning for Sequence Tagging**:
+   - "Active Learning for Sequence Tagging with Deep Pre-trained Models," *EACL 2021*
+
+### Pipeline Error Propagation (Related but NOT AL)
+Found extensive work on **error propagation in cascaded NLP pipelines**, but without active learning:
+
+1. **"Tackling Error Propagation through Reinforcement Learning"**, *EACL 2017*
+   - URL: https://aclanthology.org/E17-1064/
+   - Error propagation in sequential NLP tasks; uses RL to mitigate
+
+2. **"A Dual-layer CRFs Based Joint Decoding Method for Cascaded Segmentation"**, *IJCAI 2007*
+   - Joint decoding of cascaded sequence segmentation and labeling
+
+3. **Pipeline approaches**: "Traditional pipeline approaches yield to error propagation and prohibit joint training/decoding between subtasks"
+
+### Sequential Multi-Task Learning
+- **"One Network to Solve Them All: A Sequential Multi-Task Joint Learning Network"**, *arXiv 2105.06653*
+  - Sequential multi-task joint learning (train combined end-to-end pipeline)
+  - Not focused on active learning
+
+### Vietnamese Pipeline
+- Vietnamese NLP pipelines typically follow: word segmentation → POS tagging → dependency parsing
+- Error propagation is a known problem: "once some words are wrongly segmented, the subsequent POS tagging and parsing will also make mistakes"
+- Joint models proposed to address this (e.g., graph-based joint WS + DP)
+
+**Assessment**: ❌ **NO PUBLISHED WORK** specifically on sequential/pipeline active learning for WS → POS → DP. This represents a **clear research gap**. While multi-task AL exists (jointly selecting for multiple tasks), and cascaded pipeline error propagation is studied, there is no prior work on **active learning specifically for sequential NLP task pipelines** where each task's AL informs the next.
+
+**Research Gap**: A three-task sequential AL pipeline (WS → POS → DP) would be **novel** if properly executed.
+
+**Sources**:
+- [Multi-Task Learning in NLP Survey](https://arxiv.org/abs/2109.09138)
+- [Tackling Error Propagation through RL](https://aclanthology.org/E17-1064/)
+- [Multi-task AL for Transformers (TACL 2022)](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00515/113664/)
+
+---
+
+## 7. Solo Annotator Treebank Construction: Quality Assurance
+
+**Claim**: "Solo annotator treebank construction: What quality assurance methods are used? What self-consistency rates are typical?"
+
+**Verification**: ⚠️ **SOLO ANNOTATION NOT RECOMMENDED; TYPICAL QA USES MULTIPLE ANNOTATORS**
+
+**Evidence**:
+
+### Industry Best Practices: Multiple Annotators
+
+1. **Inter-Annotator Agreement (IAA) is Standard**:
+   - Treebank projects typically employ **multiple annotators** with IAA measurement
+   - "To ensure quality, treebank projects rely on a metric called inter-annotator agreement (IAA)"
+   - **Single-annotator accuracy measurements miss a critical signal**: whether the annotation task itself is well-defined and learnable
+
+2. **Vietnamese Treebank (VTB) Example**:
+   - Paper: "Ensuring annotation consistency and accuracy for Vietnamese treebank," *Language Resources and Evaluation*, 2017
+   - URL: https://link.springer.com/article/10.1007/s10579-017-9398-3
+   - **Process**:
+     - Two annotators annotated 600 texts (~8,000 sentences) as preliminary annotations
+     - Nine rounds of measurements to calculate accuracy and IAA
+     - **Inter-annotator agreement: >90%** for all three layers (word segmentation, POS, syntax)
+   - **Support**: Automatic labeling tools for pre-processing + tree editor for manual annotation
+
+3. **Gold Standard Quality Control**:
+   - Gold standard items inserted randomly at **5-10% rate** into annotator's queue
+   - Annotators unable to distinguish gold items from regular tasks (authentic performance measurement)
+
+4. **Iterative Guideline Refinement**:
+   - Annotators trained on draft guidelines
+   - Provide feedback about difficult constructions
+   - Guidelines revised based on feedback
+   - Multiple rounds of IAA measurement
+
+### Universal Dependencies Practices
+
+5. **English Web Treebank (EWT)**:
+   - "All the dependency annotations have been **single-annotated**"
+   - "A limited portion of them have been **double-annotated** with interannotator agreement at approximately **96%**"
+   - Subsequent correction done to improve consistency
+
+6. **Urdu UD Treebank**:
+   - **Total IAA for dependency relations: 94.5%**
+   - **Weighted Kappa: 0.876**
+   - URL: https://link.springer.com/article/10.1007/s10579-022-09581-9
+
+### Self-Consistency Targets
+
+7. **Kappa Interpretation Scales**:
+   - **0.0-0.2**: Poor
+   - **0.2-0.4**: Fair
+   - **0.4-0.6**: Moderate
+   - **0.6-0.8**: Substantial
+   - **0.8-1.0**: Almost perfect
+   - **Threshold**: Coefficient close to **0.8 is considered reliable** for most ML projects
+
+8. **Self-Consistency for Solo Annotators**:
+   - Re-annotate 5% of sentences blind → measure agreement with original annotation
+   - Target: **>95% self-consistency** (equivalent to Kappa ~0.90-0.95)
+   - This is what the UDD-1 plan proposes: "Re-annotate 5% blind every 200 sentences → target >95% self-consistency"
+
+### Pre-annotation for Quality and Speed
+
+9. **Penn Treebank Experience**:
+   - Pre-annotation using automatic parsers **increases speed and consistency** without reducing quality
+   - "Highly recommended to use in syntactic annotation projects"
+
+**Assessment**:
+- ✅ **Multiple annotators with IAA >90-95%** is the standard practice
+- ⚠️ **Solo annotation is possible** but requires rigorous self-consistency checks (re-annotating 5-10% blind, targeting >95% agreement)
+- ✅ **Pre-annotation is recommended** to improve both speed and consistency
+
+**Recommendation for Solo Annotator QA**:
+1. Re-annotate 5-10% of sentences blind every 200-300 sentences
+2. Target **>95% self-consistency** (raw agreement) or **Kappa >0.90**
+3. Use **pre-annotation** from automatic tools to improve consistency
+4. Document **difficult cases** and iteratively refine guidelines
+5. If possible, have a second expert annotator review a sample (10-20%) for external validation
+
+**Sources**:
+- [Vietnamese Treebank IAA Study](https://link.springer.com/article/10.1007/s10579-017-9398-3)
+- [UD Vietnamese-VTB](https://universaldependencies.org/treebanks/vi_vtb/index.html)
+- [Annotator Agreement Metrics](https://cleverx.com/blog/annotator-agreement-metrics-measuring-and-maintaining-annotation-quality-at-scale)
+- [Penn Treebank Annotation Manual](https://catalog.ldc.upenn.edu/docs/LDC95T7/cl93.html)
+
+---
+
+## 8. Annotation Speed Estimates
+
+**Claim**: "Annotation speed estimates: Are 50-100 sent/day for WS, 30-50 for POS, 15-30 for DP reasonable? What do other papers report?"
+
+**Verification**: ⚠️ **MIXED - SOME ESTIMATES ARE OPTIMISTIC**
+
+**Evidence**:
+
+### 8.1. Word Segmentation (WS): 50-100 sent/day
+
+**Claim**: 50-100 sentences/day for word segmentation
+
+**Verification**: ⚠️ **OPTIMISTIC BUT POSSIBLY ACHIEVABLE WITH PRE-ANNOTATION**
+
+**Evidence**:
+- **No direct quantitative data** found for manual word segmentation annotation speed
+- **Penn Treebank POS tagging** (analogous task): **750-1,500 words/hour** for experienced annotators
+  - Source: "Building a large annotated corpus of English: the Penn Treebank," *Computational Linguistics*, 1993
+  - URL: https://catalog.ldc.upenn.edu/docs/LDC95T7/cl93.html
+- **Vietnamese sentence length**: ~20-30 syllables/sentence on average
+- **Calculation**:
+  - 750 words/hour = **12.5 words/minute**
+  - At 25 syllables/sentence: **30 sentences/hour** = **240 sent/day** (8 hours)
+  - At 3 hours/day: **90 sentences/day**
+- **With pre-annotation**: Speed likely **2-3x faster** (correction vs. annotation from scratch)
+  - Estimated: **100-200 sent/day** with pre-annotation
+
+**Assessment**: ✅ **50-100 sent/day is ACHIEVABLE** for WS with pre-annotation (correcting CRF outputs). Without pre-annotation, closer to 30-50 sent/day.
+
+### 8.2. POS Tagging: 30-50 sent/day
+
+**Claim**: 30-50 sentences/day for POS tagging
+
+**Verification**: ⚠️ **CONSERVATIVE, LIKELY FASTER WITH PRE-ANNOTATION**
+
+**Evidence**:
+- **Penn Treebank POS annotation**:
+  - **750-1,500 words/hour** for experienced annotators
+  - Source: https://catalog.ldc.upenn.edu/docs/LDC95T7/cl93.html
+- **Calculation** (Vietnamese ~30 syllables/sent):
+  - 1,000 words/hour = **16.7 words/minute**
+  - At 30 syllables/sentence: **33 sentences/hour** = **264 sent/day** (8 hours)
+  - At 3 hours/day: **100 sentences/day**
+
+**Assessment**: ⚠️ **30-50 sent/day is CONSERVATIVE**. With pre-annotation (automatic POS tagger), likely **80-150 sent/day** is achievable.
+
+### 8.3. Dependency Parsing: 15-30 sent/day
+
+**Claim**: 15-30 sentences/day for dependency parsing
+
+**Verification**: ✅ **REASONABLE TO CONSERVATIVE**
+
+**Evidence**:
+
+1. **Penn Treebank Constituent Parsing** (with pre-annotation):
+   - **Partial automation reduced average annotation time from ~10 minutes to 1.5-2 minutes per sentence**
+   - Source: "Automation of Treebank Annotation," *CoNLL 1998*
+   - At 1.5 min/sent: **40 sent/hour** = **320 sent/day** (8 hours) or **120 sent/day** (3 hours)
+
+2. **Dependency Treebank (Complex Tectogrammatical Layer)**:
+   - **6.5 minutes per sentence on average**
+   - Source: "Dependency Treebanks: Methods, Annotation Schemes and Tools"
+   - At 6.5 min/sent: **9.2 sent/hour** = **74 sent/day** (8 hours) or **28 sent/day** (3 hours)
+
+3. **Without Pre-annotation**:
+   - Fully manual parsing: **10 minutes per sentence** (Penn Treebank, before automation)
+   - At 10 min/sent: **6 sent/hour** = **48 sent/day** (8 hours) or **18 sent/day** (3 hours)
+
+4. **Active Learning Context** (UDD-1's case):
+   - AL selects **most uncertain/difficult sentences** → expect **slower than average**
+   - Pre-annotation from automatic parser available → **correction task, not annotation from scratch**
+   - Estimated range: **20-40 sent/day** (2-5 min/sent average, 3-4 hours/day)
+
+**Assessment**: ✅ **15-30 sent/day is REASONABLE** for DP annotation in an AL context with pre-annotation. Conservative but realistic for difficult sentences.
+
+### Summary Table: Annotation Speed
+
+| Task | Claimed Rate (sent/day) | Literature Support | Assessment |
+|------|------------------------:|-------------------|------------|
+| **Word Segmentation** | 50-100 | 30-90 from scratch; 100-200 with pre-annotation | ✅ **Achievable with pre-annotation** |
+| **POS Tagging** | 30-50 | 100-264 with pre-annotation | ⚠️ **Conservative; likely 80-150** |
+| **Dependency Parsing** | 15-30 | 28-120 depending on complexity and pre-annotation | ✅ **Reasonable for difficult AL-selected sentences** |
+
+**Overall Assessment**: The claimed rates are **conservative to reasonable**. With pre-annotation (which UDD-1 has), the upper ends of the ranges (or even higher) are achievable for experienced annotators. The estimates for DP are particularly reasonable given that AL selects difficult sentences.
+
+**Sources**:
+- [Building a Large Annotated Corpus of English: The Penn Treebank](https://catalog.ldc.upenn.edu/docs/LDC95T7/cl93.html)
+- [Automation of Treebank Annotation](https://aclanthology.org/W98-1207/)
+- [Dependency Treebanks: Methods and Tools](https://www.researchgate.net/publication/1960118_Dependency_Treebanks_Methods_Annotation_Schemes_and_Tools)
+
+---
+
+## 9. Overall Assessment of Paper Claims
+
+### Claims That Are Accurate ✅
+
+1. **Li et al. (2016): 40-60% arc reduction** ✅ Accurate
+2. **Shi et al. (2021): ~20-30% sentence reduction with DPP** ✅ Accurate
+3. **Zhang et al. (2023): Partial annotation + self-training** ✅ Accurate and SOTA
+4. **CRF marginal uncertainty for WS** ✅ Standard practice (Settles & Craven, 2008)
+5. **Head entropy for DP** ✅ State-of-the-art (Li et al., 2016)
+6. **DPP for batch AL** ✅ Well-established (Shi et al., 2021)
+7. **DP annotation speed: 15-30 sent/day** ✅ Reasonable
+
+### Claims That Are Approximately Accurate ⚠️
+
+1. **Hwa (2004): ~50% reduction** ⚠️ Approximately accurate (documented: 36%; related work: 60%)
+2. **WS annotation speed: 50-100 sent/day** ⚠️ Achievable with pre-annotation (optimistic without)
+3. **POS annotation speed: 30-50 sent/day** ⚠️ Conservative; likely faster (80-150)
+4. **Solo annotator QA** ⚠️ Possible but not standard; requires rigorous self-consistency checks (>95%)
+
+### Claims That Are Unsupported ❌
+
+1. **Sequential/pipeline AL for WS → POS → DP** ❌ **NO PRIOR WORK** - this would be **novel**
+
+---
+
+## 10. Recommendations for Paper
+
+### Strengthen Citations
+1. **Hwa (2004)**: Cite specific result (36% reduction) or clarify "~50%" is approximate
+2. **Sequential AL**: Acknowledge this is **novel** - no prior work exists on sequential multi-task AL pipelines
+3. **Annotation speeds**: Cite Penn Treebank studies and acknowledge these are with pre-annotation
+
+### Add Missing Citations
+1. **Settles & Craven (2008)** for CRF marginal uncertainty in sequence labeling
+2. **Sassano & Kurohashi (2010)** for sub-sentential AL in dependency parsing (complements Li et al., 2016)
+3. **Penn Treebank annotation studies** for annotation speed estimates
+
+### Highlight Novel Contributions
+The paper should emphasize that:
+1. ✨ **No prior work on sequential multi-task AL** (WS → POS → DP) - this is **novel**
+2. ✨ **No prior work on Vietnamese AL for dependency parsing** - this is a **research gap**
+3. ✨ Combining modern SOTA methods (DPP, partial annotation, self-training) in a **three-task pipeline** is **original**
+
+### Quality Assurance
+For solo annotation:
+- Commit to **5-10% blind re-annotation** every 200-300 sentences
+- Target **>95% self-consistency** (Kappa >0.90)
+- Document difficult cases and guideline evolution
+- If possible, external validation on 10-20% sample
+
+---
+
+## References & Sources
+
+### Active Learning for Dependency Parsing
+- [Sample Selection for Statistical Parsing (Hwa, 2004)](https://direct.mit.edu/coli/article/30/3/253/1841/Sample-Selection-for-Statistical-Parsing)
+- [Active Learning for Dependency Parsing with Partial Annotation (Li et al., 2016)](https://aclanthology.org/P16-1033/)
+- [Diversity-Aware Batch Active Learning for Dependency Parsing (Shi et al., 2021)](https://aclanthology.org/2021.naacl-main.207/)
+- [Data-efficient Active Learning for Structured Prediction (Zhang et al., 2023)](https://aclanthology.org/2023.findings-emnlp.865/)
+
+### Active Learning for Sequence Labeling
+- [An Analysis of Active Learning Strategies for Sequence Labeling Tasks (Settles & Craven, 2008)](https://aclanthology.org/D08-1112/)
+- [LTP: A New Active Learning Strategy for CRF-Based NER (2021)](https://link.springer.com/article/10.1007/s11063-021-10737-x)
+
+### DPP and Batch Active Learning
+- [Batch Active Learning Using Determinantal Point Processes (2019)](https://arxiv.org/abs/1906.07975)
+- [BADGE: Deep Batch Active Learning (ICLR 2020)](https://arxiv.org/abs/1906.03671)
+
+### Treebank Annotation
+- [Building a Large Annotated Corpus of English: The Penn Treebank (1993)](https://catalog.ldc.upenn.edu/docs/LDC95T7/cl93.html)
+- [Ensuring annotation consistency and accuracy for Vietnamese treebank (2017)](https://link.springer.com/article/10.1007/s10579-017-9398-3)
+- [UD Vietnamese-VTB](https://universaldependencies.org/treebanks/vi_vtb/index.html)
+- [Automation of Treebank Annotation (1998)](https://aclanthology.org/W98-1207/)
+
+### Multi-Task and Pipeline Learning
+- [Multi-Task Learning in Natural Language Processing Survey (2021)](https://arxiv.org/abs/2109.09138)
+- [Tackling Error Propagation through Reinforcement Learning (2017)](https://aclanthology.org/E17-1064/)
+- [Multi-task Active Learning for Pre-trained Transformers (TACL 2022)](https://direct.mit.edu/tacl/article/doi/10.1162/tacl_a_00515/113664/)
+
+### Vietnamese NLP
+- [BKTreebank: Building a Vietnamese Dependency Treebank (2018)](https://aclanthology.org/L18-1341/)
+- [VnCoreNLP: A Vietnamese NLP Toolkit (2018)](https://aclanthology.org/N18-5012/)
+
+---
+
+## Conclusion
+
+The proposed three-task active learning pipeline for Vietnamese makes **generally accurate claims** about prior work in active learning for NLP annotation. The cost reduction percentages, query strategies, and methodologies cited are well-supported by the literature. However:
+
+1. **Sequential/pipeline AL is novel**: There is no prior work on active learning specifically for sequential multi-task NLP pipelines (WS → POS → DP). This should be highlighted as a **novel contribution**.
+
+2. **Annotation speed estimates are reasonable**: The claimed rates (50-100 for WS, 30-50 for POS, 15-30 for DP) are conservative to achievable, especially with pre-annotation.
+
+3. **Solo annotator QA requires rigor**: While possible, solo annotation is not standard practice. Rigorous self-consistency checks (5-10% blind re-annotation, >95% agreement) are essential.
+
+4. **Citation improvements needed**: Some claims (e.g., Hwa 2004's "~50%") should be more precisely cited, and foundational work (Settles & Craven 2008) should be added.
+
+**Overall**: The paper's technical foundation is **solid and well-grounded** in the state-of-the-art active learning literature. The approach is **methodologically sound** and represents a **meaningful contribution** to Vietnamese NLP and low-resource treebank construction.

research/mwe_analysis/km_ktb-ud-test.conllu

@@ -0,0 +1 @@
+404: Not Found

src/al_score_ws.py

@@ -1,6 +1,6 @@
 # /// script
 # requires-python = ">=3.9"
-# dependencies = ["pycrfsuite"]
+# dependencies = ["python-crfsuite"]
 # ///
 """Active Learning Cycle 0: Score word segmentation uncertainty on dev+test.
 

src/build_dict_plugin.py

@@ -1,403 +0,0 @@
-# /// script
-# requires-python = ">=3.9"
-# dependencies = []
-# ///
-"""Generate dictionary plugin files for Label Studio frontend JS injection.
-
-Produces three outputs:
-1. dict_data.json — sorted NFC-normalized dictionary entries as a JSON array
-2. dict_plugin.js — floating search panel + auto-lookup on region select
-3. Patches base.html — adds <script> tag in the {% block bottomjs %} block
-
-The JS plugin:
-- Loads dictionary via fetch("/static/dict_data.json") into a Set for O(1) lookup
-- Creates a fixed-position search panel at bottom-right (toggle with 'd' key)
-- Auto-fills search from selected annotation region via Htx.annotationStore polling
-- Only activates on annotation pages (checks window.APP_SETTINGS)
-
-Usage:
-    uv run src/build_dict_plugin.py --dict path/to/dictionary.txt
-    uv run src/build_dict_plugin.py --deploy  # also deploys to LS static/templates
-"""
-
-import argparse
-import json
-import re
-import unicodedata
-from pathlib import Path
-
-DEFAULT_DICT = (
-    "/home/claude-code/projects/workspace_underthesea/tree-1/"
-    "models/word_segmentation/udd_ws_v1_1-20260211_034002/dictionary.txt"
-)
-
-LS_INSTALL_DIRS = [
-    Path("/home/claude-code/.local/share/uv/tools/label-studio/"
-         "lib/python3.12/site-packages/label_studio"),
-    Path("/home/claude-code/projects/workspace_underthesea/label-studio/.venv/"
-         "lib/python3.12/site-packages/label_studio"),
-]
-
-
-def nfc(text):
-    return unicodedata.normalize("NFC", text)
-
-
-def load_dictionary(dict_path):
-    entries = []
-    with open(dict_path, encoding="utf-8") as f:
-        for line in f:
-            line = line.strip()
-            if line:
-                entries.append(nfc(line))
-    entries.sort()
-    return entries
-
-
-def build_dict_data_json(entries):
-    return json.dumps(entries, ensure_ascii=False)
-
-
-def build_dict_plugin_js():
-    return r"""(function() {
-  "use strict";
-
-  // Only activate on annotation pages
-  if (!window.APP_SETTINGS || !window.APP_SETTINGS.project) return;
-
-  var DICT_SET = null;
-  var DICT_LOWER = null;
-  var dictLoaded = false;
-
-  // Load dictionary
-  fetch("/static/dict_data.json")
-    .then(function(r) { return r.json(); })
-    .then(function(entries) {
-      DICT_LOWER = entries.map(function(e) { return e.toLowerCase(); });
-      DICT_SET = new Set(DICT_LOWER);
-      dictLoaded = true;
-      if (countEl) countEl.textContent = entries.length.toLocaleString() + " entries";
-      console.log("[dict-plugin] Loaded " + entries.length + " dictionary entries");
-    })
-    .catch(function(err) {
-      console.error("[dict-plugin] Failed to load dictionary:", err);
-    });
-
-  // --- Create floating panel ---
-  var panel = document.createElement("div");
-  panel.id = "dict-plugin-panel";
-  panel.innerHTML = [
-    '<div class="dp-header">',
-    '  <span class="dp-title">Dictionary</span>',
-    '  <span class="dp-status" id="dp-status"></span>',
-    '  <span class="dp-count" id="dp-count"></span>',
-    '  <button class="dp-close" id="dp-close" title="Close (d)">&times;</button>',
-    '</div>',
-    '<div class="dp-search">',
-    '  <input type="text" id="dp-input" placeholder="Search dictionary..." autocomplete="off">',
-    '</div>',
-    '<div class="dp-results" id="dp-results"></div>'
-  ].join("\n");
-  document.body.appendChild(panel);
-
-  // --- Styles ---
-  var style = document.createElement("style");
-  style.textContent = [
-    '#dict-plugin-panel {',
-    '  position: fixed; bottom: 16px; right: 16px; width: 340px;',
-    '  background: #fff; border: 1px solid #ccc; border-radius: 8px;',
-    '  box-shadow: 0 4px 16px rgba(0,0,0,0.15); z-index: 10000;',
-    '  font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;',
-    '  font-size: 14px; display: flex; flex-direction: column;',
-    '  max-height: 400px; overflow: hidden;',
-    '}',
-    '#dict-plugin-panel.dp-hidden { display: none; }',
-    '.dp-header {',
-    '  display: flex; align-items: center; padding: 8px 12px;',
-    '  border-bottom: 1px solid #eee; gap: 8px;',
-    '}',
-    '.dp-title { font-weight: 600; font-size: 13px; }',
-    '.dp-status {',
-    '  font-size: 12px; padding: 1px 6px; border-radius: 3px;',
-    '  white-space: nowrap;',
-    '}',
-    '.dp-status.in-dict { background: #c8e6c9; color: #2e7d32; }',
-    '.dp-status.not-dict { background: #ffcdd2; color: #c62828; }',
-    '.dp-count { font-size: 11px; color: #999; margin-left: auto; }',
-    '.dp-close {',
-    '  background: none; border: none; font-size: 18px; cursor: pointer;',
-    '  color: #999; padding: 0 4px; line-height: 1;',
-    '}',
-    '.dp-close:hover { color: #333; }',
-    '.dp-search { padding: 8px 12px; }',
-    '.dp-search input {',
-    '  width: 100%; padding: 6px 10px; font-size: 14px;',
-    '  border: 1px solid #ddd; border-radius: 4px; outline: none;',
-    '  box-sizing: border-box;',
-    '}',
-    '.dp-search input:focus { border-color: #4CAF50; }',
-    '.dp-results {',
-    '  padding: 4px 12px 8px; overflow-y: auto; max-height: 260px;',
-    '}',
-    '.dp-section { padding: 2px 0; }',
-    '.dp-section-title {',
-    '  font-size: 11px; font-weight: 600; color: #888;',
-    '  text-transform: uppercase; padding: 2px 0;',
-    '}',
-    '.dp-match { padding: 1px 0; font-size: 13px; }',
-    '.dp-match.exact { font-weight: 600; color: #2e7d32; }',
-    '.dp-match mark { background: #fff9c4; padding: 0 1px; border-radius: 2px; }',
-    '.dp-no-results { color: #999; padding: 4px 0; font-size: 13px; }'
-  ].join("\n");
-  document.head.appendChild(style);
-
-  var input = document.getElementById("dp-input");
-  var resultsDiv = document.getElementById("dp-results");
-  var statusEl = document.getElementById("dp-status");
-  var countEl = document.getElementById("dp-count");
-  var closeBtn = document.getElementById("dp-close");
-
-  // --- Toggle visibility ---
-  var visible = true;
-  function togglePanel() {
-    visible = !visible;
-    panel.classList.toggle("dp-hidden", !visible);
-    if (visible) input.focus();
-  }
-  closeBtn.addEventListener("click", togglePanel);
-
-  // Keyboard shortcut: 'd' to toggle (only when not typing in an input)
-  document.addEventListener("keydown", function(e) {
-    if (e.key === "d" && !e.ctrlKey && !e.altKey && !e.metaKey) {
-      var tag = (e.target.tagName || "").toLowerCase();
-      if (tag !== "input" && tag !== "textarea" && !e.target.isContentEditable) {
-        e.preventDefault();
-        togglePanel();
-      }
-    }
-  });
-
-  // --- Search with debounce ---
-  var debounceTimer;
-  input.addEventListener("input", function() {
-    clearTimeout(debounceTimer);
-    debounceTimer = setTimeout(doSearch, 150);
-  });
-
-  function escHtml(s) {
-    return s.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
-  }
-
-  function highlight(text, query) {
-    var idx = text.toLowerCase().indexOf(query);
-    if (idx < 0) return escHtml(text);
-    return escHtml(text.slice(0, idx)) + "<mark>" +
-      escHtml(text.slice(idx, idx + query.length)) + "</mark>" +
-      escHtml(text.slice(idx + query.length));
-  }
-
-  function doSearch() {
-    if (!dictLoaded) return;
-    var q = input.value.trim().toLowerCase();
-    if (!q) { resultsDiv.innerHTML = ""; updateStatus(q); return; }
-
-    updateStatus(q);
-
-    var exact = [], prefix = [], substring = [];
-    var MAX = 20;
-    for (var i = 0; i < DICT_LOWER.length; i++) {
-      var e = DICT_LOWER[i];
-      if (e === q) exact.push(i);
-      else if (e.startsWith(q)) { if (prefix.length < MAX) prefix.push(i); }
-      else if (e.indexOf(q) >= 0) { if (substring.length < MAX) substring.push(i); }
-      if (prefix.length >= MAX && substring.length >= MAX && exact.length > 0) break;
-    }
-
-    var h = "";
-    if (exact.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Exact</div>';
-      for (var j = 0; j < exact.length; j++)
-        h += '<div class="dp-match exact">' + highlight(DICT_LOWER[exact[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (prefix.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Prefix (' + prefix.length + ')</div>';
-      for (var j = 0; j < prefix.length; j++)
-        h += '<div class="dp-match">' + highlight(DICT_LOWER[prefix[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (substring.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Contains (' + substring.length + ')</div>';
-      for (var j = 0; j < substring.length; j++)
-        h += '<div class="dp-match">' + highlight(DICT_LOWER[substring[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (!h) h = '<div class="dp-no-results">No matches</div>';
-    resultsDiv.innerHTML = h;
-  }
-
-  function updateStatus(query) {
-    if (!query || !DICT_SET) {
-      statusEl.textContent = "";
-      statusEl.className = "dp-status";
-      return;
-    }
-    if (DICT_SET.has(query)) {
-      statusEl.textContent = "\u2713 in dict";
-      statusEl.className = "dp-status in-dict";
-    } else {
-      statusEl.textContent = "\u2717 not found";
-      statusEl.className = "dp-status not-dict";
-    }
-  }
-
-  // --- Auto-lookup on region select ---
-  var lastRegionText = "";
-  setInterval(function() {
-    if (!dictLoaded || !visible) return;
-    try {
-      var store = window.Htx && window.Htx.annotationStore;
-      if (!store) return;
-      var sel = store.selected;
-      if (!sel) return;
-      var regions = sel.regionStore && sel.regionStore.regions;
-      if (!regions) return;
-
-      // Find selected region
-      var selectedRegion = null;
-      for (var i = 0; i < regions.length; i++) {
-        if (regions[i].selected) { selectedRegion = regions[i]; break; }
-      }
-      if (!selectedRegion) {
-        if (lastRegionText) { lastRegionText = ""; }
-        return;
-      }
-
-      var text = selectedRegion.text || "";
-      if (text && text !== lastRegionText) {
-        lastRegionText = text;
-        input.value = text;
-        doSearch();
-      }
-    } catch(e) {
-      // Silently ignore — store structure may vary
-    }
-  }, 300);
-})();
-"""
-
-
-def patch_base_html(base_html_path):
-    """Patch base.html to include dict_plugin.js in the bottomjs block.
-
-    Adds <script src="/static/dict_plugin.js"></script> before {% endblock %}
-    in the bottomjs block. Idempotent — skips if already patched.
-    """
-    content = base_html_path.read_text(encoding="utf-8")
-
-    script_tag = '<script src="/static/dict_plugin.js"></script>'
-    if script_tag in content:
-        print(f"  Already patched: {base_html_path}")
-        return False
-
-    # Find {% endblock %} that closes {% block bottomjs %}
-    # Insert the script tag before the last {% endblock %} in the file
-    pattern = r'([ \t]*)({% endblock %})\s*$'
-    match = list(re.finditer(pattern, content, re.MULTILINE))
-    if not match:
-        print(f"  WARNING: Could not find {{% endblock %}} in {base_html_path}")
-        return False
-
-    # Use the last match (the bottomjs endblock is at the end of the file)
-    m = match[-1]
-    indent = m.group(1)
-    insert_pos = m.start()
-    patched = content[:insert_pos] + indent + script_tag + "\n" + content[insert_pos:]
-
-    base_html_path.write_text(patched, encoding="utf-8")
-    print(f"  Patched: {base_html_path}")
-    return True
-
-
-def main():
-    parser = argparse.ArgumentParser(
-        description="Generate dictionary plugin for Label Studio (JS injection)"
-    )
-    parser.add_argument(
-        "--dict",
-        default=DEFAULT_DICT,
-        help=f"Path to dictionary.txt (default: {DEFAULT_DICT})",
-    )
-    parser.add_argument(
-        "--output-dir",
-        default="src",
-        help="Output directory for generated files (default: src)",
-    )
-    parser.add_argument(
-        "--deploy",
-        action="store_true",
-        help="Deploy to LS static_build + patch base.html",
-    )
-    args = parser.parse_args()
-
-    dict_path = Path(args.dict)
-    if not dict_path.exists():
-        print(f"ERROR: {dict_path} not found")
-        return
-
-    # Load dictionary
-    entries = load_dictionary(dict_path)
-    print(f"Loaded {len(entries)} dictionary entries from {dict_path}")
-
-    # Generate dict_data.json
-    output_dir = Path(args.output_dir)
-    output_dir.mkdir(parents=True, exist_ok=True)
-
-    json_path = output_dir / "dict_data.json"
-    json_content = build_dict_data_json(entries)
-    json_path.write_text(json_content, encoding="utf-8")
-    json_kb = json_path.stat().st_size / 1024
-    print(f"Written {json_path} ({json_kb:.0f} KB, {len(entries)} entries)")
-
-    # Generate dict_plugin.js
-    js_path = output_dir / "dict_plugin.js"
-    js_content = build_dict_plugin_js()
-    js_path.write_text(js_content, encoding="utf-8")
-    js_kb = js_path.stat().st_size / 1024
-    print(f"Written {js_path} ({js_kb:.0f} KB)")
-
-    # Deploy
-    if args.deploy:
-        import shutil
-        print("\nDeploying to Label Studio installations:")
-        for ls_dir in LS_INSTALL_DIRS:
-            static_dir = ls_dir / "core" / "static_build"
-            templates_dir = ls_dir / "templates"
-
-            if not static_dir.exists():
-                print(f"  SKIP: {static_dir} not found")
-                continue
-
-            # Copy dict_data.json and dict_plugin.js to static_build
-            shutil.copy2(json_path, static_dir / "dict_data.json")
-            print(f"  Deployed dict_data.json -> {static_dir}")
-
-            shutil.copy2(js_path, static_dir / "dict_plugin.js")
-            print(f"  Deployed dict_plugin.js -> {static_dir}")
-
-            # Remove old dict_plugin.html if it exists
-            old_html = static_dir / "dict_plugin.html"
-            if old_html.exists():
-                old_html.unlink()
-                print(f"  Removed old dict_plugin.html")
-
-            # Patch base.html
-            base_html = templates_dir / "base.html"
-            if base_html.exists():
-                patch_base_html(base_html)
-            else:
-                print(f"  WARNING: {base_html} not found")
-
-
-if __name__ == "__main__":
-    main()

src/dict_plugin.js

@@ -1,225 +0,0 @@
-(function() {
-  "use strict";
-
-  // Only activate on annotation pages
-  if (!window.APP_SETTINGS || !window.APP_SETTINGS.project) return;
-
-  var DICT_SET = null;
-  var DICT_LOWER = null;
-  var dictLoaded = false;
-
-  // Load dictionary
-  fetch("/static/dict_data.json")
-    .then(function(r) { return r.json(); })
-    .then(function(entries) {
-      DICT_LOWER = entries.map(function(e) { return e.toLowerCase(); });
-      DICT_SET = new Set(DICT_LOWER);
-      dictLoaded = true;
-      if (countEl) countEl.textContent = entries.length.toLocaleString() + " entries";
-      console.log("[dict-plugin] Loaded " + entries.length + " dictionary entries");
-    })
-    .catch(function(err) {
-      console.error("[dict-plugin] Failed to load dictionary:", err);
-    });
-
-  // --- Create floating panel ---
-  var panel = document.createElement("div");
-  panel.id = "dict-plugin-panel";
-  panel.innerHTML = [
-    '<div class="dp-header">',
-    '  <span class="dp-title">Dictionary</span>',
-    '  <span class="dp-status" id="dp-status"></span>',
-    '  <span class="dp-count" id="dp-count"></span>',
-    '  <button class="dp-close" id="dp-close" title="Close (d)">&times;</button>',
-    '</div>',
-    '<div class="dp-search">',
-    '  <input type="text" id="dp-input" placeholder="Search dictionary..." autocomplete="off">',
-    '</div>',
-    '<div class="dp-results" id="dp-results"></div>'
-  ].join("\n");
-  document.body.appendChild(panel);
-
-  // --- Styles ---
-  var style = document.createElement("style");
-  style.textContent = [
-    '#dict-plugin-panel {',
-    '  position: fixed; bottom: 16px; right: 16px; width: 340px;',
-    '  background: #fff; border: 1px solid #ccc; border-radius: 8px;',
-    '  box-shadow: 0 4px 16px rgba(0,0,0,0.15); z-index: 10000;',
-    '  font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Roboto, sans-serif;',
-    '  font-size: 14px; display: flex; flex-direction: column;',
-    '  max-height: 400px; overflow: hidden;',
-    '}',
-    '#dict-plugin-panel.dp-hidden { display: none; }',
-    '.dp-header {',
-    '  display: flex; align-items: center; padding: 8px 12px;',
-    '  border-bottom: 1px solid #eee; gap: 8px;',
-    '}',
-    '.dp-title { font-weight: 600; font-size: 13px; }',
-    '.dp-status {',
-    '  font-size: 12px; padding: 1px 6px; border-radius: 3px;',
-    '  white-space: nowrap;',
-    '}',
-    '.dp-status.in-dict { background: #c8e6c9; color: #2e7d32; }',
-    '.dp-status.not-dict { background: #ffcdd2; color: #c62828; }',
-    '.dp-count { font-size: 11px; color: #999; margin-left: auto; }',
-    '.dp-close {',
-    '  background: none; border: none; font-size: 18px; cursor: pointer;',
-    '  color: #999; padding: 0 4px; line-height: 1;',
-    '}',
-    '.dp-close:hover { color: #333; }',
-    '.dp-search { padding: 8px 12px; }',
-    '.dp-search input {',
-    '  width: 100%; padding: 6px 10px; font-size: 14px;',
-    '  border: 1px solid #ddd; border-radius: 4px; outline: none;',
-    '  box-sizing: border-box;',
-    '}',
-    '.dp-search input:focus { border-color: #4CAF50; }',
-    '.dp-results {',
-    '  padding: 4px 12px 8px; overflow-y: auto; max-height: 260px;',
-    '}',
-    '.dp-section { padding: 2px 0; }',
-    '.dp-section-title {',
-    '  font-size: 11px; font-weight: 600; color: #888;',
-    '  text-transform: uppercase; padding: 2px 0;',
-    '}',
-    '.dp-match { padding: 1px 0; font-size: 13px; }',
-    '.dp-match.exact { font-weight: 600; color: #2e7d32; }',
-    '.dp-match mark { background: #fff9c4; padding: 0 1px; border-radius: 2px; }',
-    '.dp-no-results { color: #999; padding: 4px 0; font-size: 13px; }'
-  ].join("\n");
-  document.head.appendChild(style);
-
-  var input = document.getElementById("dp-input");
-  var resultsDiv = document.getElementById("dp-results");
-  var statusEl = document.getElementById("dp-status");
-  var countEl = document.getElementById("dp-count");
-  var closeBtn = document.getElementById("dp-close");
-
-  // --- Toggle visibility ---
-  var visible = true;
-  function togglePanel() {
-    visible = !visible;
-    panel.classList.toggle("dp-hidden", !visible);
-    if (visible) input.focus();
-  }
-  closeBtn.addEventListener("click", togglePanel);
-
-  // Keyboard shortcut: 'd' to toggle (only when not typing in an input)
-  document.addEventListener("keydown", function(e) {
-    if (e.key === "d" && !e.ctrlKey && !e.altKey && !e.metaKey) {
-      var tag = (e.target.tagName || "").toLowerCase();
-      if (tag !== "input" && tag !== "textarea" && !e.target.isContentEditable) {
-        e.preventDefault();
-        togglePanel();
-      }
-    }
-  });
-
-  // --- Search with debounce ---
-  var debounceTimer;
-  input.addEventListener("input", function() {
-    clearTimeout(debounceTimer);
-    debounceTimer = setTimeout(doSearch, 150);
-  });
-
-  function escHtml(s) {
-    return s.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;");
-  }
-
-  function highlight(text, query) {
-    var idx = text.toLowerCase().indexOf(query);
-    if (idx < 0) return escHtml(text);
-    return escHtml(text.slice(0, idx)) + "<mark>" +
-      escHtml(text.slice(idx, idx + query.length)) + "</mark>" +
-      escHtml(text.slice(idx + query.length));
-  }
-
-  function doSearch() {
-    if (!dictLoaded) return;
-    var q = input.value.trim().toLowerCase();
-    if (!q) { resultsDiv.innerHTML = ""; updateStatus(q); return; }
-
-    updateStatus(q);
-
-    var exact = [], prefix = [], substring = [];
-    var MAX = 20;
-    for (var i = 0; i < DICT_LOWER.length; i++) {
-      var e = DICT_LOWER[i];
-      if (e === q) exact.push(i);
-      else if (e.startsWith(q)) { if (prefix.length < MAX) prefix.push(i); }
-      else if (e.indexOf(q) >= 0) { if (substring.length < MAX) substring.push(i); }
-      if (prefix.length >= MAX && substring.length >= MAX && exact.length > 0) break;
-    }
-
-    var h = "";
-    if (exact.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Exact</div>';
-      for (var j = 0; j < exact.length; j++)
-        h += '<div class="dp-match exact">' + highlight(DICT_LOWER[exact[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (prefix.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Prefix (' + prefix.length + ')</div>';
-      for (var j = 0; j < prefix.length; j++)
-        h += '<div class="dp-match">' + highlight(DICT_LOWER[prefix[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (substring.length > 0) {
-      h += '<div class="dp-section"><div class="dp-section-title">Contains (' + substring.length + ')</div>';
-      for (var j = 0; j < substring.length; j++)
-        h += '<div class="dp-match">' + highlight(DICT_LOWER[substring[j]], q) + '</div>';
-      h += '</div>';
-    }
-    if (!h) h = '<div class="dp-no-results">No matches</div>';
-    resultsDiv.innerHTML = h;
-  }
-
-  function updateStatus(query) {
-    if (!query || !DICT_SET) {
-      statusEl.textContent = "";
-      statusEl.className = "dp-status";
-      return;
-    }
-    if (DICT_SET.has(query)) {
-      statusEl.textContent = "\u2713 in dict";
-      statusEl.className = "dp-status in-dict";
-    } else {
-      statusEl.textContent = "\u2717 not found";
-      statusEl.className = "dp-status not-dict";
-    }
-  }
-
-  // --- Auto-lookup on region select ---
-  var lastRegionText = "";
-  setInterval(function() {
-    if (!dictLoaded || !visible) return;
-    try {
-      var store = window.Htx && window.Htx.annotationStore;
-      if (!store) return;
-      var sel = store.selected;
-      if (!sel) return;
-      var regions = sel.regionStore && sel.regionStore.regions;
-      if (!regions) return;
-
-      // Find selected region
-      var selectedRegion = null;
-      for (var i = 0; i < regions.length; i++) {
-        if (regions[i].selected) { selectedRegion = regions[i]; break; }
-      }
-      if (!selectedRegion) {
-        if (lastRegionText) { lastRegionText = ""; }
-        return;
-      }
-
-      var text = selectedRegion.text || "";
-      if (text && text !== lastRegionText) {
-        lastRegionText = text;
-        input.value = text;
-        doSearch();
-      }
-    } catch(e) {
-      // Silently ignore — store structure may vary
-    }
-  }, 300);
-})();

src/eval_ws_gold.py

@@ -0,0 +1,409 @@
+"""Evaluate CRF word segmentation against gold annotations.
+
+Compares silver (CRF) predictions from udd-ws-v1.1-{dev,test}.txt against
+gold corrections in gold_ws_cycle1.txt. Reports Word F1/Precision/Recall,
+per-domain breakdown, and detailed error analysis.
+
+With --model, uses the CRF model to predict directly on gold sentences
+(instead of reading from silver files). This is needed when gold has been
+merged into silver files.
+
+Usage:
+    python src/eval_ws_gold.py
+    python src/eval_ws_gold.py --model path/to/model.crfsuite
+"""
+
+import argparse
+import sys
+from collections import Counter, defaultdict
+from pathlib import Path
+
+
+def parse_bio_file(path):
+    """Parse BIO file into dict of {sent_id: [(syllable, tag), ...]}."""
+    sentences = {}
+    current_id = None
+    current_tokens = []
+
+    with open(path, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.rstrip("\n")
+            if line.startswith("# sent_id = "):
+                if current_id and current_tokens:
+                    sentences[current_id] = current_tokens
+                current_id = line.split("= ", 1)[1]
+                current_tokens = []
+            elif line.startswith("# text = "):
+                continue
+            elif line.strip() == "":
+                if current_id and current_tokens:
+                    sentences[current_id] = current_tokens
+                    current_id = None
+                    current_tokens = []
+            elif "\t" in line:
+                parts = line.split("\t")
+                if len(parts) >= 2:
+                    current_tokens.append((parts[0], parts[1]))
+
+    if current_id and current_tokens:
+        sentences[current_id] = current_tokens
+
+    return sentences
+
+
+def bio_to_words(tokens):
+    """Convert BIO token list to list of word strings."""
+    words = []
+    current = []
+    for syl, tag in tokens:
+        if tag == "B-W":
+            if current:
+                words.append("_".join(current))
+            current = [syl]
+        elif tag == "I-W":
+            current.append(syl)
+    if current:
+        words.append("_".join(current))
+    return words
+
+
+def bio_to_word_spans(tokens):
+    """Convert BIO tokens to word spans as (start_idx, end_idx) tuples."""
+    spans = []
+    start = 0
+    for i, (syl, tag) in enumerate(tokens):
+        if tag == "B-W" and i > 0:
+            spans.append((start, i))
+            start = i
+    spans.append((start, len(tokens)))
+    return spans
+
+
+def get_domain(sent_id):
+    """Extract domain from sent_id prefix."""
+    if sent_id.startswith("vlc-"):
+        return "legal"
+    elif sent_id.startswith("uvn-"):
+        return "news"
+    elif sent_id.startswith("uvw-"):
+        return "wikipedia"
+    elif sent_id.startswith("uvb-f-"):
+        return "fiction"
+    elif sent_id.startswith("uvb-n-"):
+        return "non-fiction"
+    return "unknown"
+
+
+def compute_word_metrics(silver_words, gold_words):
+    """Compute word-level precision, recall, F1.
+
+    Uses multiset intersection (same as CoNLL WS eval).
+    """
+    silver_counter = Counter(silver_words)
+    gold_counter = Counter(gold_words)
+
+    # Multiset intersection
+    tp = sum((silver_counter & gold_counter).values())
+    pred_total = len(silver_words)
+    gold_total = len(gold_words)
+
+    precision = tp / pred_total if pred_total > 0 else 0
+    recall = tp / gold_total if gold_total > 0 else 0
+    f1 = 2 * precision * recall / (precision + recall) if (precision + recall) > 0 else 0
+
+    return precision, recall, f1, tp, pred_total, gold_total
+
+
+def compute_boundary_metrics(silver_tokens, gold_tokens):
+    """Compute boundary-level (syllable tag) accuracy."""
+    assert len(silver_tokens) == len(gold_tokens), \
+        f"Length mismatch: {len(silver_tokens)} vs {len(gold_tokens)}"
+
+    correct = 0
+    total = len(silver_tokens)
+    changes = {"B→I": 0, "I→B": 0}
+
+    for (s_syl, s_tag), (g_syl, g_tag) in zip(silver_tokens, gold_tokens):
+        if s_tag == g_tag:
+            correct += 1
+        else:
+            key = f"{s_tag[0]}→{g_tag[0]}"
+            changes[key] = changes.get(key, 0) + 1
+
+    accuracy = correct / total if total > 0 else 0
+    return accuracy, correct, total, changes
+
+
+def find_differences(sent_id, silver_tokens, gold_tokens):
+    """Find specific word segmentation differences between silver and gold."""
+    diffs = []
+    silver_words = bio_to_words(silver_tokens)
+    gold_words = bio_to_words(gold_tokens)
+
+    # Build position-based word mapping
+    silver_spans = bio_to_word_spans(silver_tokens)
+    gold_spans = bio_to_word_spans(gold_tokens)
+
+    if silver_spans == gold_spans:
+        return []
+
+    # Find differing positions
+    silver_set = set(silver_spans)
+    gold_set = set(gold_spans)
+
+    only_silver = silver_set - gold_set
+    only_gold = gold_set - silver_set
+
+    # Map spans to words
+    def span_to_word(tokens, start, end):
+        return "_".join(t[0] for t in tokens[start:end])
+
+    for s in sorted(only_silver):
+        word = span_to_word(silver_tokens, s[0], s[1])
+        diffs.append(("silver", s, word))
+
+    for g in sorted(only_gold):
+        word = span_to_word(gold_tokens, g[0], g[1])
+        diffs.append(("gold", g, word))
+
+    return diffs
+
+
+def classify_error(silver_word, gold_words_at_pos):
+    """Classify error type: over-merge, over-split, boundary-shift."""
+    s_parts = silver_word.split("_")
+    if len(s_parts) > 1 and all(len(g.split("_")) < len(s_parts) for g in gold_words_at_pos):
+        return "over-merge"
+    if len(s_parts) < max(len(g.split("_")) for g in gold_words_at_pos):
+        return "over-split"
+    return "boundary-shift"
+
+
+def predict_with_model(model_path, gold):
+    """Use CRF model to predict on gold sentence syllables.
+
+    Returns dict of {sent_id: [(syllable, tag), ...]}.
+    """
+    import pycrfsuite
+    # Reuse feature extraction from al_score_ws
+    sys.path.insert(0, str(Path(__file__).parent))
+    from al_score_ws import extract_syllable_features, load_dictionary
+
+    model_dir = model_path.parent
+    dict_path = model_dir / "dictionary.txt"
+
+    tagger = pycrfsuite.Tagger()
+    tagger.open(str(model_path))
+    print(f"Model loaded: {model_path}")
+
+    dictionary = None
+    if dict_path.exists():
+        dictionary = load_dictionary(dict_path)
+        print(f"Dictionary loaded: {len(dictionary)} entries")
+
+    tag_map = {"B": "B-W", "I": "I-W"}
+    predictions = {}
+    for sid, tokens in gold.items():
+        syllables = [t[0] for t in tokens]
+        # Extract features
+        xseq = [
+            [f"{k}={v}" for k, v in extract_syllable_features(syllables, i, dictionary).items()]
+            for i in range(len(syllables))
+        ]
+        pred_tags = tagger.tag(xseq)
+        predictions[sid] = [
+            (syl, tag_map.get(tag, tag)) for syl, tag in zip(syllables, pred_tags)
+        ]
+
+    return predictions
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Evaluate WS against gold")
+    parser.add_argument("--model", type=str, default=None,
+                        help="CRF model path for direct prediction")
+    args = parser.parse_args()
+
+    base = Path("/home/claude-code/projects/workspace_underthesea/UDD-1")
+    gold_path = base / "gold_ws_cycle1.txt"
+
+    # Parse gold
+    gold = parse_bio_file(gold_path)
+    print(f"Gold sentences: {len(gold)}")
+
+    if args.model:
+        # Predict using CRF model directly
+        model_path = Path(args.model)
+        if not model_path.exists():
+            # Auto-detect latest model
+            tree1_models = base.parent / "tree-1" / "models" / "word_segmentation"
+            model_dirs = sorted(tree1_models.glob("udd_ws_v1_1-*"))
+            if model_dirs:
+                model_path = model_dirs[-1] / "model.crfsuite"
+        silver = predict_with_model(model_path, gold)
+        print(f"CRF predictions: {len(silver)} sentences")
+    else:
+        # Parse silver (dev + test) from files
+        silver_dev = parse_bio_file(base / "udd-ws-v1.1-dev.txt")
+        silver_test = parse_bio_file(base / "udd-ws-v1.1-test.txt")
+        silver = {**silver_dev, **silver_test}
+        print(f"Silver sentences loaded: {len(silver_dev)} dev + {len(silver_test)} test")
+
+    # Match gold to silver
+    matched = []
+    missing = []
+    for sid in gold:
+        if sid in silver:
+            matched.append(sid)
+        else:
+            missing.append(sid)
+
+    print(f"Matched: {len(matched)}, Missing in silver: {len(missing)}")
+    if missing:
+        print(f"  Missing: {missing}")
+
+    # === Overall metrics ===
+    total_tp = total_pred = total_gold = 0
+    total_syl_correct = total_syl = 0
+    all_changes = Counter()
+    domain_stats = defaultdict(lambda: {"tp": 0, "pred": 0, "gold": 0, "syl_correct": 0, "syl_total": 0, "n": 0})
+    all_diffs = []
+    error_types = Counter()
+
+    for sid in matched:
+        s_tokens = silver[sid]
+        g_tokens = gold[sid]
+
+        # Check syllable alignment
+        s_syls = [t[0] for t in s_tokens]
+        g_syls = [t[0] for t in g_tokens]
+        if s_syls != g_syls:
+            print(f"  WARNING: syllable mismatch in {sid}")
+            print(f"    Silver: {' '.join(s_syls[:10])}...")
+            print(f"    Gold:   {' '.join(g_syls[:10])}...")
+            continue
+
+        # Word metrics
+        s_words = bio_to_words(s_tokens)
+        g_words = bio_to_words(g_tokens)
+        p, r, f1, tp, pred, gtotal = compute_word_metrics(s_words, g_words)
+        total_tp += tp
+        total_pred += pred
+        total_gold += gtotal
+
+        # Boundary metrics
+        acc, correct, total, changes = compute_boundary_metrics(s_tokens, g_tokens)
+        total_syl_correct += correct
+        total_syl += total
+        all_changes.update(changes)
+
+        # Domain stats
+        domain = get_domain(sid)
+        ds = domain_stats[domain]
+        ds["tp"] += tp
+        ds["pred"] += pred
+        ds["gold"] += gtotal
+        ds["syl_correct"] += correct
+        ds["syl_total"] += total
+        ds["n"] += 1
+
+        # Differences
+        diffs = find_differences(sid, s_tokens, g_tokens)
+        if diffs:
+            all_diffs.append((sid, domain, diffs, s_tokens, g_tokens))
+
+    # === Print Results ===
+    print("\n" + "=" * 60)
+    print("EVALUATION: CRF Silver vs Gold (Cycle 1)")
+    print("=" * 60)
+
+    # Overall
+    overall_p = total_tp / total_pred if total_pred else 0
+    overall_r = total_tp / total_gold if total_gold else 0
+    overall_f1 = 2 * overall_p * overall_r / (overall_p + overall_r) if (overall_p + overall_r) else 0
+    syl_acc = total_syl_correct / total_syl if total_syl else 0
+
+    print(f"\n## Overall ({len(matched)} sentences)")
+    print(f"  Syllable Accuracy: {syl_acc:.4f} ({total_syl_correct}/{total_syl})")
+    print(f"  Word Precision:    {overall_p:.4f}")
+    print(f"  Word Recall:       {overall_r:.4f}")
+    print(f"  Word F1:           {overall_f1:.4f}")
+    print(f"  Boundary changes:  {dict(all_changes)}")
+    print(f"    B→I (over-merge in silver): {all_changes.get('B→I', 0)}")
+    print(f"    I→B (over-split in silver): {all_changes.get('I→B', 0)}")
+
+    # Per-domain
+    print(f"\n## Per-Domain Breakdown")
+    print(f"  {'Domain':<14} {'N':>4} {'Syl Acc':>8} {'P':>7} {'R':>7} {'F1':>7}")
+    print(f"  {'-'*14} {'-'*4} {'-'*8} {'-'*7} {'-'*7} {'-'*7}")
+    for domain in ["legal", "news", "wikipedia", "fiction", "non-fiction"]:
+        ds = domain_stats[domain]
+        if ds["n"] == 0:
+            continue
+        dp = ds["tp"] / ds["pred"] if ds["pred"] else 0
+        dr = ds["tp"] / ds["gold"] if ds["gold"] else 0
+        df1 = 2 * dp * dr / (dp + dr) if (dp + dr) else 0
+        dacc = ds["syl_correct"] / ds["syl_total"] if ds["syl_total"] else 0
+        print(f"  {domain:<14} {ds['n']:>4} {dacc:>8.4f} {dp:>7.4f} {dr:>7.4f} {df1:>7.4f}")
+
+    # Error analysis
+    print(f"\n## Error Analysis ({len(all_diffs)} sentences with differences)")
+
+    merge_errors = []  # silver merged, gold split
+    split_errors = []  # silver split, gold merged
+
+    for sid, domain, diffs, s_tokens, g_tokens in all_diffs:
+        s_spans = set(bio_to_word_spans(s_tokens))
+        g_spans = set(bio_to_word_spans(g_tokens))
+
+        only_silver = s_spans - g_spans
+        only_gold = g_spans - s_spans
+
+        def span_word(tokens, s, e):
+            return "_".join(t[0] for t in tokens[s:e])
+
+        for span in only_silver:
+            word = span_word(s_tokens, span[0], span[1])
+            n_syls = span[1] - span[0]
+            # Check if this span overlaps with multiple gold spans (over-merge)
+            overlapping_gold = [g for g in only_gold if g[0] < span[1] and g[1] > span[0]]
+            if overlapping_gold and n_syls > 1:
+                gold_words = [span_word(g_tokens, g[0], g[1]) for g in overlapping_gold]
+                merge_errors.append((sid, domain, word, gold_words))
+
+        for span in only_gold:
+            word = span_word(g_tokens, span[0], span[1])
+            n_syls = span[1] - span[0]
+            overlapping_silver = [s for s in only_silver if s[0] < span[1] and s[1] > span[0]]
+            if overlapping_silver and n_syls > 1:
+                silver_words = [span_word(s_tokens, s[0], s[1]) for s in overlapping_silver]
+                split_errors.append((sid, domain, word, silver_words))
+
+    print(f"\n### Over-merge errors (silver merged what gold splits): {len(merge_errors)}")
+    for sid, domain, silver_word, gold_words in sorted(merge_errors, key=lambda x: x[1]):
+        print(f"  [{domain:>12}] {sid}: {silver_word} → {' | '.join(gold_words)}")
+
+    print(f"\n### Over-split errors (silver split what gold merges): {len(split_errors)}")
+    for sid, domain, gold_word, silver_words in sorted(split_errors, key=lambda x: x[1]):
+        print(f"  [{domain:>12}] {sid}: {' | '.join(silver_words)} → {gold_word}")
+
+    # Summary of all differences per sentence
+    print(f"\n## All Differences (sentence-level)")
+    for sid, domain, diffs, s_tokens, g_tokens in sorted(all_diffs, key=lambda x: x[1]):
+        s_words = bio_to_words(s_tokens)
+        g_words = bio_to_words(g_tokens)
+        s_set = set(s_words)
+        g_set = set(g_words)
+
+        s_only = Counter(s_words) - Counter(g_words)
+        g_only = Counter(g_words) - Counter(s_words)
+        if s_only or g_only:
+            print(f"\n  [{domain}] {sid}")
+            if s_only:
+                print(f"    Silver only: {dict(s_only)}")
+            if g_only:
+                print(f"    Gold only:   {dict(g_only)}")
+
+
+if __name__ == "__main__":
+    main()

src/fetch_ws_sentences.py

@@ -37,6 +37,11 @@ TONED_VOWELS = set(
 # All Vietnamese diacritical characters (toned vowels + base vowels ă, â, ê, ô, ơ, ư, đ)
 VIET_DIACRITICS = TONED_VOWELS | set('ăâêôơưđĂÂÊÔƠƯĐ')
 
+# Characters unique to Vietnamese (not in French/Portuguese/other Latin scripts)
+# ă/ơ/ư and their toned variants distinguish Vietnamese from French (which shares â, ê, ô, é, è, à)
+VIET_ONLY_CHARS = set('ăắằẳẵặơớờởỡợưứừửữựđ'
+                      'ĂẮẰẲẴẶƠỚỜỞỠỢƯỨỪỬỮỰĐ')
+
 
 # ============================================================================
 # Shared text cleaning
@@ -223,11 +228,12 @@ def base_valid(sent):
         # Digit glued to Vietnamese text (e.g., "59Sẹo")
         if tone_count >= 1 and re.search(r'\d[a-zA-ZĐđÀ-ỹ]', token):
             return False, sent
-    # Language detection with Vietnamese diacritics fallback
+    # Language detection with Vietnamese-only character fallback
     if lang_detect(sent) != "vi":
-        # Fallback: accept if sentence has 3+ Vietnamese diacritical characters
-        viet_char_count = sum(1 for c in sent if c in VIET_DIACRITICS)
-        if viet_char_count < 3:
+        # Fallback: accept only if sentence has Vietnamese-only chars (ă, ơ, ư, đ and variants)
+        # French shares â, ê, ô, é, è, à with Vietnamese, so those are not sufficient
+        vn_only_count = sum(1 for c in sent if c in VIET_ONLY_CHARS)
+        if vn_only_count < 1:
             return False, sent
     return True, sent
 

src/fix_ws_errors.py

@@ -4,10 +4,12 @@
 # ///
 """Fix known word segmentation errors in UDD-1.1 BIO files.
 
-Four fix passes:
+Seven fix passes:
 1.   Split cross-boundary merges (uppercase mid-token signals)
 1.5  Split long tokens (5+ syllables) via vocab-based greedy decomposition
 2.   Merge always-split compounds (dictionary compounds + inconsistent forms)
+2.5  Split foreign word merges (Latin-script tokens without Vietnamese diacritics)
+2.75 Split proper name boundary merges (uppercase→lowercase transitions within words)
 3.   Validate BIO invariants
 
 Usage:
@@ -16,6 +18,7 @@ Usage:
 """
 
 import argparse
+import re
 import sys
 from collections import Counter, defaultdict
 from os.path import dirname, isfile, join
@@ -53,6 +56,32 @@ MERGE_TERMS = {
     ("thuê", "khoán"),           # 62 split vs 2 single
     ("hòa", "giải"),             # 53 split vs 30 single
     ("bốc", "hàng"),             # 35 split vs 1 single
+    # ---- Cycle 1 gold corrections: new compound merges ----
+    ("ủy", "ban"),               # committee
+    ("lính", "thú"),             # soldier
+    ("mu", "rùa"),               # turtle shell
+    ("trêu", "ghẹo"),           # tease
+    ("sương", "mai"),            # morning dew
+    ("mái", "nhà"),              # roof
+    ("nghiến", "răng"),          # gnash teeth
+    ("nheo", "nheo"),            # squint
+    ("dơn", "dớt"),              # pale/sickly
+    ("xua", "tay"),              # wave hand
+    ("nói", "gở"),               # say unlucky things
+    ("bơi", "chó"),              # dog paddle
+    ("người", "thương"),         # beloved
+    ("chăn", "lợn"),             # pig herding
+    ("khay", "trà"),             # tea tray
+    ("đồng", "tự"),              # homograph
+    ("tại", "ngũ"),              # in service (military)
+    ("hành", "chánh"),           # administration
+    ("lượng", "tử"),             # quantum
+    ("tích", "lũy"),             # accumulate
+    ("siêu", "máy", "tính"),     # supercomputer
+    ("đường", "thẳng"),          # straight line
+    ("đầm", "đuôi", "cá"),      # fishtail dress
+    ("như", "điên"),              # like crazy
+    ("tẩy", "chay"),             # boycott
 }
 
 # Build index for efficient longest-match lookup: {length: [term, ...]}
@@ -196,6 +225,21 @@ def build_split_vocab(all_sentences, min_count=5):
     }
 
 
+def build_viet_syllables(all_sentences, min_count=50):
+    """Build set of common Vietnamese syllables for foreign word filtering.
+
+    Counts individual syllables across all sentences and returns those
+    appearing at least `min_count` times (lowercased). These are used to
+    distinguish Vietnamese multi-syllable words (like "kinh doanh") from
+    truly foreign tokens (like "Max Planck").
+    """
+    counts = Counter()
+    for sent in all_sentences:
+        for syl in sent["syllables"]:
+            counts[syl.lower()] += 1
+    return {syl for syl, c in counts.items() if c >= min_count}
+
+
 # ============================================================================
 # Fix passes
 # ============================================================================
@@ -386,6 +430,201 @@ def fix_merge_compounds(syllables, tags):
     return new_tags, changes
 
 
+def _is_latin_no_vietnamese(s):
+    """Check if a string is purely Latin-script without Vietnamese diacritics.
+
+    Returns True for ASCII Latin (a-z, A-Z, 0-9, hyphen) and common Latin
+    extensions BUT NOT Vietnamese-specific characters (ă, â, đ, ê, ô, ơ, ư
+    and their tone marks).
+    """
+    # Vietnamese diacritics pattern: any character with Vietnamese-specific marks
+    vietnamese_chars = re.compile(
+        r'[àáảãạăắằẳẵặâấầẩẫậèéẻẽẹêếềểễệìíỉĩịòóỏõọôốồổỗộơớờởỡợ'
+        r'ùúủũụưứừửữựỳýỷỹỵđÀÁẢÃẠĂẮẰẲẴẶÂẤẦẨẪẬÈÉẺẼẸÊẾỀỂỄỆÌÍỈĨỊ'
+        r'ÒÓỎÕỌÔỐỒỔỖỘƠỚỜỞỠỢÙÚỦŨỤƯỨỪỬỮỰỲÝỶỸỴĐ]'
+    )
+    if vietnamese_chars.search(s):
+        return False
+    # Must contain at least one Latin letter
+    return bool(re.search(r'[a-zA-Z]', s))
+
+
+# Known foreign proper names that should stay merged (whitelist)
+FOREIGN_NAME_WHITELIST = {
+    "beethoven", "homer", "odysseus", "cecelia", "ahern", "holly",
+    "hideoshi", "gurth", "euler", "hilbert", "rydberg", "bohr",
+    "frankael-zermelo", "giambattista", "valli", "dachau",
+    "habsburg", "newton", "einstein", "darwin", "shakespeare",
+}
+
+
+def fix_foreign_words(syllables, tags, viet_syllables):
+    """Pass 2.5: Split foreign word merges.
+
+    Detects multi-syllable tokens where ALL syllables are Latin-script only
+    (no Vietnamese diacritics) AND none of the syllables are common Vietnamese
+    syllables. Each such foreign syllable becomes its own word (B-W).
+
+    Args:
+        syllables: list of syllable strings.
+        tags: list of BIO tag strings.
+        viet_syllables: set of common Vietnamese syllables (lowercase) for
+            filtering out false positives like "kinh doanh".
+
+    Returns (new_tags, list of change descriptions).
+    """
+    new_tags = list(tags)
+    changes = []
+
+    # Reconstruct word spans
+    word_spans = []
+    current_start = 0
+    for i in range(len(tags)):
+        if tags[i] == "B-W" and i > 0:
+            word_spans.append((current_start, i))
+            current_start = i
+    word_spans.append((current_start, len(tags)))
+
+    for start, end in word_spans:
+        n_syls = end - start
+        if n_syls < 2:
+            continue
+
+        # Check if ALL syllables are Latin-only (no Vietnamese diacritics)
+        all_latin = all(_is_latin_no_vietnamese(syllables[j]) for j in range(start, end))
+        if not all_latin:
+            continue
+
+        # Check if ANY syllable is a common Vietnamese syllable → skip
+        has_viet = any(
+            syllables[j].lower() in viet_syllables for j in range(start, end)
+        )
+        if has_viet:
+            continue
+
+        # Check whitelist: if the whole token is a known name, skip
+        token_lower = " ".join(syllables[start:end]).lower()
+        if token_lower in FOREIGN_NAME_WHITELIST:
+            continue
+
+        # Split: make each syllable its own word
+        word_before = " ".join(syllables[start:end])
+        for j in range(start + 1, end):
+            new_tags[j] = "B-W"
+
+        parts = [syllables[j] for j in range(start, end)]
+        changes.append(f"split-foreign \"{word_before}\" → {' + '.join(repr(p) for p in parts)}")
+
+    return new_tags, changes
+
+
+# Vietnamese institutional compound prefixes that should NOT be split by
+# the name-boundary pass. Lowercased prefix tuples → the compound is legitimate.
+# First-syllable prefixes that start Vietnamese institutional compounds.
+# Any multi-syllable word starting with one of these + lowercase continuation
+# is likely a legitimate compound, not a proper name boundary error.
+NAME_BOUNDARY_WHITELIST_S1 = {
+    "ủy",      # Ủy ban (nhân dân / thường vụ / ...)
+    "viện",    # Viện kiểm sát / Viện nghiên cứu
+    "tổng",    # Tổng giám đốc / Tổng thư ký / ...
+    "nhà",     # Nhà khoa học / Nhà xuất bản / Nhà đầu tư
+    "phòng",   # Phòng thí nghiệm
+    "cảng",    # Cảng hàng không
+    "xuất",    # Xuất nhập khẩu
+    "sách",    # Sách (title compounds)
+    "thuế",    # Thuế thu nhập
+    "cây",     # Cây lương thực
+    "nói",     # Nói tóm lại
+    "bộ",      # Bộ luật dân sự / Bộ Tài chính
+    "đại",     # Đại hội đồng
+    "lò",      # Lò phản ứng
+    "ngay",    # Ngay lập tức
+    "việc",    # Việc làm ăn
+    "vùng",    # Vùng kinh tế
+    "sân",     # Sân vận động
+    "tiểu",    # Tiểu văn hóa
+    "trang",   # Trang thiết bị
+    "tết",     # Tết dương lịch
+    "thuyết",  # Thuyết sinh vật học
+    "điểm",    # Điểm nóng chảy
+    "lý",      # Lý thuyết
+    "hệ",      # Hệ tiên đề
+}
+
+
+def fix_proper_name_boundary(syllables, tags, vocab):
+    """Pass 2.75: Split proper name boundary merges.
+
+    Detects multi-syllable tokens where uppercase syllables are followed by
+    a lowercase common word (in vocab). Pattern: [Uppercase...][lowercase_common]
+    → split before the lowercase word.
+
+    This catches cases like "Tống_tiêu_diệt" → "Tống" + "tiêu_diệt" where
+    a proper name is merged with the following verb/noun.
+
+    Skips known Vietnamese institutional compounds (NAME_BOUNDARY_WHITELIST_PREFIXES).
+
+    Returns (new_tags, list of change descriptions).
+    """
+    new_tags = list(tags)
+    changes = []
+
+    # Reconstruct word spans
+    word_spans = []
+    current_start = 0
+    for i in range(len(tags)):
+        if tags[i] == "B-W" and i > 0:
+            word_spans.append((current_start, i))
+            current_start = i
+    word_spans.append((current_start, len(tags)))
+
+    for start, end in word_spans:
+        n_syls = end - start
+        if n_syls < 3:
+            # Need at least 3 syllables: Name + common_word(s)
+            continue
+
+        # Check whitelist: if the first syllable is a known institutional prefix, skip
+        if syllables[start].lower() in NAME_BOUNDARY_WHITELIST_S1:
+            continue
+
+        # Find the transition point: last uppercase syllable before lowercase
+        # Look for pattern: [Title/Upper...][lower...]
+        # where the lowercase portion forms a known vocab word
+        split_pos = None
+        for j in range(start + 1, end):
+            curr_syl = syllables[j]
+            prev_syl = syllables[j - 1]
+
+            # Transition: previous is title/upper, current is lowercase
+            if (prev_syl and prev_syl[0].isupper() and
+                    curr_syl and not curr_syl[0].isupper()):
+                # Check if remaining syllables form a vocab word
+                remaining = " ".join(s.lower() for s in syllables[j:end])
+                if remaining in vocab:
+                    split_pos = j
+                    break
+                # Also check if just 2 syllables from here form a vocab word
+                if end - j >= 2:
+                    two_syl = " ".join(s.lower() for s in syllables[j:j+2])
+                    if two_syl in vocab:
+                        split_pos = j
+                        break
+
+        if split_pos is None:
+            continue
+
+        word_before = " ".join(syllables[start:end])
+        new_tags[split_pos] = "B-W"
+        word_parts = bio_to_words(syllables[start:end], new_tags[start:end])
+        changes.append(
+            f"split-name-boundary \"{word_before}\" → "
+            f"{' + '.join(repr(p) for p in word_parts)}"
+        )
+
+    return new_tags, changes
+
+
 def validate_sentence(syllables, tags):
     """Pass 3: Validate BIO invariants.
 
@@ -420,23 +659,29 @@ def generate_report(all_stats, output_path=None):
     # Summary table
     lines.append("## Summary")
     lines.append("")
-    lines.append("| File | Cross-boundary splits | Long token splits | Compound merges | Validation errors |")
-    lines.append("|------|----------------------:|------------------:|----------------:|------------------:|")
+    lines.append("| File | Cross-boundary | Long token | Compound merges | Foreign splits | Name boundary | Validation errors |")
+    lines.append("|------|---------------:|-----------:|----------------:|---------------:|--------------:|------------------:|")
     total_splits = 0
     total_long_splits = 0
     total_merges = 0
+    total_foreign = 0
+    total_name_boundary = 0
     total_errors = 0
     for fname, stats in all_stats.items():
         n_splits = stats["n_cross_boundary"]
         n_long = stats["n_split_long"]
         n_merges = stats["n_merge"]
+        n_foreign = stats.get("n_foreign", 0)
+        n_name_boundary = stats.get("n_name_boundary", 0)
         n_errors = stats["n_validation_errors"]
         total_splits += n_splits
         total_long_splits += n_long
         total_merges += n_merges
+        total_foreign += n_foreign
+        total_name_boundary += n_name_boundary
         total_errors += n_errors
-        lines.append(f"| {fname} | {n_splits:,} | {n_long:,} | {n_merges:,} | {n_errors:,} |")
-    lines.append(f"| **TOTAL** | **{total_splits:,}** | **{total_long_splits:,}** | **{total_merges:,}** | **{total_errors:,}** |")
+        lines.append(f"| {fname} | {n_splits:,} | {n_long:,} | {n_merges:,} | {n_foreign:,} | {n_name_boundary:,} | {n_errors:,} |")
+    lines.append(f"| **TOTAL** | **{total_splits:,}** | **{total_long_splits:,}** | **{total_merges:,}** | **{total_foreign:,}** | **{total_name_boundary:,}** | **{total_errors:,}** |")
     lines.append("")
 
     # Merge term frequency across all files
@@ -477,6 +722,34 @@ def generate_report(all_stats, output_path=None):
                 lines.append(f"- ... and {len(stats['split_long_examples']) - 30} more")
             lines.append("")
 
+    # Foreign word split examples
+    lines.append("## Foreign Word Split Examples")
+    lines.append("")
+    for fname, stats in all_stats.items():
+        examples = stats.get("foreign_examples", [])
+        if examples:
+            lines.append(f"### {fname}")
+            lines.append("")
+            for ex in examples[:30]:
+                lines.append(f"- {ex}")
+            if len(examples) > 30:
+                lines.append(f"- ... and {len(examples) - 30} more")
+            lines.append("")
+
+    # Name boundary split examples
+    lines.append("## Name Boundary Split Examples")
+    lines.append("")
+    for fname, stats in all_stats.items():
+        examples = stats.get("name_boundary_examples", [])
+        if examples:
+            lines.append(f"### {fname}")
+            lines.append("")
+            for ex in examples[:30]:
+                lines.append(f"- {ex}")
+            if len(examples) > 30:
+                lines.append(f"- ... and {len(examples) - 30} more")
+            lines.append("")
+
     report = "\n".join(lines)
 
     if output_path:
@@ -491,13 +764,15 @@ def generate_report(all_stats, output_path=None):
 # Main
 # ============================================================================
 
-def process_file(filepath, vocab=None, sentences=None, dry_run=False):
+def process_file(filepath, vocab=None, viet_syllables=None, sentences=None, dry_run=False):
     """Process a single BIO file: apply fixes, optionally write back.
 
     Args:
         filepath: Path to BIO file.
         vocab: Set of known 2-4 syllable words for long-token splitting.
-            If None, Pass 1.5 is skipped.
+            If None, Pass 1.5 and 2.75 are skipped.
+        viet_syllables: Set of common Vietnamese syllables for foreign word
+            filtering. If None, Pass 2.5 is skipped.
         sentences: Pre-parsed sentences (avoids re-parsing if already loaded).
         dry_run: If True, report changes without modifying files.
 
@@ -516,9 +791,13 @@ def process_file(filepath, vocab=None, sentences=None, dry_run=False):
     n_cross_boundary = 0
     n_split_long = 0
     n_merge = 0
+    n_foreign = 0
+    n_name_boundary = 0
     n_validation_errors = 0
     cross_boundary_examples = []
     split_long_examples = []
+    foreign_examples = []
+    name_boundary_examples = []
     merge_term_counts = Counter()
 
     for sent in sentences:
@@ -546,6 +825,20 @@ def process_file(filepath, vocab=None, sentences=None, dry_run=False):
             term = ch.split('"')[1] if '"' in ch else ch
             merge_term_counts[term.lower()] += 1
 
+        # Pass 2.5: Split foreign word merges
+        if viet_syllables is not None:
+            tags, fw_changes = fix_foreign_words(syls, tags, viet_syllables)
+            n_foreign += len(fw_changes)
+            for ch in fw_changes:
+                foreign_examples.append(f"[{sent['sent_id']}] {ch}")
+
+        # Pass 2.75: Split proper name boundary merges
+        if vocab is not None:
+            tags, nb_changes = fix_proper_name_boundary(syls, tags, vocab)
+            n_name_boundary += len(nb_changes)
+            for ch in nb_changes:
+                name_boundary_examples.append(f"[{sent['sent_id']}] {ch}")
+
         # Pass 3: Validate
         errors = validate_sentence(syls, tags)
         n_validation_errors += len(errors)
@@ -559,10 +852,12 @@ def process_file(filepath, vocab=None, sentences=None, dry_run=False):
         sum(1 for t in s["tags"] if t == "B-W") for s in sentences
     )
 
-    print(f"  Cross-boundary splits: {n_cross_boundary:,}")
-    print(f"  Long token splits:     {n_split_long:,}")
-    print(f"  Compound merges:       {n_merge:,}")
-    print(f"  Validation errors:     {n_validation_errors:,}")
+    print(f"  Cross-boundary splits:    {n_cross_boundary:,}")
+    print(f"  Long token splits:        {n_split_long:,}")
+    print(f"  Compound merges:          {n_merge:,}")
+    print(f"  Foreign word splits:      {n_foreign:,}")
+    print(f"  Name boundary splits:     {n_name_boundary:,}")
+    print(f"  Validation errors:        {n_validation_errors:,}")
     print(f"  Words: {total_words_before:,} → {total_words_after:,} "
           f"(Δ{total_words_after - total_words_before:+,})")
     assert total_syllables_before == total_syllables_after, \
@@ -582,9 +877,13 @@ def process_file(filepath, vocab=None, sentences=None, dry_run=False):
         "n_cross_boundary": n_cross_boundary,
         "n_split_long": n_split_long,
         "n_merge": n_merge,
+        "n_foreign": n_foreign,
+        "n_name_boundary": n_name_boundary,
         "n_validation_errors": n_validation_errors,
         "cross_boundary_examples": cross_boundary_examples,
         "split_long_examples": split_long_examples,
+        "foreign_examples": foreign_examples,
+        "name_boundary_examples": name_boundary_examples,
         "merge_term_counts": merge_term_counts,
         "words_before": total_words_before,
         "words_after": total_words_after,
@@ -630,6 +929,9 @@ def main():
     vocab = build_split_vocab(all_sents)
     print(f"\nBuilt split vocab: {len(vocab):,} entries "
           f"(2-4 syllable words with count >= 5)")
+    viet_syllables = build_viet_syllables(all_sents)
+    print(f"Built Vietnamese syllable set: {len(viet_syllables):,} entries "
+          f"(syllables with count >= 50)")
 
     # Phase 3: Process each file
     all_stats = {}
@@ -638,6 +940,7 @@ def main():
         _, stats = process_file(
             path,
             vocab=vocab,
+            viet_syllables=viet_syllables,
             sentences=all_sentences_by_file[path],
             dry_run=args.dry_run,
         )
@@ -655,11 +958,15 @@ def main():
     total_splits = sum(s["n_cross_boundary"] for s in all_stats.values())
     total_long = sum(s["n_split_long"] for s in all_stats.values())
     total_merges = sum(s["n_merge"] for s in all_stats.values())
+    total_foreign = sum(s.get("n_foreign", 0) for s in all_stats.values())
+    total_name_boundary = sum(s.get("n_name_boundary", 0) for s in all_stats.values())
     total_errors = sum(s["n_validation_errors"] for s in all_stats.values())
     print(f"\n{'='*50}")
     print(f"TOTAL: {total_splits:,} cross-boundary splits, "
           f"{total_long:,} long token splits, "
           f"{total_merges:,} compound merges, "
+          f"{total_foreign:,} foreign word splits, "
+          f"{total_name_boundary:,} name boundary splits, "
           f"{total_errors:,} validation errors")
     if args.dry_run:
         print("(dry run — no files modified)")

src/ls_config_ws.xml

@@ -1,8 +1,6 @@
 <View>
   <Header value="Word Segmentation: Select each word as a span. Multi-syllable words = one span."/>
   <Header value="Sentence: $sent_id | Rank: $rank" size="4"/>
-  <Header value="Dictionary: green = in dict, red = NOT in dict" size="5"/>
-  <HyperText name="dict_view" value="$dict_html"/>
   <Labels name="label" toName="text" choice="single">
     <Label value="W" background="#4CAF50" hotkey="w"/>
     <Label value="WH" background="#2196F3" hotkey="h"/>

src/ls_import_ws.py

@@ -12,9 +12,8 @@ When --model is provided, computes per-span confidence scores from CRF
 marginal probabilities. Each word span gets score = min confidence across
 its syllables, so words with any uncertain boundary get a low score.
 
-When --dict is provided (standalone or auto-detected from model dir):
-- dict_html: inline styled spans (green = in dict, red = NOT in dict)
-- meta.dict: dict status shown in Region Details when a span is selected
+When --dict is provided (standalone or auto-detected from model dir),
+adds meta.dict to each span (shown in Region Details when selected).
 
 Usage:
     uv run src/ls_import_ws.py [--validate]
@@ -23,7 +22,6 @@ Usage:
 """
 
 import argparse
-import html
 import json
 import sys
 import unicodedata
@@ -214,60 +212,6 @@ def add_dict_meta(spans, dictionary):
                 span["meta"] = meta
 
 
-def build_dict_html(spans, text, dictionary):
-    """Build inline HTML with colored spans showing dictionary status.
-
-    Multi-syllable words are colored green (in dict) or red (not in dict).
-    Single-syllable words are shown without color.
-
-    Args:
-        spans: list of span dicts from bio_to_spans()
-        text: the sentence text
-        dictionary: set of NFC-normalized lowercase dictionary entries
-
-    Returns:
-        HTML string for the HyperText panel
-    """
-    parts = []
-    prev_end = 0
-
-    for span in spans:
-        start = span["value"]["start"]
-        end = span["value"]["end"]
-        span_text = span["value"]["text"]
-        n_syllables = span_text.count(" ") + 1
-
-        # Add any gap text (spaces between spans)
-        if start > prev_end:
-            parts.append(html.escape(text[prev_end:start]))
-
-        escaped = html.escape(span_text)
-        if n_syllables > 1:
-            normalized = nfc(span_text.lower().strip())
-            in_dict = normalized in dictionary
-            if in_dict:
-                parts.append(
-                    f'<span style="background:#c8e6c9;padding:1px 3px;'
-                    f'border-radius:3px">{escaped}</span>'
-                )
-            else:
-                parts.append(
-                    f'<span style="background:#ffcdd2;padding:1px 3px;'
-                    f'border-radius:3px;text-decoration:underline;'
-                    f'text-decoration-color:red">{escaped}</span>'
-                )
-        else:
-            parts.append(escaped)
-
-        prev_end = end
-
-    # Trailing text
-    if prev_end < len(text):
-        parts.append(html.escape(text[prev_end:]))
-
-    return '<div style="font-size:15px;line-height:1.8;padding:4px 0">' + " ".join(parts) + "</div>"
-
-
 def validate_spans(spans, text, syllables):
     """Validate that spans cover every character in text without gaps/overlaps."""
     text_len = len(text)
@@ -403,46 +347,33 @@ def main():
     tsv_rows = parse_tsv(args.tsv)
     print(f"Loaded {len(tsv_rows)} rows from {args.tsv}")
 
-    # Parse BIO files (only dev and test — train sentences not in top-500)
-    bio_data = {}
-    for split in ("dev", "test"):
+    # Parse BIO files and build text-based index
+    bio_by_text = {}
+    for split in ("dev", "test", "train"):
         bio_path = root / f"udd-ws-v1.1-{split}.txt"
         if not bio_path.exists():
-            print(f"WARNING: {bio_path} not found, skipping")
             continue
         sentences = parse_bio_file(bio_path)
-        bio_data[split] = sentences
         print(f"Loaded {len(sentences)} sentences from {bio_path}")
+        for sent in sentences:
+            key = nfc(" ".join(sent["syllables"]))
+            bio_by_text[key] = sent
+
+    print(f"Text index: {len(bio_by_text)} unique sentences")
 
     # Build tasks
     tasks = []
     errors = []
 
     for row in tsv_rows:
-        split = row["file"]
-        sent_idx = row["sent_idx"]
+        tsv_text = nfc(row["text"])
 
-        if split not in bio_data:
-            errors.append(f"Rank {row['rank']}: split '{split}' not loaded")
-            continue
-        if sent_idx >= len(bio_data[split]):
-            errors.append(
-                f"Rank {row['rank']}: sent_idx {sent_idx} >= "
-                f"{len(bio_data[split])} sentences in {split}"
-            )
+        sent = bio_by_text.get(tsv_text)
+        if sent is None:
+            errors.append(f"Rank {row['rank']}: text not found in BIO files")
             continue
 
-        sent = bio_data[split][sent_idx]
-        text = nfc(" ".join(sent["syllables"]))
-
-        # Verify text matches TSV
-        tsv_text = nfc(row["text"])
-        if text != tsv_text:
-            errors.append(
-                f"Rank {row['rank']}: text mismatch:\n"
-                f"  BIO: {text!r}\n  TSV: {tsv_text!r}"
-            )
-            continue
+        text = tsv_text
 
         # Compute per-syllable confidence if CRF model available
         confidences = None
@@ -471,20 +402,15 @@ def main():
         # Task-level prediction score = 1 - AL uncertainty score
         pred_score = round(1.0 - row["score"], 6)
 
-        # Add dictionary metadata to spans + build inline HTML
-        dict_html = ""
+        # Add dictionary metadata to spans
         if dictionary:
             add_dict_meta(spans, dictionary)
-            dict_html = build_dict_html(spans, text, dictionary)
 
         task = {
             "data": {
                 "text": text,
                 "sent_id": sent["sent_id"],
                 "rank": row["rank"],
-                "file": split,
-                "sent_idx": sent_idx,
-                "dict_html": dict_html,
             },
             "predictions": [{
                 "model_version": "silver_crf_v1.1",

src/merge_gold_silver.py

@@ -0,0 +1,140 @@
+# /// script
+# requires-python = ">=3.9"
+# dependencies = []
+# ///
+"""Merge gold annotations into fixed silver BIO files.
+
+Replaces silver sentences with gold annotations where sent_id matches.
+Produces merged BIO files for CRF training.
+
+Usage:
+    uv run src/merge_gold_silver.py
+    uv run src/merge_gold_silver.py --dry-run
+"""
+
+import argparse
+import sys
+from os.path import dirname, isfile, join
+
+
+def parse_bio_file(filepath):
+    """Parse BIO file into list of dicts with keys: sent_id, text, syllables, tags."""
+    sentences = []
+    current = {"sent_id": "", "text": "", "syllables": [], "tags": []}
+
+    with open(filepath, "r", encoding="utf-8") as f:
+        for line in f:
+            line = line.rstrip("\n")
+            if line.startswith("# sent_id = "):
+                current["sent_id"] = line.split("= ", 1)[1]
+                continue
+            if line.startswith("# text = "):
+                current["text"] = line.split("= ", 1)[1]
+                continue
+            if line.startswith("#"):
+                continue
+            if not line:
+                if current["syllables"]:
+                    sentences.append(dict(current))
+                    current = {"sent_id": "", "text": "", "syllables": [], "tags": []}
+                continue
+            parts = line.split("\t")
+            if len(parts) == 2:
+                current["syllables"].append(parts[0])
+                current["tags"].append(parts[1])
+
+    if current["syllables"]:
+        sentences.append(dict(current))
+
+    return sentences
+
+
+def write_bio_file(sentences, filepath):
+    """Write sentences back to BIO format."""
+    with open(filepath, "w", encoding="utf-8") as f:
+        for sent in sentences:
+            f.write(f"# sent_id = {sent['sent_id']}\n")
+            f.write(f"# text = {sent['text']}\n")
+            for syl, tag in zip(sent["syllables"], sent["tags"]):
+                f.write(f"{syl}\t{tag}\n")
+            f.write("\n")
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Merge gold annotations into silver BIO files."
+    )
+    parser.add_argument(
+        "--dry-run", action="store_true",
+        help="Report what would change without modifying files"
+    )
+    args = parser.parse_args()
+
+    base_dir = dirname(dirname(__file__))
+    gold_path = join(base_dir, "gold_ws_cycle1.txt")
+
+    silver_files = {
+        split: join(base_dir, f"udd-ws-v1.1-{split}.txt")
+        for split in ("train", "dev", "test")
+    }
+
+    # Check files exist
+    if not isfile(gold_path):
+        print(f"ERROR: Gold file not found: {gold_path}", file=sys.stderr)
+        sys.exit(1)
+    for split, path in silver_files.items():
+        if not isfile(path):
+            print(f"ERROR: Silver file not found: {path}", file=sys.stderr)
+            sys.exit(1)
+
+    # Parse gold
+    gold_sentences = parse_bio_file(gold_path)
+    gold_by_id = {s["sent_id"]: s for s in gold_sentences}
+    print(f"Gold sentences: {len(gold_sentences)}")
+
+    # Process each silver file
+    total_replaced = 0
+    for split, path in silver_files.items():
+        silver_sentences = parse_bio_file(path)
+        replaced = 0
+
+        for i, sent in enumerate(silver_sentences):
+            sid = sent["sent_id"]
+            if sid in gold_by_id:
+                gold = gold_by_id[sid]
+                # Verify syllable alignment
+                if sent["syllables"] != gold["syllables"]:
+                    print(f"  WARNING: syllable mismatch for {sid}, skipping")
+                    continue
+                # Replace tags with gold
+                silver_sentences[i]["tags"] = gold["tags"]
+                replaced += 1
+
+        print(f"  {split}: {replaced} sentences replaced with gold "
+              f"(out of {len(silver_sentences)} total)")
+        total_replaced += replaced
+
+        if not args.dry_run and replaced > 0:
+            write_bio_file(silver_sentences, path)
+            print(f"    Written: {path}")
+
+    print(f"\nTotal: {total_replaced} gold sentences merged into silver")
+
+    if args.dry_run:
+        print("(dry-run mode, no files modified)")
+    else:
+        # Verify symlinks in tree-1 still point to the right files
+        tree1_dir = join(base_dir, "..", "tree-1", "datasets", "udd_ws_v1_1")
+        if isfile(join(tree1_dir, "train.txt")):
+            import os
+            for split in ("train", "dev", "test"):
+                link = join(tree1_dir, f"{split}.txt")
+                if os.path.islink(link):
+                    target = os.readlink(link)
+                    print(f"  tree-1 symlink {split}.txt → {target}")
+                else:
+                    print(f"  tree-1 {split}.txt is not a symlink (direct file)")
+
+
+if __name__ == "__main__":
+    main()