Chinese Constituency Parsing
January 11, 2021 · View on GitHub
Background
Given a sentence, consituency parsing produces a parse tree whose internal nodes are constituents and whose leaf nodes are words.
Example
Input:
柴犬是一种像精灵一样的犬种。
Output:
(IP (NP-SBJ (NN 柴犬)) (VP (VC 是) (NP-PRD (QP (CD 一) (CLP (M 种))) (DVP (IP (VP (PP (P 像) (NP (NN 精灵))) (VP (VA 一样)))) (DEV 的)) (VP (VA 犬种)))) (PU 。))
Standard Metrics
- Exact match (EM): the percentage of predicted parse trees that match the ground truth exactly.
- F1 score of constituents in the predicted parse tree.
- Labeled precision (LP): precision of constituents in the predicted parse tree.
- Labeled precall (LR): recall of constituents in the predicted parse tree.
Chinese Tree Bank Datasets.
- Released by LDC. Requires LDC licence to acquire the datasets
- Link: https://verbs.colorado.edu/chinese/ctb.html
- First used for consituency parsing by Liu and Zhang (2017).
- Prior works have adopted the preprocessing in distance-parser, which selects a subset of CTB 8.0 that corresponds to CTB 5.1.
| Dataset | # sentences (train) | # sentences (dev) | # sentences (test) |
|---|---|---|---|
| CTB 5.1 | 17,544 | 352 | 348 |
Metrics
EM, F1, LP and LR can be calculated using the Evalb tool.
Results
| System | EM | F1 | LP | LR | code |
|---|---|---|---|---|---|
| Liu and Zhang (2017) | 44.94 | 91.81 | - | - | Github |
| Zhou and Zhao (2019) | - | 92.18 | 92.33 | 92.03 | Github |
| Mrini et al. (2020) | - | 92.64 | 93.45 | 91.85 | Github |
| Yang and Deng (2020) | 49.72 | 93.59 | 93.80 | 93.40 | Github |
Suggestions? Changes? Please send email to chinesenlp.xyz@gmail.com