rustling 0.8.0

A blazingly fast library for computational linguistics
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132

# Benchmarks

This directory contains benchmarking scripts to compare Rustling (Rust + PyO3) against other Python packages with similar functionalities.

**GitHub**: https://github.com/jacksonllee/rustling/tree/main/benchmarks

## Directory Structure

```
benchmarks/
├── README.md
├── run_chat.py        # CHAT parsing benchmark (Rustling vs pylangacq)
├── run_conllu.py      # CoNLL-U parsing benchmark (Rustling vs conllu)
├── run_elan.py        # ELAN parsing benchmark (Rustling vs pympi-ling)
├── run_textgrid.py    # TextGrid parsing benchmark (Rustling vs pympi-ling)
├── run_hmm.py         # HMM benchmark (Rustling vs hmmlearn)
├── run_lm.py          # Language model benchmark (Rustling vs NLTK)
├── run_wordseg.py     # Word segmentation benchmark (Rustling vs wordseg)
├── run_perceptron_pos_tagger.py  # POS tagger benchmark (Rustling vs NLTK PerceptronTagger)
├── update_readme.py   # Update benchmark tables in README files
└── common/
    ├── __init__.py
    └── data.py        # Shared HKCanCor data loader
```

## Data Sources

Most benchmarks use the **HKCanCor** corpus (~10K Cantonese sentences with POS tags), loaded via pycantonese. The shared data loader in `common/data.py` converts the corpus into the format each benchmark needs:

- **Tagging**: tagged sentences `[(word, tag), ...]` for training, untagged word lists for testing
- **Word segmentation**: word tuples for training, concatenated strings for testing
- **HMM**: word sequences (tags stripped) for unsupervised Baum-Welch EM training and Viterbi decoding
- **Language models**: word sequences (tags stripped)

The CoNLL-U benchmark uses the **UD_English-EWT** treebank (English Universal Dependencies data), auto-downloaded to `~/.rustling/ud-english-ewt/`.

The ELAN benchmark uses the **CantoMap** corpus (Cantonese conversation data with ELAN annotations), auto-downloaded to `~/.rustling/cantomap/`.

The TextGrid benchmark uses TextGrid files generated from the CantoMap ELAN data via `rustling.elan.ELAN.to_textgrid_files()`, cached at `~/.rustling/cantomap_textgrid/`.

## Prerequisites

```bash
# Build Rustling (from repo root)
uv run maturin develop --release

# Install benchmark dependencies
uv sync --group benchmarks
```

### Comparison Libraries

| Benchmark | Comparison Library |
|-----------|--------------------|
| CHAT Parsing | [pylangacq](https://pylangacq.org/) |
| CoNLL-U Parsing | [conllu](https://github.com/EmilStenstrom/conllu/) |
| ELAN Parsing | [pympi-ling](https://pypi.org/project/pympi-ling/) Eaf |
| TextGrid Parsing | [pympi-ling](https://pypi.org/project/pympi-ling/) TextGrid |
| HMM | [hmmlearn](https://hmmlearn.readthedocs.io/) CategoricalHMM |
| Word Segmentation | [wordseg](https://pypi.org/project/wordseg/) |
| POS Tagging | [NLTK](https://www.nltk.org/) PerceptronTagger |
| Language Models | [NLTK](https://www.nltk.org/) nltk.lm |

All benchmarks degrade gracefully if a comparison library is not installed.

## Results

Benchmarked against Python implementations from NLTK, wordseg (v0.0.5),
pylangacq (v0.19.1), hmmlearn (v0.3.3), pympi-ling (v1.70.2), and conllu (v6.0.0).

| Component | Task | Speedup | vs. |
|---|---|---|---|
| **Language Models** | Fit | **11x** | NLTK |
|  | Score | **2x** | NLTK |
|  | Generate | **86--107x** | NLTK |
| **Word Segmentation** | LongestStringMatching | **9x** | wordseg |
| **POS Tagging** | Training | **5x** | NLTK |
|  | Tagging | **17x** | NLTK |
| **HMM** | Fit | **14x** | hmmlearn |
|  | Predict | **0.9x** | hmmlearn |
|  | Score | **5x** | hmmlearn |
| **CHAT Parsing** | Reading from a ZIP archive | **30x** | pylangacq |
|  | Reading from strings | **35x** | pylangacq |
|  | Parsing utterances | **15x** | pylangacq |
|  | Parsing tokens | **8x** | pylangacq |
| **ELAN Parsing** | Parse single file | **4x** | pympi-ling |
|  | Parse all files | **17x** | pympi-ling |
| **TextGrid Parsing** | Parse single file | **3x** | pympi-ling |
|  | Parse all files | **8x** | pympi-ling |
| **CoNLL-U Parsing** | Parse from strings | **15x** | conllu |
|  | Parse from files | **15x** | conllu |

---

## Running Benchmarks

Each script supports `--quick` (fewer iterations), `--export FILE` (JSON output), and `--quiet`:

```bash
python benchmarks/run_chat.py
python benchmarks/run_conllu.py
python benchmarks/run_elan.py
python benchmarks/run_textgrid.py
python benchmarks/run_hmm.py
python benchmarks/run_wordseg.py
python benchmarks/run_perceptron_pos_tagger.py
python benchmarks/run_lm.py
```

## Updating Benchmark Tables

After running benchmarks with `--export`, update the performance table in `benchmarks/README.md`:

```bash
python benchmarks/run_chat.py --export benchmarks/.results/chat.json
python benchmarks/run_conllu.py --export benchmarks/.results/conllu.json
python benchmarks/run_elan.py --export benchmarks/.results/elan.json
python benchmarks/run_textgrid.py --export benchmarks/.results/textgrid.json
python benchmarks/run_hmm.py --export benchmarks/.results/hmm.json
python benchmarks/run_wordseg.py --export benchmarks/.results/wordseg.json
python benchmarks/run_perceptron_pos_tagger.py --export benchmarks/.results/tagger.json
python benchmarks/run_lm.py --export benchmarks/.results/lm.json

python benchmarks/update_readme.py --from-json benchmarks/.results/
```

## Tips

- Use `--release` when building Rustling for accurate benchmarks: `maturin develop --release`
- Close other applications to reduce noise
- Run multiple times to verify consistency
- Use `--quiet` with `--export` for machine-readable output only