README.md

---
license: mit
datasets:
- lc_quad
---

This repo contains a custom tokenizer for SPARQL. It is a SentencePieceBPE tokenizer trained on lc_quad. Here is an example.

Original query:
```
SELECT ?answer WHERE { wd:Q825946 wdt:P371 ?X . ?X wdt:P2048 ?answer}
```

Result from default T5 tokenizer (just as an example):
```
['▁', 'SEL', 'ECT', '▁', '?', 'ans', 'wer', '▁W', 'HER', 'E', '▁', '{', '▁', 'w', 'd', ':', 'Q', '82', '59', '46', '▁',
  'w', 'd', 't', ':', 'P', '37', '1', '▁', '?', 'X', '▁', '.', '▁', '?', 'X', '▁', 'w', 'd', 't', ':', 'P', '20', '48',
  '▁', '?', 'ans', 'wer', '}']
```

Result from this tokenizer:
```
['▁SELECT', '▁?answer', '▁WHERE', '▁{', '▁wd:Q8', '259', '46', '▁wdt:P371', '▁?X', '▁.', '▁?X', '▁wdt:P2048', '▁?answer', '}']
```

# How to use

```python
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("InfAI/sparql-tokenizer")
tokenizer.tokenize("SELECT ?answer WHERE { wd:Q825946 wdt:P371 ?X . ?X wdt:P2048 ?answer}")
```

```
['▁SELECT', '▁?answer', '▁WHERE', '▁{', '▁wd:Q8', '259', '46', '▁wdt:P371', '▁?X', '▁.', '▁?X', '▁wdt:P2048', '▁?answer', '}']
```

```python
tokenizer("SELECT ?answer WHERE { wd:Q825946 wdt:P371 ?X . ?X wdt:P2048 ?answer}")
```

```
{'input_ids': [441, 444, 431, 422, 606, 1388, 720, 1791, 456, 418, 456, 3657, 444, 185], 'token_type_ids': [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], 'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]}
```