Few-shot Learning with Multilingual Language Models
Introduction
In this work, we train a family of multilingual generative language models, dubbed XGLM, on a balanced corpus covering a diverse set of languages, and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters sets new state of the art in few-shot learning on more than 20 representative languages, outperforming GPT-3 of comparable size in multilingual commonsense reasoning (+7.4 accuracy points for 0-shot, +9.4 for 4-shot) and natural language inference (+5.4 for 0-shot, +5.4 for 4-shot). We have included a model card of XGLM for transparency and accountability.
Data and Languages
XGLM models are trained on a new multilingual corpus extracted from CommonCrawl (CC100-XL), a significantly larger multilingual dataset covering 68 Common Crawl (CC) snapshots (from Summer 2013 to March/April 2020 consisting of 134 languages. The detailed languages and data statistics are reported in the paper (Table A.1).
Pre-trained models
| Model | Layers | Model Dim | FFN Dim | Languages | Download |
| ----------- | ------ | --------- | ------- | ------------------------ | ------------------------------------------------------------ |
| XGLM 564M | 24 | 1024 | 4096 | trained on 30 languages | xglm.564M.tar.gz |
| XGLM 1.7B | 24 | 2048 | 8192 | trained on 30 languages | xglm.1.7B.tar.gz |
| XGLM 2.9B | 48 | 2048 | 8192 | trained on 30 languages | xglm.2.9B.tar.gz |
| XGLM 7.5B | 32 | 4096 | 16384 | trained on 30 languages | xglm.7.5B.tar.gz |
| XGLM 4.5B | 48 | 2048 | 16384 | trained on 134 languages | xglm.4.5B.tar.gz |
Pre-training Data Format
Our models were pre-trained with data in the following format (i.e. paragraphs are separated with new lines and documents were separated with double new lines).
``` ... # X0: number of tokens in para0 of doc0 ... # Y0: number of tokens in para1 of doc0
... # X1: number of tokens in para0 of doc1 ... # Y1: number of tokens in para1 of doc1
... ```
Fairseq's preprocessing replaces newlines with the end-of-sentence symbol (`). As a result, the models never saw newline characters during pretraining and the same preprocessing should be run prior to few-shot inference to maximize performance. For example, our language model scoring function hasreplace_newlines_with_eos` argument to trigger this preprocessing:
``` from fairseq.models.transformer_lm import TransformerLanguageModel
model_dir = 'path_to_decompressed_tar_gz_dir' lm = TransformerLanguageModel.from_pretrained(model_dir, bpe='sentencepiece')
text = """First paragraph of the first document. Second paragraph of the first document.
First paragraph of the second document. """ tokens = lm.score(text, replace_newlines_with_eos=True)['tokens'] assert '\n' not in lm.decode(tokens) # no newlines were encoded ```
Evaluation
Example (COPA)
The following snippet show how to evaluate our models on the Choice of Plausible Alternatives (COPA) task, using examples in English, Chinese and Hindi.
data_samples = {
'en': [
{
"premise": "I wanted to conserve energy.",
"choice1": "I swept the floor in the unoccupied room.",
"choice2": "I shut off the light in the unoccupied room.",
"question": "effect",
"label": "1"
},
{
"premise": "The flame on the candle went out.",
"choice1": "I blew on the wick.",
"choice2": "I put a match to the wick.",
"question": "cause",
"label": "0"
}
],
'zh': [
{
"premise": "我想节约能源。",
"choice1": "我在空着的房间里扫了地板。",
"choice2": "我把空房间里的灯关了。",
"question": "effect",
"label": "1"
},
{
"premise": "蜡烛上的火焰熄灭了。",
"choice1": "我吹灭了灯芯。",
"choice2": "我把一根火柴放在灯芯上。",
"question": "cause",
"label": "0"
}
],
'hi': [
{
"premise": "M te vle konsève enèji.",
"choice1": "Mwen te fin baleye chanm lib la.",
"choice2": "Mwen te femen limyè nan chanm lib la.",
"question": "effect",
"label": "1"
},
{
"premise": "Flam bouji a te etenn.",
"choice1": "Mwen te soufle bouji a.",
"choice2": "Mwen te limen mèch bouji a.",
"question": "cause",
"label": "0"
}
]
}
In this example, we format the examples use the non-verbal prompts {premise}\n{choice1} and {premise}\n{choice2}, which are shared by all three languages.
``` from fairseq.models.transformer_lm import TransformerLanguageModel
model_dir = 'path_to_decompressed_tar_gz_dir' lm = TransformerLanguageModel.from_pretrained(model_dir, bpe='sentencepiece') lm = lm.eval() lm = lm.half() lm = lm.cuda()
def get_logprobs(prompt): import re prompt = re.sub('\n+' , '\n', prompt) # collapse repeated newlines, which indicate separate documents return lm.score(prompt, replace_newlines_with_eos=True)['positional_scores']
Zero-shot evaluation for the Choice of Plausible Alternatives (COPA) task.
A return value of 0 indicates that the first alternative is more plausible,
while 1 indicates that the second alternative is more plausible.
def COPA_eval(prompt, alternative1, alternative2): lprob1 = get_logprobs(prompt + "\n" + alternative1).sum() lprob2 = get_logprobs(prompt + "\n" + alternative2).sum() return 0 if lprob1 > lprob2 else 1
for lang in ['en', 'zh', 'hi']: for idx, example in enumerate(data_samples[lang]): predict = COPA_eval(example["premise"], example["choice1"], example["choice2"]) print(f'{lang}-{idx}', predict, example['label'])
en-0 1 1
en-1 0 0
zh-0 1 1
zh-1 0 0
hi-0 1 1
hi-1 0 0
```
XStoryCloze
We release XStoryCloze, a new multilingual dataset intended for few-shot evaluation, alongside this paper. XStoryCloze consists of professional translation of the validation split of the English StoryCloze dataset (Spring 2016 version) to 10 other languages. It is opensourced under CC BY-SA 4.0, the same license as the English StoryCloze.
You can download the dataset via this link.
| Language | ar | es | eu | hi | id | my | ru | sw | te | zh | | ---------- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | ---- | | Train size | 360 | 360 | 360 | 360 | 360 | 360 | 360 | 360 | 360 | 360 | | Eval size | 1511 | 1511 | 1511 | 1511 | 1511 | 1511 | 1511 | 1511 | 1511 | 1511 |
Please refer to the dataset doc for more information.
Visit Official Website
https://github.com/facebookresearch/fairseq/tree/main/examples/xglm
