Modelpublic

susnato/ernie-m-large_pytorch

sourceHugging Faceapache-2.0updated 4y agoView on Hugging Face

5likes30downloads

Model Card

Disclaimer :- I don't own the weights of `ernie-m-large` neither did I train the model. I only converted the model weights from paddle to pytorch(using the scripts listed in files).

The real(paddle) weights can be found here.

The rest of the README is copied from the same page listed above,

![paddlenlp-banner](https://github.com/PaddlePaddle/PaddleNLP)

PaddlePaddle/ernie-m-base

Ernie-M

ERNIE-M, proposed by Baidu, is a new training method that encourages the model to align the representation of multiple languages with monolingual corpora, to overcome the constraint that the parallel corpus size places on the model performance. The insight is to integrate back-translation into the pre-training process by generating pseudo-parallel sentence pairs on a monolingual corpus to enable the learning of semantic alignments between different languages, thereby enhancing the semantic modeling of cross-lingual models. Experimental results show that ERNIE-M outperforms existing cross-lingual models and delivers new state-of-the-art results in various cross-lingual downstream tasks.

We proposed two novel methods to align the representation of multiple languages:

Cross-Attention Masked Language Modeling(CAMLM): In CAMLM, we learn the multilingual semantic representation by restoring the MASK tokens in the input sentences. Back-Translation masked language modeling(BTMLM): We use BTMLM to train our model to generate pseudo-parallel sentences from the monolingual sentences. The generated pairs are then used as the input of the model to further align the cross-lingual semantics, thus enhancing the multilingual representation.

[image]

Benchmark

XNLI

XNLI is a subset of MNLI and has been translated into 14 different kinds of languages including some low-resource languages. The goal of the task is to predict testual entailment (whether sentence A implies / contradicts / neither sentence B).

Model	en	fr	es	de	el	bg	ru	tr	ar	vi	th	zh	hi	sw	ur	Avg
Cross-lingual Transfer
XLM	85.0	78.7	78.9	77.8	76.6	77.4	75.3	72.5	73.1	76.1	73.2	76.5	69.6	68.4	67.3	75.1
Unicoder	85.1	79.0	79.4	77.8	77.2	77.2	76.3	72.8	73.5	76.4	73.6	76.2	69.4	69.7	66.7	75.4
XLM-R	85.8	79.7	80.7	78.7	77.5	79.6	78.1	74.2	73.8	76.5	74.6	76.7	72.4	66.5	68.3	76.2
INFOXLM	86.4	80.6	80.8	78.9	77.8	78.9	77.6	75.6	74.0	77.0	73.7	76.7	72.0	66.4	67.1	76.2
ERNIE-M	85.5	80.1	81.2	79.2	79.1	80.4	78.1	76.8	76.3	78.3	75.8	77.4	72.9	69.5	68.8	77.3
XLM-R Large	89.1	84.1	85.1	83.9	82.9	84.0	81.2	79.6	79.8	80.8	78.1	80.2	76.9	73.9	73.8	80.9
INFOXLM Large	89.7	84.5	85.5	84.1	83.4	84.2	81.3	80.9	80.4	80.8	78.9	80.9	77.9	74.8	73.7	81.4
VECO Large	88.2	79.2	83.1	82.9	81.2	84.2	82.8	76.2	80.3	74.3	77.0	78.4	71.3	80.4	79.1	79.9
ERNIR-M Large	89.3	85.1	85.7	84.4	83.7	84.5	82.0	81.2	81.2	81.9	79.2	81.0	78.6	76.2	75.4	82.0
Translate-Train-All
XLM	85.0	80.8	81.3	80.3	79.1	80.9	78.3	75.6	77.6	78.5	76.0	79.5	72.9	72.8	68.5	77.8
Unicoder	85.6	81.1	82.3	80.9	79.5	81.4	79.7	76.8	78.2	77.9	77.1	80.5	73.4	73.8	69.6	78.5
XLM-R	85.4	81.4	82.2	80.3	80.4	81.3	79.7	78.6	77.3	79.7	77.9	80.2	76.1	73.1	73.0	79.1
INFOXLM	86.1	82.0	82.8	81.8	80.9	82.0	80.2	79.0	78.8	80.5	78.3	80.5	77.4	73.0	71.6	79.7
ERNIE-M	86.2	82.5	83.8	82.6	82.4	83.4	80.2	80.6	80.5	81.1	79.2	80.5	77.7	75.0	73.3	80.6
XLM-R Large	89.1	85.1	86.6	85.7	85.3	85.9	83.5	83.2	83.1	83.7	81.5	83.7	81.6	78.0	78.1	83.6
VECO Large	88.9	82.4	86.0	84.7	85.3	86.2	85.8	80.1	83.0	77.2	80.9	82.8	75.3	83.1	83.0	83.0
ERNIE-M Large	89.5	86.5	86.9	86.1	86.0	86.8	84.1	83.8	84.1	84.5	82.1	83.5	81.1	79.4	77.9	84.2

Cross-lingual Named Entity Recognition

—datasets：CoNLI

Model	en	nl	es	de	Avg
Fine-tune on English dataset
mBERT	91.97	77.57	74.96	69.56	78.52
XLM-R	92.25	78.08	76.53	69.60	79.11
ERNIE-M	92.78	78.01	79.37	68.08	79.56
XLM-R LARGE	92.92	80.80	78.64	71.40	80.94
ERNIE-M LARGE	93.28	81.45	78.83	72.99	81.64
Fine-tune on all dataset
XLM-R	91.08	89.09	87.28	83.17	87.66
ERNIE-M	93.04	91.73	88.33	84.20	89.32
XLM-R LARGE	92.00	91.60	89.52	84.60	89.43
ERNIE-M LARGE	94.01	93.81	89.23	86.20	90.81

Cross-lingual Question Answering

—datasets：MLQA

Model	en	es	de	ar	hi	vi	zh	Avg
mBERT	77.7 / 65.2	64.3 / 46.6	57.9 / 44.3	45.7 / 29.8	43.8 / 29.7	57.1 / 38.6	57.5 / 37.3	57.7 / 41.6
XLM	74.9 / 62.4	68.0 / 49.8	62.2 / 47.6	54.8 / 36.3	48.8 / 27.3	61.4 / 41.8	61.1 / 39.6	61.6 / 43.5
XLM-R	77.1 / 64.6	67.4 / 49.6	60.9 / 46.7	54.9 / 36.6	59.4 / 42.9	64.5 / 44.7	61.8 / 39.3	63.7 / 46.3
INFOXLM	81.3 / 68.2	69.9 / 51.9	64.2 / 49.6	60.1 / 40.9	65.0 / 47.5	70.0 / 48.6	64.7 / 41.2	67.9 / 49.7
ERNIE-M	81.6 / 68.5	70.9 / 52.6	65.8 / 50.7	61.8 / 41.9	65.4 / 47.5	70.0 / 49.2	65.6 / 41.0	68.7 / 50.2
XLM-R LARGE	80.6 / 67.8	74.1 / 56.0	68.5 / 53.6	63.1 / 43.5	62.9 / 51.6	71.3 / 50.9	68.0 / 45.4	70.7 / 52.7
INFOXLM LARGE	84.5 / 71.6	75.1 / 57.3	71.2 / 56.2	67.6 / 47.6	72.5 / 54.2	75.2 / 54.1	69.2 / 45.4	73.6 / 55.2
ERNIE-M LARGE	84.4 / 71.5	74.8 / 56.6	70.8 / 55.9	67.4 / 47.2	72.6 / 54.7	75.0 / 53.7	71.1 / 47.5	73.7 / 55.3

Cross-lingual Paraphrase Identification

—datasets：PAWS-X

Model	en	de	es	fr	ja	ko	zh	Avg
Cross-lingual Transfer
mBERT	94.0	85.7	87.4	87.0	73.0	69.6	77.0	81.9
XLM	94.0	85.9	88.3	87.4	69.3	64.8	76.5	80.9
MMTE	93.1	85.1	87.2	86.9	72.0	69.2	75.9	81.3
XLM-R LARGE	94.7	89.7	90.1	90.4	78.7	79.0	82.3	86.4
VECO LARGE	96.2	91.3	91.4	92.0	81.8	82.9	85.1	88.7
ERNIE-M LARGE	96.0	91.9	91.4	92.2	83.9	84.5	86.9	89.5
Translate-Train-All
VECO LARGE	96.4	93.0	93.0	93.5	87.2	86.8	87.9	91.1
ERNIE-M LARGE	96.5	93.5	93.3	93.8	87.9	88.4	89.2	91.8

Cross-lingual Sentence Retrieval

—dataset：Tatoeba

Model	Avg
XLM-R LARGE	75.2
VECO LARGE	86.9
ERNIE-M LARGE	87.9
ERNIE-M LARGE（ after fine-tuning）	93.3

Citation Info

text

@article{Ouyang2021ERNIEMEM,
  title={ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora},
  author={Xuan Ouyang and Shuohuan Wang and Chao Pang and Yu Sun and Hao Tian and Hua Wu and Haifeng Wang},
  journal={ArXiv},
  year={2021},
  volume={abs/2012.15674}
}

Disclaimer :- I don't own the weights of ernie-m-large neither did I train the model. I only converted the model weights from paddle to pytorch(using the scripts listed in files).