Cross-Encoder for Natural Language Inference(NLI) for Japanese
[!NOTE] Considering the results of the JNLI evaluation result, we recommend using akiFQC/bert-base-japanese-v3_nli-jsnli-jnli-jsick for natural language inference in Japanese.
This model was trained using SentenceTransformers Cross-Encoder class. This model is based on tohoku-nlp/bert-base-japanese-v3.
Training Data
The model was trained on the JSNLI datasets. For a given sentence pair, it will output three scores corresponding to the labels: {0:"entailment", 1:"neutral", 2:"contradiction}.
Usage
Pre-trained models can be used like this:
from sentence_transformers import CrossEncoder
model = CrossEncoder('akiFQC/bert-base-japanese-v3_nli-jsnli')
scores = model.predict([('男はピザを食べています', '男は何かを食べています'), ('黒いレーシングカーが観衆の前から発車します。', '男は誰もいない道を運転しています。')])
#Convert scores to labels
label_mapping = ['entailment', 'neutral', 'contradiction',]
labels = [label_mapping[score_max] for score_max in scores.argmax(axis=1)]
Usage with Transformers AutoModel
You can use the model also directly with Transformers library (without SentenceTransformers library):
from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
model = AutoModelForSequenceClassification.from_pretrained('cross-encoder/nli-deberta-v3-base')
tokenizer = AutoTokenizer.from_pretrained('cross-encoder/nli-deberta-v3-base')
features = tokenizer(['男はピザを食べています', '黒いレーシングカーが観衆の前から発車します。'], ['男は何かを食べています', '男は誰もいない道を運転しています。'], padding=True, truncation=True, return_tensors="pt")
model.eval()
with torch.no_grad():
scores = model(**features).logits
label_mapping = ['contradiction', 'entailment', 'neutral']
labels = [label_mapping[score_max] for score_max in scores.argmax(dim=1)]
print(labels)
Zero-Shot Classification
This model can also be used for zero-shot-classification:
from transformers import pipeline
classifier = pipeline("zero-shot-classification", model='akiFQC/bert-base-japanese-v3_nli-jsnli')
sent = "Appleは先程、iPhoneの最新機種について発表しました。"
candidate_labels = ["技術", "スポーツ", "政治"]
res = classifier(sent, candidate_labels)
print(res)
Benchmarks
JGLUE-JNLI validation set accuracy: 0.770