This is a Japanese sentence-LUKE model.
日本語用Sentence-LUKEモデルです。
日本語Sentence-BERTモデルと同一のデータセットと設定で学習しました。
手元の非公開データセットでは、日本語Sentence-BERTモデルと比べて定量的な精度が同等〜0.5pt程度高く、定性的な精度は本モデルの方が高い結果でした。
事前学習済みモデルとしてstudio-ousia/luke-japanese-base-liteを利用させていただきました。
推論の実行にはSentencePieceが必要です（pip install sentencepiece、。

使い方

from transformers import MLukeTokenizer, LukeModel<br /> import torch<br /> class SentenceLukeJapanese:<br /> def __init__(self, model_name_or_path, device=None):<br /> self.tokenizer = MLukeTokenizer.from_pretrained(model_name_or_path)<br /> self.model = LukeModel.from_pretrained(model_name_or_path)<br /> self.model.eval()<br /> if device is None:<br /> device = "cuda" if torch.cuda.is_available() else "cpu"<br /> self.device = torch.device(device)<br /> self.model.to(device)<br /> def _mean_pooling(self, model_output, attention_mask):<br /> token_embeddings = model_output[0] #First element of model_output contains all token embeddings<br /> input_mask_expanded = attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()<br /> return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(input_mask_expanded.sum(1), min=1e-9)<br /> @torch.no_grad()<br /> def encode(self, sentences, batch_size=8):<br /> all_embeddings = []<br /> iterator = range(0, len(sentences), batch_size)<br /> for batch_idx in iterator:<br /> batch = sentences[batch_idx:batch_idx + batch_size]<br /> encoded_input = self.tokenizer.batch_encode_plus(batch, padding="longest",<br /> truncation=True, return_tensors="pt").to(self.device)<br /> model_output = self.model(**encoded_input)<br /> sentence_embeddings = self._mean_pooling(model_output, encoded_input["attention_mask"]).to('cpu')<br /> all_embeddings.extend(sentence_embeddings)<br /> return torch.stack(all_embeddings)<br /> MODEL_NAME = "sonoisa/sentence-luke-japanese-base-lite"<br /> model = SentenceLukeJapanese(MODEL_NAME)<br /> sentences = ["暴走したAI", "暴走した人工知能"]<br /> sentence_embeddings = model.encode(sentences, batch_size=8)<br /> print("Sentence embeddings:", sentence_embeddings)<br />