BERT-KPE

• BERT-KPE - Joint Keyphrase Chunking and Salience Ranking with BERT
  • Methodology

BERT-KPE - Joint Keyphrase Chunking and Salience Ranking with BERT

Methodology

$$文档： D = \{w_1, \cdots, w_i, \cdots, w_n\} \\ BERT表示文档： H = BERT(D) = \{h_1, \cdots, h_i, \cdots, h_n \} \\ ngram \,的 \, phrase : c_i^k = w_{i:i+k-1} \\ CNN \,表示 \, ngram: g_i^k = CNN^k \{h_i, \cdots, h_{i+k-1} \}$$

• a chunking network： 识别出有意义的n-grams， 直接采用全连接接网络+softmax判断该n-gram $c_i^k$ 是否为合适的chunk

$$p(c_i^k=y_i^k) = softmax(Linear(g_i^k))$$

• a ranking network：为 phrase 评分 $$f(c_i^k, D) = Linear(g_i^k)$$