Automatic Generation of the Draft Procuratorial Suggestions Based on an Extractive Summarization Method: BERTSLCA

Table 4

ROGUE-L comparisons between different methods.

Method\value	P	R	F1
LEAD	18.3	13.3	15.4
NEURALSUM	19.1	14.6	16.6
TextRank	18.2	13.9	15.8
BERTSUM + FC	18.7	26.7	22.0
BERTSUM + Transformer	21.1	27.3	23.8
BERTSUM + LSTM	19.4	27.1	22.6
BERTSUM + BiLSTM + attention	19.5	28.0	23.0
BERTSUM + CNN	19.9	27.8	23.2
BERTSUM + BiLSTM + CNN	20.8	27.3	23.7
Proposed method	22.6	31.1	26.2

Note. The value of the proposed method is the average value after 10-fold cross-validation test layer (8.47), BERTSUM with CNN (0.7), and BERTSUM with Transformer (1.47). And at the final step, our method also achieves the lowest loss, 0.46. All this points to one fact, the BERTSUM-based methods are pretrained with large corpora providing weight that fits the downstream task. As a result, the models exhibit better document representation and the losses of training are small at the beginning. Our method leverages BiLSTM and CNN with attention mechanism exhibits higher feature extraction ability which benefits the document summarization task.