Abstract: In order to solve the problem that new slot values are difficult to be extracted from user input in task-oriented dialogue, a feature fusion model based on bidirectional long and short-term memory network is proposed. The negative sample addition technique was used to alleviate the problem of new slot value recognition. The attention mechanism was introduced to extract the weight features between words in user input. The residuals were used to link the multi-layer stacked LSTM networks to alleviate the loss of information in the process of feature extraction. The fused features were input into the conditional random field to screen out the reasonable labeling sequence. In this paper, we tested the model on DSTC3 dataset. The experimental results show that the recognition rates of this model on new slot values, known slot values and total data sets are 70.73%, 91.56% and 87.11%, respectively. The performance of this model is significantly better than that of baseline method.