Contract net protocol (CNP) and its extensions (CNCP, ICNP) have been widely applied to task allocation in muti-agent system (MAS). However, when being applied to large-scale MAS (LSMAS), they expose shortcomings such as low efficiency and large information amounts. To solve this problem, an improvement to CNCP was proposed in this paper, In which a method for deciding the bidding threshold of participant was designed, so does a general definition formula of participant's DoA (degree of availability) and its weight coefficient, then a new bids evaluation function was given. At last, three groups of experiments were conducted to simulate the process of task allocation in Large-Scale MAS, each of them run both CNCP and the improved CNCP, results show that the improvement proposed in this paper can enhance the performance of CNCP in LSMAS effectively.