A membership inference attack (MIA) finds out if the given data is used for model training. Currently, MIA is still struggling with unstable attack effectiveness in text classification tasks. This paper proposes three novel MIA methods: multi-model coordination attack (MCA), adaptive threshold attack (ATA), and strong hypothesis attack (SHA). MCA improves the transferability between the target and shadow model with a new data partitioning and weighting scheme. ATA optimizes the bias to determine the threshold between members and nonmembers more accurately. SHA is the first to expose the risk of privacy leakage in text-specific pre-processing. Experiments results on different datasets and target models show that our proposed methods achieve higher attack successful rate than those stare-of-the-art methods and have better stability.