Distributed arithmetic coding (DAC) is similar to syndrome coding, in the sense that message sequences sharing the same interval can be considered a coset of the space of the source sequences, and the codeword is the index of the coset. In this paper, the minimum Hamming distance of cosets is studied and it is proved that such a distance is as small as one. By only allowing the sequences with a large Hamming distance to overlap in the same interval, an improved DAC scheme is proposed. Simulation results show that, for equiprobable memoryless sources, this approach outperforms DAC in terms of decoding error rate at the same coding cost. In addition, at small sequence length, the decoding error rate of the proposed scheme is lower than that of distributed source coding based on low-density parity-check codes for highly correlated sources.