Current reputation systems are facing the inflation problem, which renders reputation systems to lose information and sometimes even cause misunderstandings. To address this problem, we propose a data-driven approach that combines natural language processing techniques with the conditional logit model for reputation deflation. We consider multiplicative long short-term memory neural networks (mLSTM) to predict sentiment scores from the feedback content. The mLSTM was pre-trained on 82.83 million unique reviews. We conduct experiments on one of the largest online labor marketplaces, Freelancer.com. We focus on comparing ratings and predicted sentiment scores in the online labor market. The results show that our proposed model can estimate deflated reputation information effectively. In addition, the estimated sentiment score is a quality disclosure signal, and has a better effect on the market outcome than the inflated reputation rating.

当前的信誉系统正面临通货膨胀问题，这使信誉系统失去信息，有时甚至会引起误解。为了解决这个问题，我们提出了一种数据驱动的方法，该方法将自然语言处理技术与条件Logit模型相结合，以降低声誉。我们考虑乘性长短期记忆神经网络（mLSTM），以根据反馈内容预测情绪得分。对mLSTM进行了8283万条独特评论的预培训。我们在最大的在线劳动力市场之一Freelancer.com上进行实验。我们专注于比较在线劳动力市场中的评级和预测的情绪得分。结果表明，我们提出的模型可以有效地估计紧缩的信誉信息。此外，