As social media becomes a major part of everyday life, analyzing language on social media is becoming a potentially fruitful approach to discover depression patients. At a time when depression is starting to be taken seriously, sentiment analysis through social media should be studied further so that people can be treated in the early stages of depression. With the intention to detect depression behavior through studying language on social media, we built a classification model for depression detection by training on a Tweets dataset from Shen et al 2017. After optimizing hyper-parameters including learning rate and embedding dimension, the language classification model achieves the test accuracy of 98.94% and an F1 score of 99.04% which is higher than the best performance of 85% F1-measure achieved by Shen’s work. These results show that the method is effective and can be used in a wider range of unlabeled data to locate potential depressed users.

随着社交媒体成为日常生活的重要组成部分，分析社交媒体上的语言已成为发现抑郁症患者的一种富有成果的方法。在开始严重关注抑郁症的时候，应进一步研究通过社交媒体进行的情绪分析，以便可以在抑郁症的早期阶段对患者进行治疗。为了通过在社交媒体上学习语言来发现抑郁行为，我们通过对Shen等2017年的Tweets数据集进行训练，建立了抑郁检测的分类模型。优化了学习率和嵌入维度等超参数后，语言分类模型达到了98.94％的测试准确度和99.04％的F1分数，这比Shen的工作所达到的85％的F1测量的最佳性能要高。