Machine Learning Approach for Suicide and Depression Identification with Corrected Unsupervised Labels

EasyChair Preprint 15684

Abstract

It can be life-saving to identify suicidal thoughts in depressed people early on so that they can receive the necessary medical care and support. Novel NLP research endeavours to categorise, based on provided text, whether an individual is clinically healthy or suicidal. But there haven’t been any significant efforts to distinguish between suicidal ideation and depression, which is a different and significant clinical challenge. Web query data has become a promising alternative because EHR data suicide notes, and other verified sources are hard to come by. Internet sources like Reddit are credible even in a clinical setting because they provide an anonymous environment that encourages candid symptom disclosure. Nevertheless, web-scraped labels resulting from online datasets also contain inherent noise, which makes a noise-removal procedure necessary to enhance performance. Consequently, we propose SDCNL, a deep neural network method for classifying suicide versus depression. Our algorithm is trained using online content, and we suggest a novel unsupervised label correction technique that does not require prior knowledge of the noise distribution to verify and correct noisy labels, in contrast to previous work. Our thorough testing of numerous deep word embedding models and classifiers demonstrates the potent performance of SDCNL as a novel and clinical solution for a difficult issue. Another use cases that we have considered in this research are : financial news dataset, Vietnam news agency, credit default swap. We have derived to final results for all the four use cases as mentioned above, majorly developing from the use case of suicide/depression dataset by building and testing the framework and then testing them on other three datasets.

Keyphrases: Credit default, Financial News, Natural Language Processing, Suicide/Depression, deep learning, news agency, noisy labels, online content, unsupervised learning

@booklet{EasyChair:15684,
  author    = {Meenal Badki},
  title     = {Machine Learning Approach for Suicide and Depression Identification with Corrected Unsupervised Labels},
  howpublished = {EasyChair Preprint 15684},
  year      = {EasyChair, 2025}}