Temporal Language Modelling For Early Depression Risk Detection

Abstract

This requires the development of models that go beyond snapshot-based classifications to account for the temporal evolution of language use. In this paper, we propose a temporal trajectory learning pipeline using contextual language embeddings to represent the sequential evolution of user posts, which uses a GRU-based recurrent temporal layer to learn the evolution of depression risk over time. We use a patience-based decision policy to make early risk decisions, which are typically based on separate thresholds. We use various metrics, including time-to-detection statistics, the proposed ERDE metric in compliance with the benchmark, and the usual classification metrics. We test the proposed temporal trajectory learning pipeline using the official CLEF 2017 eRisk benchmark as the primary evaluation corpus, as well as the Dreaddit corpus as a secondary generalization test. We perform an ablation analysis using three different baseline models: TF-IDF + Logistic Regression, LSTM-only, and BERT-Final-only. We find that the proposed model performs best on the eRisk 2017 benchmark, obtaining AUC= 0.8852, ERDE_50 = 0.0775 in compliance with the benchmark and outperforms the LSTM-only variant by 0.37 AUC and the BERT-Final-only variant by 0.14 AUC.