Prolonged Grief Disorder (PGD) — characterised by persistent yearning, identity disruption, and functional impairment following bereavement — was formally recognised in ICD-11 and DSM-5-TR in 2022 and is estimated to affect between 10-13% of bereaved individuals globally, representing millions of people whose grief has become a clinical disorder requiring intervention. Despite its clinical significance PGD remains severely underdiagnosed. Natural language processing (NLP) tools have shown promise for early detection of depression and suicidality in social media text, raising the question of whether similar approaches could support PGD identification. However a fundamental prerequisite — whether existing NLP tools can even distinguish grief language from depression language — has never been examined. This study investigates whether existing NLP emotion datasets adequately represent grief language, whether current sentiment and semantic NLP tools can distinguish grief from depression in social media text, and whether a grief-specific psycholinguistic lexicon grounded in ICD-11 PGD clinical criteria can provide a foundation for future bereavement-sensitive NLP applications. Grief comprised only 0.18% of the GoEmotions dataset — 16.9 times less represented than sadness — revealing a structural data barrier to bereavement NLP model development. VADER sentiment analysis could not distinguish grief from depression (p=0.640). Three-group validation revealed that LIWC2007 identifies grief by exclusion — detecting absent depression markers — scoring similarly in grief and non-depressed text on 13 of 20 categories. GriefLex v1.0 identifies grief by inclusion — detecting present bereavement markers — correctly discriminating grief from non-depressed text on 13 of 14 categories. GriefLex v1.0 is proposed as the first domain-specific psycholinguistic lexicon for bereavement applications, grounded in ICD-11 PGD clinical criteria and introducing five previously unoperationalised categories with no equivalent in existing psycholinguistic frameworks.
Misri et al. (Tue,) studied this question.