Am I No Good? Towards Detecting Perceived Burdensomeness and Thwarted Belongingness from Suicide Notes

Soumitra Ghosh, Asif Ekbal, Pushpak Bhattacharyya

TL;DR

This paper introduces a novel multitask framework (TEMF) for automatically detecting two critical interpersonal risk factors for suicide – Perceived Burdensomeness (PB) and Thwarted Belongingness (TB) – directly from suicide notes. They also release CoMCEASE-v2.0, a new code-mixed (Hindi-English) suicide note corpus, to address the scarcity of multilingual resources in this domain.

What Problem Does This Paper Address?

Suicide remains a significant global health crisis, and understanding the underlying psychological factors is crucial for prevention. The interpersonal theory of suicide posits that suicidal desire arises from feelings of Perceived Burdensomeness (PB) – the belief that one is a burden to others – and Thwarted Belongingness (TB) – the feeling of isolation and lack of connection. While suicide notes offer a unique window into an individual's state of mind, analyzing them for these complex psychological constructs is challenging. Existing research often focuses on English-only datasets and lacks methods to handle code-mixed languages prevalent in multilingual societies. Furthermore, the scarcity of annotated suicide note datasets hinders the development of robust automated detection systems.

Methods

The authors propose a Temporal and Emotion-assisted Multitask Framework (TEMF). This framework leverages a combination of pre-trained models and novel architectural components:

  • Input Embeddings: They utilize both GloVe (English) and fastText (Hindi) word embeddings to capture the semantics of code-mixed text. Emotion and temporality class labels are also embedded.
  • Document Encoder: A BERT (base) model is used to generate a high-level contextual representation of the entire suicide note.
  • Sentence Encoder: A Transformer Encoder (TE) is employed to capture local semantic information and long-term dependencies within individual sentences. Positional encodings are added to account for word order.
  • Attention Module: This module fuses the document-level context from BERT with sentence-level representations from the TE using additive attention. Crucially, it also incorporates separate attention mechanisms for temporal and emotional information, allowing the model to focus on relevant words related to these aspects.
  • High-level Document Abstraction: The sentence-level representations, enriched with temporal and emotional context, are further processed by another Transformer Encoder to capture global document context. Max-pooling is applied to abstract this information.
  • Task-specific Dense Layers and Output Layers: The abstracted document representation is fed into task-specific dense layers for PB and TB detection, culminating in softmax activation for classification.
  • Unified Loss Function: The model is trained using a unified loss function that combines categorical cross-entropy losses for PB and TB with a differential loss to minimize the discrepancy between BERT and TE representations, encouraging a shared feature space.

Datasets

  • CEASE-v2.0: The benchmark English suicide note dataset, annotated with 15 fine-grained emotions at the sentence level. The authors manually annotated this dataset for PB and TB at the note level and for Temporal Orientation (TPO) at the sentence level.
  • CoMCEASE-v2.0: A newly introduced, manually translated code-mixed (Hindi-English) suicide note corpus. This dataset is based on CEASE-v2.0 and is annotated with temporal orientation, PB, and TB labels.

Key Insights and Ideas

  • Multilingualism in Suicide Notes: The paper highlights the prevalence of code-mixed language in suicide notes and the necessity of developing models that can handle it, especially in multilingual regions.
  • Temporal and Emotional Cues are Crucial: The research strongly suggests that incorporating temporal orientation and emotional information significantly enhances the detection of PB and TB from suicide notes.
  • Fusion of Document and Sentence Context: The TEMF framework effectively combines global document context (from BERT) with local sentence-level information (from TE) through an attention mechanism, leading to richer representations.
  • Code-Mixed Corpus Value: The creation of CoMCEASE-v2.0 addresses a critical gap in resources for analyzing suicide notes in code-mixed languages, enabling more inclusive research.

Key Results

  • The proposed TEMF system outperforms several state-of-the-art baselines on both the CEASE-v2.0 and CoMCEASE-v2.0 datasets for PB and TB detection.
  • On the CEASE-v2.0 dataset, TEMF achieved improvements of 2.86% for PB and 3.08% for TB over the next best models.
  • On the CoMCEASE-v2.0 dataset, TEMF achieved the highest F1-scores, with improvements of 2.34% for PB and 4.12% for TB over the next best models.
  • Ablation studies showed that removing either temporal or emotional information significantly degraded performance, with temporal information removal having a more pronounced negative impact.
  • The model demonstrated improved performance with increasing context length, particularly outperforming baselines on the TB task.

Limitations and Future Work

  • Error Analysis: The authors identified several error categories:
  • Inconsistent grammar rules: Informal writing styles in suicide notes can make syntactic information difficult to parse.
  • Insufficient context: Lack of broader context can hinder accurate detection.
  • Sarcastic undertone: Capturing sarcasm, especially with limited context, remains a challenge.
  • Future Work: The authors suggest exploring other potential pathways that increase the ability to commit suicide, such as duration, intensity, and frequency of PB, TB, and acquired capability for self-injury. They also plan to make their code and data available for research purposes.

Summary updated

Abstract

The World Health Organization (WHO) has emphasized the importance of significantly accelerating suicide prevention efforts to fulfill the United Nations' Sustainable Development Goal (SDG) objective of 2030. In this paper, we present an end-to-end multitask system to address a novel task of detection of two interpersonal risk factors of suicide, Perceived Burdensomeness (PB) and Thwarted Belongingness (TB) from suicide notes. We also introduce a manually translated code-mixed suicide notes corpus, CoMCEASE-v2.0, based on the benchmark CEASE-v2.0 dataset, annotated with temporal orientation, PB and TB labels. We exploit the temporal orientation and emotion information in the suicide notes to boost overall performance. For comprehensive evaluation of our proposed method, we compare it to several state-of-the-art approaches on the existing CEASE-v2.0 dataset and the newly announced CoMCEASE-v2.0 dataset. Empirical evaluation suggests that temporal and emotional information can substantially improve the detection of PB and TB.
BibTeX
@inproceedings{ghosh2022am,
  author = {Soumitra Ghosh and Asif Ekbal and Pushpak Bhattacharyya},
  title = {{Am I No Good? Towards Detecting Perceived Burdensomeness and Thwarted Belongingness from Suicide Notes}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/704},
}
Figure