Integrating topic modeling and word embedding to characterize violent deaths

Alina Arseniev-Koehler, Susan D. Cochran, Vickie M. Mays, Kai-Wei Chang, Jacob Gates Foster

TL;DR

This paper introduces Discourse Atom Topic Modeling (DATM), a novel method that integrates topic modeling and word embeddings to uncover latent themes in text data. DATM allows for a more nuanced understanding of complex datasets, demonstrated here by characterizing the gendered nature of violent death narratives in the US National Violent Death Reporting System (NVDRS).

What Problem Does This Paper Address?

Large-scale text data, such as those found in public health surveillance systems, hold immense potential for uncovering patterns and insights. However, traditional text analysis methods often struggle to extract meaningful information from these unstructured narratives. Specifically, while topic modeling can identify themes and word embeddings can capture semantic relationships, integrating these approaches to provide a richer, more interpretable understanding of complex phenomena like violent deaths has been a challenge. The authors aim to bridge this gap by developing a method that can both identify latent topics and represent them within a semantic space, allowing for the exploration of nuanced relationships, such as gendered patterns, that might be missed by existing methods.

Methods

The core of the paper is the introduction of Discourse Atom Topic Modeling (DATM). DATM combines two key techniques:

  • K-SVD for Topic Identification: DATM uses K-SVD, a sparse dictionary learning algorithm, to identify a set of "discourse atoms" within a pre-trained word embedding space. These atom vectors are interpreted as latent topics, where words closest to an atom vector are representative of that topic.
  • Latent Variable Model for Text Mapping: A Latent Variable Model is employed to map observed text (sentences or documents) into the semantic space. This allows for the inference of a "gist" vector representing the context. By finding the closest discourse atom to this gist vector, DATM can assign topics to text segments.

This approach allows DATM to represent both words and topics within a shared semantic space, enabling the analysis of semantic dimensions (like gender) in relation to identified topics.

Datasets

The primary dataset used in this study is the National Violent Death Reporting System (NVDRS), collected from 2003 to 2017. This dataset contains narratives from law enforcement and medical examiner/coroner reports for over 300,000 violent deaths in the US. The authors also mention using the WordSim-353 Test and the family section of the Google Analogy test for evaluating the quality of their trained word embeddings.

Key Insights and Ideas

  • Bridging Topic Models and Word Embeddings: DATM offers a theoretically grounded way to integrate the strengths of topic modeling (identifying latent themes) and word embeddings (capturing semantic meaning and relationships).
  • Interpretable Topics in Semantic Space: By representing topics as vectors in the same space as word embeddings, DATM allows for the direct application of semantic analysis techniques (like gender bias detection) to the identified topics.
  • Uncovering Hidden Patterns in Violent Death Narratives: The NVDRS, despite its rich narrative data, has been underutilized. DATM demonstrates its potential to reveal nuanced patterns, such as gendered language associated with specific causes of death.
  • Gendered Language in Violent Death: The study found that topics identified by DATM exhibit distinct gendered meanings, and the prevalence of these topics in narratives correlates with the gender of the victim, suggesting that language used to describe violent deaths is gendered.

Key Results

  • 225 Latent Topics Identified: DATM successfully identified 225 interpretable latent topics within the NVDRS narratives, many of which were not captured by the structured variables in the dataset.
  • Gendered Topics: The analysis revealed distinct gendered associations for many topics. For example, "long guns" (rifles and shotguns) was identified as a masculine topic, while "sedative and pain medications" was identified as feminine.
  • Correlation between Topic Gender and Victim Gender: A strong correlation (Spearman ρ = 0.69, p<0.0001) was found between the gendered meaning of topics and their prevalence in narratives of female versus male victims.
  • Topic Prevalence Analysis: Logistic regression analysis showed that topics like "rifles and shotguns" were more common in narratives of male decedents, while "sedative and pain medications" were more common in narratives of female decedents. This held true even after controlling for other demographic factors.

Limitations and Future Work

  • Topic Interpretability: While DATM aims for interpretable topics, the authors acknowledge that manual assignment of topic labels is subjective.
  • Computational Cost: Training word embeddings and applying K-SVD can be computationally intensive, especially for very large corpora.
  • Generalizability: The study focused on violent death narratives. Further research is needed to assess the generalizability of DATM to other domains and text types.
  • Future Directions: The authors suggest that DATM can be used to explore other latent semantic dimensions (e.g., social class), to analyze temporal trends in topic prevalence, and to integrate with other analytical methods for a more comprehensive understanding of complex phenomena. They also highlight the potential for DATM to uncover actionable insights for public health interventions and research.

Summary updated

Abstract

There is an escalating need for methods to identify latent patterns in text data from many domains. We introduce a new method to identify topics in a corpus and represent documents as topic sequences. Discourse Atom Topic Modeling draws on advances in theoretical machine learning to integrate topic modeling and word embedding, capitalizing on the distinct capabilities of each. We first identify a set of vectors ("discourse atoms") that provide a sparse representation of an embedding space. Atom vectors can be interpreted as latent topics: Through a generative model, atoms map onto distributions over words; one can also infer the topic that generated a sequence of words. We illustrate our method with a prominent example of underutilized text: the U.S. National Violent Death Reporting System (NVDRS). The NVDRS summarizes violent death incidents with structured variables and unstructured narratives. We identify 225 latent topics in the narratives (e.g., preparation for death and physical aggression); many of these topics are not captured by existing structured variables. Motivated by known patterns in suicide and homicide by gender, and recent research on gender biases in semantic space, we identify the gender bias of our topics (e.g., a topic about pain medication is feminine). We then compare the gender bias of topics to their prevalence in narratives of female versus male victims. Results provide a detailed quantitative picture of reporting about lethal violence and its gendered nature. Our method offers a flexible and broadly applicable approach to model topics in text data.
BibTeX
@inproceedings{arsenievkoehler2022integrating,
  author = {Alina Arseniev-Koehler and Susan D. Cochran and Vickie M. Mays and Kai-Wei Chang and Jacob Gates Foster},
  title = {{Integrating topic modeling and word embedding to characterize violent deaths}},
  booktitle = {Proceedings of the National Academy of Sciences},
  year = {2022},
  doi = {10.1073/pnas.2108801119},
}
Figure