TL;DR
This paper introduces a novel computational framework to formally analyze "style" in personal narratives, moving beyond subjective descriptions to identify quantifiable linguistic patterns. By combining systemic functional linguistics with sequence analysis, they demonstrate how these patterns can reveal psychological insights, particularly in the context of trauma narratives.
What Problem Does This Paper Address?
Authors use language to convey subjective experiences, and "style" is a key aspect of this. However, there's a lack of formal, systematic methods to analyze these stylistic choices in personal narratives. Existing approaches are often qualitative or lack operational tools. This paper aims to bridge this gap by creating a computational framework that can automatically extract and analyze linguistic patterns, linking them to psychological states and offering potential applications in therapy and authorship analysis.
Methods
The core of the approach is a sequence-based framework. First, they leverage language models (specifically Llama 3.1 8B Instruct) to automatically extract linguistic features from narratives, focusing on processes (actions, mental states, verbal acts, states of being) and their associated participants and circumstances. These extracted features are then converted into symbolic sequences.
The paper then employs sequence analysis techniques, drawing inspiration from computational biology. They focus on identifying substrings (contiguous blocks of symbols) within these sequences. Cosine similarity is used to measure the stylistic similarity between narratives, treating sequences as vectors. Hierarchical agglomerative clustering is applied to group similar narratives, and representative sequences are identified to summarize cluster characteristics.
Datasets
The primary dataset used is the DreamBank corpus, a collection of thousands of dream narratives from the United States. For their case study, they specifically analyze the "viet" series, which consists of narratives from a Vietnam War veteran with post-traumatic stress disorder. They also created a "norm" dataset by sampling ten narratives from each of five series (blind, ed, izzy, merri, viet) to establish a comparative baseline.
Key Insights and Ideas
- Style as quantifiable linguistic patterns: The paper argues that style isn't just an abstract concept but can be represented by recurring patterns in linguistic choices, particularly in how subjective experiences are encoded.
- Bridging linguistics and computation: By integrating Systemic Functional Linguistics (SFL) with computational sequence analysis, they create a powerful tool for analyzing text at a deeper, meaning-driven level.
- Psychological insights from linguistic patterns: The analysis of the "viet" series suggests that specific linguistic patterns (e.g., a prevalence of verbal processes over mental ones) can correlate with psychological states, potentially offering diagnostic markers or therapeutic targets.
- Automation is key: Leveraging large language models for feature extraction automates a previously labor-intensive process, enabling large-scale analysis of narrative corpora.
Key Results
The case study on the "viet" series revealed distinctive patterns in process types. For instance, the veteran's narratives showed a 40% higher likelihood of using verbal processes and a 40% lower likelihood of using mental processes compared to the norm. Clustering of "viet" narratives identified two main patterns: one dominated by action processes and another more balanced between action and state processes. These findings are discussed in relation to potential psychological implications, such as how combat trauma might influence cognitive and emotional processing.
Limitations and Future Work
The authors acknowledge that their language model-based annotation can introduce inaccuracies, and future work should focus on improving model performance and validating annotations with expert data. The current connections between linguistic patterns and psychological states are correlational and descriptive, and further validation with clinical assessments is needed to strengthen claims about therapeutic utility.
Future work includes:
- Expanding the framework to include more linguistic features (participants, circumstances, tense, aspect).
- Exploring subsequences in addition to substrings for more flexible pattern detection.
- Applying complexity science measures to quantify sequence redundancy and gain deeper psychological insights.
- Developing style-conditioned narrative generation for therapeutic applications.
- Enhancing authorship profiling by identifying signature stylistic patterns.
Summary updated