Formalizing Style in Personal Narratives

AI summary

Overview

Research area: Natural Language Processing / computational stylistics, at the intersection of systemic functional linguistics, formal language theory (sequence analysis), and computational psychology.

Technical level: Intermediate. The formal machinery (alphabets, words, substrings, subsequences, cosine similarity, hierarchical clustering) is introduced in plain terms, but readers will get the most out of it with some familiarity with basic formal language concepts and narrative-analysis terminology.

Scope: The paper proposes and demonstrates a sequence-based framework that defines personal-narrative style as recurring patterns in the linguistic choices authors make when encoding subjective experience, applied to hundreds of DreamBank dream narratives with a focused case study on a Vietnam War veteran with post-traumatic stress disorder.

What This Paper Is About

Personal narratives are stories authors build to make meaning of their experiences, and the way they choose language — their style — carries information about their subjective experience. Despite long traditions in stylometry and stylistics, the authors argue there is no formal framework for systematically analyzing these stylistic choices. The goal of the paper is to define style formally as patterns in sequences of linguistic choices, to show how those patterns can be extracted automatically using language models and sequence analysis, and to demonstrate that the resulting patterns can be given psychological interpretations.

Key Contributions

1. A sequence-based framework defining style as patterns in sequences of linguistic choices grounded in systemic functional linguistics, where narratives are mapped to symbolic words over alphabets derived from the transitivity system (processes, participants, circumstances).
2. A methodology for automatically identifying patterns using sequence analysis: extracting linguistic features with language models at scale, then comparing substring distributions, computing cosine similarity, and applying hierarchical agglomerative clustering to find representative sequences.
3. A case study on dream narratives across five DreamBank series (blind, ed, izzy, merri, viet) plus a 720-narrative norm baseline, showing how pattern analysis can reveal psychological insights and support therapeutic applications.
4. An automated systemic functional linguistics pipeline, which the authors describe as the first attempt to automate analysis from systemic functional linguistics using language models, eliminating manual annotation campaigns.

Main Findings

• Automated extraction was reliable on gold-standard material. Fifty clauses randomly sampled from fully analysed examples in Banks (2019) and Halliday et al. (2014) were annotated; predicted processes, participants and circumstances were all correctly matched to the references.
• The veteran's narratives are verbal-heavy and mental-poor. Relative to the norm baseline, the presence of verbal processes is 40% more likely in the viet series (odds ratio 1.4), while the presence of mental processes is 40% less likely (odds ratio 0.6).
• A preference for staying in the same process type. The viet series shows high odds ratios for the substrings verbal.verbal (2.00) and verbal.verbal.verbal (1.75), indicating the veteran tends to remain in a verbal process rather than switching.
• Some patterns appear in the baseline but not in the veteran's narratives. For example, mental.state.mental.mental occurs in the norm but does not appear in viet.
• Two distinct stylistic templates emerged from clustering. The silhouette score reached its maximum with substrings up to length three and a two-cluster partition. The first representative sequence (savamasasaaamaaasavvvaaaaaaavssaaaaa) shows dominance of action processes (23 occurrences), minimal mental processes (2), covers 274 sequences, and has an average distance of 8 to other sequences in its cluster. The second (sssssavaavssvsavvvvsmasasaasasaamaamvmsss) shows a more balanced pattern with action (13) and state (16) processes and few mental processes (4), covers 179 sequences, and has an average distance of 21.
• Corpus-scale substring statistics vary widely by series. For viet, distinct substring counts and the counts of substrings with significant differences versus the norm run from size 1 (4 distinct, 3 significant) through size 9 (4773 distinct, 13 significant); izzy reaches 8568 distinct substrings at size 9, while ed peaks at 1790 (size 8) and merri at 3610 (size 9).
• The reported psychological link is descriptive, not diagnostic. The authors suggest that framing experience mainly through action and verbal processes rather than mental ones may align with literature on how combat trauma affects cognitive and emotional processing (American Psychiatric Association, 2013), but state that further research would be needed to confirm this interpretation.

Methodology in Plain English

• Build a feature system. Drawing on Halliday's transitivity system, the authors treat language as a set of meaningful choices. They categorise processes into four types — action, mental, verbal, and state — each with associated participants (for example, Actor/Affected for action, Senser/Phenomenon for mental).
• Turn narratives into strings. Each dream narrative is split into sentences with the SpaCy segmentation model en_core_web_trf-3.8.0, then into clauses using a language model with in-context examples (dependent and independent clauses are not distinguished). Linguistic features are extracted with Llama 3.1 8B Instruct, an instruction-tuned autoregressive model with 8 billion parameters, which the authors note was the best-performing 8B open-weights model at the time of their experiments. Each clause becomes a symbol, so a narrative becomes a word such as amv, and the set of available choices becomes an alphabet, e.g. Σ_process = {action, mental, verbal, state}. Experiments ran on a Tesla V100 32GB for 80 hours.
• Look for recurring patterns. They search for substrings (contiguous blocks) and note that the framework extends naturally to subsequences (non-contiguous, order-preserving), which formal language theory calls factors and subwords. The analogy to computational biology is explicit: replacing nucleotides with linguistic features lets the authors apply sequence-analysis tooling.
• Compare against a baseline and test significance. A norm was constructed from a random sample of ten narratives from each DreamBank series, giving 720 dream narratives. Substring proportions in each series versus the norm were compared with a Fisher exact test (Holm–Bonferroni correction), with significant results reported as odds ratios at p < 0.05 for substring sizes one through three.
• Measure similarity and cluster. Sequences are compared as frequency vectors of substrings using cosine similarity (ranging 0 to 1); the worked example s1 = amvma and s2 = amma yields a cosine similarity of 0.836. Hierarchical agglomerative clustering with Ward linkage and cosine similarity groups similar narratives; the authors state that no number of clusters needs to be specified because the method produces a full dendrogram. Representative sequences were chosen as the sequence with the minimal average distance to all others in its cluster, covering about 80% of sequences.

Why This Matters

The paper's main value is offering an operational bridge between qualitative, theory-rich accounts of personal expression (phenomenology, Granger's philosophy of style, Dilts' modelling of strategies) and quantitative text analysis. Instead of measuring what is said, it targets how subjective experience is constructed linguistically, complementing existing stylometry and stylistics.

Real-world applications the paper points to:

• Therapeutic narrative work. Identifying low- or high-odds substrings could serve as diagnostic markers or guide interventions as therapists and patients restructure narratives, potentially supporting trauma processing.
• Authorship profiling. Inferring demographic and psychological characteristics from text, with uses in forensic linguistics, literary attribution, online security, and bot detection, by finding signature substrings that characterise an author's way of constructing narratives.
• Style-conditioned narrative generation. Reversing the mapping — generating narratives from sequences of choices — to control fine-grained linguistic attributes, extending controllable text generation beyond high-level attributes like formality and sentiment toward applications such as cognitive reframing of negative thoughts.
• Scalable dream and narrative research. Replacing labour-intensive manual coding systems with language-model extraction that scales to thousands of narratives without additional annotation effort.

Industry relevance sits mainly in mental-health and wellbeing technology (therapist-in-the-loop tools and text-transformation systems), forensic and security text analytics, and any application that needs interpretable, controllable stylistic control over generated text rather than only content control.

Future Directions

• Extend beyond process types. The current case study is a proof of concept using only process types; the authors plan fine-grained analysis including participants, circumstances, and functional-role features such as the duration and completeness of processes (tense and aspect) and the concreteness of participants, using larger alphabets formed by cartesian products.
• Move from substrings to subsequences and complexity measures. Non-contiguous pattern detection is flagged as more flexible, and complexity measures such as the Lempel-Ziv measure could quantify redundancy: low complexity may indicate fixed mental representations or recurring cognitive schemas relevant in trauma, while higher complexity may suggest greater cognitive flexibility.
• Validate the psychological claims. The authors state the linguistic-to-psychological connections are correlational and descriptive, and that validation against clinical assessments would be necessary to strengthen claims about therapeutic utility; they also note that future work could evaluate the models against more expert annotations.
• Inverse mapping and generation. Exploring the inverse process of generating narratives from sequences of choices opens style-conditioned narrative generation for therapeutic interventions or literary writing. The framework also invites incorporation of alternative similarity measures (Jaccard similarity, Euclidean distance) and better language models, which the authors say can replace Llama 3.1 8B without altering the essence of the approach.

Target Audience

This paper will benefit computational linguists and NLP researchers working on stylometry, controllable text generation, or interpretable text analysis; researchers in dream research and computational psychology; therapists and clinician-researchers interested in how narrative reconstruction is reflected in language; and formal-language-theory readers curious about applied sequence analysis on natural language. It is also relevant to anyone designing applied systems that need to characterise how a person writes about their own experience, given the authors' explicit caution that the framework is intended for exploratory research or therapist-in-the-loop settings, not fully autonomous clinical deployment.

Authors’ abstract

Personal narratives are stories authors construct to make meaning of their experiences. Style, the distinctive way authors use language to express themselves, is fundamental to how these narratives convey subjective experiences. Yet there is a lack of a formal framework for systematically analyzing these stylistic choices. We present a novel approach that formalizes style in personal narratives as patterns in the linguistic choices authors make when communicating subjective experiences. Our framework integrates three domains: functional linguistics establishes language as a system of meaningful choices, computer science provides methods for automatically extracting and analyzing sequential patterns, and these patterns are linked to psychological observations. Using language models, we automatically extract linguistic features such as processes, participants, and circumstances. We apply our framework to hundreds of dream narratives, including a case study on a war veteran with post-traumatic stress disorder. Analysis of his narratives uncovers distinctive patterns, particularly how verbal processes dominate over mental ones, illustrating the relationship between linguistic choices and psychological states.