The lexical divide: propositive modes and non-agentic attitudes define the progressive left in Chile

The study investigates how psychological traits related to agency, locus of control, and modal attitudes differentiate political orientations within Chile’s party system, particularly contrasting the emerging New Left (Frente Amplio) with the traditional left and right. Motivated by global cases like Sanders’ movement in the U.S. and Podemos in Spain—where individual sense of agency correlates with political engagement—the authors ask how such traits manifest in Chile. Traditional survey methods are prone to biases (social desirability, acquiescence, extreme responding), so the paper leverages psychological language analysis to infer attitudes from linguistic structure rather than content. Set during Chile’s 2015–2016 constituent process, the research examines texts produced in self-convened local meetings (ELAs), using them as markers of ideological orientation due to high internal agreement within groups. The central research question: How do linguistic indicators of agency and modal attitudes differ across progressive left, traditional left, and right-wing communities in Chile?

The paper situates its inquiry within work linking stable psychological differences (e.g., openness, locus of control, agency) to political participation and ideology. It builds on psychological language analysis, which treats language use as a reflection of mental states, extending prior applications in neuropsychology and political behavior. It critiques the prevalent “words as attention” paradigm, advocating a shift to structural linguistic features (agency, modality, process types). The study also draws on theories of political agency in the digital age and systems worldview moderating civic engagement. In Chilean context, historical cleavages and coalition dynamics (Concertación/Nueva Mayoría vs. Chile Vamos) and the rise of the New Left (Frente Amplio) provide the political backdrop. Rights are framed via Vasak’s three generations to interpret co-occurrence clusters. Together, these strands motivate using text-based, structural markers to approximate psychological constructs relevant to ideology.

Data: Text justifications from Chile’s 2015–2016 constituent process self-convened local meetings (ELAs). Values dataset: 46,660 rows, 971,849 words, avg 20.8 words per sentence. Rights dataset: 45,094 rows, 925,317 words, avg 20.5 words per sentence. Each row includes concept name (value/right), ELA id, commune, a three-level agreement indicator (agreement, partial, disagreement), and the justification text. Most ELAs achieved consensus: 90.6% agreement for Values (8.9% partial, 0.5% disagreement) and 92.1% for Rights (7.4% partial, 0.5% disagreement). The study uses the finalized, systematized datasets provided by the official team, with added normalized text, syntagm decomposition, and modal attitude variables.

1. Identification of ideological communities via co-occurrence networks: An M×N incidence matrix A=a(i,j) (ELAs × concepts) is constructed separately for Values and Rights. Distances between concept pairs are computed using the phi-correlation coefficient; significance is assessed (chi-square with 1 d.f.). If significant at 95%, a link is added with weight 1−d (Onnela distance). Community detection uses the Louvain algorithm to identify clusters of frequently co-selected concepts. Modularity: Values 0.56; Rights 0.44.

2. Linguistic feature extraction: Spanish texts are analyzed using LIWC2015 and Stanford CoreNLP (POS Tagger, Dependency Parser). Key features:

• Agency: Categorizes sentences as Low (passive/reflexive passive with “se”, modal existential periphrases like “debe haber/existir”), Medium (active voice, third-person agents such as state, government, universities), High (active voice, first-person singular/plural). Extraction focuses on verbal constructions, identifying passive markers and person via verb conjugation.
• Modal attitude: Factual (descriptive/factive), Evaluative (judgment/opinion), Propositive (normative, deontic modals like should/would). Variable constructed via manual annotation and automatable NLP rules.
• Type of process (systemic functional): Material (doings), Relational (being), Existential (existence). Main verb extracted and classified. Additional features: Text length (control), Categorical–Dynamic Index (CDI) as proxy for education/scholarly aptitude, semiotic Multidimensionality (degree of incorporating multiple social dimensions; noted operational issues capturing equifunctionality rather than true multidimensionality).

3. Discrete choice modeling: Multinomial Probit (primary) and Logit (robustness) models estimate the probability that a text belongs to a given ideological cluster from its linguistic features and controls. Results reported as Relative Risk Ratios (RRR). Sentence-level independent variables include mean age, CDI, text length, agency (low/medium/high), dimensionality (single/multi), modal attitude (factual/evaluative/propositive), type of process (material/relational/existential). Model specifics and tests in Supplementary Material.

Validation: Bootstrapping over 100 random samples to prevent dominance by highly popular concepts (sampling up to 100 observations per concept). Alternative community detection algorithms and link thresholds yield consistent clusters. Randomized null models (1,000 shuffled matrices) show networks differ significantly from random pairings.

• Network-derived ideological clusters: • Cluster A (Progressive left): In Values and Rights, emphasizes second-generation rights and most third-generation rights (cultural heritage, environment, animal rights), with values around justice/equality. • Cluster C (Traditional left-wing): Focus on second-generation rights (education, healthcare, social security) and values like justice and equality. • Cluster D (Traditional right-wing): Dominated by first-generation (negative) rights (property, freedom, civil liberties) and values like patriotism, rule of law, free enterprise. • Cluster B (Evangelical groups): Appears only in Values, with conservative/evangelical themes (e.g., freedom of worship, heterosexual marriage family); closest to Cluster D in Rights but absent as a distinct cluster there. Modularity: Values 0.56; Rights 0.44.

• Participation consensus: High agreement in group decisions supports treating group texts as markers for ideology: Values 90.6% agreement (8.9% partial, 0.5% disagreement); Rights 92.1% agreement (7.4% partial, 0.5% disagreement).

• Demographic/education patterns: Younger participants and those with higher CDI (proxy for education) are more likely associated with Cluster A, consistent with generational shifts toward post-materialist values.

• Modal attitudes and agency by cluster: • Cluster A: Strongly propositive (normative) stance; higher likelihood of material process verbs (notably “guarantee”); greater use of passive voice (lower agency), indicating normative demands with less explicit assignment of responsible actors; longer and more multidimensional texts. • Cluster C: Predominantly evaluative attitudes, especially on values like Democracy and Respect; frequent first-person plural (high agency) in Values, signaling collective responsibility; third-person subjects when discussing rights, often identifying the State as responsible for provisioning second-generation rights. • Cluster D: Factual attitudes and third-person usage when defining concepts (objective stance); switches to evaluative attitudes in Rights to emphasize the fundamental nature of first-generation rights (e.g., property, freedom of conscience).

• Robustness: Clusters align with real-world political coalitions (Frente Amplio, traditional left, Chile Vamos). Results are stable across alternative algorithms/thresholds and differ significantly from randomized networks.

The study addresses its central question by demonstrating systematic differences in linguistic markers of agency and modal attitudes across Chile’s ideological communities. The progressive left’s propositive and comparatively non-agentic style contrasts with the traditional left’s evaluative and collectively agentic framing in values (and state-centric agency in rights), while the right adopts factual definitional stances and evaluative emphasis on first-generation rights. These findings substantiate the role of psychological constructs—agency, locus of control, and normative orientation—in shaping ideological expression. Methodologically, the work illustrates that structural linguistic features extracted from deliberative group texts can reliably proxy psychological dispositions and political orientation, offering a complementary tool to surveys that may suffer from response biases. The alignment of clusters with real-world coalitions and validation against randomized baselines reinforce the approach’s credibility. The implications extend to understanding how different political groups frame responsibilities (self/collective/state) and legitimize rights (negative vs. positive), informing theories of political psychology and communication in digital-era collective deliberation.

This paper integrates network science and text analytics to map ideological communities in Chile’s 2015–2016 constitutional deliberations and to link linguistic structure to psychological constructs of agency and modality. It identifies three major clusters mirroring Chile’s political landscape—progressive left, traditional left, and right—and shows distinctive attitudinal styles: propositive and less agentic (progressive left), evaluative and collectively agentic (traditional left), and factual with evaluative emphasis on first-generation rights (right). The approach demonstrates that group-generated deliberative texts can serve as valid proxies for psychological and ideological profiling, offering an alternative to bias-prone surveys. Future work should: (i) analyze individual-level data to sharpen inference; (ii) develop a unified theoretical framework for psychological language approaches; (iii) enhance operationalization of constructs like multidimensionality by incorporating semantic understanding; and (iv) extend and compare across contexts and time, accounting for linguistic variation and more complex sentence structures beyond main verbs.

• Data aggregation and representativeness: Texts are group-level justifications from voluntary ELAs, likely homogeneous but not representative of the population; aggregate data may dilute effects in regressions.
• Ideological coverage: Evangelical cluster appears only in Values and was excluded from some analyses for consistency, potentially omitting a distinct conservative subgroup.
• Linguistic operationalization: Agency detection relies on main verb and Spanish passive markers (e.g., “se”); complex sentences and multiple verbs may obscure agency assignment. Dictionaries (e.g., Halliday verb lists) may be incomplete; regional/sociocultural language variations may affect results.
• Modal attitude annotation: While definable and partly automatable, it may be sensitive to context and subtle pragmatic cues.
• Multidimensionality metric: Operationalization captured equifunctionality rather than true multidimensionality due to reliance on syntactic rules without semantic comparison.
• Temporal and cross-context generalizability: Language use changes over time and across contexts; findings are specific to Chilean Spanish and the 2015–2016 process, limiting generalization without adaptation.