Universal attractors in language evolution provide evidence for the kinds of efficiency pressures involved

The paper investigates how efficiency pressures shape linguistic coding, distinguishing online (speaker-specific, contextual) and offline (conventionalized, population-level) effects. Online efficiency involves trade-offs among articulatory effort, processing, and planning, where predictability can trigger reduction but ambiguity and incremental processing impose costs. Offline effects arise via conventionalization of efficient variants over time, under both individual and sociocultural constraints. The authors argue that a missing ingredient for an integrative predictive theory is the notion of universal attractors: cross-linguistically favored states that languages tend to evolve toward and not away from. They hypothesize an attractor in the domain of verbal subject person-number indexing, characterized by (i) preferred absolute lengths for each person-number combination and (ii) a preference for cumulative (non-compositional) person-number coding. They test this using a global sample and diachronic comparisons with reconstructed proto-languages.

Background work highlights efficiency as a central organizing principle (e.g., Gibson et al., 2019) and distinguishes online signal reduction effects (Aylett & Turk, 2004, 2006; Pluymaekers et al., 2005) from offline conventionalization (Kirby, 2001; Pierrehumbert, 2001; Diessel, 2007; Seyfarth, 2014). Zipf’s Law of Abbreviation and information-theoretic accounts link frequency/informativity to shorter forms (Zipf, 1935; Piantadosi et al., 2011; Bentz & Ferrer-i-Cancho, 2016). Trade-offs among articulatory, processing, and planning costs can produce mismatches between predictability and cue length (Jaeger & Tily, 2010; Bornkessel-Schlesewsky & Schlesewsky, 2014; Seyfarth, 2014). Languages develop context-independent cues to forestall ambiguity (Malchukov, 2008; Seržant, 2019). The concept of attractors stems from dynamic models (Norton, 1995) and has been linked to diachronic typological tendencies (e.g., Greenberg, 1966; Bybee, 1988, 2006, 2008; Maslova, 2000, 2004; Cristofaro, 2012, 2014; Dunn et al., 2011). Prior work also notes frequency asymmetries across person/number (Greenberg, 1966; Bybee, 1985; Karlsson, 1986) and the Uniform Information Density perspective (Aylett & Turk, 2004; Levy & Jaeger, 2007; Coupé et al., 2019). On compositionality vs cumulative coding and efficiency-lexicon complexity trade-offs, see Kirby (2001), Christiansen & Chater (2008), Kemp & Regier (2012), Kemp et al. (2018), Xu et al. (2020).

Data and sampling: A manually compiled database of verbal subject person-number indexes for intransitive verbs in the morphologically unmarked (typically present) tense, excluding dual. Six categories per language: 1SG, 2SG, 3SG, 1PL, 2PL, 3PL. The sample includes 383 languages from 53 families across all six macro-areas (Eurasia, North and South America, Australia, Africa, Oceania). Fifteen families contribute 10–50 languages each to control for language- and family-specific effects; very large families are split into subfamilies (e.g., Nuclear Trans New Guinea split; Afroasiatic split into Semitic and other; Atlantic-Congo represented by Bantu). Diachronic component: reconstructed proto-language indexes for 15 (sub)families (e.g., Proto-Indo-European, Proto-Athabaskan, Proto-Semitic, Proto-Salishan, Proto-Muskogean, Proto-Bantu, Proto-Dravidian, Proto-Oceanic, Proto-Mayan, Proto-Uralic, Proto-Sogeram, Proto-Awyu-Dumut, Proto-Worroran, Proto-Kiranti, Proto-rGyalrongic), based on authoritative sources. Families without consensual reconstructions were excluded from diachronic analysis. Measurement: Index length operationalized as number of segments, proxied by number of letters (exceptions noted for French/English orthography); long segments counted as 1.5. All analyses conducted in R. Models: - Attractor length estimation: Poisson mixed-effects regression with person and number as fixed effects and clade as a random effect; formula: index length ~ person*number + (1|clade). Goodness assessed against a model neglecting person/number (Fisher exact test) and by significance of predictors. - Diachronic movement toward attractor: For each proto vs. modern form, compute whether the modern length is closer to the estimated attractor than the proto length (or retained if already near). Binary outcome (moved toward/within attractor vs. not). Logistic mixed-effects model with person and number as fixed effects and clade as random effect; formula: movement ~ person*number + (1|clade). - Compositionality preference: For each person, classify diachronic change across four categories: no compositionality, compositionality disappears, compositionality remains, compositionality appears. Then model dispreference vs. preference by merging categories (dispreferred: no compositionality + disappears; preferred: remains + appears). Logistic mixed-effects model with person and proto compositionality as fixed effects and clade as random effect; formula: compositionality(modern) ~ person*compositionality(proto) + (1|clade). Interpretation links attractor properties to efficiency pressures at articulatory, processing, planning, memory, and informativity stages.

- Existence of an attractor for index lengths: Across languages, person-number index lengths exhibit low dispersion and are tightly predicted by person and number. In the Poisson mixed-effects model, both person and number are statistically significant predictors; estimated absolute average lengths per person-number combination define the attractor (Figure 2). - Diachronic convergence toward attractor: Logistic mixed-effects modeling shows high probabilities that modern languages move toward or remain within attractor lengths across all person-number combinations. Probabilities are extremely high for singular forms (~90–100%) and lower but still high for plurals (~65–90%); the singular vs. plural difference is statistically significant (Figure 3). No significant bias by person. - Directional adjustments: Families with proto-forms shorter than the attractor tend to lengthen (e.g., Uralic singulars: 1SG m, 2SG n, 3SG Ø lengthened in modern Uralic; plural forms that were already near attractor remained). Families with overly long proto-forms shorten (e.g., Proto-Indo-European 2SG -e-si, 3 segments, shortened to 1.57 segments on average in modern Indo-European; Proto-Mayan 1/2 persons ~2.5 segments shortened to ~2, while 1/2 plurals ~2 segments lengthened to ~2.64). Indices near attractor remain stable (e.g., 1SG in Sogeram, Athabaskan, Semitic). - Compositionality dispreferred: Diachronic evidence shows a strong long-run avoidance of compositional index coding. The logistic model predicts an extremely high probability (>95%) of non-compositional (cumulative) coding across persons (Figures 4–5). Families without proto compositionality (e.g., Indo-European) do not develop it; some with proto compositionality (e.g., Awyu-Dumut) reduce it. - Frequency-structure alignment: Person-number frequency asymmetries align with length asymmetries consistent with Zipf’s Law of Abbreviation: third person is most frequent and shortest; singulars more frequent and shorter than plurals. Example from Russian National Corpus (oral, 216,112 words): 1SG 26% (2,276), 2SG 5% (471), 3SG 69% (6,021); 1PL 15% (601), 2PL 23% (926), 3PL 62% (2,493); singular total 69% (8,768) vs. plural 31% (4,020). Both frequency asymmetries (3rd vs. others; singular vs. plural) are significant (p=0.002, χ²). - Zero coding not favored: Despite high frequency of 3SG, zeros are not generally preferred; some lineages even replace inherited 3SG zero with overt segments (e.g., Finnic replacing Proto-Uralic 3SG zero).

Findings support the hypothesis that language change is channeled by universal attractors in verbal subject indexing. The attractor exhibits structured absolute lengths by person and number and favors cumulative, non-compositional coding. These outcomes reflect interacting efficiency pressures: (i) articulatory efficiency explains shorter forms for more frequent categories (3rd person, singular) per Zipf; yet pure articulation is constrained by (ii) processing and planning efficiency, which disfavor zero markers due to potential ambiguity and increased processing load in incremental communication. Obligatory indexing and non-zero cues reduce planning uncertainty and facilitate processing. Longer plural forms, beyond what would be minimally needed for disambiguation, align with constant information flow/Uniform Information Density: less expected meanings (plurals) benefit from longer forms that distribute information more uniformly over time, aiding both production and comprehension. The robust cross-linguistic preference for cumulative coding indicates that, in high-frequency grammatical domains, processing ease outweighs lexicon simplicity and memory/learnability costs. Thus, for items used extremely often, storing and retrieving six atomic forms is more efficient overall than maintaining compositional structure that imposes higher processing costs during production and comprehension.

The study establishes a universal attractor state in verbal subject person-number indexing, characterized by specific absolute lengths per person-number combination and a strong preference for cumulative (non-compositional) coding. Diachronic analyses show consistent movement toward or stability within this attractor across diverse families, via reduction, enlargement, or retention as needed. These patterns reveal the relative strengths of efficiency pressures: articulatory and processing/planning efficiency, along with constant information flow, dominate, whereas minimizing lexicon size and memory costs is weaker in this high-frequency domain. Future work should complement cross-linguistic comparative findings with experimental and psycholinguistic evidence to directly test processing and planning trade-offs and to probe generalization to other grammatical domains.

- Diachronic scope constrained by availability of consensual reconstructions: only 15 (sub)families included; 38 families excluded from diachronic analysis. - Measurement proxy: index length approximated by letter counts; long segments assigned 1.5; orthographic proxies may not fully capture phonetic segment length across languages. - Domain restrictions: only intransitive verbs, morphologically unmarked (typically present) tense; dual excluded; potential effects in other verb classes/tenses not assessed. - Potential language-specific processes (reduction, analogy, reanalysis) may introduce idiosyncrasies, though sampling and modeling aimed to average these out. - Cross-linguistic comparative evidence without direct experimental validation; authors note the need for experimental support. - Zero vs. overt coding and compositionality judgments rely on descriptive sources; cross-linguistic comparability may vary.