# Compositional processing in the recognition of Chinese compounds: Behavioural and computational studies.

- **DOI:** [10.3758/s13423-025-02668-8](https://doi.org/10.3758/s13423-025-02668-8)
- **Rating:** Sound (3.4 out of 5)
- **Verdict:** CAOSS predicts Chinese compound behavior across three tasks, extending compositional modeling beyond Germanic languages, but the observational design leaves the proposed cognitive mechanism unresolved.
- **Framework version:** v5.10
- **Pipeline version:** v0.4
- **Evaluation date:** 2026-11-09
- **Evaluation page:** https://www.nabu.science/paper/10.3758%2Fs13423-025-02668-8

## Key claims

1. These results suggest that a compositional process is involved in Chinese compound processing, even in tasks that do not explicitly require meaning combination.
   - Support: Tentative
   - Type: Primary finding
   - Basis: observational model–behavior associations do not isolate composition from other semantic explanations
   - Location: Abstract
2. [The CAOSS-derived proximity measure] predicted sensibility judgements on novel compounds (Study 1), lexical decision latencies for rejecting novel compounds (Study 2), and lexical decision latencies for recognising existing compounds (Study 3).
   - Support: Tentative
   - Type: Secondary
   - Basis: aggregate observational associations establish prediction within the analyzed data rather than a direct effect of semantic integration
   - Location: Abstract; Table 3
3. the CAOSS model provides a good description of compositional processing in Chinese, and that it outperforms an additive model in which compound representations are derived by adding standalone vectors for the constituent characters together.
   - Support: Supported
   - Type: Secondary
   - Basis: lower AIC across all three analyses supports comparative fit, but held-out predictive superiority is not established
   - Location: Discussion
4. Our computational work further suggests that readers can learn the meaningful functions of constituent characters through exposure to whole compounds
   - Support: Tentative
   - Type: Secondary
   - Basis: computational learning and behavioral fit do not directly test human acquisition through exposure
   - Location: Discussion, concluding paragraph
5. However, our findings reveal significant proximity effects for both characters, suggesting both characters are important during compound processing.
   - Support: Supported
   - Type: Secondary
   - Basis: both-position coefficients are significant across tasks, within an observational design that does not isolate constituent mechanisms
   - Location: Discussion

## Strengths

- **Three tasks extend a cross-linguistic test** The study tests a framework developed mainly for Germanic languages against Chinese sensibility judgments, novel-compound rejection, and existing-compound recognition. This addresses whether the model remains informative under greater variability in compound structure and constituent meaning.
  - Dimension: contribution
  - Location: Research Aims; Studies 1–3
- **Large item sets support association estimates** The analyses cover 1,478, 22,553, and 15,145 items, with random intercepts for constituent characters. Table 3 reports semipartial effect sizes alongside significance tests, making the small proximity contributions visible.
  - Dimension: methodological rigour
  - Location: Analysis Scheme Using Behavioural Measures; Tables 2–3
- **Position-sensitive modeling receives an explicit comparison** CAOSS is compared with a simple additive representation in every study and receives lower AIC throughout. The contrast directly tests the proposed value of representing constituent position rather than merely combining standalone meanings.
  - Dimension: contribution
  - Location: Method, Compositional Models; Table 2

## Limitations

- **Semantic alternatives leave the mechanism unresolved** The proximity associations are not separated from character-frequency, ambiguity, or non-compositional semantic explanations. They therefore support model–behavior relationships more directly than a specific routine composition mechanism.
  - Dimension: methodological rigour
  - Location: Method, Probing Semantic Composition and Analysis Scheme; Table 3
- **Comparative fit lacks held-out predictive validation** The model comparison uses AIC on the analyzed outcomes without a held-out predictive test. Separation between CAOSS training compounds and the existing compounds tested in Study 3 is not established.
  - Dimension: methodological rigour
  - Location: Method, Compositional Models and Analysis Scheme; Table 2; Study 3
- **Mechanistic conclusions exceed the observational evidence** The Discussion describes integration ease as impacting behavior and composition as routine, although no human composition or learning process is directly manipulated. The interpretation also leaves population differences and alternative semantic accounts insufficiently developed.
  - Dimension: reporting
  - Location: Discussion, paragraph 2 and concluding paragraph

## Reliability checks

Caveats (4 of 4 checks).

- **Results check:** Noted. 4 reported values do not reconcile, but the key claims do not rest on them.
  - Minor reporting notes (4):
    - Descriptive: "all the main effects were significant (p values < .0001; see Table 3)" (Study 1, Results) vs "C2 proximity0.7920.2712.92.00350.62" (Table 3, Study 1 row 'C2 proximity'). The Study 1 text states that every main-effect p-value is below .0001, whereas Table 3 reports C2 proximity at .0035, which remains significant. → Conclusion holds
    - Descriptive: "leaving 22,808 words (drop rate = 9.8%)" (Study 3, Material and Design) vs "leaving 15,145 compounds (drop rate = 32.2%) in the analysis" (Study 3, Material and Design (second exclusion step)). The reduction from 22,808 to 15,145 is approximately 33.6% of the preceding sample, rather than the reported 32.2%. → Conclusion holds
    - Descriptive: "leaving 23,807 novel compounds (drop rate = 5.1%)" (Study 2, Material and Design) vs "leaving 22,553 novel compounds (drop rate = 3.8%) in the analysis" (Study 2, Material and Design (second exclusion step)). The reduction from 23,807 to 22,553 is approximately 5.3% of the preceding sample, rather than the reported 3.8%. → Conclusion holds
    - Descriptive: "We randomly selected 1,500 novel compounds from Tse et al. (2017)" (Study 1, Material and Design) vs "1 (n = 1,478)" (Table 2, Study 1 row). Table 2 analyzes 22 fewer items than the 1,500 selected in Study 1, without a stated explanation for the reduction. → Effect on result unclear
- **Reporting and transparency:** Clean. Study 1 ethics approval, availability of data and scripts, and non-preregistration are explicitly declared. The apparent frequency-101 threshold is a footnote-formatting artifact, not an established reporting conflict.
  - Ethics approval: declared · Preregistration: not stated · Data availability: declared · Code availability: declared
- **Reference integrity:** Noted
  - 76 references in manuscript
  - 64 of 72 checkable references found in an index
  - 4 are books, websites or datasets — counted, but not index-checkable
  - Could not be verified automatically — queued for manual review:
    - Arcodia GF Montermini F Boyé G Hathout N Chinese: A language of compound words Selected Proceedings of the 5th Décembrettes: Morphology in Toulouse 2007 vol. 1617 Cascadilla Press 79 90 Arcodia, G. F. (2007). Chinese: A language of compound words. In F. Montermini, G. Boyé, & N. Hathout (Eds.), Selected Proceedings of the 5th Décembrettes: Morphology in Toulouse (vol. 1617, pp. 79–90). Cascadilla Press. https://www.lingref.com/cpp/decemb/5/abstract1617.html (Crossref bibliographic top candidate ratio 0.43 — no plausible match)
    - Ceccagno, A. (2016). Chinese neologisms: Word-formation strategies in Chinese. In S. Chan, F. Li Wing Yee, & J. W. Minett (Eds.), The Routledge encyclopedia of the Chinese language (pp. 227–241). Routledge. https://hdl.handle.net/11585/572812 (Crossref bibliographic top candidate ratio 0.49 — no plausible match)
    - Dinu G Pham N. T. Baroni M DISSECT—DIStributional SEmantics Composition Toolkit Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics: System Demonstrations 2013 Association for Computational Linguistics 31 36 Dinu, G., Pham, N. T., & Baroni, M. (2013). DISSECT—DIStributional semantics composition toolkit. Proceedings of the 51st Annual Meeting of the Association for Computational Linguistics: System Demonstrations (pp. 31–36). Association for Computational Ling (Ambiguous: title ratio 0.50 via crossref_bibliographic sits between the divergence (0.30) and match (0.70) thresholds; too close to call either way, so it is recorded as unresolved rather than judged.)
    - Guevara ER A regression model of adjective-noun compositionality in distributional semantics Proceedings of the 2010 Workshop on Geometrical Models of Natural Language Semantics 2010 Association for Computational Linguistics 33 37 Guevara, E. R. (2010). A regression model of adjective-noun compositionality in distributional semantics. Proceedings of the 2010 Workshop on Geometrical Models of Natural Language Semantics (pp. 33–37). Association for Computational Linguistics. https://aclanthology.o (Ambiguous: title ratio 0.52 via crossref_bibliographic sits between the divergence (0.30) and match (0.70) thresholds; too close to call either way, so it is recorded as unresolved rather than judged.)
    - Li, P., & Zhou, G. (2012). Employing morphological structures and sememes for Chinese event extraction. Proceedings of COLING 2012 (pp. 1619–1634). The COLING 2012 Organizing Committee. https://aclanthology.org/C12-1099 (Crossref bibliographic top candidate ratio 0.26 — no plausible match)
    - 64. Spalding TL Gagné CL Mullaly A Ji H Oslen S Relation-based interpretation of noun-noun phrases: A new theoretical approach New impulses in word-formation (Linguistische Berichte Sonderheft 17) 2010 Helmut Buske Verlag 283 315 Spalding, T. L., Gagné, C. L., Mullaly, A., & Ji, H. (2010). Relation-based interpretation of noun-noun phrases: A new theoretical approach. In S. Oslen (Ed.), New impulses in word-formation (Linguistische Berichte Sonderheft 17) (pp. 283–315). Helmut Buske Verlag. (Ambiguous: title ratio 0.53 via crossref_bibliographic sits between the divergence (0.30) and match (0.70) thresholds; too close to call either way, so it is recorded as unresolved rather than judged.)
    - 71. Turian J Ratinov L Bengio Y Word representations: A simple and general method for semi-supervised learning Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics 2010 Association for Computational Linguistics 384 394 Turian, J., Ratinov, L., & Bengio, Y. (2010). Word representations: A simple and general method for semi-supervised learning. Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics (pp. 384–394). Association for C (Ambiguous: title ratio 0.60 via crossref_bibliographic sits between the divergence (0.30) and match (0.70) thresholds; too close to call either way, so it is recorded as unresolved rather than judged.)
    - 72. Wu JT Liu IM Liu IM Chen H-C Chen MJ A data base system about the psychological features of Chinese characters and words Cognitive aspects of the Chinese language 1988 Asian Research Service 171 186 Wu, J. T., & Liu, I. M. (1988). A data base system about the psychological features of Chinese characters and words. In I. M. Liu, H.-C. Chen, & M. J. Chen (Eds.), Cognitive aspects of the Chinese language (pp. 171–186). Asian Research Service. (Ambiguous: title ratio 0.58 via crossref_bibliographic sits between the divergence (0.30) and match (0.70) thresholds; too close to call either way, so it is recorded as unresolved rather than judged.)
  - References not verified automatically are resolved manually within 24 hours.

## Post-publication record

- **Status:** Clean
- **Summary:** No retraction or expression of concern found in Crossref (includes Retraction Watch).
- **Last checked:** 2026-11-09
