PaperlayerView exact revision on arXiv

Source-linked AI summary

From Parallels to Divergence: Language Models' Learning of Action Nominal Constructions

Iskar Deng, Shane Steinert-Threlkeld

arXiv:2610.22468v1cs.CL

TL;DR

The paper asks how language models’ preferences across grammatical configurations compare with natural-language typology, extending evidence beyond isolated phenomena. It uses shared semantic input and computational grammars to construct 96 synthetic languages and evaluate models on grammatical minimal pairs. Models align relatively well with typology on basic clause configurations but diverge on complement systems and ANC strategies, a partial match that may reflect pressures not captured by controlled experiments.

  • Problem

    Comparing model learning across natural languages makes grammatical effects difficult to isolate because languages also differ in corpus source, training-set size, and lexical distributions.

    Source: §2.1 · p. 2

  • Method

    The paper combines shared semantic input from natural corpora with computational grammars to create 96 target languages and evaluates model learning using targeted minimal pairs.

    Source: §6 · p. 5, §8 · p. 8

  • Results

    Models reproduce typological tendencies relatively well for basic clause configurations, but diverge on complement systems and ANC strategies, also learning SVO + Deranking and Deranking + NOMN, rare or unattested in the natural-language sample.

    Source: §8 · p. 8

  • Takeaways & Limitations

    The basic-clause parallels suggest general learning pressures may help shape typological tendencies, while complex configurations may also reflect constructional functions and language-use pressures not fully captured by controlled experiments.

    Source: §8 · p. 8

    The shared semantic input comes from an English corpus, and the study does not test other source languages or control for construction distributions.

    Source: Limitations · p. 9

Abstract

from arXiv · show

Action nominal constructions (ANCs) express events in nominal form, with their argument marking following patterns found in clauses or ordinary noun phrases. In this paper, we investigate how different grammatical configurations affect language models' learning of ANCs and related constructions, and how these learning preferences compare with typological distributions in natural languages. We use a grammar-engineered method for synthetic corpus construction that combines shared semantic input from natural corpora with computational grammars. We construct 96 target languages varying in clause word order, noun-phrase word order, alignment, complement system, and ANC strategy. For each language, we train GPT-2-small models from scratch and evaluate them using targeted minimal pairs covering 34 grammatical phenomena. The results show typological parallels in basic clause configurations but divergences in complement systems and ANC strategies, with models also learning combinations rare or unattested in natural languages. This partial alignment suggests that natural language distributions may reflect both general learning pressures and functional demands not fully captured by model learning experiments.

1 Introduction

The paper extends comparisons of language-model learning preferences beyond individual grammatical phenomena to relationships among action nominal constructions (ANCs) and related constructions. Across 96 synthetic languages, models align with typology on basic clause configurations but diverge on more complex complement and ANC configurations.

1 Introduction

  • Comparing typological alignment and divergence can help distinguish tendencies that may arise from LM learning pressures from those requiring additional explanatory factors.The paper notes that natural-language distributions may reflect factors beyond those captured by model learning experiments.Source: §1 · p. 1
  • ANCs extend prior work on individual phenomena by enabling comparisons across constructions that share some structural properties but differ in others.ANCs express events in nominal form and may pattern with clauses, clausal complements, or noun phrases in argument marking, predicate form, and internal word order.Source: §1 · p. 1
  • The study compares model preferences with natural-language typology using synthetic languages that vary interrelated grammatical configurations.It combines shared semantic input derived from natural corpora with target grammars, constructing 96 languages and evaluating GPT-2-small models on targeted minimal pairs covering 34 phenomena.Source: §1 · pp. 1–2
  • Models reproduce typological tendencies relatively well for basic clause configurations, but only partially for complement systems and ANC strategies.They also learn SVO + Deranking and Deranking + NOMN, combinations rare or unattested in the natural-language sample.Source: §1 · p. 2

2 Background

Controlled synthetic corpora help isolate grammatical effects, but existing construction methods trade off naturalistic input distributions against control over coherent grammatical systems. The paper frames ANCs as constructions with both nominal and clausal properties, whose argument encoding varies across strategies.

2.1 Synthetic Corpus Experiments

  • Synthetic-language experiments control for differences across natural-language corpora, allowing grammatical properties to be examined more directly.Corpus perturbation is limited by structures in its source data, while artificial grammars offer parameter control but often rely on researcher-specified distributions and frequencies.Source: §2.1 · p. 2
  • The proposed grammar-engineered method derives shared semantic representations from natural corpora and realizes them through target grammars.It aims to coordinate control over ANC properties connected to alignment and complement systems while preserving relatively natural predicate–argument distributions.Source: §2.1 · p. 2

2.2 Action Nominal Constructions

  • ANCs express events in nominal form, occur in nominal positions, and may retain verbal argument selection or use nominal encoding patterns.Their semantic functions differ from those of main clauses and clausal complements, even when they occupy similar argument positions.Source: §2.2 · pp. 2–3
  • The four ANC strategies range from clause-like SENT marking to noun-phrase-like NOMN marking, with POSS-ACC and ERG-POSS combining nominal and clausal resources.SENT retains finite-clause marking; POSS-ACC marks S and A as genitive-like while P retains object marking; ERG-POSS groups S and P under genitive-like marking; NOMN uses genitive-like marking for S and A and ordinary NP-dependent marking for P.Source: §2.2 · p. 3
  • ANC strategies describe encoding when arguments are overt, not a requirement that every possible argument appear.An argument may remain unexpressed when recoverable from context or irrelevant to the communicative goal.Source: §2.2 · p. 3

3 Grammatical Configurations

The target grammars cross five typological parameters to define 96 configurations. Together, these settings determine word order, argument marking, and predicate form across clauses, complements, and ANCs.

3 Grammatical Configurations

  • Five parameters—clause word order, NP word order, alignment, complement system, and ANC strategy—yield 96 grammatical configurations.The parameters jointly determine word order, argument marking, and predicate form across constructions.Source: §3 · p. 3
  • Complement systems distinguish Balancing predicates with finite-clause-like form from Deranking predicates with reduced, nonfinite form.In the experiment, Balancing predicates take -s and Deranking predicates take -ing.Source: §3 · p. 4
  • ANCs take -ing predicates, while ANC strategy determines internal argument marking and word order depends jointly on clause order, NP order, and ANC strategy.The ANC as a whole receives case or genitive marking according to its external role; argument omission is permitted inside ANCs but not in main or complement clauses.Source: §3 · p. 4
  • The example target configuration combines SVO, GN, NOM-ACC, Balancing, and NOMN.These settings specify clause and NP order, alignment, complement type, and ANC strategy in one configuration.Source: §3 · p. 4

4 Corpus Generation

The corpus-generation pipeline uses English predicate–argument content to create shared semantic representations, then realizes those meanings through 96 target grammars. Grammar engineering and a shared lexicon support parallel corpora with varied grammatical settings.

4 Corpus Generation

  • The pipeline maps natural English predicate–argument content into shared semantic representations and realizes the same meanings through 96 target grammars.The source is the 100M-word BabyLM 2026 Strict corpus; the resulting corpora preserve some source lexical distributions and word-form regularities.Source: §4 · p. 4
  • Grammar Matrix choices files encode each target language’s settings, which are used to generate HPSG grammars compiled with ACE.A shared lexicon is inserted into both the pseudo-English grammar and all target grammars.Source: §4 · p. 4
  • The pseudo-English corpus is parsed into MRS representations of predicates and argument relations, which target grammars realize according to their morphosyntactic settings.When an input yields multiple candidates, output selection can follow rules such as target ANC word order.Source: §4 · p. 5

5 Model Training

The models were trained from scratch on generated corpora for 96 target languages, using three seeded runs per language and a fixed 70,000-step schedule.

5 Model Training

  • Each of the 96 target languages had three GPT-2-small models trained from scratch on approximately 4.1 million generated sentences.The runs differed in parameter initialization and data shuffling, while sharing the same training configuration.Source: §5 · p. 5
  • All models were trained for 70,000 steps, and evaluation used each model’s final checkpoint.The schedule corresponds to approximately 2.59–2.96 corpus-equivalent epochs because tokenized corpus sizes vary across languages.Source: §5 · p. 5
  • The three runs per language used different random seeds.Their differences were limited to parameter initialization and data shuffling.Source: §5 · p. 5

6 Evaluation

The evaluation uses targeted minimal pairs to test grammatical regularities across 34 phenomena, then compares overall and ANC-focused accuracy across target languages. Models learned many target-grammar phenomena, with a mean accuracy of 88.4%.

6 Evaluation

  • The evaluation tests 34 phenomena in six groups, covering basic clauses, noun phrases, complements, and ANC structure and marking.The ANC-focused groups also test argument omission, including whether internal arguments can be omitted in ANCs but not in main or complement clauses.Source: §6 · p. 5
  • For each phenomenon, templates and perturbation rules generate GOOD/BAD minimal pairs, with 100 pairs retained per phenomenon and target language.English source sentences are generated first, transformed into target-language sentences, and then perturbed; pairs are balanced across templates and rules.Source: §6 · p. 5
  • Accuracy is the proportion of pairs where the model assigns the GOOD sentence a higher length-normalized log probability than the BAD sentence.Mean accuracy is calculated across all 34 phenomena and separately across the 15 ANC-related phenomena.Source: §6 · pp. 5–6
  • 88.4% mean accuracy across all 34 phenomena, 96 configurations, and three random seeds shows that models learned many target-grammar phenomena.Basic clause word order and finite verb form were near-perfect, while valency and some argument-marking phenomena were harder; Table 4 summarizes targeted minimal-pair results.Source: §6 · p. 6, Table 4 · p. 6
  • Overtly marked roles were generally easier to learn than zero-marked roles, while a local possessor–head sequence could attract models despite lacking a complete grammatical analysis.The examples contrast ERG-ABS performance on Transitive A marking with NOM-ACC performance on Transitive P marking, and suggest sensitivity to surface heuristics.Source: §6 · p. 6

7 Discussion

The discussion compares model preferences with natural-language counts across clause alignment, complement systems, and ANC strategies. Alignment is strongest for basic clause configurations, while complement and ANC patterns show partial correspondence.

7 Discussion

  • The discussion compares natural-language counts with model accuracies averaged over the remaining grammatical parameters.It focuses on three parameter interactions.Source: §7 · p. 6

7.1 Clause Word Order and Alignment

  • SVO and SOV favor NOM-ACC, whereas VOS favors ERG-ABS; the strongest alignment difference occurs under SVO, where frame overlap is also greatest.The learned preferences match natural-language frequencies in the sample, although the VOS sample is small.Source: §7.1 · pp. 6–7, Table 5 · p. 7

7.2 Word Order and Complement System

  • Balancing yields slightly higher overall accuracy under SVO, while Deranking performs better under SOV; models also learn the rare SVO + Deranking combination relatively well.The proposed explanation is that SVO word order and Deranking’s nonfinite forms may jointly cue argument roles and clause boundaries.Source: §7.2 · p. 7

7.3 Complement System and ANC Strategy

  • ANC strategy preferences match typology only partly: SENT performs better with Deranking, but NOMN also performs well with Deranking despite being unattested in the sample.Shared semantic input may miss functional pressures associated with differing ANC uses, including content complements and process or manner readings.Source: §7.3 · pp. 7–8, Table 7 · p. 8

8 Conclusion

The paper introduces grammar-engineered synthetic corpora that combine shared semantic input with computational grammars. Models’ preferences align with typology in basic clause configurations but diverge for complement systems and ANC strategies.

8 Conclusion

  • The grammar-engineered method combines shared semantic input from natural corpora with computational grammars.It retains some lexical and semantic properties of the source corpus while coordinating control over interrelated grammatical properties.Source: §8 · p. 8
  • Across 96 target languages, the experiments compare model learning of ANCs and related constructions with natural-language typology.Source: §8 · p. 8
  • Model preferences parallel typology in basic clause configurations but diverge for complement systems and ANC strategies, including SVO + Deranking and Deranking + NOMN.These combinations are rare or unattested in the natural-language sample.Source: §8 · p. 8
  • Constructional functions and communicative and production demands may also shape distributions that controlled learning experiments do not fully capture.The paper proposes that these pressures may help explain why complex complement and ANC configurations diverge from model preferences.Source: §8 · p. 8

Limitations

The experiments have design and scope limitations, while the appendix describes four ANC strategies that differ in how they mark participants. The strategy range extends from clause-like marking to nominal-style marking.

Limitations

  • English-derived semantic input preserves source-language construction distinctions and co-occurrence patterns, which may favor configurations closer to English; this potential bias was not tested with other sources or controlled distributions.The authors specifically note NOMN as a potentially favored configuration.Source: Limitations · p. 9
  • Crossing typological parameters can change multiple cues across constructions, so the design compares configurations and interactions but does not isolate each formal cue’s contribution.Source: Limitations · p. 9
  • The experiments do not test semantic or contextual effects on ANC choice, coexistence of multiple ANC strategies in one language, or variation across model architectures, scales, training-data sizes, and tokenizers.Source: Limitations · p. 9

A ANC Strategies

  • Argument encoding describes how an ANC marks the relation between an action nominal and its participants, through dependent marking, head marking, or word order.The study focuses primarily on dependent marking.Source: A ANC Strategies · p. 11

A.1 Sentential ANCs

  • SENT ANCs retain finite-clause dependent marking internally, although they remain distinct from finite clauses in external distribution and some syntactic properties.Korean illustrates retained case marking alongside nominal external distribution.Source: §A.1 · p. 11

A.2 Possessive-Accusative ANCs

  • POSS-ACC ANCs recode S and A as genitive-like dependents while P retains finite-clause object marking.The Selkup example contrasts genitive A with accusative P.Source: §A.2 · p. 11

A.3 Ergative-Possessive ANCs

  • ERG-POSS ANCs mark S and P genitive-like and mark A separately, creating an ergative-like alignment unlike the accusative-like alignment of POSS-ACC.Russian illustrates instrumental A and genitive P in an ANC.Source: §A.3 · p. 12

A.4 Nominal ANCs

  • NOMN ANCs align most closely with ordinary noun phrases: S and A are genitive-like, while P uses a dependent type available in noun phrases.NOMN includes Double-Possessive and Possessive-Adnominal subtypes; when only one argument is overt, their surface distinction is limited.Source: §A.4 · p. 12

B Grammatical Specifications

The target grammars vary five typological parameters that jointly determine word order, argument marking, and predicate form. The corpus pipeline applies these settings to shared lexical and semantic material to generate target-language corpora.

B Grammatical Specifications

  • Five parameters—clause word order, noun-phrase word order, alignment, complement system, and ANC strategy—jointly determine target-language word order, argument marking, and predicate form.Their combinations yield 96 configurations.Source: B Grammatical Specifications · pp. 12–13

B.1 Finite Clauses

  • Finite clauses vary across SOV, SVO, and VOS orders and NOM-ACC or ERG-ABS alignment, with the distinct argument overtly marked and the paired arguments zero-marked.The same overt marker, -ca, is used in both alignment systems.Source: §B.1 · p. 13, Table 9 · p. 13

B.2 Noun Phrases

  • Noun phrases vary between genitive-before-head (GN) and head-before-genitive (NG), with each nominal head restricted to one direct genitive dependent.A genitive dependent may itself contain another genitive relation.Source: §B.2 · p. 13

B.3 Complement System

  • The complement-system setting distinguishes balancing complements with finite -s verbs from deranking complements with nonfinite -ing verbs; embedded argument marking follows the target language’s finite-clause alignment.Overt complementizers are omitted by default, so complementizer presence is not manipulated.Source: §B.3 · p. 14, §B.4 · p. 14, Table 10 · p. 14

B.5 ANC Case Marking

  • Four ANC strategies determine argument marking: SENT preserves finite-clause marking, POSS-ACC assigns S/A genitive-like marking, ERG-POSS assigns S/P genitive-like marking, and NOMN assigns S/A genitive-like and P oblique-like marking.Each experimental target language receives exactly one ANC strategy.Source: §B.5 · p. 14, Table 11 · p. 14

B.6 ANC Word Order

  • ANC-internal order combines NP order for possessive-like arguments with finite-clause head–complement direction for non-possessive arguments; when A and P fall on the same side, P is closer to the nominalized predicate.The strategy determines which arguments occupy the possessive-like versus non-possessive positions.Source: §B.6 · pp. 14–15

C.1 English Parsing and Pseudo-English

  • The English-processing pipeline extracts selected predicate–argument structures and nominal modifiers, then builds a shared lexicon and pseudo-English representations for grammar-based generation.It retains intransitive, transitive, clausal-complement, and nominal-predicate copular constructions; only intransitive and transitive verbs are nominalization sources.Source: §C.1 · pp. 15–16

C.2.1 Choices Files and Grammars

  • Automatically generated Grammar Matrix choices files encode each parameter combination, receive the shared lexicon, and are compiled into grammars used to parse pseudo-English and generate target corpora.The pseudo-English grammar uses an overt complementizer and a dedicated -nmz marker, whereas target grammars use -ing for nominalization.Source: §C.2.1 · p. 16

C.2.2 Grammar Patching

The pipeline patches ANC grammars to realize split argument orders that the library cannot directly encode, then uses MRS as shared semantic input for generation across target grammars. Grammar-engineering tools support this parsing and generation workflow.

C.2.2 Grammar Patching

  • The ANC library’s single word-order profile cannot represent some split ANC orders, so grammar patching routes possessive-like and non-possessive arguments through separate ordering paths.The patch permits the split patterns defined by the experiment.Source: §C.2.2 · pp. 16–17
  • The patch can overgenerate alternative A/P orders, so postgeneration selection retains the candidate matching the configuration’s effective ANC transitive order.Some cases of accidental attachment ambiguity instead receive a reproducible choice using a fixed random seed.Source: §C.2.2 · p. 17, §C.3 · p. 18

C.3 MRS Parsing and Surface Selection

  • The pseudo-English grammar parses shared pseudo-English into an MRS bank, which the 96 target grammars realize according to their own morphosyntactic settings.Multiple MRS analyses or surface candidates can arise; ambiguous one-argument nominalizations default to the S-like candidate.Source: §C.3 · pp. 17–18

C.4.1 Head-Driven Phrase Structure Grammar

  • HPSG represents words and phrases as typed feature structures encoding syntactic, morphosyntactic, valence, and semantic information, supporting interactions among grammatical modules.Source: §C.4.1 · p. 18

C.4.2 Minimal Recursion Semantics

  • MRS represents predicate–argument structure and scope constraints, serving as a shared semantic interface so different grammars can realize the same event structure in different surface forms.Source: §C.4.2 · p. 18

C.4.3 The Grammar Matrix

  • The Grammar Matrix turns typological settings specified in a choices file into HPSG grammar source files for compilation, parsing, and generation.Source: §C.4.3 · p. 18

C.4.4 The ANC Library

  • The ANC library derives nominal forms from verbal entries while preserving event semantics and argument relations, then adjusts valence to the selected strategy’s requirements.It allows configurable nominal and verbal properties, argument realization, and separate constraints for transitive and intransitive nominalizations.Source: §C.4.4 · pp. 18–19

C.4.5 DELPH-IN Tools

  • ACE parses strings into MRS and generates strings from MRS using compiled HPSG grammars, while PyDelphin connects these operations to the experimental pipeline.Source: §C.4.5 · p. 19

D.1 Training Corpus

The models train on language-specific corpora generated from a shared English-derived source split. A shared GPT-2 tokenizer segments the synthetic forms, and tokenized corpus sizes vary across languages.

D.1 Training Corpus

  • The 90/5/5 train/dev/test split supplies the training lexicon and generates 96 target-language training corpora of about 4.1 million sentences each; models train only on their corresponding corpus.Source: §D.1 · p. 19

D.2 Tokenization

  • The shared GPT-2 byte-level BPE tokenizer can leave morphological markers identifiable as subword cues, though token boundaries need not match morpheme boundaries; training corpora contain approximately 24.2–27.6 million tokens each.Mean corpus size was approximately 25.4 million tokens, and a fixed training schedule covers different numbers of corpus-equivalent epochs across languages.Source: §D.2 · pp. 19–20

D.3 Training Configuration

This section covers the controlled minimal-pair evaluations and model training setup, including tests of verb forms and order, valency, and intransitive genitive marking. Reported results show ceiling performance on transitive A–V and V–P order and weaker performance on verb valency; the intransitive genitive-marker passage describes test construction but reports no result.

D.3 Training Configuration

  • Each target language used three GPT-2-small models trained from scratch for 70,000 steps, with different random seeds.The approximately 124-million-parameter models received about 2.59–2.96 corpus-equivalent epochs and required about 1.5 hours each on one NVIDIA GeForce RTX 4060 GPU.Source: §D.3 · p. 20

E.1.1 Source Sentence Generation; E.1.2 Valency Phenomena; E.1.3 Case Marking between Clausal Complements and ANCs; 1.7 Transitive A–V Order

  • Evaluation pairs were generated from shared English source sentences, filtered through the semantic pipeline, and selected as 100 GOOD–BAD pairs per phenomenon and target language.Perturbations targeted constituent order, morphology, valency, and case marking; valency pairs drew on BLiMP and MegaAcceptability, while complement/ANC tests exchanged construction-specific markers. The A–V test swaps the A span with the V+P chunk.Source: §E.1.1 · p. 20, §E.1.2 · p. 21, §E.1.3 · p. 21

E.1.4 BAD Sentence Parsability; E.2 Per-Phenomenon Evaluation

  • BAD parse rate supplements minimal-pair accuracy by tracking BAD sentences that remain parsable under an alternative grammatical analysis.The appendix reports construction details and results for 34 phenomena in six evaluation groups; its heatmaps show three-seed mean accuracy alongside grammar-based BAD parse rates.Source: §E.1.4 · pp. 21–22, §E.2 · p. 22

1.1 Intransitive Finite Verb Form; 1.3 Intransitive S–V Order

  • Finite intransitive verb form averaged 99.4% accuracy and S–V order 99.99%, with zero BAD parses in both tests.The reported minima were 95.3% for verb form and 99.3% for order; both tests reached 100% maximum accuracy.Source: Figure 4 · p. 22, §1.1 · p. 22, Figure 6 · p. 23

1.2 Intransitive S Marker

  • Intransitive S marking averaged 87.7% accuracy, with a 51.7% minimum and BAD parse rates of 3.0–4.0%.The weakest configuration, SOV/NG + ERG-ABS, averaged 59.7%; errors were linked to interaction between alignment and word order, including confusion between genitive possessor and transitive argument marking.Source: Figure 5 · p. 22, b. BAD: Doctorge nurseca waits · pp. 22–23

1.4 Transitive Finite Verb Form

  • Transitive finite verb form averaged 99.7% accuracy, with zero BAD parses across grammars.Remaining errors were mostly in SVO grammars with Deranking, consistent with V-ing being more common in deranking than balancing systems.Source: Figure 7 · p. 23, §1.4 · p. 23

1.5 Transitive A Marker

  • Transitive A marking reached 96.6% mean accuracy, but accuracy and BAD parse rates varied asymmetrically by alignment.NOM-ACC averaged 93.2% accuracy and a 7.3% BAD parse rate; ERG-ABS averaged 100% accuracy and an 18.1% BAD parse rate. In SOV/GN + NOM-ACC, the weakest configuration scored 81.9%; some ERG-ABS BAD parses required semantically unnatural reanalysis.Source: §1.5 · pp. 23–24, Figure 8 · p. 24

1.6 Transitive P Marker

  • Transitive P marking reached 97.9% mean accuracy, with NOM-ACC at 100% and ERG-ABS at 95.8%.NOM-ACC had a higher mean BAD parse rate (16.7%) than ERG-ABS (5.7%); SOV/NG + ERG-ABS was weakest at 84.6% accuracy. Genitive reanalyses could make some BAD sentences parsable while models still preferred GOOD sentences.Source: Figure 9 · p. 24, §1.6 · p. 24

1.7 Transitive A–V Order

  • All 96 grammars achieved 100% accuracy and a 0 BAD parse rate on transitive A–V order.The authors report that models learned this order relation robustly.Source: §1.7 · p. 25, Figure 10 · p. 25

1.8 Transitive V–P Order

  • All 96 grammars achieved 100% accuracy on transitive V–P order, with a mean BAD parse rate of 0.08%.The authors report that models learned the V–P order relation robustly.Source: §1.8 · p. 25, Figure 11 · p. 25

1.9 Intransitive V Valency

  • Intransitive V-valency accuracy averaged 70.4% (range 59.7%–81.3%), while the BAD parse rate was 50.0% in every grammar.SVO grammars averaged 76.6% accuracy, versus 68.1% for SOV and 66.4% for VOS; other parameters had small effects, and the authors identify lexical argument structure as difficult for small models trained on small corpora.Source: §1.9 · p. 26

1.10 Transitive V Valency

  • Transitive V-valency accuracy averaged 76.2% (range 67.7%–86.7%), and the BAD parse rate was 93.0% across grammars.SVO grammars averaged 79.4% accuracy, compared with 74.4% for SOV and 74.8% for VOS; the authors note that accuracy reflects both valency sensitivity and preferences between grammatical alternatives because most BAD verbs occurred in transitive frames during training.Source: §1.10 · p. 26, Figure 13 · p. 26

2.1 Intransitive Genitive Marker

  • The intransitive genitive-marker test changed the possessor’s -ge marker by deleting it or replacing it with -ca in the BAD sentence.The GOOD sentence was a finite intransitive clause with a subject NP containing one genitive possessor; no result is provided in these passages.Source: §2.1 · p. 26

b. BAD: Student father laughs

This section examines models’ learning of genitive and complement patterns, as well as ANC verb forms, internal word order, and argument marking. Models learned several patterns well, but performance varied or was weaker for embedded complement argument marking, complement-verb valency, and some ANC argument-marker phenomena, often where surface forms allowed an alternative parse.

b. BAD: Student father laughs; 2.2 Transitive Genitive Marker

  • Across 96 configurations, intransitive and transitive genitive marking averaged 89.3% and 91.6% accuracy, with mean BAD parse rates of 16.0% and 0.1%, respectively.Intransitive marking was weakest mainly under SOV and VOS configurations; transitive marking averaged 95.3% in SVO, versus 87.9% in SOV and 91.5% in VOS.Source: b. BAD: Student father laughs · p. 26, Figure 14 · p. 27, §2.2 · p. 27

2.3 Genitive Possessor–Head Order

  • Genitive possessor–head order averaged 94.9% accuracy; performance was higher in GN than NG grammars (98.3% vs. 91.6%) and in SVO than VOS grammars (100% vs. 88.6%).Some errors reflected plausible reanalyses of the perturbed sequence as part of another argument, rather than a straightforward grammatical mistake.Source: §2.3 · pp. 27–28, Figure 16 · p. 28

3.1 Clausal Complement Verb Form

  • Clausal-complement verb form averaged 98.6% accuracy, with only a slight difference between Balancing (98.7%) and Deranking grammars (98.4%).One Balancing + ERG-ABS + SENT configuration had a 100% BAD parse rate but 96.7% accuracy, as the altered embedded sequence could be reanalyzed as an ANC; models still preferred the GOOD sentence.Source: §3.1 · p. 28, Figure 17 · p. 28

3.2 Clausal Complement S Marker

  • Embedded S marking was difficult, averaging 75.2% accuracy; SVO grammars reached 87.6%, compared with 63.8% for SOV, while ERG-ABS reached 83.6% versus 66.7% for NOM-ACC.Accuracy and BAD parse rate diverged: the weakest SOV/NG + NOM-ACC case had 35.1% accuracy despite a 15.9% BAD parse rate, while some parsable ANC reanalyses retained high accuracy.Source: §3.2 · p. 29, Figure 18 · p. 29

3.3 Clausal Complement A Marker

  • Embedded A marking averaged 83.9% accuracy, with ERG-ABS at 99.7% and NOM-ACC at 68.0%; within NOM-ACC, SOV/GN reached only 39.7%.In NOM-ACC, added -ge could create a local genitive-like surface sequence; ERG-POSS and SVO performed better, at 95.1% and 93.5%, respectively.Source: §3.3 · pp. 29–30, Figure 19 · p. 30

3.4 Clausal Complement P Marker

  • Embedded P marking averaged 97.7% accuracy; NOM-ACC was near-perfect, while ERG-ABS averaged 95.4%, and the weakest configuration scored 79.7%.The error pattern resembled embedded A marking, but interference was weaker because an embedded P is often inanimate and less plausible as a genitive possessor.Source: §3.4 · pp. 30–31, Figure 20 · p. 30

3.5 Clausal Complement–Matrix Verb Order

  • Models learned clausal-complement–matrix verb order perfectly, averaging 100.0% accuracy across configurations, with a mean BAD parse rate of 0.4%.The maximum BAD parse rate was 4.0%.Source: §3.5 · p. 31, Figure 21 · p. 31

3.6 Clausal Complement CV Valency; 4.1 ANC Verb Form

  • Clausal-complement verb valency averaged 74.9% accuracy, while the mean BAD parse rate was 76.4%, indicating many BAD verbs also occurred with clausal complements in training.The accompanying ANC verb-form test contrasts nonfinite V-ing with finite V-s in ANCs serving as matrix A or P; no result for that test appears in the supplied text.Source: §3.6 · pp. 31–32, Figure 22 · p. 32, §4.1 · p. 32

4.1 ANC Verb Form

  • 91.9% mean accuracy shows models generally learned that verbs inside ANCs take nonfinite forms.Balancing + SENT was weakest at 81.6% mean accuracy; finite forms in clausal complements and SENT ANCs made BAD sentences more clause-like.Source: Figure 23 · p. 32, §4.1 · p. 32

4.2 ANC S Marker

  • 82.1% mean accuracy on ANC S marking masks a sharp strategy contrast: SENT reached 40.7%, versus 93.9%–98.2% for the other strategies.Under SENT, zero-marked ANC-internal S matches finite-clause S, so adding -ge can yield a natural-looking genitive NP despite being ungrammatical in the target grammar.Source: §4.2 · pp. 32–33, Figure 24 · p. 33

4.3 ANC A Marker

  • 68.7% mean accuracy makes ANC A marking harder than ANC S marking, with POSS-ACC and NOMN outperforming SENT and ERG-POSS.The strategy means were 80.8%, 85.0%, 54.1%, and 54.8%, respectively; ERG-POSS + ERG-ABS reached 34.9% accuracy despite a 0.0% BAD parse rate, and the reason remained unclear.Source: Figure 25 · p. 33, §4.3 · pp. 33–34

4.4 ANC P Marker

  • 90.1% mean accuracy shows models generally learned ANC P marking, with NOM-ACC outperforming ERG-ABS.Across strategies, ERG-POSS, NOMN, SENT, and POSS-ACC reached 97.6%, 96.2%, 86.0%, and 80.5% mean accuracy, respectively; POSS-ACC + ERG-ABS errors often formed a parsable genitive NP.Source: Figure 26 · p. 34, §4.4 · p. 34

4.5 ANC S–V Order

  • 90.3% mean accuracy shows models generally learned ANC S–V order, but performance depended on word-order profile and strategy.Non-SENT configurations had 71.8% accuracy and a 100% BAD parse rate in SOV/NG, and 75.2% accuracy with a 100% BAD parse rate in VOS/GN; alternative analyses from argument omission also made some BAD sentences natural.Source: §4.5 · pp. 34–35, Figure 27 · p. 35

4.6 ANC A–V Order

  • 93.1% mean accuracy shows models generally learned ANC A–V order across the 96 configurations.Accuracy ranged from 57.7% to 100%, while the mean BAD parse rate was 16.7%; errors clustered where ANC-internal and finite-clause A–V orders conflicted.Source: Figure 28 · p. 35, §4.6 · pp. 35–36

4.7 ANC V–P Order

  • 92.9% mean accuracy shows models generally learned ANC V–P order across the 96 configurations.Accuracy ranged from 51.3% to 100%, with a 13.9% mean BAD parse rate; one weak configuration allowed a genitive NP to be reanalyzed as the S of a clausal complement.Source: Figure 29 · p. 36, §4.7 · p. 36

E.2.5 ANC External

Accuracy on ANC external-marker tests varies by marker and grammatical configuration. Some alignment and word-order combinations perform well, while others remain difficult.

5.1 ANC External S Marker

  • 62.9% mean accuracy for external S marking; SVO/NOM-ACC/non-SENT averaged 89.0%, versus 59.1% for the remaining configurations.The authors say the error source is unclear; a BAD sentence with two -ge markers was still preferred.Source: §5.1 · pp. 36–37, Figure 30 · p. 37

5.2 ANC External A Marker

  • 82.7% mean accuracy for external A marking; ERG-ABS averaged 99.0%, compared with 66.4% for NOM-ACC.Models learned overt external A marking more robustly than zero marking under NOM-ACC.Source: Figure 31 · p. 37, §5.2 · p. 37

5.3 ANC External P Marker

  • 91.4% mean accuracy for external P marking; NOM-ACC reached 98.4%, versus 84.4% for ERG-ABS, with SOV/NG + ERG-ABS at 55.8%.The especially difficult SOV/NG case may reflect ANC-internal arguments separating the ANC head from the matrix verb.Source: §5.3 · pp. 37–38, Figure 32 · p. 38

5.4 ANC External Genitive Marker

  • 75.9% mean accuracy for external genitive marking; ERG-POSS and NOMN averaged 84.1% and 79.2%, while SENT and POSS-ACC averaged 69.1% and 71.3%.Only 0.24% of ANC occurrences are genitive possessors; distributional imbalance is one possible explanation for the pattern.Source: Figure 33 · p. 38, §5.4 · p. 38

E.2.6 ANC Argument Omission

Models varied in their ability to distinguish ANC argument omission from finite-clause patterns. Performance also differed across ANC tests and random seeds, though seed disagreement was generally low.

6.1 ANC Omitted S

  • 98.5% mean accuracy on the bare intransitive ANC test, with a mean BAD parse rate of 11.0%.The models learned that bare intransitive ANCs can function as matrix arguments and that the ANC head takes a nonfinite form.Source: Figure 34 · p. 39, §6.1 · p. 39

6.2 ANC Omitted A

  • The omitted-A test contrasts a transitive ANC retaining only P with a finite transitive clause requiring both A and P.The BAD version replaces the ANC head with a finite transitive predicate and changes the ANC-internal P marker.Source: §6.1 · p. 39

6.1 ANC Omitted S

  • 75.8% mean accuracy on the omitted-S test; Deranking averaged 92.2%, compared with 59.3% for Balancing.BAD parses can recast a transitive structure missing an argument as an intransitive structure with one overt argument; some errors occurred without a BAD parse.Source: Figure 35 · p. 39, §6.1 · p. 39

6.3 ANC Omitted P

  • 83.9% mean accuracy on omitted-P configurations; Deranking averaged 92.5% versus 75.3% for Balancing, and NOM-ACC averaged 93.9% versus 73.8% for ERG-ABS.The BAD parse rate ranged from 0.0% to 100%, with a mean of 16.7%.Source: Figure 36 · p. 40, §6.3 · p. 40

6.4 ANC Omitted A+P

  • 90.5% mean accuracy when both ANC arguments were omitted; Deranking averaged 94.5%, compared with 86.5% for Balancing.Balancing + ERG-ABS was weakest at 78.9%; the paper attributes greater ease to the more salient violation of omitting both finite-predicate arguments.Source: Figure 37 · p. 40, b. BAD: Manager watchs cleans · p. 40

E.3 Variation across Seeds

  • Across 34 × 96 conditions, mean seed disagreement was 0.061, with a median of 0.020 and a 90th percentile of 0.160.The few phenomena with higher disagreement also tended to have lower overall accuracy.Source: §E.3 · p. 41
Loading 2610.22468v1…