Network statistics on early English Syntax: Structural criteria

• The paper presents an experimental design for a more extensive study of syntax networks through language acquisition.Source: §1 · p. 1
• The planned broader work concerns the ontogeny of syntax networks during language acquisition.Source: §1 · p. 1
• The paper functions as a design-stage contribution rather than the complete extensive study.Source: §1 · p. 1

1 Introduction

• The paper uses network methods to study syntax acquisition while framing networks as representations of observed syntactic relations rather than direct access to syntactic machinery.Its approach is related to Dependency Grammar and Word-Grammar but uses its own criteria and a constituency-based graph projection.Source: §1 · p. 3, §1.1 · p. 3

1.1 Different abstractions, different questions: Syntax and Statistical Physics

• Network analysis examines global language dynamics and organization rather than the detailed rules generating individual sentences.The paper presents this as a different abstraction from syntax, intended to complement questions about grammar.Source: §1.1 · pp. 3–4

1.2 Aims

• The study develops descriptive criteria for identifying structure in early child grammars and building syntactic networks from children’s utterances.These criteria address the absence of adult-style concepts such as grammaticality or convergence in early child language.Source: §1.2 · p. 4

2.1 From syntax to networks: what we win and what is lost

• The network perspective shifts attention from sentence-generation rules to global profiles of syntactic relations and system-wide dynamics.The paper infers global performance patterns from naturalistic data while assuming relevant competence abilities.Source: §2.1 · p. 5

2.2 Syntactic Networks

• Lexical items form network nodes, while syntactic rules are mapped to directed links through an additional structural projection.The node mapping is straightforward, but defining links requires explicit criteria.Source: §2.2 · p. 5, §2.2.1 · p. 6

2.2.1 From syntactic relations to links

• Syntax generates hierarchical and long-range relations that a linear network representation cannot capture explicitly, so the network analysis remains an approximation.The approximation is deliberate because the study targets global system patterns rather than every local syntactic property.Source: §2.2.1 · p. 5
• The annotation begins by identifying unlabeled constituent structure, distinguishing complements and heads, and detecting finite verbs.The procedure then projects constituent structure into lexical dependencies.Source: §2.2.1 · p. 6, Figure 1 · p. 7
• Dependencies run from complements to heads, including links between internal and external phrase heads to recover merging order.This projection is reversible in the figure’s workflow, allowing the tree to be rebuilt from dependency relations.Source: §2.2.1 · p. 6, Figure 1 · p. 7
• Heads are selected as the semantically most relevant items, with finite verbs designated as sentence heads.These rules provide the main lexical criteria for converting constituent structures into graph links.Source: §2.2.1 · p. 6
• The criteria aim to remain relatively nonrestrictive, allowing arbitrarily long sentences while preserving finite observed datasets and compatibility with theoretical infinite generativity.The authors present this as a way to accommodate children’s developing syntax without committing to one linguistic school.Source: §2.2.1 · p. 6
• The approach treats networks as approximations for global evolutionary features, not replacements for decision rules, constituent hierarchy, or movement.This distinguishes the paper’s use of networks from views that treat the network abstraction as ontological.Source: §2.2.1 · p. 7

3 Data

• The dataset comprises the first 11 Peter Corpus sections, spanning ages 1 year 9 months to 2 years 4 months.The intervals are approximately two or three weeks, and the corpus is intended to capture global properties despite the system’s small size.Source: §3 · p. 8
• The study is naturalistic, using spontaneous child free-speech recordings from the CHILDES Database rather than responses to controlled experimental situations.The selected corpus is from Bloom’s Peter data and includes conversations with adults.Source: §3 · p. 8
• The selected period begins after the lexical spurt and ends when children’s syntactic structures become complex enough for comparison with adult structures.The authors note that more abrupt changes may occur in later acquisition stages, which are outside the analysis.Source: §3 · p. 8
• Child productions are cleaned by discarding adult imitations and non-structured utterances, using explicit criteria reported in the following section.The utterances are interpreted in their discourse context to support this cleaning process.Source: §3 · pp. 8–9
• The analysis does not treat networks as substitutes for syntactic decision rules because hierarchy and movement are not properly represented by graph-theoretic abstraction.This limitation constrains what can be inferred from the resulting networks.Source: §3 · p. 8

• Structure is identified when a child production contains at least one lexico-thematic relation between elements.More complex relations involving functional words appear later, making syntactic structure easier to identify.Source: §4 · p. 9
• Words are used as fundamental units, but this choice is acknowledged as intuitive and rather arbitrary.A morpheme-based analysis might reveal more similarities across language-acquisition processes.Source: §4 · p. 9

• Onomatopoeias and unintelligible transcriptions are discarded to keep the corpus data transparent.Onomatopoeias accompanying lexical items are treated as nonexistent in analysis.Source: §4.1 · p. 10
• The example “I want ta write the choo choo train” is reduced to “I want ta write the train” by omitting the onomatopoeia.Source: §4.1 · p. 10
• Productions with no identified structure are represented as isolated lexical nodes without links.This preserves lexical items while avoiding unsupported syntactic relations.Source: §4.1 · p. 10
• Imitations are removed when discourse context suggests that unusually complex utterances may reproduce untranscribed adult speech.The decision reflects uncertainty about whether the utterance represents the child's own structured production.Source: §4.1 · p. 10

4.2 Accepted Productions

• Structured substrings are accepted separately when an utterance lacks a global structure but contains internally structured strings.Source: §4.2 · pp. 10–11

4.2.1 Phrases and missing arguments

• In the pre-functional stage, thematic relations between verbs and arguments define syntax despite missing determiners, agreement, subjects, and other complex structure.Subject omission is associated with imperatives or unresolved English subject-marking parameters; productions such as “open the box” without the determiner remain analyzable, whereas *open the lacks the required argument relation.Source: §4.2.1 · p. 11
• Phrasal verbs are treated as structured utterances because their intrinsic complexity and adult discontinuous realizations support analyzing the verb–functional-word relation.Source: §4.2.1 · p. 11

4.2.2 To be verb

• Missing to be does not eliminate structure when a semantically meaningful predication remains, so copular omissions are treated like missing functional words.Conversely, forms such as “I’m write too” may represent emerging inflection or an irrelevant auxiliary and are treated as single finite sentences when appropriate.Source: §4.2.2 · pp. 11–12

4.2.3 Infra-specification and semantic extension of lexical items

• Lexical substitutions and pronoun or possessive errors may remain structured, but highly ambiguous semantic extensions are rejected unless context supports a syntactic relation.Wrong pronouns can be associated with a weak case system, and immediate self-correction supports treating them as structured rather than unstructured.Source: §4.2.3 · pp. 13–14

4.2.4 First functional particles

• Functional particles are analyzed occurrence by occurrence because schwa and similar forms can develop from phonological fillers into syntactically meaningful elements.The interpretation is cautious because the schwa’s phonological versus functional status remains unsettled, while Veneziano & Sinclair describe intermediate grammatical-morpheme development.Source: §4.2.4 · pp. 14–15
• The particle ta is treated as an intermediate form functioning as a preposition, linking the emergence of functional particles to the development of English syntax.Its rare occurrence at the expected preposition-to position appears at the beginning of the functional stage, suggesting transition from phonological to functional use.Source: §4.2.4 · p. 16

4.2.5 Duplication of functional words

• Duplicated prepositions are interpreted as evidence of prepositional-phrase structure, but duplicated determiners are excluded from structural analysis.In “Look at in there,” the second preposition is independent; in “One that screwdriver,” one is excluded and only [that screwdriver]SD is retained.Source: §4.2.5 · pp. 16–17

4.2.7 Negation Structures

• Examples such as “Peter 7 No ride a bike” and “Peter 8 in the bag no” are treated as unsuitable for assigning no to a structure.The authors state that no should not be considered a member of any structure in such situations.Source: §4.2.7 · p. 18
• The analysis excludes productions when no syntactic relation between the negation operator and another lexical item can be established reliably.A parametric hypothesis might account for the operator’s structural position, but the authors reject these productions for reliability.Source: §4.2.7 · p. 18
• “Peter 9 No in this box” suggests that the child knows English negation ordering, but the authors still avoid identifying no as part of a larger sentence structure.They regard this identification as risky, as in the preceding cases.Source: §4.2.7 · p. 19

• The criteria identify and characterize syntactic structures in early child language.Source: §5 · p. 19
• Identified structures are projected into word-word dependencies to annotate the corpus.Source: §5 · p. 19
• Manual annotation supports accuracy and management of ambiguous situations, using the Dependency Grammar Annotator.The annotator works with XML files whose internal structure is described through the paper's example.Source: §5 · p. 19

• The measure is presented as close to classical MLU, which commonly measures utterance length in words or morphemes.Source: §6 · p. 19
• Average structure size ⟨S⟩ is computed from the lexical-item counts of syntactically unrelated structures within utterances.For example, sizes 3 and 1 yield an utterance average of 2; single words count as size-1 structures.Source: §6 · p. 19
• The measure is averaged over all utterances to assess whether production size contains information about grammatical complexity.Its evolution may also be related to working-memory limitations.Source: §6 · p. 19

7 Building the Network

• The network contains lexical items as nodes and syntactic relations between words as directed links, with each word and link represented only once.The resulting graph may include isolated nodes and is represented by an adjacency matrix of size n^2 for n distinct words.Source: §7 · p. 20

7.1 Measures

• Child syntactic graphs display a clear Giant Connected Component from the beginning, containing most connected words active in productions.Its size or relative size provides a statistical indicator of network connectivity.Source: §7.1 · p. 20
• The analysis computes degree and average degree over the largest component, where degree counts a word’s syntactic relations with neighboring words.Source: §7.1 · p. 21
• Average path length measures shortest-path accessibility, while clustering measures the likelihood that neighboring words are connected and form triangles.The average clustering coefficient is computed over node neighborhoods in the corpus graph.Source: §7.1 · p. 21
• The framework also compares networks with random and small-world expectations and quantifies hub connectivity using Pearson’s assortativity coefficient.Negative coefficients indicate disassortativity, whereas positive coefficients indicate assortativity.Source: §7.1 · pp. 21–22

8 Example

• The example begins with a Corpus Peter conversation and selects Peter’s productions for structural analysis and network construction.Source: §8 · p. 22, §8.1 · pp. 22–23

8.2 Selected Productions and Analysis

• A PERL routine extracts child productions from the corpus, producing strings such as “xxx telephone go right there” and “put in there.”Source: §8.2 · pp. 23–24

8.2.1 XML Format to be read by DGAanotator

• The extracted strings are converted into XML so the DGA Annotator can process them.The paper illustrates this format with the production “put in there.”Source: §8.2.1 · p. 24, Figure 2 · p. 25

8.2.2 Selection of valid strings, annotation and computation of S

• The analysis rejects “xxx” and “0” as lexical items before applying the annotation criteria and computing average structure size.Source: §8.2.2 · pp. 25–26
• After annotation, the corpus is sampled to generate the word set and graph, with stronger languages recommended for graph parameters and mathematical computations.Source: §8.2.2 · pp. 25–26