Introduction
In the early stages of language acquisition, typically between the ages of 1.5 and 3, children often produce "telegraphic" speech—sentences that lack functional elements or required arguments. A particularly striking phenomenon in English-speaking children is the frequent omission of subjects (e.g., "Want cookie" instead of "I want cookie"). While adults in languages like English are required to include subjects, toddlers omit them approximately 55% of the time, compared to a much lower 9% omission rate for objects.
This discrepancy has fueled a long-standing debate in linguistics: is the "null subject" phenomenon a result of an immature grammar (the grammatical parameter hypothesis), or is it a consequence of limited cognitive resources (the processing overload hypothesis)? This paper introduces FELICITY, a computational model designed to test whether simple resource constraints can explain these asymmetric omission patterns.
The FELICITY Model
FELICITY is a generative model of sentence production based on a modular architecture that moves from semantics to syntax and finally to phonetics. The core of the model is a processing overload metric that assigns a computational "cost" to various stages of sentence construction:
- Lexical Lookup: The cost of retrieving a word from the mental lexicon.
- Structural Building: The cost of merging phrases into a syntactic tree (based on X-bar theory).
- Phonological Retrieval: The cost of accessing the phonetic representation of the words.
Crucially, the model builds phrases from the bottom up. In this hierarchy, the verb first merges with its complement (the object) to form a Verb Phrase (VP) before it merges with the specifier (the subject). If the cumulative cost of these operations exceeds the model's predefined "processing limit," the model halts and outputs only the fragments currently held in its buffer.
Methodology
To validate the model, its performance was compared against longitudinal transcripts of three children—Adam, Eve, and Sarah—from the CHILDES database. The model's processing limit was calibrated to match the Mean Length of Utterance (MLU) observed in these real-world datasets, allowing for a direct comparison between simulated and actual developmental data.
Key Findings
The Impact of Construction Order
The most significant insight from the FELICITY simulations is that the higher rate of null subjects in English emerges naturally from the order of operations. Because the model constructs the object phrase before the subject, the processing budget is often exhausted by the time the model attempts to integrate the subject. This structural priority explains why objects are preserved more frequently than subjects in early speech.
Structural Sensitivity
The model successfully predicted several nuanced patterns observed in child language:
- Clause Type: Null subjects are more frequent in main (matrix) clauses than in embedded ones.
- Linear Position: Omissions are more common at the beginning of a sentence than following a "wh-" word (e.g., who or what).
Developmental Trajectory
As the simulated "processing limit" increases—representing the cognitive maturation of a child—the percentage of null subjects decreases. Eventually, the rate of subject omission drops to match the low level of object omission, mirroring the developmental curves seen in the CHILDES data.
Cross-Linguistic Implications
While processing limits effectively explain English data, they do not immediately account for languages like Italian or Chinese, where null subjects are either grammatically licensed or discourse-driven. The paper suggests a hybrid account: children likely possess the correct grammatical settings for their specific language, but their limited processing capacity forces them to prune sentences. By toggling specific grammatical parameters (such as "Discourse Oriented" for Chinese) while maintaining the same processing constraints, FELICITY can potentially simulate the unique omission patterns of diverse languages.
Limitations and Future Directions
As a work in progress, FELICITY currently focuses on basic sentence structures. Future iterations of the model will need to address more complex grammatical features, including:
- Tense and Agreement: How children learn to mark verbs for time and person.
- Subject-Auxiliary Inversion: The transition from "I can go?" to "Can I go?"
- Weight Refinement: Further empirical work is required to fine-tune the specific costs assigned to lexical vs. structural operations to ensure the model's predictive power extends beyond the current datasets.
Conclusion
The FELICITY model demonstrates that many characteristics of early child language, specifically the lopsided omission of subjects, can be explained by the interaction between a standard grammatical architecture and finite processing resources. This suggests that toddlers may know more about grammar than their "telegraphic" speech initially suggests.
Summary updated