TL;DR
This paper introduces a computational procedure designed to infer phonological rules from surface linguistic data. It tackles the challenge of how language learners might deduce grammatical structures, proposing a method that systematically identifies rule contexts and their ordering, even when these are not directly observable.
What Problem Does This Paper Address?
The core problem this paper addresses is a fundamental question in linguistics and cognitive science: how do humans acquire the complex phonological rules of their language? Specifically, it focuses on the "language acquisition device" – the innate linguistic knowledge that allows a learner to go from raw linguistic input (surface forms) to a complete grammar. The paper assumes that "universal grammar" provides a set of principles and constraints that limit the possible grammars a language can have. However, even with these constraints, there can be multiple grammars consistent with the observed data. This leads to the question of how a learner chooses among these possibilities, especially when some aspects of the grammar (like the precise context for a rule) are not explicitly present in the surface data. The paper aims to formalize a "discovery procedure" that can perform this deduction.
Methods
The paper proposes a discovery procedure implemented in Lisp. It operates on a set of "surface paradigms" – essentially, lists of words that share roots and affixes. The procedure is designed to infer a set of ordered phonological rules.
The core of the method involves two main phases:
- Data Analysis (Superficial): This phase identifies potential phonological alternations (where a sound changes) and their contexts. For a given alternation (e.g., 'a' becomes 'b'), it determines two sets of contexts:
- $C_o$: Contexts where the alternation did occur.
-
$C_a$: Contexts where the alternation did not occur (i.e., where 'a' appeared instead of 'b'). The procedure then calculates $S_c$, the set of all possible contexts that simultaneously match all $C_o$ contexts and none of the $C_a$ contexts. This calculation is described as efficient and does not require enumerating all possible contexts.
-
Rule Inference and Ordering: This phase takes the identified contexts and attempts to build an ordered set of rules. It hypothesizes that one of the observed alternations corresponds to the "most surfacy" rule (the last rule applied). The procedure then "undoes" this rule to create intermediate forms, making previously opaque contexts transparent. It iteratively applies this process to identify subsequent rules and their orderings. The procedure systematically explores different rule orderings, testing each one against the data.
A key aspect is how it handles "weakly data determined" choices. When multiple contexts or rule orderings are consistent with the data, the procedure doesn't arbitrarily pick one. Instead, it suggests that an "evaluation metric" (part of universal grammar) would make this choice, potentially favoring more specific or linguistically motivated contexts.
Datasets
Not discussed in the paper. The paper uses hypothetical examples and refers to "some Japanese data" for illustrative purposes regarding rule ordering complexity, but no specific datasets are provided or described.
Key Insights and Ideas
- Formalizing Language Acquisition: The paper moves beyond abstract principles of universal grammar to propose a concrete, computational "discovery procedure" for phonological rules. This makes the process of language acquisition more tangible and testable.
- Context Determination: A significant contribution is the efficient algorithm for determining the precise phonetic or phonological context that triggers a sound change, even when that context isn't fully transparent in the surface data.
- Rule Ordering as a Discovery Problem: The paper argues that rule ordering, a notoriously complex area in phonology, can be approached as part of the discovery process. It shows that even with opaque contexts (where a later rule obscures the environment for an earlier one), a systematic procedure can still infer the correct order.
- Weak Data Determination and Evaluation Metrics: The procedure highlights situations where data alone is insufficient to pick a unique grammar. This reinforces the idea that an innate "evaluation metric" must play a role in language learning, guiding choices based on principles beyond mere empirical adequacy.
- Deductive vs. Enumerative Procedures: The author advocates for deductive discovery procedures that build hypotheses step-by-step, rather than brute-force enumeration of all possible grammars, which is deemed inefficient and less insightful.
Key Results
The paper presents a detailed algorithm for determining rule contexts and a framework for exploring rule orderings. While no specific quantitative results are provided (due to the lack of formal datasets), the author claims the procedure is "linear time with respect to the size of the data" for the context determination phase, which is a desirable efficiency property. The discussion of Japanese data suggests that the procedure, if applied to real data, could generate a large number of rule orderings when rules are not strongly determined by the data, highlighting a potential area for improvement in capturing linguistic insights.
Limitations and Future Work
- Restricted Rule Type: The procedure is explicitly designed for a "restricted class of phonological rules," specifically rules of the form $a \rightarrow b / C$ (a segment 'a' becomes 'b' in context 'C'). It does not handle rules involving epenthesis (insertion) or deletion, nor does it inherently deal with alternations that are not visible on the surface.
- Data Format: The procedure relies on "surface paradigms" and assumes that root and suffix morphemes are already identified, which is acknowledged as a potentially non-trivial task.
- Rule Ordering Complexity: While the paper demonstrates that ordered rules are not in principle unlearnable, the proposed procedure for exploring orderings can be inefficient (generating $n!$ solutions for $n$ unordered rules). The author suggests that a more sophisticated approach is needed to "localize the weak data determinacy" in rule ordering, perhaps by partitioning orderings into equivalence classes.
- Future Work: The author suggests that the identification of segments involved in alternations and their contexts could form the basis of a different rule order discovery procedure. Further research is needed to refine the handling of unordered rules and to integrate the "evaluation metric" more explicitly into the discovery process.
Summary updated