Learning a Second Language

Steven L. Lytinen, Carol E. MoonAAAI 1988

Abstract

In the late 1980s, natural language processing systems faced a significant hurdle: the rigidity of hand-coded grammars. Researchers at the University of Michigan addressed this by developing IMMIGRANT, a system designed to learn the grammar of a second language (L2) through instructions provided in a first language (L1). By utilizing a unification-based representation, IMMIGRANT treats the acquisition of grammatical rules as a specialized form of parsing, allowing for the seamless integration of new constraints and the structural reordering of existing rules.

The Representation Bottleneck

Traditional natural language systems of the era relied on context-free rules (e.g., S -> NP VP). While effective for static parsing, these rules make learning from instruction difficult. Because word order is implicit in the sequence of the rule, a simple instruction like "In German, the verb goes at the end" requires a complex mapping process to determine which part of the rule to rewrite.

To overcome this, the authors proposed using Unification Grammar, where linguistic knowledge is stored as Directed Acyclic Graphs (DAGs). This format is used consistently across the system's internal grammar, its lexicon, and its parsing output.

Methods: Learning as Unification

The core innovation of IMMIGRANT is that it treats learning a rule almost exactly like parsing a sentence. When the system receives a grammatical instruction, it parses that instruction into a DAG and attempts to unify it with its existing knowledge base.

Explicit Word Order and Borders

One of the key features of this representation is the use of explicit "border" links (LBORDER and RBORDER). By making word order an explicit property of the graph rather than an implicit sequence, the system can modify the position of elements without disrupting their functional relationships (such as subject-verb-object roles).

Figure 1: A Directed Acyclic Graph (DAG) representing a standard English-like sentence structure where the verb (V) precedes the object (NP).

Handling Rule Conflicts

IMMIGRANT categorizes new rules into two types:

  1. Constraint-addition rules: These add new requirements (e.g., case agreement) that do not contradict existing ones. These are simply merged into the existing DAG.
  2. Reordering rules: These directly contradict the existing L1 word order.

When a reordering rule causes a unification failure, the system identifies the conflicting adjacency links. Because it recognizes the instruction as a "reordering" type, it gives precedence to the new information, effectively "rewiring" the DAG to reflect the syntax of the new language.

Figure 2: The modified DAG after a reordering instruction, demonstrating the verb (V) moved to the final position, typical of certain German constructions.

Key Insights and Results

The system operates on the Interlanguage Hypothesis, assuming the L2 is identical to the L1 (English) until instructed otherwise. This allows English to serve as a default template that is incrementally refined.

In evaluations involving approximately 15 curated examples—specifically focusing on English and German relative clauses and prepositional case agreement—IMMIGRANT successfully demonstrated the ability to:

  • Internalize English instructions about German word order.
  • Automatically generate functional parsing rules for the new language.
  • Maintain structural flexibility that traditional context-free grammars lack.

Limitations and Future Directions

While IMMIGRANT proved the viability of instruction-based learning, the authors noted several limitations. The system was tested on a small, hand-crafted feature set and relied on a "chicken and egg" scenario where a grammar is needed to parse the instructions that build the grammar. Future work suggested integrating examples alongside instructions to help the system "relax" constraints more efficiently and expanding the system to handle lexical semantics in addition to syntax.

Summary updated

BibTeX
@inproceedings{lytinen1988learning,
  author = {Steven L. Lytinen and Carol E. Moon},
  title = {{Learning a Second Language}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1988},
}

Loading related content...

Figure