Abstract
This paper describes a tactical generator for Turkish, a language characterized by its free constituent order. In Turkish, the arrangement of sentence elements shifts according to the information structure—specifically the topic, focus, and background—of the intended message. While a default Subject-Object-Verb (SOV) order exists, constituents are almost freely movable depending on discourse constraints. To manage this flexibility, the system utilizes a recursively structured finite state machine (FSM) implemented as a right-linear grammar backbone within the GenKit environment. Morphological realization is handled by an external component that performs concrete morpheme selection and manages morphographemic processes.
Introduction
In Natural Language Generation (NLG), the tactical phase is responsible for transforming a structured internal representation of a thought into a linear string of words. For languages like English, this process follows relatively rigid word-order rules. However, Turkish presents a unique challenge as a "free constituent order" language. While its default order is SOV, constituents can be reordered to signal pragmatic nuances: the Topic (what the sentence is about), the Focus (the new or emphasized information), and the Background (contextual information).
The challenge for a tactical generator is twofold: it must handle the vast number of possible permutations for a single sentence and choose the pragmatically appropriate permutation based on the discourse context, rather than simply producing any grammatically valid string.
System Architecture and Methodology
The generator is built upon a Pseudo Unification Grammar using the GenKit system, developed at Carnegie Mellon University’s Center for Machine Translation. The implementation relies on three core components:
1. Case-Frame Input and Information Structure
The system accepts "case-frames" as input—feature structures that define the verb, its arguments (subject, object, etc.), and adjuncts (time, place). Crucially, the input includes a CONTROL feature that specifies the information structure, identifying which elements serve as the Topic, Focus, or Background.
2. Recursive Finite State Machine (FSM)
The core of the grammar is a right-linear backbone that functions as an FSM. This machine traverses states representing different parts of the sentence. To handle complexity, the FSM is recursive: if an argument is a nested clause, the FSM calls itself to generate that sub-sentence before returning to the main clause level. This architecture allows the generator to "save" the Focus element and ensure it is placed in the critical pre-verbal position.
3. Morphological Interface
Turkish is highly agglutinative, meaning it adds multiple suffixes to a single root to convey grammatical meaning. To manage this, the tactical generator produces abstract morphological features (e.g., +PAST, +3SG, +ACC). These features are then passed to an external morphological generator, built with Xerox Two-Level Tools, which handles the final surface-form generation and morphographemic processes.
Key Insights
- Morphology as an Enabler: The research highlights that Turkish can maintain free word order because its rich morphology explicitly signals grammatical roles. Case markings ensure that the relationship between words remains clear regardless of their position in the sentence.
- The Pre-verbal Focus: A fundamental rule of Turkish syntax is that the Focus element must immediately precede the verb. The FSM architecture is specifically designed to enforce this constraint.
- Default vs. Marked Order: To avoid non-determinism (generating multiple valid sentences for a single input), the system defaults to standard SOV order in the absence of specific discourse instructions.
- Phrase-Internal Rigidity: While sentence-level order is free, the author notes that phrase-internal order is quite rigid. For example, Turkish Noun Phrases follow a strict 11-step hierarchy from specifiers to the head noun.
Evaluation and Results
The implementation resulted in a comprehensive grammar consisting of 273 rules. The system was evaluated using hand-crafted test cases and technical documentation from computer manuals to ensure it could handle the precise language required for professional translation.
Key results include:
- Qualitative Success: The generator successfully produced complex sentences with multiple embedded clauses, such as: "Adam kadının kitap okuduğunu zannetti" (The man thought that the woman was reading a book).
- Pragmatic Accuracy: By manipulating the
CONTROLfeatures, the system demonstrated it could generate different versions of the same event to reflect different emphases (e.g., distinguishing between "Ali gave the book to Ahmet" and "It was Ali who gave the book..."). - Efficiency: The right-linear grammar backbone allowed the system to avoid the computational overhead often associated with more complex formalisms like Combinatory Categorial Grammar (CCG), while still maintaining the necessary flexibility for Turkish syntax.
Limitations and Future Work
While robust, the current model has limitations. It assumes that phrases remain contiguous, meaning it cannot yet handle "discontinuous constituents" where a noun phrase is split across a sentence—a phenomenon found in informal or poetic Turkish. Additionally, the system focuses on scrambling within a single clause; "long-distance scrambling," where a constituent moves from an embedded clause into the main sentence, remains a goal for future development. Finally, the system is strictly tactical; the "strategic" decision of which information structure to use remains the responsibility of a higher-level discourse component.
Summary updated