Bilingual Hebrew-English Generation of Possessives and Partitives: Raising the Input Abstraction Level

Yael Dahan Netzer, Michael ElhadadACL 1999 · DOI

Introduction: The "Goldilocks" Problem in Bilingual NLG

In Natural Language Generation (NLG), the "syntactic realization" component is the engine that transforms structured input into fluent sentences. A central challenge in this field is determining the optimal level of abstraction for the input. If the input is too "shallow" (close to surface syntax), it becomes difficult to use for bilingual systems because it forces language-specific choices too early. Conversely, if the input is too abstract (purely semantic), the generator must perform excessive computational work to resolve grammatical details.

Researchers at Ben Gurion University have addressed this "Goldilocks" problem by developing a more abstract representation for complex noun phrases (NPs). Their work focuses on generating grammatically correct possessives and partitives in both Hebrew and English from a single, unified input structure.

Methodology: Abstracting Syntactic Markers

The authors performed a contrastive analysis of how Hebrew and English handle complex NPs and modified two existing realization components: SURGE (for English) and HUGG (for Hebrew). The core innovation involves moving away from specifying language-specific syntactic markers, such as the English "'s" or "of," and instead utilizing functional features:

  • possessor: Identifies the possessor entity without prescribing the syntactic form (e.g., genitive vs. prepositional).
  • ref-set (Reference Set): Defines the group from which a subset is taken in partitive constructions.

By raising the abstraction level, the system allows language-specific grammars to handle syntactic differences automatically through internal preference ranking and constraint satisfaction.

Key Innovations

Syntactic Competition in Noun Phrases

A major hurdle in NP generation is "syntactic competition," where different semantic relations vie for the same syntactic slot. In Hebrew, for example, the presence of a classifier (e.g., "Shabat" in "Shabat dress") prevents the use of the "construct state" (smixut) for a possessor. The abstract input allows the grammar to resolve this "tug-of-war" locally based on the specific constraints of the target language.

Figure 3: Feature structure representing a noun phrase with both a classifier ('Sabat') and a possessor ('yalda/girl'), illustrating syntactic competition.

Recursion in Partitives

Partitives are unique due to their recursive nature (e.g., "ten of all of the demonstrators"). The authors introduced the ref-set feature to handle this recursive logic without hard-coding specific function words like "of." This allows the English grammar to decide on the appropriate prepositional structure while the Hebrew grammar determines the correct genitive or prepositional marker.

Figure 4: Abstract representation of a partitive noun phrase using the ref-set feature to define a subset of a larger group.

Results and Future Directions

The system successfully generated fluent complex NPs in both languages from 100 sample input structures. Qualitatively, the grammars correctly handled divergences, such as English requiring an "of-construct" while Hebrew requires a "free genitive" (Sel).

Despite these successes, challenges remain in determining the "unmarked" or most natural-sounding choice when multiple grammatical versions are technically correct. Additionally, mapping definiteness defaults between Hebrew and English remains a complex task. Future work will focus on refining these default behaviors and expanding the model to cover more complex clausal structures.

Summary updated

BibTeX
@inproceedings{netzer1999bilingual,
  author = {Yael Dahan Netzer and Michael Elhadad},
  title = {{Bilingual Hebrew-English Generation of Possessives and Partitives: Raising the Input Abstraction Level}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {1999},
  doi = {10.3115/1034678.1034708},
}
Figure