TL;DR
This paper proposes a universal lexicon architecture where language-specific variations in syntax are explained by the presence or absence of specific morphemes, rather than arbitrary rules. This approach aims to constrain the complexity of lexical entries and improve machine translation by unifying word senses across languages.
What Problem Does This Paper Address?
Building comprehensive lexicons for natural language processing, especially for machine translation (MT), is a significant challenge. Traditional approaches often treat word senses and their associated syntactic behaviors as arbitrary and language-specific. This leads to a massive combinatorial explosion of possible lexical entries, making lexicons difficult to construct, maintain, and use effectively for translation. The paper tackles the problem of "arbitrariness" in lexical information, particularly concerning argument structure, word sense, and paraphrasability, arguing that this arbitrariness can be significantly reduced and explained by a more principled, syntax-driven approach.
Methods
The authors propose a generative lexicon architecture based on Hale and Keyser's work on lexical syntax and Pustejovsky's ideas on event structure. The core idea is to replace language-specific syntactic rules with a universal set of "basic building blocks" (lexical categories like N, A, V, P) and "rules of composition" (like X-bar projection and Head Movement). Language-specific variations are then explained by the presence or absence of certain morphemes within this universal framework. This means that the underlying syntactic machinery for generating argument structures is fixed across languages, and differences arise from the "morphological component" of the lexicon. They illustrate this with the "Spray/Load" alternation, showing how verbs that don't participate in this alternation in one language might do so in another due to the presence or absence of specific incorporated prepositions or aspectual morphemes.
Datasets
Not discussed in the paper. The paper focuses on theoretical linguistic analysis and computational modeling rather than empirical evaluation on large datasets. Examples are drawn from existing linguistic literature and specific language constructions.
Key Insights and Ideas
- Arbitrariness is Morphological: The paper's central claim is that much of the perceived arbitrariness in lexical behavior across languages is not inherent to the meaning but rather a consequence of the presence or absence of specific morphemes in the morphological component of the lexicon.
- Universal Syntactic Core: The generative machinery for producing argument structures is proposed to be universal and fixed across languages, simplifying the core computational model.
- Morphemes as Language Parameters: Language-specific parameters are reinterpreted as the presence or absence of lexical material (morphemes) in the morphological component.
- Explaining Nonalternation: The paper provides a principled explanation for why certain verbs participate in alternations (like Spray/Load) in one language but not another, attributing it to incorporated prepositions or aspectual morphemes that "freeze" the verb's argument structure.
- Implications for MT: This unified approach to lexical representation, by constraining word senses and their syntactic behavior, offers a path towards more robust and efficient interlingua-based machine translation systems.
Key Results
The paper presents a theoretical framework and demonstrates its explanatory power through linguistic examples. It shows how the proposed architecture can account for:
- The shared properties of the Spray/Load alternation across Japanese, Hindi, Bengali, Greek, and English.
- The "nonalternation" of certain verbs in specific languages, explaining it through the incorporation of prepositions or aspectual morphemes.
- The potential for a more constrained and less arbitrary lexicon, reducing the number of necessary word senses and diacritics for MT.
Quantitative results are not presented, as the paper is primarily theoretical and analytical.
Limitations and Future Work
- Survey Completion: The authors acknowledge that their survey of verbal alternations is incomplete and plan to extend it to cover approximately 150 types of alternations identified by Levin (1993), and to include more languages.
- Computational Implementation: While the paper outlines an architecture, a full computational implementation and evaluation are future steps.
- MT Application: The paper suggests manageable applications to machine translation but does not present a fully developed MT system based on this architecture.
- Pragmatic Considerations: The paper touches on pragmatic interpretations (e.g., the "baby covered" example) but suggests that compositional behavior of morphemes is the primary driver of different alternation paradigms, implying a need for further integration of pragmatic factors.
Summary updated