[Paper Review] Verb Semantics and Lexical Selection

Zhibiao Wu, Martha Palmer|ArXiv.org|Jun 22, 1994

Natural Language Processing Techniques460 citations

TL;DR

This paper proposes a multi-domain conceptual representation for verbs based on shared semantic concepts across cognitive domains, enabling accurate lexical selection in machine translation even without exact dictionary matches. By measuring conceptual similarity and relaxing selection restrictions, the system achieves 99.45% accuracy on concrete verb usages and 88.8% on metaphorical uses, significantly outperforming traditional transfer-based MT systems.

ABSTRACT

This paper will focus on the semantic representation of verbs in computer systems and its impact on lexical selection problems in machine translation (MT). Two groups of English and Chinese verbs are examined to show that lexical selection must be based on interpretation of the sentence as well as selection restrictions placed on the verb arguments. A novel representation scheme is suggested, and is compared to representations with selection restrictions used in transfer-based MT. We see our approach as closely aligned with knowledge-based MT approaches (KBMT), and as a separate component that could be incorporated into existing systems. Examples and experimental results will show that, using this scheme, inexact matches can achieve correct lexical selection.

Motivation & Objective

• To address the limitations of transfer-based machine translation systems that rely on exhaustive verb pair listings and rigid selection restrictions.
• To improve lexical selection accuracy in MT by modeling verb semantics through shared conceptual domains rather than fixed argument constraints.
• To enable inexact matches for novel or metaphorical verb usages by introducing a similarity measure based on conceptual representation.
• To demonstrate that conceptual similarity can outperform strict selection restrictions in resolving ambiguous verb translations.
• To integrate the proposed representation into existing MT systems as a modular component compatible with knowledge-based MT approaches.

Proposed method

• Represent each verb as a set of concepts across multiple conceptual domains (e.g., physical change, force, motion), forming a multi-domain semantic representation.
• Define a similarity measure between verb senses based on the overlap and proximity of their conceptual components in a hierarchical ontology.
• Use selection restrictions not as hard constraints but as probabilistic indicators, with conceptual similarity used to resolve ambiguity when constraints are violated.
• Apply an extended selection process that relaxes constraints and selects the most semantically similar target verb based on conceptual proximity.
• Integrate the system into the UNICON MT framework, using a conceptual hierarchy to interpret unknown verb arguments and guide lexical choice.
• Leverage domain-specific knowledge (e.g., that 'price' can be 'fixed' or 'changed in value') to map abstract or metaphorical usages to appropriate target verbs.

Experimental results

Research questions

• RQ1 Can a multi-domain conceptual representation of verbs improve lexical selection accuracy in machine translation beyond rigid selection restrictions?
• RQ2 To what extent can conceptual similarity between verb senses enable correct lexical selection in the absence of exact dictionary matches?
• RQ3 How effective is the proposed method in handling metaphorical or non-concrete verb usages that violate standard argument selection restrictions?
• RQ4 Can the system achieve high accuracy on novel or rare verb usages not covered in traditional bilingual dictionaries?
• RQ5 How does relaxing selection constraints while prioritizing conceptual similarity affect performance on ambiguous verb translations?

Key findings

• The system achieved a 99.45% correct lexical selection rate on 154 sentences with concrete objects after encoding conceptual meanings for unknown verb arguments.
• On a test set of 116 sentences including metaphors and non-concrete objects, the system reached 88.8% accuracy after applying the extended selection process with relaxed constraints.
• The extended selection process improved accuracy by 13.8 percentage points compared to the baseline, demonstrating the value of similarity-based matching.
• The method successfully translated metaphorical uses such as 'break the language barrier' and 'hit a price peak' by mapping abstract meanings to appropriate Chinese verbs like 'da-po' and 'da-dao'.
• The system outperformed TranStar, a commercial transfer-based MT system, which achieved only 19.1% accuracy on the same test set of 157 'break' sentences.
• The results show that conceptual similarity should take precedence over strict selection restrictions, especially in handling context-dependent or novel verb usages.