TL;DR
This paper introduces a novel "hybridization" approach to quantifier scoping in natural language processing, arguing it's superior to traditional "integration of specialists" (IS) methods. By combining lexical and syntactic factors before determining scope, the system achieves greater precision and flexibility in handling complex ambiguities.
What Problem Does This Paper Address?
Natural language is rife with ambiguity, and a significant source of this is "quantifier scope." Consider the sentence "Some target was hit by every arrow." This could mean either that there was a specific target that all arrows hit, or that each arrow hit some target (which might be different for each arrow). Determining the correct scope is crucial for accurate meaning representation.
Traditional systems (IS) tackle this by having individual "specialists" for different factors (e.g., a "lexical specialist" for quantifier strength, a "syntactic specialist" for embedded phrases). These specialists assign scores, which are then integrated. The paper argues this modular approach struggles with interactions between factors, leading to suboptimal or incorrect scope assignments in complex cases. For instance, a PP specialist might not know how a quantifier's tendency to scope changes based on the quantifiers in the head noun phrase it modifies.
Methods
The paper proposes a "hybridization" approach, contrasting it with the "integration of specialists" (IS). Instead of separate specialists, hybridization conjoins lexical and syntactic factors first to create a single function (e.g., flex-syn). This function directly computes scope preferences based on the combined input.
The implemented system, the PAKTUS Scoping Module (PSM), uses a database to store these hybrid rules. The database is structured hierarchically:
- RULEGRPS: Top-level groups based on grammatical roles (e.g., subject-direct object).
- RULEGRP25: Sub-groups with conditions on syntactic properties (e.g., passive voice, adverb position).
- RULE112: Specific rules with conditions on quantifier pairs and their properties (e.g., "monotone decreasing").
The scoping algorithm first determines horizontal scope (within a clause) and then vertical scope (for embedded structures like PPs or relative clauses). A default "natural" (left-to-right) order is assumed, and rules then impose preferences to override this default. The system prioritizes rules that represent strong preferences and avoids undoing previously established preferences.
Datasets
The paper mentions a "body of data (150 pages, described in the appendix of Chien 1992)" which subsumes existing consensus on lexical and syntactic scoping influences. This data was used to build the database for the PSM system. No specific benchmark datasets for evaluation are detailed, but the system's performance is illustrated with example sentences.
Key Insights and Ideas
- Hybridization vs. Integration: The core idea is that combining factors before judgment (hybridization) is more powerful than integrating judgments from isolated specialists (IS). This allows for capturing nuanced interactions between different linguistic cues.
- Modularity Trade-off: While IS emphasizes modularity, the paper argues that for quantifier scoping, this modularity comes at the cost of precision. Hybridization sacrifices some modularity for better performance.
- Database Design: The hierarchical database structure (RULEGRPS, RULEGRP25, RULE112) efficiently stores and retrieves complex, combined rules, allowing for staged testing of conditions.
- Default Preference: Using a default "natural" scope order simplifies the system by eliminating rules that simply confirm this default, focusing instead on exceptions.
- Generalization: The system can generalize rules based on syntactic information alone or combined with partial lexical information, reducing the number of explicit rules needed.
Key Results
- The system can determine preferred scope orders for quantified phrases.
- The paper provides an example of a sentence with multiple quantifiers ("A person in each house on both streets saw several men who were robbing some banks") for which PSM returns a scope order in 0.7 seconds.
- Another example sentence ("Rarely did a park supervisor serving several districts in two counties assign everyone many trees with no large branches on some limb which might fall on a passerby") yields 4 scopings in 1.283 seconds.
- The hybridization approach is claimed to recover scope intuitions more directly than IS.
Limitations and Future Work
- Semantic and Pragmatic Factors: The current system deliberately excludes semantic and pragmatic factors, focusing only on syntactic and lexical influences. The paper acknowledges that incorporating these would be a significant challenge.
- Hybridization vs. Integration for Semantics: A key open question is whether semantic/pragmatic factors should also be hybridized or integrated with the existing system. Full hybridization of all factors could lead to intractable complexity.
- Modularity of Data Updates: While the hybrid approach offers greater precision, updating the database with new lexical or syntactic factors is not modular, posing an engineering challenge.
- Future Work: The authors suggest that the system could be extended to handle semantic and pragmatic factors, and that the translator to logical form (currently not implemented) would be the next step.
Summary updated