TL;DR
This paper introduces a Disambiguation Module (DM) that automatically analyzes dictionary definitions of the form "to VERB with NP" to identify the semantic relationship between the verb and its prepositional phrase. By leveraging lexical resources and semantic heuristics, the DM aims to improve lexical knowledge bases and contribute to resolving prepositional phrase attachment ambiguities in natural language processing.
What Problem Does This Paper Address?
Building comprehensive lexical databases like IBM's COMPLEX requires not only gathering information from various sources but also interpreting and disambiguating it. A significant challenge lies in understanding the semantic relationships within dictionary definitions themselves. Specifically, this paper focuses on definitions of the form "to VERB with NP," where the preposition "with" can indicate a variety of semantic roles (e.g., instrument, manner, accompaniment). Disambiguating these relationships is crucial for correctly understanding word senses and their connections, which in turn aids in resolving ambiguities in broader natural language processing tasks, such as prepositional phrase attachment. For instance, in the sentence "The player hit the ball with the bat," it's syntactically ambiguous whether "with the bat" modifies "hit" or "ball." By understanding the semantic role of "with the bat" in dictionary definitions, we can better resolve such ambiguities in free text.
Methods
The core of the paper is the Disambiguation Module (DM), which is an extension of an earlier prototype. The DM operates in two main stages:
- Parsing and Semantic Categorization: The DM first parses the definition string using the PEG parser to identify the VERB and NP-head. It then retrieves definitions for these words from Webster's Seventh New Collegiate Dictionary (W7) and parses them. Heuristics are applied to determine the semantic categories of the VERB (e.g., change, affliction) and the NP-head (e.g., implement, body part, human). These heuristics rely on lexical and syntactic conditions. For example, a noun is marked as +INSTRUMENT if its definition contains terms like "instrument," "implement," "device," "tool," or "weapon," or if it's post-modified by "used as a weapon."
- Semantic Relation Identification: Once semantic categories are established, the DM attempts to identify the semantic relation between the VERB and the NP-head, drawing from a set of five identified senses of "with": USE, MANNER, ALTERATION, CO-AGENCY/PARTICIPATION, and PROVISION. These relations are identified using specific heuristics. For instance, a USE relation is often identified when the NP-head is +INSTRUMENT (e.g., "fish with a hook"). The DM also considers links between the definitions of the VERB and NP-head, though these are noted as rare. The paper also discusses a sixth, purely syntactic function, PHRASAL, for cases like "charge with a crime."
The DM prioritizes verb-based interpretations over noun-based ones in cases of conflict.
Datasets
- Training Corpus: 170 verb definitions of the form "to VERB with NP" (148 transitive, 22 intransitive) randomly extracted from the letters 'a' and 'b' of Webster's Seventh New Collegiate Dictionary (W7).
- Testing Corpus: 132 definitions of the form "to VERB with NP" not previously processed.
- Lexical Resource: Webster's Seventh New Collegiate Dictionary (W7) is used for definitions. The Longman Dictionary of Contemporary English (LDOCE) is also referenced for verb subcategorization information.
Key Insights and Ideas
- Semantic Roles of "with": The paper identifies and categorizes five distinct semantic relations that "with" can convey in "to VERB with NP" constructions (USE, MANNER, ALTERATION, CO-AGENCY/PARTICIPATION, PROVISION). This provides a structured way to understand the preposition's function.
- Leveraging Dictionary Definitions for NLP: The core idea is that dictionary definitions, despite their inherent structure, can serve as valuable resources for disambiguation tasks in broader NLP. This approach differs from methods relying solely on hand-built knowledge bases.
- Heuristics for Semantic Categorization: The development of specific, rule-based heuristics to identify semantic categories of verbs and nouns based on their dictionary definitions is a key methodological contribution.
- Interplay of Verb and Noun Semantics: The paper highlights that disambiguating the meaning of "with" depends on the semantic categories of both the verb and the NP-head, and sometimes on the relationship between their definitions.
- Bridging Syntactic and Semantic Ambiguity: The authors argue that resolving semantic ambiguity in dictionary definitions can directly inform the resolution of syntactic ambiguity (like PP attachment) in free text.
- Human Judgement Validation: The study includes a crucial step of validating the DM's output against human judgments, demonstrating the program's alignment with human linguistic intuition.
Key Results
- Coverage: The DM achieved a coverage of 78.2% on the training corpus and 83.3% on the testing corpus, meaning it assigned at least one interpretation to this percentage of strings.
- Correct Interpretations: The DM correctly assigned the semantic relation for 113 out of 170 strings (66.5%) in the training corpus and 75 out of 132 strings (56.8%) in the testing corpus.
- Partial Success: In 15 cases (training) and 13 cases (testing), the DM assigned two interpretations, one of which was correct. This is considered a partial success.
- Reliability: The COR-TO-INC RATIO (correct interpretations to incorrect ones) was 113/5 (85%) for training and 75/22 (68%) for testing. When including partially correct interpretations, the reliability increased significantly (96.2% and 80% respectively).
- Human Validation: The DM's judgements coincided with human judgements in 51 out of 56 cases (91%), suggesting the heuristics are well-aligned with human linguistic intuition.
Limitations and Future Work
- Ambiguity Resolution: The DM sometimes assigns multiple interpretations, and distinguishing between them or selecting the most appropriate one remains a challenge, especially when frequency information for word senses is unavailable.
- Idiosyncratic Ambiguities: Some ambiguities are highly specific and cannot be resolved by general heuristics, requiring further study.
- Overlapping Cases: The current "final preference" heuristic for verb-based interpretations can lead to the loss of information when both verb-based and noun-based links are genuinely applicable (e.g., "spatter with a discoloring substance").
- PHRASAL Interpretation Conflicts: Conflicts between PHRASAL and semantic interpretations are difficult to resolve systematically as PHRASAL is verb-specific.
- Future Work:
- Addressing multiple conjuncts in coordinate structures.
- Developing more sophisticated heuristics to handle idiosyncratic ambiguities.
- Investigating the implications of the verb-based preference heuristic on overlapping cases.
- Further fine-tuning of heuristics based on testing results.
- Studying the significance of unclear and overlapping cases in human judgments.
- Expanding the semantic categories and the number of prepositions and definition types handled.
Summary updated