Abstract
Understanding entailment—the logical consequence of a statement—is a fundamental challenge in Natural Language Processing (NLP). This paper presents a novel method for identifying "implicative verbs" (e.g., manage) versus "non-implicative verbs" (e.g., want) by exploiting grammatical tense constraints. We demonstrate that verbs whose tense is temporally locked to their complements exhibit high "tense agreement" scores, which serve as a reliable proxy for semantic entailment. Our results show that verbs with high tense agreement scores correlate with human-judged entailment in 91% of cases, providing a data-driven alternative to manually curated linguistic lists.
Introduction
In NLP, distinguishing between sentences like "She managed to fix the bug" and "She wanted to fix the bug" is crucial. The former entails that the bug was fixed, while the latter does not. Verbs that carry this entailment property are known as implicative verbs. Traditionally, identifying these verbs relied on static, manually curated lists. However, the fluid nature of language means many verbs behave implicatively in specific contexts without being formally classified as such.
This research introduces a "linguistic fingerprint" for implicativity based on temporal constraints. We hypothesize that implicative verbs are temporally "locked" to their complements (e.g., one cannot "manage to fix it tomorrow" if the managing happened yesterday), whereas non-implicative verbs allow for temporal divergence (e.g., "planned to fix it tomorrow").
Methodology
To quantify these temporal constraints, we developed a Tense Agreement Score. The methodology involved three primary steps:
- Construction Extraction: We extracted all "Verb + to + Verb" constructions (e.g., refused to go) from the massive Annotated Gigaword corpus.
- Temporal Tagging: Using the Stanford Temporal Tagger, we identified and resolved time-referring expressions (e.g., "last night," "next year") within the sentences to past, present, or future tenses.
- Score Calculation: For each verb, we calculated the frequency with which the main verb's tense matched the complement's temporal expression. A high score suggests the verb is implicative.
Results and Analysis
The tense agreement score successfully separated known implicative verbs from non-implicative ones. As shown in the data, verbs like venture, forget, and manage achieved high scores, while hope, plan, and want scored significantly lower.
One notable finding was the behavior of the verb decide. While traditionally not always considered strictly implicative, it received a higher score than expected. Qualitative analysis suggests that in common usage, "I decided to leave" is almost always interpreted by listeners as "I left," demonstrating the model's ability to capture de facto implicativity.
To validate these scores, we compared them against crowdsourced human judgments of entailment. The correlation was striking: verbs with a tense agreement score greater than 0.6 were judged to entail their complements 91% of the time, whereas those with scores below 0.2 only signaled entailment 30% of the time.
Discussion and Conclusion
Our findings suggest that simple grammatical constraints like tense agreement can act as powerful proxies for complex semantic properties. This method allows for the automated discovery of implicative behavior across large datasets, bypassing the limitations of manual curation.
Future work will address current limitations, such as the noise inherent in automated temporal tagging and the model's current inability to distinguish between positive and negative entailment (polarity). For instance, both manage (positive) and fail (negative) exhibit high tense agreement, but they imply opposite outcomes. Integrating these temporal constraints into distributional word representations could significantly enhance the performance of modern NLP systems on entailment tasks.
Summary updated