[Paper Review] Monotonicity Marking from Universal Dependency Trees

Zeming Chen, Qiyue Gao|arXiv (Cornell University)|Apr 17, 2021

Topic Modeling25 references4 citations

TL;DR

This paper presents a novel system that automatically annotates monotonicity polarity (monotone, antitone, or none) on dependency trees using Universal Dependencies (UD), outperforming existing tools like NatLog and ccg2mono. By applying a rule-based polarization algorithm to UD parse trees, the system achieves state-of-the-art accuracy on both token-level and sentence-level monotonicity annotation, demonstrating superior robustness and coverage on a benchmark dataset.

ABSTRACT

Dependency parsing is a tool widely used in the field of Natural language processing and computational linguistics. However, there is hardly any work that connects dependency parsing to monotonicity, which is an essential part of logic and linguistic semantics. In this paper, we present a system that automatically annotates monotonicity information based on Universal Dependency parse trees. Our system utilizes surface-level monotonicity facts about quantifiers, lexical items, and token-level polarity information. We compared our system's performance with existing systems in the literature, including NatLog and ccg2mono, on a small evaluation dataset. Results show that our system outperforms NatLog and ccg2mono.

Motivation & Objective

• To develop an automatic polarity marking system that leverages Universal Dependency (UD) parse trees instead of CCG trees for monotonicity reasoning.
• To improve the accuracy and robustness of monotonicity annotation in natural language inference by using surface-level facts about quantifiers, lexical items, and token-level polarity.
• To enable broader application of monotonicity reasoning across diverse text genres and languages by relying on high-accuracy UD parsing.
• To provide a more explainable and formally precise alternative to black-box neural models in natural language inference.

Proposed method

• Convert input sentences into Universal Dependency (UD) parse trees using a standard UD parser.
• Transform the UD parse tree into a binary tree structure to simplify rule application.
• Apply a lexicon of polarization rules that assign monotone (↑), antitone (↓), or neutral (=) polarity to each node based on syntactic and semantic cues.
• Use a two-phase algorithm: first marking polarities from leaves to root, then propagating them from root to leaves to ensure consistency.
• Leverage known surface-level monotonicity facts—such as the behavior of quantifiers (e.g., 'all', 'some'), negation, and determiners—to guide polarity assignment.
• Validate the system against gold-standard annotations using precision, recall, and F1-score on both all tokens and key tokens (content words, determiners, numbers).

Experimental results

Research questions

• RQ1 Can monotonicity polarity be accurately inferred from Universal Dependency parse trees rather than CCG parse trees?
• RQ2 How does the performance of a UD-based polarity marking system compare to existing systems like NatLog and ccg2mono on standard evaluation benchmarks?
• RQ3 To what extent does using UD parsing improve the robustness and coverage of monotonicity annotations across diverse linguistic phenomena?
• RQ4 Can a rule-based system using surface-level syntactic and semantic cues achieve higher accuracy than previous approaches in monotonicity annotation?

Key findings

• Our system achieved the highest token-level accuracy on all tokens (96.5) and key tokens (96.5), outperforming both NatLog and ccg2mono.
• On sentence-level evaluation, our system achieved 87.5% accuracy on all tokens and 89.2% on key tokens, surpassing both baselines.
• Our system demonstrated superior robustness, with F1-scores of 97.4 (all tokens) and 89.9 (key tokens), significantly higher than NatLog and ccg2mono.
• All instances correctly annotated by ccg2mono were also correctly annotated by our system, indicating broader coverage of linguistic phenomena.
• Our system achieved 97.6% recall and 97.4% F1-score for antitone polarity on key tokens, showing strong performance on challenging cases.
• The precision and recall for all three polarity labels (monotone, antitone, none) were consistently higher in our system than in NatLog and ccg2mono, confirming improved robustness.