Enhanced Noun-Noun Compound Interpretation through Textual Enrichment

Bingyang Ye, Jingxuan Tu, James PustejovskyEMNLP 2025

TL;DR

Large Language Models (LLMs) struggle to interpret the nuanced semantic relationships in noun-noun compounds (NNCs) because these relationships are often implicit. This paper introduces a textual enrichment framework that enhances prompts with explicit event-oriented descriptions and contextual cues to disambiguate polysemous compounds, leading to significant improvements in NNC interpretation accuracy across various LLMs.

What Problem Does This Paper Address?

Interpreting noun-noun compounds (NNCs) like "olive oil" or "court approval" is a persistent challenge for Large Language Models (LLMs). The semantic relationship between the modifier (e.g., "olive") and the head noun (e.g., "oil") is rarely explicitly stated. Existing benchmarks often frame this as a multiple-choice task with predefined relation descriptions. However, this approach suffers from two main limitations:

  1. Description Vagueness: The template-based relation descriptions are often too generic and lack the specific eventive connection between the modifier and head, making it difficult for LLMs to pinpoint the correct meaning.
  2. Polysemy: A significant portion of NNCs have multiple plausible interpretations (e.g., "court approval" can mean approval by a court or approval for a court). Without contextual constraints, LLMs struggle to disambiguate these meanings, leading to noisy supervision and capped performance.

This paper aims to address these limitations by enriching the input prompts to LLMs, providing them with more explicit semantic and pragmatic information without requiring any model parameter updates.

Methods

The paper proposes a dual-faceted textual enrichment framework:

  1. Relation Description Enrichment: This facet aims to combat description vagueness. It leverages a noun-noun compound relation taxonomy to generate concrete, event-oriented descriptions for each candidate relation. These enriched descriptions explicitly surface the hidden event connecting the modifier and head, making the intended meaning clearer. For example, instead of a vague description for "art class," it might generate something like "a class that teaches art."
  2. Conditioned Context Enrichment: This facet addresses polysemy. It builds upon Generative Lexicon (GL) theory, which views noun meanings as structured by qualia roles (Formal, Constitutive, Agentive, Telic). The framework first identifies potentially polysemous compounds by analyzing qualia-role binding. For each plausible reading, it generates a "condition cue"—a minimal, relation-agnostic state variable (e.g., TRASH_IN_BAG = yes/no). These cues are then used to create short context sentences that explicitly instantiate the condition, helping LLMs disambiguate between different meanings.

Datasets

The authors evaluate their framework on two existing datasets:

  • LEXICALIZED: Contains 668 NNCs spanning nine relations.
  • NOVEL: A set of 124 novel NNCs created by replacing head or modifier nouns with their hypernyms from WordNet. This dataset is used to test the models' generalization capabilities.

They also create new conditioned datasets:

  • CONDITIONED-LEXICALIZED (C-LEX): LEXICALIZED NNCs augmented with polysemy and condition cues.
  • CONDITIONED-NOVEL (C-NOVEL): NOVEL NNCs augmented with condition cues.

Key Insights and Ideas

  • Implicit Events are Key: The core insight is that the vagueness in NNC interpretation stems from the implicit nature of the event connecting the modifier and head. Explicitly surfacing these events dramatically improves LLM performance.
  • Polysemy is a Major Bottleneck: A significant portion of NNCs are ambiguous. Simply providing more descriptive relation labels isn't enough; explicit contextual constraints are needed to guide LLMs to the correct interpretation.
  • Qualia Theory as a Disambiguation Tool: Leveraging Generative Lexicon theory and qualia roles provides a principled way to identify potential ambiguities in NNCs and generate meaningful condition cues.
  • Parameter-Efficient Improvement: The textual enrichment approach offers a way to significantly boost LLM performance on NNC interpretation without requiring model retraining or parameter updates, making it a cost-effective strategy.
  • Generalization to Novel Compounds: The framework's effectiveness on the NOVEL dataset suggests that the learned semantic understanding is generalizable and not just memorization of specific examples.

Key Results

  • Substantial Accuracy Gains: The proposed textual enrichment framework consistently improves accuracy across three LLM families (Llama-2, Mistral, GPT-40) on both LEXICALIZED and NOVEL datasets.
  • Relation Description Enrichment: Led to substantial improvements on LEXICALIZED, with GPT-40 reaching the top score.
  • Conditioned Context Enrichment: Significantly boosted performance on NOVEL datasets, especially in the 3-shot setting, indicating the importance of context for disambiguation.
  • Compound-Level and Reading-Level Improvements: The framework showed gains at both granular (reading) and holistic (compound) levels.
  • Parameter-Efficient Gains: Lightweight, open-source models (Llama-2, Mistral) saw particularly pronounced improvements, suggesting the enrichment helps them overcome limitations without needing massive scaling.
  • High Accuracy of Automated Pipeline: The automated pipeline for generating condition cues achieved an end-to-end accuracy of 84.4%, greatly reducing human annotation effort.

Limitations and Future Work

  • Limited to Two Readings: The current framework handles at most two alternative readings per compound. Extending it to handle more complex, multi-way ambiguities is a direction for future work.
  • English-Centric: The study is limited to English NNCs. Future work should investigate the transferability of the framework to languages with different compounding strategies (e.g., German, Chinese).
  • Dataset Size: The initial application was on the LEXICALIZED set, which is modest in size. The authors plan to apply their framework to much larger resources to demonstrate scalability.

Summary updated

BibTeX
@inproceedings{ye2025enhanced,
  author = {Bingyang Ye and Jingxuan Tu and James Pustejovsky},
  title = {{Enhanced Noun-Noun Compound Interpretation through Textual Enrichment}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure