Semantic Modelling of Adjective-Noun Collocations Using FrameNet

Yana Strakatova, Erhard HinrichsACL 2019

TL;DR

This paper proposes using FrameNet's semantic frame theory to model the meaning of adjective-noun collocations, arguing it's a more robust approach than traditional statistical methods. They demonstrate this by analyzing German collocations, showing how FrameNet's "Frame Elements" can capture the nuanced semantic relationships between adjectives and nouns, even for abstract concepts.

What Problem Does This Paper Address?

Collocations, like "black coffee" or "heavy rain," are word combinations where one word (the collocate) is restricted in its choice by the other (the base). While common, these expressions pose challenges for language learners and computational linguistics due to their partial lexicalization and often non-literal meanings. Existing methods for identifying collocations primarily focus on statistical co-occurrence, providing little insight into their semantic properties. This paper aims to bridge this gap by proposing a semantic modeling framework for adjective-noun collocations, specifically focusing on German.

Methods

The authors leverage Frame Semantics, a theory that posits word meanings are understood in relation to conceptual structures called "frames." They specifically utilize FrameNet, a lexical database that annotates English words with semantic frames and their constituent parts, known as Frame Elements (FEs). The core idea is to map the semantic relationship between an adjective and a noun in a collocation to these FEs.

The process involves:

  1. Identifying collocations: Using the German wordnet GermaNet and the DWDS dictionary with the Wortprofil tool to extract adjective-noun collocations.
  2. Mapping to FrameNet: Translating German nouns to their English FrameNet equivalents and then analyzing how the German adjectives map to the FEs of the evoked semantic frame.
  3. Analyzing concrete vs. abstract nouns: Investigating how the framework applies differently to concrete nouns (e.g., "chocolate") and abstract nouns (e.g., "friend," "fear").

The paper argues that FrameNet's FEs are flexible enough to capture various semantic roles, such as degree, manner, or relationship, that adjectives play in modifying nouns, even when the adjective's meaning is not literal (e.g., "heavy rain" referring to intensity, not weight).

Datasets

  • German wordnet GermaNet: Used for identifying semantic fields of nouns.
  • DWDS (Digitales Wörterbuch der deutschen Sprache): A German electronic dictionary used for collocation extraction.
  • Wortprofil: A collocation extraction tool used in conjunction with DWDS.
  • MRC Psycholinguistic Database: Used to estimate the concreteness of nouns.
  • English FrameNet: The primary resource for semantic frame information, adapted for German collocations through translation.

No new datasets or benchmarks were introduced in this paper.

Key Insights and Ideas

  • Frame Semantics as a unifying framework: FrameNet's concept of semantic frames and Frame Elements provides a powerful and generalizable way to model the semantic relationships in adjective-noun collocations, transcending specific semantic fields.
  • Abstract nouns are richer: Abstract nouns tend to have a greater number and variety of collocates compared to concrete nouns, highlighting the complexity of their semantic modeling.
  • Beyond literal meaning: FrameNet's FEs effectively capture non-literal meanings of adjectives in collocations (e.g., "heavy" in "heavy rain" as intensity, "black" in "black coffee" as a type).
  • Potential for cross-lingual application: The success in applying English FrameNet to German collocations suggests the framework's adaptability for other languages.
  • FrameNet's limitations: While powerful, FrameNet's current structure may lack specific FEs for certain nuanced meanings, requiring potential extensions.

Key Results

The paper demonstrates through several examples that FrameNet's FEs can successfully model the semantic relations in German adjective-noun collocations. For instance:

  • For concrete nouns like "Schokolade" (chocolate), FEs like TYPE and DESCRIPTOR are used to explain adjectives like "schwarz" (black) and "dunkel" (dark), referring to cocoa content rather than literal color.
  • For abstract nouns like "Freund" (friend), FEs like DEGREE, DURATION, and RELATIONSHIP are used to model adjectives like "eng" (close) and "alt" (old).
  • For emotions like "Angst" (fear), FEs like DEGREE, CIRCUMSTANCES, and MANNER are employed to capture the nuances of adjectives like "panisch" (panic) and "existenziell" (existential).

The study found that abstract nouns generally exhibit more diverse collocates and require a richer set of FEs for accurate modeling.

Limitations and Future Work

The authors acknowledge that FrameNet, while extensive, is still under development and may not always have the precise FEs needed to capture all aspects of a noun's meaning as expressed by its collocates. They highlight cases like the noun "Zukunft" (future) where existing FEs in the "Alternatives" frame might not fully account for temporal proximity or evaluative descriptors.

Future work includes:

  • Scaling up: Preparing a larger dataset of German adjective-noun collocations (over 2000) to systematically examine FrameNet's coverage across various semantic fields.
  • FrameNet extension: Investigating the need to add new FEs to existing frames to better accommodate specific collocational meanings.
  • Inter-annotator agreement study: Conducting a study to assess the reliability of FrameNet annotations for collocations.
  • Cross-lingual comparison: Extending the study to other languages like Russian and English to compare FrameNet's approach with other lexical frameworks (e.g., Mel'čuk's Lexical Functions).

Summary updated

BibTeX
@inproceedings{strakatova2019semantic,
  author = {Yana Strakatova and Erhard Hinrichs},
  title = {{Semantic Modelling of Adjective-Noun Collocations Using FrameNet}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2019},
}
Figure