Identifying Semantic Relations and Functional Properties of Human Verb Associations

Sabine Schulte im Walde, Alissa MelingerEMNLP 2005

TL;DR

This paper explores how human-generated verb associations can enrich our understanding of semantic verb relations and functional properties, offering a valuable complement to existing lexical databases like GermaNet. By analyzing a large dataset of human responses, the authors identify a broader range of semantic relations and noun functions, crucial for improving NLP applications like verb clustering and thesaurus extraction.

What Problem Does This Paper Address?

Natural Language Processing (NLP) tasks often rely on understanding the semantic relationships between words, particularly verbs. While existing lexical databases like GermaNet (a German counterpart to WordNet) capture "classical" semantic relations (synonymy, antonymy, hypernymy), they may not fully encompass the richness of human semantic understanding. Specifically, the paper addresses the gap in understanding the full spectrum of verb-verb relations and the functional roles nouns play in relation to verbs, as perceived by humans. This limitation hinders NLP systems' ability to accurately model verb semantics, impacting tasks like verb clustering, thesaurus extraction, and text summarization. The authors propose that human associations, collected through experiments, can reveal these underrepresented relations and functions.

Methods

The study employs a multi-stage analysis of human verb associations:

  1. Morpho-Syntactic Analysis: An empirical grammar model is used to assign part-of-speech tags (noun, verb, adjective, adverb) to the elicited noun associates. This step helps in distinguishing and quantifying the types of responses.
  2. Semantic Verb Relations Analysis: For verb-verb pairs, the authors leverage GermaNet to identify "classical" semantic relations. Crucially, they also analyze pairs not found in GermaNet, hypothesizing these represent missing links or non-classical relations (e.g., temporal order, cause, consequence).
  3. Syntax-Semantic Noun Functions Analysis: For noun associates, the empirical grammar model is used to determine the syntactic functions (e.g., direct object, subject) these nouns play in relation to the target verb. This analysis aims to uncover conceptual roles that contribute to distributional verb descriptions.
  4. Window Look-up: To assess the contextual relevance of identified semantic relations and noun functions, a window look-up is performed on a large newspaper corpus. This checks how often associated verbs or nouns co-occur within a defined window.

Datasets

  • Human Verb Associations Dataset: Collected through a web experiment where 299 native German speakers provided associations to 330 German verbs. Each trial elicited an average of 5.16 associates.
  • GermaNet: A German lexical semantic taxonomy used to identify "classical" semantic relations between verb pairs.
  • German Newspaper Corpus: A 35-million-word corpus (used for training the empirical grammar model) and a 200-million-word corpus (used for window look-up analysis).

Key Insights and Ideas

  • Human associations reveal a broader semantic landscape: A significant portion of verb-verb pairs elicited by humans do not fit into GermaNet's "classical" relations, suggesting these databases are incomplete.
  • "No relation" in GermaNet is informative: Pairs not found in GermaNet often represent potential new semantic relations (e.g., near-synonymy, antonymy) or non-classical relations like temporal order, cause, and consequence, providing a basis for extending lexical resources.
  • Nouns play diverse functional roles: Beyond direct syntactic roles (like direct objects), noun associates capture scene-related information and world knowledge, which are not always explicitly captured by verb subcategorization.
  • Window co-occurrence complements syntactic analysis: The window look-up demonstrates that while syntactic functions are important, nouns that encode scene information or world knowledge also frequently co-occur with verbs, suggesting the value of integrating window-based approaches.
  • Verb frequency influences association types: The rate of verb and adverb responses is positively correlated with target verb frequency, while noun and adjective responses are inversely correlated.

Key Results

  • Semantic Relations: 37% of verb-verb pairs had identifiable semantic relations in GermaNet. Of the remaining, 9% were attributed to part-of-speech confusion, 54% showed "no relation" despite both verbs being in GermaNet, and 9% were unknown cases. Non-classical relations like temporal order, cause, and consequence were found to be prevalent.
  • Part-of-Speech Distribution: Noun associates were the most frequent (62%), followed by verbs (25%), adjectives (11%), and adverbs (2%).
  • Noun Functions: Only 28% of noun associates were identified as frame-slot fillers by the statistical grammar. However, the analysis revealed that a range of linguistic functions were activated by more than 1% of associates, suggesting these are candidates for defining distributional verb descriptions.
  • Window Co-occurrence: For paradigmatically related verb pairs found in GermaNet, co-occurrence rates within windows of 5, 20, and 50 words were 85%, 95%, and 97%, respectively. For pairs with unspecified relations, these rates were lower (61%, 74%, 79%).

Limitations and Future Work

  • GermaNet Incompleteness: The study acknowledges that GermaNet is not exhaustive, and a significant portion of "no relation" cases might be due to un-coded relations.
  • Ambiguity in Part-of-Speech Tagging: Some noun associates were ambiguous between different part-of-speech tags, leading to frequency splits.
  • Future Work: The authors plan to establish a set of non-classical verb-verb relations, explore dependencies between feature descriptions and verb relations, and investigate how the varied distributions of associate types across semantic classes can enhance automatic class acquisition. They also aim to further investigate the integration of window-based approaches with function-based approaches for distributional verb descriptions.

Summary updated

BibTeX
@inproceedings{walde2005identifying,
  author = {Sabine Schulte im Walde and Alissa Melinger},
  title = {{Identifying Semantic Relations and Functional Properties of Human Verb Associations}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2005},
}
Figure