580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How can ontology- and taxonomy-based models enhance semantic similarity judgment in structured lexical databases?
This research theme focuses on leveraging structured lexical knowledge bases such as WordNet and domain-specific ontologies to assess semantic similarity. It examines how taxonomic relationships (e.g., hypernym/hyponym, meronym/holonym), information content, and edge-weighted path lengths within ontological hierarchies can quantify semantic proximity more closely aligned with human judgment. This is important because it grounds semantic similarity in well-curated resources and formalizes semantic distance metrics, facilitating applications in information retrieval, word sense disambiguation, and semantic search where lexical resources exist.
Key finding: Introduces a novel edge-counting model in the WordNet taxonomy that incorporates different weights for synonymy, hypernymy, and meronymy links, achieving a high correlation (0.921) with human similarity judgments on a... Read more
Key finding: Presents SemSim^p, a parametric semantic similarity approach that uses information content weighting in ontologies and compares it across six established semantic similarity methods on large annotated datasets derived from... Read more
Key finding: Assesses four ontology weighting strategies, divided into intensional (ontology structure based) and extensional (also leveraging resource annotation data), within a semantic search engine framework. The study empirically... Read more
Key finding: Critically analyzes existing WordNet-based similarity tests (WBST) and identifies their insufficiency for robust semantic similarity function evaluation. Proposes a more demanding test framework targeting Polish nouns that... Read more
Key finding: Provides a comprehensive survey of WordNet- and Roget's thesaurus-based semantic similarity methods, emphasizing path-length and information content approaches within taxonomic hierarchies. Discusses the utility of... Read more
2. What role do distributional and corpus-based models play in capturing semantic similarity, particularly for weakly related or dissimilar concepts?
This theme explores the use of statistical and distributional semantics, including semantic networks derived from co-occurrence information, latent semantic analysis, and lexico-syntactic pattern mining, to capture semantic similarity across words and texts. It is particularly concerned with modeling similarity judgments for weakly related or even dissimilar concepts, where knowledge-based resources offer limited coverage. Understanding these patterns is key for modeling human-like similarity judgments, expanding semantic coverage beyond taxonomic relations, and enhancing tasks such as semantic priming, episodic memory modeling, and robust semantic vector space embeddings.
Key finding: Demonstrates through four experiments that human similarity judgments among weakly related or apparently unrelated concepts are systematic and stable, contradicting the notion that such similarities are arbitrary. Introduces... Read more
Key finding: Develops a comprehensive Polish semantic priming dataset, revealing graded semantic priming effects and a linear relationship between semantic relatedness strength and reaction times. Evaluates semantic spaces against human... Read more
Key finding: Introduces PatternSim, a novel corpus-based semantic similarity measure derived from lexicosyntactic patterns extracted via finite-state transducers. Achieves a correlation up to 0.739 with human judgments without relying on... Read more
Key finding: Constructs Word Association Spaces (WAS) by applying singular value decomposition and multidimensional scaling on large free association norm datasets. Shows that WAS representations capture latent associative semantic... Read more
Key finding: Focuses on semantic similarity modeling within constrained, dynamic environments such as IoT and edge computing by proposing a new distributional profile model leveraging search engine data. Evaluates this model against... Read more
3. How can semantic similarity be operationalized and evaluated in applied NLP tasks involving human interpretability and textual entailment?
This theme addresses approaches combining semantic similarity measures with supervised learning and task evaluation to maximize interpretability and to support downstream applications such as semantic textual similarity (STS), textual entailment, and semantic search. It covers feature-rich models integrating lexical, syntactic, and semantic similarity metrics, and discusses the design and assessment of tests and benchmarks to evaluate how well computational similarity mimics human judgments, including the development of datasets and test tasks focusing on human interpretability and graded semantic equivalence.
Key finding: Compares three semantic similarity computation methods—cosine similarity with TF-IDF, cosine similarity with word embeddings, and soft cosine similarity with word embeddings—on short text similarity tasks. Finds that cosine... Read more
Key finding: Presents a supervised system combining lexical (bag-of-words cosine similarity), syntactic (BLEU score over base-phrases), and semantic similarity metrics (preservation of named entities and predicate-argument alignment)... Read more
Key finding: Develops an STS system leveraging textual entailment techniques and eight WordNet-based word-to-word similarity metrics to derive semantic similarity between sentence pairs. The system models text similarity as a function of... Read more
Key finding: Implements semantic textual similarity using distributional word representations constructed from large corpora via Random Indexing and Latent Semantic Analysis, including a novel vector permutation technique embedding... Read more
Key finding: Develops and validates a Semantic Similarity Judgment (SSJ) test specifically for Persian action verbs and non-action nouns to assess lexical-semantic processing deficits in brain injury patients. Uses expert semantic ratings... Read more