580 California St., Suite 400
San Francisco, CA, 94104
Key research themes
1. How do domain-specific string-based and ontology-informed methods improve biomedical sentence similarity measurement with reproducibility?
This research theme centers on the development and evaluation of sentence similarity methods tailored to the biomedical domain, emphasizing the importance of reproducible experimental setups. Due to the highly specialized vocabulary, complex syntactic structures, and abundant acronyms in biomedical texts, conventional sentence similarity models often underperform. The theme addresses the integration of string-based techniques and ontology-based semantic methods, the impact of preprocessing stages and Named Entity Recognition (NER) tools on method performance, and the establishment of reproducible resources and protocols to enhance experimental rigor and comparability.
Key finding: Introduces LiBlock, a novel aggregated string-based sentence similarity measure that significantly outperforms all evaluated state-of-the-art machine learning models and most ontology-based methods on multiple biomedical... Read more
Key finding: Proposes a hybrid similarity measure combining word embeddings with named-entity based semantic similarity, addressing semantic variation and noise in short biomedical texts. The approach achieves enhanced performance in... Read more
Key finding: Demonstrates that weighting sentence similarity calculations by noun phrase (NP) importance, as opposed to standard term frequency, significantly improves semantic similarity measures applicable to text categorization and... Read more
2. What are the impacts of lexical, syntactic, and semantic features, combined with vector-based distributional representations, on general domain sentence similarity?
This theme investigates the multifaceted role of lexical overlap, syntactic structure, and semantic frame alignment in modeling sentence similarity within general or cross-domain corpora. It explores supervised machine learning integration of these heterogeneous features, the utilization of distributional semantic models such as Random Indexing and Latent Semantic Analysis, and the embedding of syntactic and semantic information directly into vector representations (e.g., via vector permutations or recursive autoencoders). The research seeks to identify complementary strengths of diverse features to improve semantic textual similarity estimates beyond lexical matching alone.
Key finding: Employs a supervised regression model combining lexical metrics (word overlap and cosine similarity), syntactic similarity via BLEU scores over base-phrases, and semantic similarity based on named entity preservation and... Read more
Key finding: Utilizes distributional semantic spaces constructed via Random Indexing, Latent Semantic Analysis, and an innovative vector permutation method to inject syntactic information into representations. Results show that combining... Read more
Key finding: Compares multiple word embedding methodologies — recursive autoencoders, eigenword spectral methods, and selector generalizations — to generate word-level similarities that are aggregated at the sentence level through... Read more
3. How do hybrid and deep learning approaches integrating lexical relationships and sentence structure advance sentence similarity estimation?
This research focus explores the development of hybrid methodologies that combine deep learning models (e.g., CNNs, RNNs, BERT) with lexical knowledge-based techniques (e.g., WordNet) to measure semantic similarity between sentences. It emphasizes the need to incorporate lexical relationships, syntactic structures, word order, and semantic nuances such as determiners and negations. The goal is to improve similarity measures by capturing compositional semantic phenomena beyond simple lexical overlap, with particular attention to datasets and tasks where paraphrase detection and nuanced semantic differences are critical.
Key finding: Proposes a hybrid sentence similarity measurement method combining deep learning architectures (CNN, RNN, BERT) with lexical relationship analysis based on WordNet, integrating cosine similarity on embedding vectors and... Read more
Key finding: Introduces a challenging evaluation dataset emphasizing semantic differences stemming from word order swaps and determiner replacements in paraphrase detection. Results reveal that compositional distributional semantics... Read more
Key finding: Demonstrates that raw similarity scores between sentence pairs can be misleading for textual entailment prediction and proposes a supervised method that jointly learns non-linear transformations of multiple lexical similarity... Read more