[Paper Review] Calculating the similarity between words and sentences using a lexical database and corpus statistics

Atish Pawar, Vijay Mago|arXiv (Cornell University)|Feb 15, 2018

Topic Modeling18 references63 citations

TL;DR

The paper proposes an edge-based semantic similarity method using WordNet and domain-specific corpus statistics to compute word and sentence similarity, achieving high correlations on standard benchmarks (Pearson ~0.8753 for words and ~0.8794 for sentences).

ABSTRACT

Calculating the semantic similarity between sentences is a long dealt problem in the area of natural language processing. The semantic analysis field has a crucial role to play in the research related to the text analytics. The semantic similarity differs as the domain of operation differs. In this paper, we present a methodology which deals with this issue by incorporating semantic similarity and corpus statistics. To calculate the semantic similarity between words and sentences, the proposed method follows an edge-based approach using a lexical database. The methodology can be applied in a variety of domains. The methodology has been tested on both benchmark standards and mean human similarity dataset. When tested on these two datasets, it gives highest correlation value for both word and sentence similarity outperforming other similar models. For word similarity, we obtained Pearson correlation coefficient of 0.8753 and for sentence similarity, the correlation obtained is 0.8794.

Motivation & Objective

• Improve semantic similarity measurement by combining lexical database structure with corpus statistics.
• Disambiguate word senses to improve accuracy in similarity calculations.
• Integrate word-level similarity with sentence structure to form robust sentence similarity measures.
• Demonstrate domain adaptivity through information content and corpus-driven statistics.

Proposed method

• Use WordNet to compute word similarity via shortest path distance and hierarchical information with an exponential decay function.
• Apply word sense disambiguation (max similarity) to select appropriate synsets for comparison.
• Incorporate hyponymy/hypernymy via hierarchical distance scaling with a hyperbolic function g(h).
• Optionally incorporate word information content from a domain corpus to make the measure domain-specific.
• Form dynamic semantic vectors for sentences by aligning words across sentences and computing vector magnitudes for similarity.
• Introduce a zeta normalization based on benchmark similarity values to scale the final sentence similarity.
• Provide an optional word order similarity component to account for syntactic arrangement when needed.

Experimental results

Research questions

• RQ1 How can WordNet-based edge distances be combined with corpus statistics to measure word similarity more accurately?
• RQ2 Can a dynamic, domain-adapted semantic representation improve sentence similarity beyond fixed-vocabulary approaches?
• RQ3 What is the impact of word sense disambiguation on the accuracy of word and sentence similarity in this framework?
• RQ4 How does the proposed method perform against Rubenstein & Goodenough benchmarks and mean human similarity datasets?

Key findings

• Word similarity achieving Pearson correlation of 0.8753 on Rubenstein & Goodenough benchmark.
• Sentence similarity achieving correlation of 0.8794 on the same benchmark.
• The method outperforms several prior models on standard benchmarks and mean human similarity datasets.
• The approach demonstrates robustness when integrated with domain-specific corpus statistics.