Exploiting Symmetry in Relational Similarity for Ranking Relational Search Results

Abstract

Abstract. Relational search is a novel paradigm of search which focuses on the similarity between semantic relations. Given three words (A, B, C) as the query, a relational search engine retrieves a ranked list of words D, where a word D ∈ D is assigned a high rank if the relation between A and B is highly similar to that between C and D. However, if C and D has numerous co-occurrences, then D is retrieved by existing relational search engines irrespective of the relation between A and B. To overcome this problem, we exploit the symmetry in relational similarity to rank the result set D. To evaluate the proposed ranking method, we use a benchmark dataset of Scholastic Aptitude Test (SAT) word analogy questions. Our experiments show that the proposed ranking method improves the accuracy in answering SAT word analogy questions, thereby demonstrating its usefulness in practical applications.

Cross-Language Latent Relational Search between Japanese and English Languages Using a Web Corpus

“…Therefore, to achieve high precision, the relational similarity between (A , B) and (C, D) should be measured using a well-defined method such as [Bollegala et al 2009a;Turney 2006], in which the relation between C and D is represented by lexical patterns that are in the same sentence with the pair (C, D). Goto et al [2010] propose a method for improving performance of latent relational search by exploiting the symmetries of relational similarity. Because the degree of relational similarity between (A , B) and (C, D) is similar to that between (B, A) and (D, C), Goto et al [2010] propose to incorporate the score of D in the reversed query In our previous work on monolingual latent relational search ], we propose a method for indexing entity pairs and relations to perform latent relational search in high speed.…”

“…Goto et al [2010] propose a method for improving performance of latent relational search by exploiting the symmetries of relational similarity. Because the degree of relational similarity between (A , B) and (C, D) is similar to that between (B, A) and (D, C), Goto et al [2010] propose to incorporate the score of D in the reversed query In our previous work on monolingual latent relational search ], we propose a method for indexing entity pairs and relations to perform latent relational search in high speed. Moreover, we apply the method in Bollegala et al [2009b] to recognize paraphrased lexical patterns in the same language.…”

“…-We propose a multi-lingual indexing method to index entity pairs and relationships in multiple languages and to utilize identical surface forms of named entities across languages for cross-language relational similarity measuring. -Following previous work on relational similarity measuring algorithms [Bollegala et al 2009a[Bollegala et al , 2009bTurney 2005Turney , 2006] and monolingual latent relational search Goto et al 2010;Kato et al 2009], we represent the semantic relations between two entities in an entity pair by lexical patterns of the context surrounding the two entities. We propose a novel two-phase clustering algorithm to capture the semantic similarity of paraphrased lexical patterns across two languages, Japanese and English.…”

“…There are several methods for answering latent relational search queries Goto et al 2010;Halskov and Barriere 2008;Kato et al 2009]. However, these methods focus on monolingual latent relational search, as they represent the semantic relations between two entities in an entity pair by terms or lexico-syntactic patterns from the context surrounding the two entities and compare them in the same language.…”

“…With enormous numbers of entities and relations extracted from the Web, relying only on traditional keyword-based Web search engines is inadequate to fulfill users' information needs. This motivates us to explore the latent relational search approach Goto et al 2010;Kato et al 2009], in which the retrieval process is based on semantic relations between entities, rather than only on keywords.…”