FEED2SEARCH: a framework for hybrid-molecule based semantic search

ABSTRACT

Adopting semantic technologies has proven several benefits for enabling a better representation of the data and empowering reasoning capabilities over it. However, there are still unresolved issues, such as the shift from heterogeneous documents to semantic data models and the representation of search results. Thus, in this paper, we introduce a novel FramEwork for hybrid molEcule-baseD SEmantic SEARCH, entitled FEED2SEARCH, which facilitates Information Retrieval over a heterogeneous document corpus. We first propose a semantic representation of the corpus, which automatically generates a semantic graph covering both structural and domain-specific aspects. Then, we propose a query processing pipeline based on a novel data structure for query answers, extracted from this graph, which embeds core information together with structural-based and domain-specific context. This provides users with interpretable search results, helping them understand relevant information and track cross document dependencies. A set of experiments conducted using real-world construction projects from the Architecture, Engineering and Construction (AEC) industry shows promising results, which motivates us to further investigate the effectiveness of our proposal in other domains.

KEYWORDS:

• Heterogeneous document corpus
• Ontology
• Tightly coupled semantic graph
• Semantic Information Retrieval

Disclosure statement

No potential conflict of interest was reported by the author(s).

Notes

1 Available at http://spider.sigappfr.org/linkedmdr/

2 For the sake of simplicity, we only focus on the image metadata with different levels of precision, its related visual features and semantic descriptors.

3 A sample of the ifcOWL ontology concepts with simplified properties and relations.

4 A Search Engine Results Page (SERP) is the list of results that a search engine returns in response to a user query.

5 At start time, the input value $Input(v_j)$ is set to the initial activation value of $v_j$, i.e. 1 for start nodes and 0 for others.

6 $Output(v_j)=\frac{Input(v_j)-Inpu{t}_{min}}{Inpu{t}_{max}-Inpu{t}_{min}}$ where $Inpu{t}_{min}$, $Inpu{t}_{max}$ are respectively the minimum and the maximum values of all nodes' inputs in the graph.

7 For ease of presentation, the indices of the hybrid molecules follow those of their hybrid edges

8 Available at http://spider.sigappfr.org/linkedmdr/

9 Available at: http://spider.sigappfr.org/linkedmdr/lmdr-annotator/

10 Available at https://protegewiki.stanford.edu/wiki/NavigOWL.

11 Excluding the pluggable Domain-Specific Layer as we focus on the invariant part of LinkedMDR across domains.

12 Unlike the case of the effectiveness evaluation of LMDR Annotator, here we equally emphasize false negatives and positives as it is usually considered in IR applications

13 We consider that the relevance of a hybrid molecule is the relevance of the information that the hybrid edge (the core of the molecule) holds.

14 https://bim4ren.eu/

15 https://www.iesve.com/

16 https://www.wisebim.fr/

Additional information

Funding

This study was co-funded by Univ. Pau & Pays Adour - LIUPPA, and the French institute Nobatek/INEF4 for the Energy Transition of the building.