Towards Holistic Contextual Understanding in Search: the Context-Aware Search Engine (CASE) Framework

Information retrieval (IR) systems generally start with a user-provided query that guides the search algorithm. The internal information structure across matching documents is used for relevance ranking. IR engines have been developed and optimized for several types of information sources, such as structured, semi-structured, and unstructured documents. These engines typically focus on the content within the documents being indexed and the structure of the overall information. The user’s context (the purpose of the query) is not an active consideration for the IR engine, making the user try various adjustments to the keywords in order to find the information they need. In real-world applications, (i) context manifests as an awareness of aspects surrounding the query such as geography, roles, capabilities, and other component attributes, and (ii) unfortunately often is not directly embedded in the metadata of the information within the documents being indexed. For this to be achieved, an IR engine needs three capabilities: an ability to index the information effectively, an ability to comprehend the search query, and the ability to understand the context of the search query. However, many recent techniques are only capable of deriving intrinsic context from domain-specific corpora and have yet to be integrated with extrinsic context such as a query’s location, or any other aspect. Moreover, existing search techniques are often general and sometimes discrete in nature. We present here a novel end-to-end Context-Aware Search Engine, named CASE, which provides three key features: (i) automating the indexing of documents by using a state-of-the-art contextual keyphrase extraction framework; (ii) understanding a query with both the intrinsic and extrinsic contexts via meta models and query-based knowledge graphs; and (iii) achieving robustness using a combined relevance ranking score. The results of our experiments demonstrate that CASE outperforms other baseline methods when compared to the machine reading comprehension (MS MARCO) dataset.