AliMe Guide: This article explains the construction of knowledge graphs in the e-commerce domain from the perspective of demand analysis and systematic building, detailing the entire conceptual system formed during the process. It also emphasizes the significant efforts made by algorithm, engineering, product, operation, and outsourcing teams to gradually refine the platform architecture and review processes through continuous collaboration.
Authors: Yu Kun, Xi Qie, Yuan Shang, Hong Lang, Zi Yin, Jiu Yue
1. Background
The e-commerce cognitive graph has been evolving since its launch in June 2017, gradually forming a relatively complete e-commerce data cognitive system through continuous exploration from practice to systematization..
In the context of the group’s continuous expansion of business boundaries, the demand for data interconnectivity is becoming increasingly urgent, as it is the foundation for cross-domain search discovery, shopping guidance, and interaction, and is essential for allowing users to “browse” effectively. However, before this, we need to analyze the current problems.
1.1 Problems
The more complex data application scenarios are not limited to traditional e-commerce; we are now facing new retail, multi-language, and complex online-offline shopping scenarios. The data used often exceeds the previous textual scope and tends to have the following characteristics:
A vast amount of unstructured internet data is scattered across various sources, primarily represented in unstructured text form. The current category system has done extensive work from a product management perspective, yet it still only covers the tip of the iceberg of a large amount of data, which is far from sufficient to truly understand user needs.
Full of noise: Unlike traditional text analysis, most of the data within the group consists of queries, titles, comments, and guides. Due to user habits and merchant demands, this data exhibits significantly different syntactic structures compared to ordinary text, and there is a lot of noise and dirty data due to profit motives, making it extremely difficult to genuinely discover and structure user needs.
Multi-modal and multi-source: With the expansion of the group’s business, the current search recommendations not only encompass textual information within products but also utilize a large amount of video and image content. How to integrate data from various sources and associate multi-modal data is a significant challenge in data construction.
Data is scattered and cannot interconnect: From the current state of product system construction, different departments, due to rapid business development, often need to maintain their own cpv systems, which is crucial for subsequent product management and search. However, due to the different industry attributes of application scenarios, for example, the