Semantics-Aware Hybrid Encoder for Forum Post Relevance
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for recommending relevant conversations in online forums rely on keyword matching and subject line comparison, failing to capture the meaning of complex queries and often suggesting irrelevant posts, leading to poor customer experience.
Innovation Solution
A computer-implemented method that encodes the context of conversations from both the body and subject of posts, using a 3-level hierarchical ranking mechanism and a semantics-oriented hybrid-search technique to provide more relevant recommendations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If keyword matching or frequency based matching is used to match user queries to documents, then the matching process is simple and fast, but the meaning of complex queries cannot be captured and relevant documents are not retrieved accurately
Solution Approach 1:
The patent segments the post into two distinct parts: subject line and body content. It applies different encoding strategies to each part - the subject is encoded using a subject encoder while the body is encoded using a body encoder. This segmentation allows the system to handle the limited information in subject lines separately from the contextual information in bodies, resolving the contradiction between simple matching and accurate meaning capture.
Solution Approach 2:
The patent introduces a hierarchical two-dimensional structure for encoding: the first dimension is the subject line encoding, and the second dimension is the body content encoding. By stacking these encodings hierarchically, the system captures both the topical information from subjects and the contextual information from bodies, achieving accurate relevance matching without oversimplification.
2Loss of information
If only the subject line is used to recommend semantically similar posts, then the recommendation process is simple, but the limited information in subject lines fails to capture the exact issue and useful information from post bodies
Solution Approach 1:
The patent merges the subject encoding and body encoding into a unified hierarchical representation. The subject encoder processes the subject line to extract topical information, while the body encoder processes the post body to extract contextual information. These two encodings are combined hierarchically to form a comprehensive representation that retains all useful information from both parts of the post.
Solution Approach 2:
The patent introduces a hierarchical two-dimensional structure for encoding: the first dimension is the subject line encoding, and the second dimension is the body content encoding. By stacking these encodings hierarchically, the system captures both the topical information from subjects and the contextual information from bodies, achieving accurate relevance matching without oversimplification.
3Speed
If text snippets are used for search instead of full text matching, then search speed is improved, but the text snippets do not match the user query and relevance is lost
Solution Approach 1:
The patent replaces traditional mechanical text matching mechanisms with semantic encoding mechanisms. Instead of comparing raw text snippets or keywords, the system uses neural network encoders to transform subjects and bodies into semantic representations. This substitution enables the system to capture the meaning of queries rather than just matching surface-level text, maintaining accuracy while enabling efficient semantic search.
Data Source
AI summary
A method of finding online relevant conversing posts, comprises receiving, by a web server serving an online forum, a query post from an inquirer using the online forum, computing a contextual similarity score between each conversing post of a set of conversing posts with a query post, wherein the contextual similarity score is computed between the body of each of conversing posts and of the query post, wherein N1 conversing posts with a highest contextual similarity score are selected; computing a fine grained similarity score between the subject of the query post and of each of the N1 conversing posts, wherein N2 conversing posts with a highest fine grained similarity score are selected; and boosting the fine grained similarity score of the N2 conversing posts based on relevance metrics, wherein N3 highest ranked conversing posts are selected as a list of conversing posts most relevant to the query post.