Semantics-aware hybrid encoder for improved related conversations

A computer-implemented system for finding, in an online forum, conversing posts relevant to a query post, comprising:

• a subject encoding module that calculates a subject embedding vector of a subject of a query post received by a web server serving an online forum and subject embedding vectors of a set of conversing posts previously posted to the online forum, wherein each of the query post and the set of conversing posts includes a subject and a body, wherein a user wants to find other conversing post in the online forum that are relevant to the query post;
• a fine grained relevance scoring module that calculates a fine grain similarity score between the subject embedding vectors of the query post and the set of conversing posts, and that selects N2 documents from the set of conversing post with a highest fine grained relevance scorer with respect to the query post;
• a boosting module that boosts the fine grain similarity score of at least some of the N2 selected conversing posts based on one or more relevance metrics and selects N3 boosted conversing posts with a highest boosted fine grain similarity score from the N2 selected conversing posts as a list of conversing posts most relevant to the query post, wherein N3<N2;
• and
• a display device wherein the N3 boosted conversing post are displayed to the user by the web server,
• wherein the fine grain similarity score of the N2 selected conversing posts is boosted by a weighted sum of the one or more relevance metrics of each of the N2 selected conversing posts, wherein each metric is weighted by its relevance to the selected conversing post.

A computer-implemented method of retrieving online relevant conversing posts, comprising:

• receiving, by a web server serving an online forum, a query post from an inquirer using the online forum, wherein the online forum facilitates conversing posts from users on subjects that are relevant and irrelevant to the query post;
• computing, by a contextual similarity scoring module, a contextual similarity score between each conversing post of a set of conversing posts in the online forum with the query post, wherein the query post and each conversing post of the set of conversing posts includes a subject and a body, wherein the contextual similarity score is computed between the body of each of the set of conversing posts and the body of the query post, wherein N1 conversing posts of the set of conversing posts with a highest contextual similarity score are selected;
• and
• computing, by a fine grained similarity scoring module, a fine grained similarity score between an embedding of a subject of the query post and embeddings of a subject of each of the N1 selected conversing posts by applying a computer-implemented multi-lingual classifier to the subject of the query post and each of the N1 selected conversing posts where embeddings are obtained from the subject of the query post and from each of the N1 selected conversing posts, and calculating a similarity between the embedding of the subject of the query post and of each conversing post of the N1 selected conversing posts, wherein N2 conversing posts of the N1 selected conversing posts with a highest fine grained similarity score are selected, wherein N2<N1, wherein the N2 selected conversing posts are those conversing posts of the set of N1 conversing posts in the online forum that are relevant to the query post.