Automated news ranking and recommendation system

Patent metadata

Field	Value
Publication number	US-11334949-B2
Application number	US-202016779434-A
Country	US
Kind code	B2
Filing date	Jan 31, 2020
Priority date	Oct 11, 2019
Publication date	May 17, 2022
Grant date	May 17, 2022

Abstract

A framework for an automated news recommendation system for financial analysis. The system includes the automated ingestion, relevancy, clustering, and ranking of news events for financial analysts in the capital markets. The framework is adaptable to any form of input news data and can seamlessly integrate with other data used for analysis like financial data.

First claim

What is claimed is:

1. A computer-implemented method for ranking news articles in a news recommendation system, the method comprising:
creating a portfolio for a user, wherein the portfolio comprises subscriptions of the user to different entities;
ingesting news articles associated with the subscriptions of the user from a plurality of news sources;
converting each news articles into a one-hot vector based on named entities extracted from each news article;
determining a number of pairwise distances for the news articles;
clustering the news articles based on the number of pairwise distances;
selecting a representative news article for each cluster based on news publication date and news source significance;
determining a vector for each representative news article by modeling characteristics of word use and change on word use across linguistic context, wherein the vector presents semantic of each representative news article;
merging clusters with similar semantic based on the vectors determined for the representative news articles of clusters;
forming a set of ranked clusters using a machine learning model, including:
ranking news articles within each cluster based on trustworthiness and linking volume for each news source of news articles;
ranking clusters within the set of ranked clusters based on cluster size;
storing relational information between the set of ranked clusters, news stories of the news articles, and the subscriptions of the user to a database;
digitally presenting the set of ranked clusters in a graphical user interface of the subscription-based news system;
receiving, in response to a user input, feedback on the set of ranked clusters from the user; and
adjusting the machine learning model based on the feedback from the user.

2. The computer-implemented method of claim 1 , wherein ranking news articles within each cluster further comprises:
determining a news ranking score for each news article, wherein the news ranking score is a weighted sum of the entity relevance score and the source relevance score; and
ranking news articles within each cluster sequentially by publication date and the news ranking score.

3. The computer-implemented method of claim 2 , further comprising:
determining an entity relevance score for each news article, wherein the entity relevance score is based on a relevance of the news article to an entity mentioned in the news article.

4. The computer-implemented method of claim 3 , wherein determining the entity relevance score further comprises:
identifying main entities in each article using the machine learning model to independently derives features from the title, description, and content of each article, wherein identifying the main entries includes:
generating a first set of independent features by phrase matching of the title to identify a main entity; and
generating a second set of independent features by natural language processing of the description and content to identify the main entity, including n-gram modeling that counts a number of tokens that match with each n-grams of an entity name, and then weights the counts exponentially.

5. The computer-implemented method of claim 2 , further comprising:
determining a source relevance score for each news article, wherein the source relevance score is based on an empirically determined influence of a news source relative to influences of other news sources.

6. The computer-implemented method of claim 5 , wherein determining the source relevance score further comprises:
building a mapping from news sources to domains, including grouping News articles by news sources and extracting domains from the URLs of news articles;
for each news source, sequentially looking up the domains in a database of website popularity by frequency until a match is made; and
removing news articles published by news sources that rank below a popularity threshold in the database of website popularity.

7. The computer-implemented method of claim 1 , wherein ranking clusters within the set of ranked clusters further comprises:
determining a clustering ranking score, including:
identifying a maximum news ranking score for all the news within the cluster;
summing the weighted cluster size; and
ranking the clusters sequentially by update date, clustering ranking score, and cluster size.

8. A news ranking system, the system comprising:
a bus system;
a storage device connected to the bus system, wherein the storage device stores program instructions; and
a number of processors connected to the bus system, wherein the number of processors execute the program instructions:
to create a portfolio for a user, wherein the portfolio comprises subscriptions of the user to different entities;
to ingest news articles associated with the subscriptions of the user from a plurality of news sources;
to convert each news articles into a one-hot vector based on named entities extracted from each news article;
to determine a number of pairwise distances for the news articles;
to cluster the news articles based on the number of pairwise distances;
to select a representative news article for each cluster based on news publication date and news source significance;
to determine a vector for each representative news article by modeling characteristics of word use and change on word use across linguistic context, wherein the vector presents semantic of each representative news article;
to merge clusters with similar semantic based on the vectors determined for the representative news articles of clusters;
and
to form a set of ranked clusters, including:
to rank news articles within each cluster based on trustworthiness and linking volume for each news source of news articles;
to rank clusters within the set of ranked clusters based on cluster size;
to store relational information between the set of ranked clusters, news stories of news articles, and the subscriptions of the user to a database;
to digitally present the set of ranked clusters in a graphical user interface of the subscription-based news system;
to receive, in response to a user input, feedback on the set of ranked clusters from the user; and
to adjust the machine learning model based on the feedback from the user.

9. The news ranking system of claim 8 , wherein the number of processors further execute the program instructions:
to determine a news ranking score for each news article, wherein the news ranking score is a weighted sum of the entity relevance score and the source relevance score; and
to rank news articles within each cluster sequentially by publication date and the news ranking score.

10. The news ranking system of claim 9 , wherein the number of processors further execute the program instructions:
to determine an entity relevance score for each news article, wherein the entity relevance score is based on a relevance of the news article to an entity mentioned in the news article.

11. The news ranking system of claim 10 , wherein in determining the entity relevance score, the number of processors further execute the program instructions:
to identify main entities in each article using the machine learning model to independently derives features from the title, description, and content of each article, wherein identifying the main entries includes:
generating a first set of independent features by phrase matching of the title to identify a main entity; and
generating a second set of independent features by natural language processing of the description and content to identify the main entity, including n-gram modeling that counts a number of tokens that match with each n-grams of an entity nam

Assignees

• S&P Global Inc

Inventors

• Kim Lisa
• Ma Zhiqiang
• Bang Grace
• Wang Chong
• Singh Himani
• Kociuba Russell
• Pomerville Steven
• Liu Xiaomo

Classifications

• G06Q40/06 Primary

  Asset management; Financial planning or analysis · CPC title

• G06Q30/0282 Primary

  Rating or review of business operators or products · CPC title

• G06F18/24133

  Distances to prototypes · CPC title

• G06N3/045

  Combinations of networks · CPC title

• G06F18/24143

  Distances to neighbourhood prototypes, e.g. restricted Coulomb energy networks [RCEN] · CPC title