CATEGORY-BASED CONTENT RECOMMENDATION
Techniques for category-based content recommendation are described. Some embodiments provide a content recommendation system (“CRS”) configured to recommend content items (e.g., Web pages, images, videos) that are related to specified categories. In one embodiment, the CRS processes content items to determine entities referenced by the content items, and to determine categories related to the referenced entities. The determined entities and/or categories may be part of a taxonomy that is stored by the CRS. Then, in response to a received request that indicates a category, the CRS determines and provides indications of one or more content items that each have a corresponding category that matches the indicated category. In some embodiments, at least some of these techniques are employed to implement a category-based news service.
This application is a continuation of U.S. application Ser. No. 14/231,238, entitled “CATEGORY-BASED CONTENT RECOMMENDATION,” filed Mar. 31, 2014, which is a continuation of U.S. application Ser. No. 13/286,778, entitled “CATEGORY-BASED CONTENT RECOMMENDATION,” filed Nov. 1, 2011, now U.S. Pat. No. 8,725,739 issued May 13, 2014, which claims the benefit of U.S. Provisional Application No. 61/408,965, entitled, “CATEGORY-BASED CONTENT RECOMMENDATION,” filed Nov. 1, 2010, all of which are incorporated herein by reference in their entireties.
TECHNICAL FIELDThe present disclosure relates to methods, techniques, and systems for category-based content recommendation and, more particularly, to methods, techniques, and systems for recommending content items, such as news stories, that reference entities in an indicated category.
BACKGROUNDVarious approaches for automated categorization (or classification) of texts into predefined categories exist. One approach to this problem uses machine learning: a general inductive process automatically builds a classifier by learning, from a set of pre-classified documents that are represented as vectors of key terms, the characteristics of the categories. Various machine-learning techniques may be employed. In one approach, for each category, a set of human-labeled examples are collected as training data in order to build classifiers, such as Decision Tree classifiers, Naive Bayes classifiers, Support Vector Machines, Neural Networks, or the like. A separate classifier typically must be built for each new category. Such approaches also may not scale well when processing a large quantity of documents. For example, to add a new category, a new classifier may need to be built. Then, every document may need to be run through the resulting classifier.
In addition, various approaches to providing computer-generated news Web sites exist. One approach aggregates headlines from news sources worldwide, and groups similar stories together. The stories are grouped into a handful of broad, statically defined categories, such as Business, Sports, Entertainment, and the like. In some approaches, the presentation of news items may be customized, such as by allowing users to specify keywords to filter news items. However, such a keyword-based approach to customization may be limited because it can be difficult or impossible to express higher-order concepts with simple keywords. For example, if a user wishes to obtain articles about NBA basketball players, the term “NBA” may yield an over-inclusive result set, by including many articles that do not mention any basketball players. On the other hand, the terms “NBA basketball player” may yield an under-inclusive result set, by not including articles that do not include the specified keywords but that do mention some NBA basketball player by name.
Embodiments described herein provide enhanced computer- and network-based methods and systems for recommending content and more particularly, recommending content items in specified categories. Example embodiments provide a content recommendation system (“CRS”) configured to recommend content items such as articles, documents, videos, advertisements, product information, software applications/modules, and the like. Recommending content items may include determining content items that are in an indicated category, based on whether entities referenced by those content items are in or otherwise related to the indicated category. In some embodiments, the content items are news items that include timely information about recent or important events. Example news items may be documents that include news stories or reports, press releases, video news reporting, radio news broadcasts, or the like.
In some embodiments, the CRS indexes a corpus of content items by determining semantic information about the content items. Determining semantic information may include determining entities that are referenced by the content items. A content item may reference an entity by naming or describing the entity. Entities may include people, places (e.g., locations), organizations (e.g., political parties, corporations, groups), events, concepts, products, substances, and the like. For example, the sentence “Sean Connery starred as James Bond,” names the entities Sean Connery and James Bond. Table 1, below, includes a list of example entity types. Fewer or more entity types may be available.
Determining semantic information may further include determining categories that are associated with the determined entities. In some embodiments, entities may further be associated with (e.g., related to) one or more categories (also called “facets”). Example facets include actor, politician, athlete, nation, drug, sport, automobile, and the like. The entities and/or categories may be arranged in a semantic network, such as a taxonomic graph that includes relations between various categories and/or entities. In one embodiment, each entity is associated with at least one category, via an is-a relationship (e.g., Sean Connery is-a Actor), and each category may be associated with one or more other categories (e.g., an Actor is-a Person). The CRS may exploit the relationships or links in the semantic network to determine and store relevant categories associated with a particular content item. Table 2, below, includes a list of example categories/facets. Fewer, greater, or different categories may be incorporated or utilized.
In some embodiments, the CRS provides a search and discovery facility that is configured to recommend content items that include or otherwise are related to one or more specified categories. In one embodiment, the CRS receives an indication of a category, such as via a Web-based search interface. In response, the CRS determines (e.g., finds, selects, obtains, identifies) one or more content items that reference or are otherwise related to the specified category. The CRS may then rank or order the selected content items, such that more relevant content items appear before less relevant content items. The CRS then provides indications of the selected content items, such as by storing, transmitting, or forwarding the selected content items.
1. Overview of Category-Based Content Recommendation in One EmbodimentThe process begins at block 102, where it builds a repository of entities and concepts. In one embodiment, building the repository may include automatically identifying entities by processing structured or semi-structured data, such as may be obtained Wikipedia, Techcrunch, or other public or private data repositories, knowledge bases, news feeds, and the like. In other embodiments, unstructured text documents or other content items (e.g., audio data) may be processed to identify entities. As noted above, entities may be organized into taxonomic hierarchies, based on taxonomic relations such as is-a, part-of, member-of, and the like. In some embodiments, the entities are also associated with properties. Taxonomic paths and/or properties may be extracted from structured and semi-structured sources (e.g., Wikipedia). An example taxonomic hierarchy is illustrated with respect to
At block 104, the process determines a ranked list of entities for each content item in a corpus of content items. In some embodiments, the process uses entity tagging and disambiguation to link references to entities that occur in the text of a content item to entries in the repository of entities and concepts generated at block 102. Then, for each content item, the process determines a ranked list of entities, ordered by their importance and relevance to the main subject/topic of the content item.
At block 106, the process assigns categories to each content item in the corpus of content items, based on the ranked list of entities determined at block 104. The categories may be or include any node or path in a semantic network and/or a taxonomic graph, or any properties that may be shared by a group of entities (e.g., Pac-10 conference teams, University of Washington Huskies football players, left-handed baseball pitchers, rookie football quarterbacks). The assigned categories may be based on groups of entities or entity types, grouped based on their taxonomic paths and/or any selected properties. Assigning categories to a content item may further include storing the determined categories in an inverted index or other type of data structure for efficient retrieval at a later time. In other embodiments, the process assigns categories to only some of the content items in the corpus, leaving content items without an assigned category. This may occur for various reasons, such as because no category can be determined, for performance reasons (e.g., only recent content items are processed), or the like.
At block 108, the process determines content items that are relevant to an indicated category. In one embodiment, the CRS provides a search engine facility that can answer queries requesting content items related to one or more specified categories. Thus, determining relevant content items may include finding content items that match or are otherwise related to at least one of the specified categories. The determined content items may be ranked by factors such as source credibility, popularity of the topic, recency, or the like. The determined content items may then be provided (e.g., transmitted, sent, forwarded, stored), such as in response to a received search query or other request.
Additional related information may also or instead be determined that this time, including producing a summarization of the specified category by producing a list of current and popular entities in the specified category. For example, given the category “NCAA basketball”, the process may determine and return teams and players that are most popular in the news recently. Such additional information may then also be provided in response to a received search query or other request.
At block 110, the process determines popular and/or emerging categories. In one embodiment, the CRS is configured to provide a feed, stream, or other dynamic collection, such as may be part of an automated news service, where news items are organized by popular and/or emerging categories. Popular and/or emerging categories may be automatically determined based on popularity ranking and detection of emerging categories, by aggregating popularity measures of entities and documents that belong to each category. The source of entity popularity measures may be page views, user clicks, recent (e.g., last day/week) mentions in content items, Wikipedia traffic, Twitter references, and the like.
In addition, the CRS may also determine popular and emerging entities within a category. For example, for a category such as NBA Basketball, the CRS may determine entities (e.g., players, teams, coaches) that have recently received significant press coverage, such as NBA basketball teams (and their players) that are playing each other in a current championship series or other event that has recently (e.g., within the last day, week, month) received additional press coverage.
The CRS may identify, organize, and/or present content items in other or additional ways. For example, within a given category, the CRS may identify content items that reference or are related to particular events. For example, in a sports category, the CRS may automatically identify, group, and/or present content items about various sports-related events, such as injuries within particular sports (e.g., injuries suffered by football players), awards (e.g., player of the year), player-team engagements (e.g., contract renewals), or the like. Or, in an entertainment category, the CRS may automatically identify content items about common gossip events, such as romances, break-ups, movie openings, or the like.
2. Functional Elements of an Example Content Recommendation SystemThe content ingester 211 receives and indexes content from various content sources 255, including sources such as Web sites, Blogs, news feeds, video feeds, and the like. The content ingester 211 may also receive content from non-public or semi-public sources, including subscription-based information services, access-controlled social networks, and the like. The content ingester 211 provides content information, including data included within content items (e.g., text, images, video) and meta-data about content items (e.g., author, title, date, source), to the entity and relationship identifier 212. The content information may be provided directly (as illustrated) and/or via some intermediary, such as the content index 217a.
The entity and relationship identifier 212 determines semantic information about content items obtained from the various content sources 255, and stores the determined information in the data store 217. More specifically, the entity and relationship identifier 212 receives content information from the content ingester 211 and identifies entities and relationships that are referenced therein. Various automatic and semi-automatic techniques are contemplated for identifying entities within content items. In one embodiment, the identifier 212 uses natural language processing techniques, such as parts of speech tagging and relationship searching, to identify sentence components such as subjects, verbs, and objects, and to identify and disambiguate entities. Example relationship searching technology, which uses natural language processing to determine relationships between subjects and objects in ingested content, is described in detail in U.S. Pat. No. 7,526,425, filed Dec. 13, 2004, and entitled “METHOD AND SYSTEM FOR EXTENDING KEYWORD SEARCHING FOR SYNTACTICALLY AND SEMANTICALLY ANNOTATED DATA” issued on Apr. 28, 2009, and example entity recognition and disambiguation technology is described in detail in U.S. patent application Ser. No. 12/288,158, filed Oct. 15, 2008, and entitled “NLP-BASED ENTITY RECOGNITION AND DISAMBIGUATION,” both of which are incorporated herein by reference in their entireties. Amongst other capabilities, the use of relationship searching, enables the CRS 200 to establish second order (or greater order) relationships between entities and to store such information in the data store 217.
For example, given a sentence such as “Sean Connery starred in Goldfinger,” the identifier 212 may identify “Sean Connery” as the sentence subject, “starred” as the sentence verb (or action), and “Goldfinger” as the sentence object, along with the various modifiers present in the sentence. These parts-of-speech components of each sentence, along with their grammatical roles and other tags may be stored in the relationship index 217c, for example as an inverted index as described in U.S. Pat. No. 7,526,425. As part of the indexing process, the CRS recognizes and disambiguates entities that are present in the text. Indications of these disambiguated entities are also stored with the sentences information, when the sentence contains uniquely identifiable entities that the CRS already knows about. These entities are those that have been added previously to the entity store 217b. In some cases, the indexed text contains subjects and objects that indicate entities that are not necessarily known or not yet disambiguated entities. In this case the indexing of the sentence may store as much information as it has in index 217c, but may not refer to a unique identifier of an entity in the entity store 217b. Over time, as the CRS encounters new entities, and in some cases with the aid of manual curation, new entities are added to the entity store 217b. In the above example, “Sean Connery” and “Goldfinger” may be unique entities already known to the CRS and present in the entity store 217b. In this case, their identifiers will be stored along with the sentence information in the relationship index 217c. The identified verbs also define relationships between the identified entities. These defined relationships (e.g., stored as subject-action-object or “SAO” triplets, or otherwise) are then stored in the relationship index 217c. In the above example, a representation of the fact that the actor Sean Connery starred in the film Goldfinger would be added to the relationship index 217c. In some embodiments, the process of identifying entities may be at least in part manual. For example, entities may be provisionally identified by the identifier 212, and then submitted to curators (or other humans) for editing, finalization, review, and/or approval.
The content index 217a associates content items with one or more entities and categories, and vice versa, in order to support efficient searches such as searches for content items having a particular entity or for categories associated with a particular content item. For example, given an entity or category, the CRS 200 may provide a list of content items that reference that facet. In addition, given an indication of a content item, the CRS may provide a list of entities or categories referenced by that content item.
The entity store 217b is a repository of entities (e.g., people, organization, place names, products, events, things), concepts, and other semantic information. In at least some embodiments, the entities in the entity store 217b are related such that they form a semantic network, taxonomy, or graph. The entities in the entity store 217b are associated with categories/facets. The categories themselves are organized into one or more taxonomies based on taxonomic relations such as is-a, part-of, member-of, and the like. In addition, entities are associated with certain properties, such as name and aliases, a unique identifier, types and facets, descriptions, and the like. Entities may also have type/facet-specific properties. For example, for a sports athlete, common properties may include: birth place, birth date, sports teams, player positions, awards, and the like. Note that some of the properties are relational, that is, the property value may itself be another entity in the entity store 217b. For example, the team property for an athlete may be link to a sports team entity in the entity store 217b, and vice versa. Thus, the entities in the entity store 217b are interconnected through the property links, creating a semantic network or graph. Certain taxonomic relations are represented as such property links (e.g., the “member-of” relation for the players-team relation, and team-league relation in the sports domain). In some embodiments, the entities, their taxonomic paths and/or properties are extracted from one or more structured and semi-structured sources (e.g., Wikipedia). In other embodiments, the process of identifying entities may be at least in part manual. For example, entities may be provisionally identified by the content ingester 211, and then submitted to humans for editing, finalization, review, and/or approval.
The category identifier 213 determines category-related semantic information about content items obtained from the various content sources 255, and stores the determined information in the data store 217. More specifically, the category identifier 213 receives content information from the content ingester 211 and entity and relationship information from the entity and relationship identifier 212, and determines categories associated with the entities referenced by content items. The category identifier 213 thus associates categories with content items, and store such associations in the data store 217, such as by annotating content items stored in the content index 217a. The category identifier 213 may perform other or additional category-related functions, such as identifying popular or trending categories, summarizing categories by determining popular entities in the category, or the like.
The content recommender 214 provides indications of content items in response to a request received from a user 202 or a device operated by the user 202. In one embodiment, the content recommender 214 provides an interface (e.g., a Web-based interface, an application program interface) that receives requests/queries that specify one or more categories. In response, the content recommender 214 determines content items that are related to at least one of the one or more categories, and provides (e.g., transmits, sends, forwards) indications of the determined content items. In another embodiment, the content recommender 214 operates in a “push” model, where it provides a stream or feed of content items related to one or more categories.
The optional other content recommender 215 provides recommendations of other types of content obtained from or provided by third-party services/sources. In some embodiments, the recommender 215 may query third-party services to retrieve other media types (e.g., videos, podcasts, social media messages) that may not be included in the content index 217a. In one embodiment, the recommender 215 may, given a specified category, automatically construct a query adapted for a third-party information/content service by taking the top entities (e.g., top three) from a list of current and popular entities for the specified category.
In addition, although the described techniques for content recommendation are illustrated primarily with respect to textual content, other types of content are contemplated. In one embodiment, the CRS 200 may utilize at least some of the described techniques to perform or facilitate the category-based recommendation of content items based on other types of content, including advertisements, audio (e.g., music), video, images, and the like.
In some embodiments, the CRS 200 is configured to ingest video streams (e.g., live streaming of sports games) in a similar fashion. In particular, the CRS 200 may obtain text content from the stream via either closed captions or speech recognition. Then, the CRS 200 analyzes the obtained text content as discussed above, such that the CRS 200 can provide category-based recommendations for such content items as well.
Furthermore, the described techniques are not limited to the specific architecture shown in
Although the techniques of category-based content recommendation and the CRS are generally applicable to any type of content item, the phrase “content item” is used generally to refer to or imply any type of information and/or data, regardless of form or purpose. For example, a content item may be in textual or binary format, or a content item may be a news item, a report, an image, an audio source, a video stream, a code module (e.g., an application, an executable), an online activity (e.g., to purchase a good or service), or the like. Essentially, the concepts and techniques described are applicable to any category-based recommendation system. Also, although certain terms are used primarily herein, other terms could be used interchangeably to yield equivalent embodiments and examples. For example, the term “category” and “facet” are used interchangeably. Other terms for category may include “class,” “property-based set,” or the like. In addition, terms may have alternate spellings which may or may not be explicitly mentioned, and all such variations of terms are intended to be included.
Example embodiments described herein provide applications, tools, data structures and other support to implement a content recommendation system to be used for recommending content items, such as news items, that belong to a particular category. Other embodiments of the described techniques may be used for other purposes, including for category-based recommendation of technology reports (e.g., reviews of computer systems, programs, games, mobile devices). In the following description, numerous specific details are set forth, such as data formats and code sequences, etc., in order to provide a thorough understanding of the described techniques. The embodiments described also can be practiced without some of the specific details described herein, or with other specific details, such as changes with respect to the ordering of the code flow, different code flows, etc. Thus, the scope of the techniques and/or functions described are not limited by the particular order, selection, or decomposition of steps described with reference to any particular routine.
3. Example Screen Displays for Category-Based News Item RecommendationThe news items area 306 includes multiple news item snippets, such as snippets 308 and 310. Each snippet provides information about a news item, such as the title of the news item, some of text from the news item, the source of the news item, or the like. For example, snippet 308 provides information about an AP Online story, including its title (“FCC opens up unused TV signals for broadband”) and some of its text (“The Federal Communications Commission is opening up unused airwaves between television stations . . . ”). Note that the news item presented via snippet 308 does not mention the term politics. Rather, the CRS has determined that various entities referenced in the news item (e.g., the FCC) are entities that are in some manner related to the Politics category. Snippet 310 provides information about a social media message, in this case a post from the Twitter micro-blogging service. Again, the underlying social media message presented via snippet 310 does not include the term politics. Instead, the CRS has determined that the referenced entity Timothy Geithner is related to the Politics category.
Although the category-based content recommendation techniques of
In
Next, the CRS ranks the tagged entities by their importance and relevance to the main subject of the content item. The ranking may be based on one or more of the following factors: number of mentions (e.g., references) of each entity in the text; positions of the mentions in the text (e.g., entities appearing in document title may be weighted more; entities appearing earlier in the text would be weighted more than the ones appearing later in the text; entities appearing in boilerplate text may be weighted less); and penalties to certain types of entities (e.g., if the publisher of the document appears in the text, it may be weighted less).
Nodes in the graph 430 also have associated levels, where the root is at the first level, categories at the level of category 434a are at the second level, categories at the level of category 436a are at the third level, and so on. When levels are counted starting from one, the level of a category is defined as one plus the number of arcs/links that must be traversed to reach the root category.
A taxonomic path is a path between one category and another in the graph 430. For example, the path connecting categories 432, 434a, 436a, and 438a form a taxonomic path that specifies the Basketball Team category as well as all of its ancestor categories up to the root of the graph 430. The path connecting categories 432, 434a, 436a, and 438a, may also be denoted textually as: Evri/Organization/Sports/Basketball_Team.
In the processing of the example content item shown in
In the illustrated example, the above ranking results due to the relative number of mentions of the entities in the news item of
In the next step, the CRS assigns categories to the content item, based on the ranked list of entities. The CRS takes the top K entities from the ranked list. In one implementation, K=5. By taking the top K entities, the CRS includes only the entities that are relatively important and relevant to the document. Then, the CRS derives categories by aggregating the common nodes in the taxonomic paths of the top entities, as well as their properties, as described below.
First, the CRS counts the third-level nodes in the taxonomic paths, which correspond to the major topic domains (e.g., Politics, Entertainment, Sports, Business, Technology, Heath, Food, etc.). In
Second, the CRS ranks the leaf nodes (facets) of the taxonomic path, based on the number of entities having the facet and/or the rank of the entity in the ranked entity list. In the above example, the ranked list of facets would be: (1) Basketball Team (because the Trail Blazers and Lakers are both basketball teams), (2) Basketball Player (because Martell Webster is a basketball player), and (3) Sports Venue (because the Staples Center is a sports venue). Those facets would be assigned to the document as facet categories.
Other or additional processing may also be performed. For example, any levels or individual nodes/paths in the taxonomy may be selected as a category, and the same entity aggregation process can then be applied. For example, the CRS may assign a category “entertainment_people” by aggregating entities that share taxonomic path [Evri/Person/Entertainment]. Similarly, by aggregating on path [Evri/Person/Sports/Athlete], the CRS would create an “athletes” category that covers players in all kinds of sports.
Besides nodes/paths in the taxonomy, the CRS may also support configurable categories based on entity properties. This allows the CRS to create categories from any group of entities that share a common property. A configuration is a set of key-value pairs, where the key is a facet, and the value is a list of properties to be checked. For example, a basketball player entity may have the following configuration: Facet (key)=Basketball Player and Properties (value)=Sports League.
During processing of a document, the CRS processes the list of entities assigned to the content item. For each entity, the CRS takes its facets, and performs a lookup against its configuration. If there is a match, the CRS obtains the property list from the configuration. Then, the CRS retrieves the entity's corresponding entry from the entity store. The CRS then processes the entity's properties available in the entity store, and determines if the entity has the property specified in the configuration. If a match is found, the CRS adds the property value as a potential category.
For example, leaf nodes in the example taxonomy of
-
- Portland Trail Blazers—Sports_League=NBA
- Los Angeles Lakers—Sports_League=NBA
- Martell Webster—Sports_League=NBA
Again, by aggregating the specified properties from the ranked list of entities, the CRS extracts the top categories (NBA basketball in this case), and assigns them to the content item, as discussed further below.
As described above, in one embodiment, certain categories are always detected by default, while others can be specified through a configuration. To decide what categories to populate in the index, human curators may select the categories manually based on their knowledge of popular topics in the news as well as scheduled events (e.g., Oscars Academy Award, March madness college basketball). Another approach includes automated popularity ranking and detection of emerging categories, by looking at popularity of individual entities and aggregating to the category level. The sources of entity popularity may include one or more of:
-
- Frequency counts in an index of news articles, as well as a corresponding timeline used in order to detect emerging entities that have rising popularity.
- User clicks on entities.
- Wikipedia page view counts (e.g., Wikipedia traffic statistics on an hourly basis are publicly available from sources such as http://stats.grok.se).
- Number of mentions in Tweets or other social media messages, obtained through the Twitter API, for example.
- Popular/trending queries from search engines such as Google and Yahoo.
In one embodiment, for the underlying information retrieval capability, the CRS uses a typical Vector Space Model (“VSM”) based system. An example of such systems is Apache Lucene. The vector space model procedure can be divided in to three stages. The first stage includes document indexing where content bearing terms are extracted from content item text and indexed/stored. The second stage includes weighting of the indexed terms to enhance retrieval of content items relevant to a user. A common weighting scheme for terms within a content item is based on the frequency of term occurrence. The last stage includes ranking the content items with respect to a query according to a similarity measure.
In one embodiment, the CRS creates a special term for each category that is to be assigned to a content item in the index. As discussed above, during the indexing of a content item, the CRS identifies the top K entities in the content item, collects the facets assigned to the top entities, and then annotate the content item with those facets using special tokens (facet.basketball.team for the category/facet Basketball_Team). The tokens are treated as single terms in the index. During search time, given a facet category Basketball_Team as input, the CRS translates it into a query term “facet.basketball.team.” Content items containing the query term can then be located and returned.
The CRS may retrieve and rank content items in the index by determining how relevant a given content item is to a user's query. In one embodiment, the more times a query term appears in a content item relative to the number of times the term appears in all the content items in the collection, the more relevant that content item is to the query. This is accomplished by using TF-IDF (term frequency-inverse document frequency) weighting in the scoring of content items. As noted above, the CRS creates artificial tokens for the categories extracted from the content item. These tokens are thus not terms originally appearing in the content item text. However, the CRS can simulate the TF weights on the category tokens, by controlling how many times each token is added to the index. The frequency may be determined based on the importance measure of the categories that are computed when ranking the categories. Another option to use a binary document vector (e.g., true/false to indicate whether the document is assigned a particular category). With this approach, the CRS takes the top K categories from the document, and consider them as equally important, such that each category token is added to the index once.
In addition, some embodiments can influence search results by “boosting.” The CRS may apply document level boosting, by specifying a boost value for a content item as it is added to the index. The boost value for each content item is computed as a combination of the item's source credibility (e.g., articles from more credible or popular sources are weighted more) and publishing date (e.g., fresher or more recent articles are weighted more).
The CRS may be configured to retrieve recently published content items from credible sources. In order to do so, the CRS may assign to each content item source a credibility score (e.g., 1 to 5), which is translated into a credibility weight. Source credibility can be assigned manually and/or computed based on some properties of a source (e.g., search engine rank of a source web page, Internet traffic to the source web page, etc.). Credibility weight can then be multiplied by recency weight to form a boost value for a content item. Recency weight may be based on the difference between the publishing date of a document and the date when ingestion started.
6. Category-Based Content Item RetrievalIn one embodiment, for a given category, the CRS returns a list of relevant and recent content items that are associated with the category in the index. The relevance ranking is based on factors such as: query matching, based on TF-IDF weighting as described above; source credibility; and/or publication date.
In one embodiment, the CRS generates a news stream for a given category by providing popular, fresh, and or breaking news items that are associated with the category. Popular or breaking news stories tend to be covered by multiple sources, get updated more frequently, and have more follow-up stories. Thus, the CRS may capture popular stories by determining its coverage in terms of the number of similar articles.
After retrieving a list of relevant news articles, the CRS may collapse the similar articles that cover the same topic into a group. The similarity between two articles is computed based on one or more of: distance between document signatures computed during index; overlap between article titles; overlap between article summaries; source URLs; and/or publisher. If the similarity measure exceeds some threshold, then two articles are considered as similar.
In one embodiment, the CRS detects similar documents by computing a signature (e.g., a hash key of fixed length) of a document summary. A document summary is a snippet of the document text that best represents the document subject and best matches the search query. To determine a document summary, the CRS may scan through the document text, and select a snippet of a pre-determined window size, by maximizing a combination of the following measures: the density of the document's top entities appearing in the snippet and overlap of the words in the snippet with the words in the document title.
In one embodiment, the logic for determining whether articles A and B are similar is as follows: if the article URLs are identical, return true; else if the Hamming distance between the two article signatures is lower than a pre-determined threshold, return true; else if overlap between article titles is higher than a threshold, and overlap between article summaries is higher than a threshold, return true; else, return false.
By using the above logic, articles that cover the same story are grouped together. In one embodiment, each group contains a head article, and underneath is zero or more similar articles. Article groups may then be ranked by various factors, including one or more of: the number of articles in each group (e.g., the larger the group size, usually the more popular the story is); publication dates of the articles, preferring recently published articles; and/or source diversity of the articles, preferring stories covered by many different sources.
As discussed above, some embodiments of the CRS may also determine popular, trending, or recent entities for a given category. In one embodiment, in addition to retrieving news articles, the CRS also produces a summary list of entities/concepts that are most popular at a current time for the given category. The popular entities may be extracted and ranked based on entity occurrences in the top articles returned for the given category. In one embodiment, the CRS iterates through the top articles in a result set, takes the top entities from each article (top entities for each article are stored in the index), and aggregates the occurrence frequency of each entity. Popular entities may be further extracted and ranked based on entity popularity scores in an entity store. From the entity store, the CRS may retrieve a list of entities that belong to the specified category, ranked by their scores. The score for each entity may be computed based on the total number of recent articles (in the content store) that mention the entity and/or user visit traffic to the entity's page (e.g., via the CRS and/or external sites such as Wikipedia).
7. Third-Party Content RetrievalAs noted above, the CRS may be configured to return, in response to a search or other request, indications of content items from external (e.g., third-party) sources. Third-party sources may include or provide content items of various media types (e.g., images, videos, audio, social media messages), some of which may not be indexed by the CRS. The CRS may be configured to retrieve results via external APIs or other retrieval facilities provided by the third-party sources. In some cases, a third-party source may not have or utilize the same category-based approach used by the CRS. In such cases, the CRS may translate categories used in the CRS to an approximation query that would retrieve from the third-party source content items similar to those that could be retrieved via the CRS itself. In one embodiment, categories may be translated into keywords used as part of a search query. However, such an approach may not provide good result coverage. This problem may be exacerbated for certain media types, such as images, videos, and social media messages (e.g., tweets), where the textual content is typically very short. For example, using keyword query “NBA Basketball” may not find videos or tweets that are about championship games between the Los Angeles Lakers and Boston Celtics, where the words “NBA” and “basketball” are not mentioned in the text, even though such content items are clearly related to NBA Basketball.
In another approach, for a given category, the CRS generates a keyword query where the keywords are based on the entities related to the given category. In particular, the CRS takes the top entities (e.g., most popular and recent) within the category, and uses the names and/or aliases of such entities to construct keyword queries against the third-party service. Thus, if the category is NBA Basketball, the CRS may construct a query that includes names of multiple (e.g., five) popular NBA Basketball entities, such as the Los Angeles Lakers, Boston Celtics, LeBron James, and the like. Queries may be enhanced using techniques similar to those mentioned in U.S. Patent Application No. 61/256,851, filed Oct. 30, 2009, and entitled “IMPROVING KEYWORD-BASED SEARCH ENGINE RESULTS USING ENHANCED QUERY STRATEGIES,” incorporated herein by reference in its entirety.
In addition, in some cases third-party sources may use descriptors or other indicators (e.g., tags) that can be mapped, translated, or otherwise related to categories utilized in the CRS. For example, a video sharing service may have “channels” or other higher-level grouping constructs that can be mapped (manually or automatically) to CRS categories. As another example, a social messaging service (e.g., Twitter) may provide streams from credible sources that can be related to one or more CRS categories. In general, such mappings may be compiled or generated through automated discovery and/or manual curation. Once a mapping has been established, the CRS can generate a query for a third-party source by translating an indicated category into a corresponding descriptor used by the third-party source.
8. Example Computing System and ProcessesNote that one or more general purpose or special purpose computing systems/devices may be used to implement the content recommendation system 510. In addition, the computing system 500 may comprise one or more distinct computing systems/devices and may span distributed locations. Furthermore, each block shown may represent one or more such blocks as appropriate to a specific embodiment or may be combined with other blocks. Also, the content recommendation system 510 may be implemented in software, hardware, firmware, or in some combination to achieve the capabilities described herein.
In the embodiment shown, computing system 500 comprises a computer memory (“memory”) 501, a display 502, one or more Central Processing Units (“CPU”) 503, Input/Output devices 504 (e.g., keyboard, mouse, CRT or LCD display, and the like), other computer-readable media 505, and network connections 506. The content recommendation system 510 is shown residing in memory 501. In other embodiments, some portion of the contents, some or all of the components of the content recommendation system 510 may be stored on and/or transmitted over the other computer-readable media 505. The components of the content recommendation system 510 preferably execute on one or more CPUs 503 and recommend content items, as described herein. Other code or programs 530 (e.g., an administrative interface, a Web server, and the like) and potentially other data repositories, such as data repository 520, also reside in the memory 501, and preferably execute on one or more CPUs 503. Of note, one or more of the components in
The content recommendation system 510 interacts via the network 550 with content sources 555, third-party applications 565, and client computing devices 560. The network 550 may be any combination of media (e.g., twisted pair, coaxial, fiber optic, radio frequency), hardware (e.g., routers, switches, repeaters, transceivers), and protocols (e.g., TCP/IP, UDP, Ethernet, Wi-Fi, WiMAX) that facilitate communication between remotely situated humans and/or devices. The devices 560 include desktop computers, notebook computers, mobile phones, smart phones, tablet computers, personal digital assistants, and the like.
In a typical embodiment, the content recommendation system 510 includes a content ingester 511, an entity and relationship identifier 512, a category identifier 513, a user interface manager 515, a content recommendation system application program interface (“API”) 516, and a data store 517. The modules 511-514 respectively perform functions such as those described with reference to modules 211-214 of
The UI manager 515 provides a view and a controller that facilitate user interaction with the content recommendation system 510 and its various components. For example, the UI manager 515 may provide interactive access to the content recommendation system 510, such that users can search for content items related to specified categories. In some embodiments, access to the functionality of the UI manager 515 may be provided via a Web server, possibly executing as one of the other programs 530. In such embodiments, a user operating a Web browser executing on one of the client devices 560 can interact with the content recommendation system 510 via the UI manager 515. For example, a user may manually submit a search for content items related to a specified category.
The API 516 provides programmatic access to one or more functions of the content recommendation system 510. For example, the API 516 may provide a programmatic interface to one or more functions of the content recommendation system 510 that may be invoked by one of the other programs 530 or some other module. In this manner, the API 516 facilitates the development of third-party software, such as user interfaces, plug-ins, news feeds, adapters (e.g., for integrating functions of the content recommendation system 510 into Web applications), and the like.
In addition, the API 516 may be in at least some embodiments invoked or otherwise accessed via remote entities, such as code executing on one of the client devices 560 or as part of one of the third-party applications 565, to access various functions of the content recommendation system 510. For example, an application on a mobile device may obtain recommended content items for a specified category via the API 516. As another example, one of the content sources 555 may push content information to the content recommendation system 510 via the API 516. The API 516 may also be configured to provide recommendation widgets (e.g., code modules) that can be integrated into the third-party applications 565 and that are configured to interact with the content recommendation system 510 to make at least some of the described functionality available within the context of other applications.
The data store 517 is used by the other modules of the content recommendation system 510 to store and/or communicate information. In particular, modules 511-516 may use the data store 517 to record various types of information, including semantic information about content items, such as entities, categories, and relationships. Although the modules 511-516 are described as communicating primarily through the data store 517, other communication mechanisms are contemplated, including message passing, function calls, pipes, sockets, shared memory, and the like.
In an example embodiment, components/modules of the content recommendation system 510 are implemented using standard programming techniques. For example, the content recommendation system 510 may be implemented as a “native” executable running on the CPU 503, along with one or more static or dynamic libraries. In other embodiments, the content recommendation system 510 may be implemented as instructions processed by a virtual machine that executes as one of the other programs 530. In general, a range of programming languages known in the art may be employed for implementing such example embodiments, including representative implementations of various programming language paradigms, including but not limited to, object-oriented (e.g., Java, C++, C#, Visual Basic.NET, Smalltalk, and the like), functional (e.g., ML, Lisp, Scheme, and the like), procedural (e.g., C, Pascal, Ada, Modula, and the like), scripting (e.g., Perl, Ruby, Python, JavaScript, VBScript, and the like), and declarative (e.g., SQL, Prolog, and the like).
The embodiments described above may also use either well-known or proprietary synchronous or asynchronous client-server computing techniques. Also, the various components may be implemented using more monolithic programming techniques, for example, as an executable running on a single CPU computer system, or alternatively decomposed using a variety of structuring techniques known in the art, including but not limited to, multiprogramming, multithreading, client-server, or peer-to-peer, running on one or more computer systems each having one or more CPUs. Some embodiments may execute concurrently and asynchronously, and communicate using message passing techniques. Equivalent synchronous embodiments are also supported. Also, other functions could be implemented and/or performed by each component/module, and in different orders, and by different components/modules, yet still achieve the described functions.
In addition, programming interfaces to the data stored as part of the content recommendation system 510, such as in the data store 517, can be available by standard mechanisms such as through C, C++, C#, and Java APIs; libraries for accessing files, databases, or other data repositories; through scripting languages such as XML; or through Web servers, FTP servers, or other types of servers providing access to stored data. The data store 517 may be implemented as one or more database systems, file systems, or any other technique for storing such information, or any combination of the above, including implementations using distributed computing techniques.
Different configurations and locations of programs and data are contemplated for use with techniques of described herein. A variety of distributed computing techniques are appropriate for implementing the components of the illustrated embodiments in a distributed manner including but not limited to TCP/IP sockets, RPC, RMI, HTTP, Web Services (XML-RPC, JAX-RPC, SOAP, and the like). Other variations are possible. Also, other functionality could be provided by each component/module, or existing functionality could be distributed amongst the components/modules in different ways, yet still achieve the functions described herein.
Furthermore, in some embodiments, some or all of the components of the content recommendation system 510 may be implemented or provided in other manners, such as at least partially in firmware and/or hardware, including, but not limited to one or more application-specific integrated circuits (“ASICs”), standard integrated circuits, controllers executing appropriate instructions, and including microcontrollers and/or embedded controllers, field-programmable gate arrays (“FPGAs”), complex programmable logic devices (“CPLDs”), and the like. Some or all of the system components and/or data structures may also be stored as contents (e.g., as executable or other machine-readable software instructions or structured data) on a computer-readable medium (e.g., as a hard disk; a memory; a computer network or cellular wireless network or other data transmission medium; or a portable media article to be read by an appropriate drive or via an appropriate connection, such as a DVD or flash memory device) so as to enable or configure the computer-readable medium and/or one or more associated computing systems or devices to execute or otherwise use or provide the contents to perform at least some of the described techniques. Some or all of the components and/or data structures may be stored on tangible, non-transitory storage mediums. Some or all of the system components and data structures may also be stored as data signals (e.g., by being encoded as part of a carrier wave or included as part of an analog or digital propagated signal) on a variety of computer-readable transmission mediums, which are then transmitted, including across wireless-based and wired/cable-based mediums, and may take a variety of forms (e.g., as part of a single or multiplexed analog signal, or as multiple discrete digital packets or frames). Such computer program products may also take other forms in other embodiments. Accordingly, embodiments of this disclosure may be practiced with other computer system configurations.
The illustrated process begins at block 602, where it processes a content item to determine multiple entities referenced by the content item. Determining the multiple entities may include identifying entities referenced by the content item, each of the determined entities being electronically represented by the content recommendation system. Determining the multiple entities may further include ranking the entities on factor such as the number/quantity of mentions in the content item, the position of the mentions of the entity in the content item, and/or penalties based on the type of the entity.
At block 604, the process determines at least one category that is associated with one of the multiple entities. The determined at least one category may be part of a taxonomy stored by the content recommendation system and may be associated with one of the multiple corresponding entities referenced by the content item, as determined at block 602. Determining the at least one category may further include selecting a predetermined number of highest ranked entities from entities ranked at block 602, and then selecting and/or aggregating categories associated with the selected entities.
At block 606, the process stores the determined multiple entities and the determined at least one category. Storing the determined entities and categories may include annotating the content item (e.g., with a token that represents and entity or category) in an index or other data structure that supports efficient retrieval of content items.
At block 608, the process determines whether there are more content items. If so, the process proceeds to block 602, else returns.
The process begins at block 702, where it receives an indication of a category. In one embodiment, receiving the indication of the category includes receiving a search query/request that specifies the category.
At block 704, the process selects a content item that has a corresponding category that matches the indicated category. The corresponding category may be associated with one or more entities that are referenced by the selected content item and that are part of a taxonomy stored by the content recommendation system. Selecting the content item may include ranking multiple content items that reference the category, based on various factors, such as term frequency, the number of times a category token was added to the content item, a credibility score of the content item, and/or a recency score of the content item.
At block 706, the process transmits an indication of the selected content item. Transmitting the indication of the content item may include transmitting an identifier (e.g., a URL) and/or information about or from the content item (e.g., article title, summary, text).
Some embodiments perform one or more operations/aspects in addition to, or instead of, the ones described with reference to the process of
The following Table defines several example entity types in an example embodiment. Other embodiments may incorporate different types.
The following Table defines several example facets in an example embodiment. Other embodiments may incorporate different facets.
All of the above U.S. patents, U.S. patent application publications, U.S. patent applications, foreign patents, foreign patent applications and non-patent publications referred to in this specification and/or listed in the Application Data Sheet, including but not limited to U.S. application Ser. No. 13/286,778, entitled “CATEGORY-BASED CONTENT RECOMMENDATION,” filed Nov. 1, 2011, and U.S. Provisional Patent Application No. 61/408,965, entitled “CATEGORY-BASED CONTENT RECOMMENDATION,” filed Nov. 1, 2010, are incorporated herein by reference, in their entireties.
From the foregoing it will be appreciated that, although specific embodiments have been described herein for purposes of illustration, various modifications may be made without deviating from the spirit and scope of this disclosure. For example, the methods, techniques, and systems for category-based content recommendation are applicable to other architectures. For example, instead of recommending textual content items, the techniques may be used to automatically recommend other types of items, such as music or other audio items, videos, applications (e.g., mobile applications), online activities, or the like. Also, the methods, techniques, and systems discussed herein are applicable to differing query languages, protocols, communication media (optical, wireless, cable, etc.) and devices (e.g., desktop computers, wireless handsets, electronic organizers, personal digital assistants, tablet computers, portable email machines, game machines, pagers, navigation devices such as GPS receivers, etc.).
Claims
1. A computer-implemented method in a content recommendation system, the method comprising:
- processing a corpus of content items to determine, for each of the content items, multiple corresponding entities referenced by the content item, each of the determined entities being electronically represented by the content recommendation system;
- determining, for each of at least some of the content items, at least one corresponding category that is part of a taxonomy represented as a graph stored by the content recommendation system and that is associated with one of the multiple corresponding entities referenced by the content item, wherein determining the at least one corresponding category includes aggregating common nodes in taxonomic paths that are associated respectively with a first determined entity and a second determined entity that are part of the graph such that the corresponding category relates the first and second determined entities in an is-a relation, a part-of relation, or a member-of relation; and
- storing, for each of the content items, the determined multiple corresponding entities and the determined at least one corresponding category.
2. The method of claim 1 wherein determining the at least one corresponding category includes traversing a path in the taxonomy stored by the content recommendation system.
3. The method of claim 1 wherein processing the corpus of content items includes ranking, for one of the content items, the determined multiple corresponding entities.
4. The method of claim 3 wherein ranking the determined multiple corresponding entities includes ranking an entity based on at least one of: the quantity of times that the entity is referenced by the content item, a position/location of a reference of the entity, or a penalty assessed based on a type of the entity.
5. The method of claim 4 wherein determining the at least one corresponding category includes selecting a predetermined number of highest ranked entities from the ranked entities.
6. The method of claim 1 wherein determining the at least one corresponding category includes ranking leaf node categories of taxonomic paths associated with the determined multiple corresponding entities, the ranking based on the quantity of entities having a particular category and/or the a rank of an entity in a ranked list of entities.
7. The method of claim 1 wherein storing the determined multiple corresponding entities and the determined at least one corresponding category includes annotating a content item entry in an index with tokens that reflect the determined at least one category.
8. The method of claim 1, further comprising:
- receiving an indication of a category;
- selecting one or more of the content items that each have a corresponding category that matches the indicated category; and
- providing indications of the selected content items.
9. The method of claim 8 wherein selecting the one or more content items includes ranking content items based on term frequency minus inverse document frequency.
10. The method of claim 8 wherein selecting the one or more content items includes ranking content items based the quantity of times a category token was added to a content item index.
11. The method of claim 8 wherein selecting the one or more content items includes ranking the one or more content items based on a credibility score determined for each content item.
12. The method of claim 8 wherein selecting the one or more content items includes ranking the one or more content items based on recency of each content item, such that more recent content items are ranked higher than less recent content items.
13. The method of claim 8 wherein selecting the one or more content items includes collapsing similar content items into groups of content items, wherein similarity between two content items is based on at least one of: distance between signatures of the two content items, amount of overlap between titles of the two content items, amount of overlap between summaries of the two content items, amount of overlap between URLs referencing the two content items, and publishers of the two content items.
14. The method of claim 8, further comprising:
- determining content items that are related to the indicated category but that are not in the corpus of content items; and
- receiving from a third-party content service indications of at least one of a video, an image, an audio file, an instant message, or a message in a social network.
15. The method of claim 14 wherein determining content items that are related to the indicated category but that are not in the corpus of content items includes generating a keyword query that includes names of popular entities in the indicated category.
16. The method of claim 1, further comprising:
- determining popular entities for an indicated category, the popular entities having recently received an increased number of references by content items in the corpus and/or having more references by content items in the corpus than other entities; and
- transmitting indications of the determined popular entities.
17. A computer-readable medium having contents that, when executed, enable a computing system to recommend content, by performing a method comprising:
- processing a corpus of content items to determine, for each of the content items, multiple corresponding entities referenced by the content item, each of the determined entities being electronically represented by the content recommendation system;
- determining, for each of at least some of the content items, at least one corresponding category that is part of a taxonomy represented as a graph stored by the content recommendation system and that is associated with one of the multiple corresponding entities referenced by the content item, wherein determining the at least one corresponding category includes aggregating common nodes in taxonomic paths that are associated respectively with a first determined entity and a second determined entity that are part of the graph such that the corresponding category relates the first and second determined entities in an is-a relation, a part-of relation, or a member-of relation; and
- providing category-based content recommendations, by: receiving an indication of a category; selecting a content item that has a corresponding category that matches the indicated category, the corresponding category associated with one or more entities that are referenced by the selected content item; and transmitting an indication of the selected content item.
18. The computer-readable medium of claim 17 wherein selecting the content item includes collapsing similar content items into groups of content items, wherein similarity between two content items is based on at least one of: distance between signatures of the two content items, amount of overlap between titles of the two content items, amount of overlap between summaries of the two content items, amount of overlap between URLs referencing the two content items, and/or publishers of the two content items.
19. A computing system configured to recommend content, comprising:
- a memory;
- a module stored on the memory that is configured, when executed, to: process a corpus of content items to determine, for each of the content items, multiple corresponding entities referenced by the content item, each of the determined entities being electronically represented by the content recommendation system; determine, for each of at least some of the content items, at least one corresponding category that is part of a taxonomy represented as a graph stored by the content recommendation system and that is associated with one of the multiple corresponding entities referenced by the content item, wherein determining the at least one corresponding category includes aggregating common nodes in taxonomic paths that are associated respectively with a first determined entity and a second determined entity that are part of the graph such that the corresponding category relates the first and second determined entities in an is-a relation, a part-of relation, or a member-of relation; receive from a search query an indication of a category; select a content item from the corpus of content items, the selected content item having a corresponding category that matches the indicated category; and transmit an indication of the selected content item.
20. The computing system of claim 19 wherein the module is a category-based news service configured to recommend news items to at least one of a personal digital assistant, a smart phone, a tablet computer, a laptop computer, and/or a third-party application.
Type: Application
Filed: Aug 13, 2018
Publication Date: Dec 20, 2018
Inventors: Jisheng Liang (Bellevue, WA), Krzysztof Koperski (Seattle, WA), Jennifer Cooper (Seattle, WA), Theodore Diamond (Seattle, WA)
Application Number: 16/102,631
