EP2438544A2

EP2438544A2 - Searching methods and devices

Info

Publication number: EP2438544A2
Application number: EP10783951A
Authority: EP
Inventors: John J. Lee; Andrew William Hogue; Daniel N. Quine; Andriy Bihun; Daniel Loreto; Randolph G. Brown; Yohann R. Coppel; John Alexander Komoroske; Craig Nevill-Manning
Original assignee: Google LLC
Current assignee: Google LLC
Priority date: 2009-06-01
Filing date: 2010-06-01
Publication date: 2012-04-11
Also published as: WO2010141502A2; KR20120038418A; AU2010256777A1; CN102460440B; CA2764157A1; CN102460440A; WO2010141502A3

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for improving search with user corrections. In one aspect, a methods performed by a data processing apparatus include the actions of receiving a value result set, accessing historical records of user corrections stored at one or more data storage devices, the historical records describing user corrections of the characterization of instance attributes by values, determining that the historical records of user corrections describe a first user correction involving a first value in the value result set, and changing a confidence parameter embodying a confidence that the first value correctly characterizes the attribute of the instance. The value result set comprises a collection of one or more values. The values are candidates for characterizing an attribute of an instance. The first value is involved in the correction as either a corrected value or an uncorrected value.

Description

SEARCHING METHODS AND DEVICES

BACKGROUND

This specification relates to improving the rankings in search results with user corrections or identification of a group of related instances, e.g., by searching a unstructured collection of electronic documents

Search is generally an automated process in which a user enters a search query and receives responsive results in a result set. The results identify content that is relevant to the search query, e.g., in a machine-readable collection of digital data stored on data storage device. An electronic document is a collection of machine-readable digital data. Electronic documents are generally individual files and are formatted in accordance with a defined format (e.g., PDF, TIFF, HTML, XML, MS Word, PCL, PostScript, or the like). An electronic document collection can be stored as digital data on one or more data storage devices. Electronic document collections can either be unstructured or structured. The formatting of the documents in an unstructured electronic document collection is not constrained to conform with a predetermined structure and can evolve in often unforeseen ways. In other words, the formatting of individual documents in an unstructured electronic document collection is neither restrictive nor permanent across the document collection. Further, in an unstructured electronic document collection, there are no mechanisms for ensuring that new documents adhere to a format or that changes to a format are applied to previously existing documents. Thus, the documents in an unstructured electronic document collection cannot be expected to share a common structure that can be exploited in the extraction of information. Examples of unstructured electronic document collections include the documents available on the Internet, collections of resumes, collections of journal articles, and collections of news articles. Documents in some unstructured electronic document collections are not prohibited from including links to other documents inside and outside of the collection.

In contrast, the documents in structured electronic document collections generally conform with formats that can be both restrictive and permanent. The formats imposed on documents in structured electronic document collections can be restrictive in that common formats are applied to all of the documents in the collections, even when the applied formats are not completely appropriate. The formats can be permanent in that an upfront commitment to a particular format by the party who assembles the structured electronic document collection is generally required. Further, users of the collections — in particular, computer programs that use the documents in the collection — rely on the documents' having the expected format. As a result, format changes can be difficult to implement. Structured electronic document collections are best suited to applications where the information content lends itself to simple and stable categorizations. Thus, the documents in a structured electronic document collection generally share a common structure that can be exploited in the extraction of information. Examples of structured electronic document collections include databases that are organized and viewed through a database management system (DBMS) in accordance with hierarchical and relational data models, as well as a collections of electronic documents that are created by a single entity for presenting information consistently. For example, a collection of web pages that are provided by an online bookseller to present information about individual books can form a structured electronic document collection. As another example, a collection of web pages that is created by server-side scripts and viewed through an application server can form a structured electronic document collection. Thus, one or more structured electronic document collections can each be a subset of an unstructured electronic document collection.

Instances are individually identifiable entities. Instances can be grouped according to their attributes. An attribute is a property, feature, or characteristic of an instance. A group of instances can be defined by one or more attributes. The instances that belong to a group are determined by the attributes that define the group. For example, the instances New York, Chicago, and Tokyo can be grouped together as cities, whereas Tokyo is excluded from a group of North American cities.

SUMMARY This specification describes technologies relating to improving search with user corrections, and technologies relating to the identification of one or more groups of related instances. In some implementations, the groups of related instance identifiers are identified by searching an unstructured collection of electronic documents, for example, the electronic documents available on the Internet. In general, one innovative aspect of the subject matter described in this specification can be embodied in methods performed by data processing apparatus that include the actions of receiving a value result set, the value result set comprising a collection of one or more values, the values being candidates for characterizing an attribute of an instance, accessing historical records of user corrections stored at one or more data storage devices, the historical records describing user corrections of the characterization of instance attributes by values, determining that the historical records of user corrections describe a first user correction involving a value in the value result set, wherein the value is involved in the correction as either a corrected value or an uncorrected value; and changing a confidence parameter embodying a confidence that the involved value correctly characterizes the attribute of the instance.

Other embodiments of this aspect include corresponding systems, apparatus, and computer programs, configured to perform the actions of the methods, encoded on computer storage devices.

These and other embodiments can each optionally include one or more of the following features. The method can include ranking the values in the value result set to reflect the changed confidence parameter and visually displaying at least a portion of the value result set on a display screen. Outputting at least the portion of the value result set can include presenting a structured presentation to a user. The structured presentation can be populated with a first value included in the value result set. The first value is the value in the value result set that is most likely to correctly characterize the instance attribute. Visually displaying at least a portion of the value result set can include displaying a candidate window that includes candidate values for characterizing an instance attribute. Changing the confidence parameter can include generating a delta value suitable for application to a scaled confidence rating. The scaled confidence rating can embody the confidence that the involved value correctly characterizes the attribute of the instance. Generating the delta value can include weighting a category of a user correction of the involved value or categorizing the user correction. Another innovative aspect of the subject matter described in this specification can be embodied in computer storage medium encoded with a computer program. The program can include instructions that when executed by data processing apparatus cause the data processing apparatus to perform operations. The operations can include receiving a description of a user correction involving a value characterizing an instance attribute, wherein the value is involved in the correction as either a corrected value or an uncorrected value, changing a confidence parameter reflecting the likelihood that the value correctly characterizes the instance attribute, and ranking a collection of candidate values that includes the value according to respective confidence parameters, including the changed a confidence parameter. Other embodiments of this aspect include corresponding systems, apparatus, and methods, configured to perform the operations performed by the data processing apparatus.

These and other embodiments can each optionally include one or more of the following features. The operations can include transmitting a description of the ranked collection of candidate values over a data communication network in response to receipt of a search query, the response to which includes an attribute value for an instance.

Receiving the description of the user correction can include receiving a description of whether that the user confirmed the correction with a source, receiving a description that the user did not change an uncorrected value after reviewing an electronic document, and receiving a description of the uncorrected value prior to the user correction and the corrected value after the user correction. Changing the confidence parameter can include categorizing the user correction and weighting the impact of the user correction on the confidence parameter according to the categorization of the user correction. Weighting the impact of the user correction can include weighting user corrections made after confirmation from a source more heavily than user corrections made without confirmation from a source or weighting more recent user corrections more heavily than earlier user corrections. Changing the confidence parameter can include changing the confidence parameter reflecting the likelihood that an corrected value correctly characterizes the instance attribute.

Another innovative aspect of the subject matter described in this specification can be embodied in systems that include a client, a correction tracker operable to interact with the client to track the user input correcting the characterizations of the instance attributes and to store descriptions of the user input in records of the user correction history, one or more data storage devices storing the records of the user correction history, and a search engine operable to interact with the one or more data storage devices to access the records of the user correction history and to change a confidence that a first value correctly characterizes a first instance attribute in response to identifying a record describing a user correction correcting the characterization of the first instance attribute. The client includes an input device, a display screen, and a digital data processing device operable to display, on the display screen, characterizations of instance attributes by values and to receive, over the input device, user input correcting characterizations of instance attributes. Other embodiments of this aspect include corresponding methods, apparatus, and computer programs, configured to perform the actions of the system elements, encoded on computer storage devices.

These and other embodiments can each optionally include one or more of the following features. The display screen can display a structured presentation under the direction of the digital data processing device, the structured presentation can associate instance attributes with values. The structured presentation can include interactive elements selectable by a user to identify an instance attribute whose characterization by a value is to be corrected. The interactive elements can include cells of the structured presentation. The structured presentation can be a deck of cards. The display screen can display a candidate window under the direction of the digital data processing device. The candidate window can present candidate corrected values for replacing an uncorrected value characterizing an instance attribute.

Another innovative aspect of the subject matter described in this specification can be embodied in methods performed by one or more data processing apparatus that include the actions of the data processing apparatus receiving a search query at a data processing apparatus, the data processing apparatus identifying groups of instance identifiers in an unstructured collection of electronic documents with the data processing apparatus, the data processing apparatus determining relevance of the groups of instance identifiers to the search query with the data processing apparatus, and the data processing apparatus scoring at least some of the instance identifiers in the groups of instance identifiers individually with the data processing apparatus; and the data processing apparatus ranking the at least some instance identifiers according to the scores with the data processing apparatus. The search query specifies attributes shared by a group of related instances. Other embodiments of this aspect include corresponding systems, apparatus, and computer programs, configured to perform the actions of the methods, encoded on computer storage devices.

These and other embodiments can each optionally include one or more of the following features. Determining the relevance of the groups of instance identifiers to the search query can include computing relevance of the groups of instance identifiers to source documents that include the groups of instance identifiers, computing likelihoods that the identified groups of instance identifiers are indeed groups of instance identifiers, and computing relevance of source documents which include the groups of instance identifiers to the search query. Identifying the groups of instance identifiers can include forming a first new query biased to identify groups, forming a second new query constrained to search compendia sources, and searching the unstructured collection of electronic documents with the received query, the first new query, and the second new query.

The method can also include the data processing apparatus rescoring the at least some instance identifiers before ranking. Scoring at least some of the instance identifiers in the groups of instance identifiers can include representing features of the instance identifiers in a vertex-edge graph and scoring the instance identifiers according to the features represented in the vertex-edge graph. The vertices in the vertex-edge graph can represent groups of instance identifiers. Respective edges in the vertex-edge graph can be weighted according to overlap between the vertices connected by the edge.. Vertices in the vertex-edge graph can represent individual instance identifiers. Respective edges in the vertex-edge graph represent features shared by the instance identifiers. A first edge in the vertex-edge graph can represent an extractor that identified a pair of vertices joined by the first edge. A first edge in the vertex- edge graph can represent other instance identifiers in potential groups where vertices joined by the first edge are found. A first edge in the vertex-edge graph can represent a class of the query that identified source document where vertices joined by the first edge are found. Scoring the instance identifiers can include identifying cliques in the vertex-edge graph. Scoring the instances identifiers can include scoring the instance identifiers using a predictive analytic tree-building algorithm. Scoring the instance identifiers using the predictive analytic tree-building algorithm can include training the predictive analytic tree-building algorithm using a group of instance identifiers of confirmed accuracy that are relevant to a search query, a set of potential groups of instance identifiers that have been identified from an unstructured collection of electronic documents, and features of the instance identifiers in the potential groups and generating a classification and regression tree. Another innovative aspect of the subject matter described in this specification can be embodied in computer storage media encoded with a computer program. The programs can include instructions that when executed by data processing apparatus cause a data processing apparatus to perform operations. The operations can include receiving a search query at a data processing apparatus, the search query specifying attributes shared by a group of related instances, searching an electronic document collection to identify identifiers of instance that are responsive to the search query, representing features of the instance identifiers in a vertex-edge graph, and scoring relevance of the instance identifiers to the search query according to the features represented in the vertex-edge graph. Other embodiments of this aspect include corresponding systems, apparatus, and methods, configured to perform the actions of the operations.

These and other embodiments can each optionally include one or more of the following features. The operations can also include identifying groups of instance identifiers in the electronic documents of the collection and determining relevance of the groups of instance identifiers to the search query. A first feature represented in the vertex-edge graph can include the relevance of the groups that include respective instance identifiers to the search query. The operations can also include identifying electronic documents available on the Internet that are relevant to the search query and extracting groups of instance identifiers from the electronic documents that are relevant to the search query. The operations can also include computing relevance of the electronic documents from which the groups of instance identifiers are extracted to the search query, computing relevance of the groups of instance identifiers to the electronic documents from which the groups of instance identifiers are extracted, and computing likelihoods that the groups of instance identifiers are groups of instance identifiers.

Identifying the groups of instance identifiers can include forming a new query biased to identify groups and searching the electronic document collection with the new query. A first edge in the vertex-edge graph can represent a class of the query that identified a pair of vertices joined by the first edge. A first edge in the vertex-edge graph can represent other instance identifiers in potential groups where vertices joined by the first edge are found.

Scoring relevance of the instance identifiers to the search query can include identifying cliques in the vertex-edge graph.

Another innovative aspect of the subject matter described in this specification can be embodied in systems that include a client device and one or more computers programmed to interact with the client device and the data storage device. The computers are programmed to perform operations comprising receiving a search query from the client device, the search query explicitly or implicitly specifying attributes of instances, searching an electronic document collection to identify identifiers of instances that may have the attributes specified by the search query, representing features of the search of the electronic document collection in a vertex-edge graph, scoring the instance identifiers that may have the attributes specified by the search query according to the features represented in the vertex-edge graph, and outputting, the client device, instructions for visually presenting at least some of the instance identifiers. Other embodiments of this aspect include corresponding methods and computer programs encoded on computer storage devices configured to perform the operations of the computers.

These and other embodiments can each optionally include one or more of the following features. Outputting the instructions can include outputting instructions for visually presenting a structured presentation at the client device and the client device is configured to receive the instructions and cause the structured presentation to be visually presented. The system can include a data storage device storing a data describing multiple groups of instances. The system can include a data storage device storing machine-readable instructions tailored to identify and extract groups of instance identifiers from electronic documents in an unstructured collection. Representing features can include representing the relevance of the groups in which the instance identifiers appear in the vertex-edge graph. Scoring the instance identifiers can include scoring the instance identifiers individually according to the relevance of the groups in which the instance identifiers appear to the search query. Scoring the instance identifiers can include identifying cliques in the vertex-edge graph. Scoring the instance identifiers can include scoring the instance identifiers according to an extractor represented in the vertex-edge graph. Scoring the instance identifiers can include scoring the instance identifiers according to a class of a query represented in the vertex-edge graph. The details of one or more implementations of the subject matter described in this specification are set forth in the accompanying drawings and the description below. Other features, aspects, and advantages of the subject matter will become apparent from the description, the drawings, and the claims.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a schematic representation of a system in which a historical record of user corrections is used to improve search for a current user.

FIG. 2 is a schematic representation of the supplementation of user correction history in the system of FIG. 1

FIGS. 3-5 are examples of structured presentations that characterize attributes of instances with values.

FIGS. 6 and 7 are flow charts of processes for improving search with user corrections. FIGS. 8-11 are schematic representations of structured presentations in which user corrections of values of instance attributes can be received. FIG. 12 is a flow chart of a process for improving search with user corrections.

FIG. 13 is a schematic representation of a user correction log.

FIG. 14 is a flow chart of a process for improving search with user corrections.

FIG. 15 is a schematic representation of an aggregated feedback data collection. FIG. 16 is a schematic representation of a weighting parameter data collection.

FIG. 17 is a flow chart of a process for improving search with user corrections.

FIG. 18 is a schematic representation of a weighting parameter data collection.

FIG. 19 is a schematic representation of a system in which a group of related instances is identified. FIG. 20 is a flow chart of a process for identifying a group of related instances.

FIG. 21 is a schematic representation of a process for identifying a group of related instances.

FIG. 22 is a flow chart of a process for identifying electronic documents relevant to a query. FIG. 23 is a schematic representation of a process for identifying electronic documents relevant to a query.

FIG. 24 is a flow chart of a process for determining the relevance of groups of instances to a search query.

FIG. 25 is a flow chart of a process for scoring instances according to relevance of groups in which instances appear.

FIG. 26 is a flow chart of a process for scoring instances according to the relevance of groups in which instances appear.

FIG. 27 is a schematic representation of a vertex-edge graph that represents features of the instances in the potential groups. FIG. 28 is a schematic representation of another vertex-edge graph that represents features of the instances in the potential groups.

FIG. 29 is a flow chart of a process for rescoring instances.

Like reference numbers and designations in the various drawings indicate like elements.

DETAILED DESCRIPTION

FIG. 1 is a schematic representation of a system 100 in which a historical record of user corrections is used to improve search for a current user. A user correction is an alteration of the characterization of an instance attribute by a value. Instances are individually identifiable entities. An attribute is a property, feature, or characteristic of an instance. For example, Tom, Dick, and Harry are instances of individuals. Each such individual has attributes such as a name, a height, a weight, and the like. As another example, city instances each have a geographic location, a mayor, and a population. As yet another example, a product instance can have a model name, a maker, and a year. The attributes of an instance can be characterized by values. The value of a particular attribute of a particular instance characterizes that particular instance. For example, the name of an individual can have the value "Tom," the population of a city can have the value "4 million," and the model name of a product can have the value "Wrangler." A user correction can also be an attempt to alter the characterization of an instance attribute by a value. User corrections are made by human users. User corrections are generally designed to correct or improve a value from the perspective of the user making the correction. A user correction can alter a value, e.g., by deleting the value, by editing the value, by refining the value, by substituting a corrected value for the uncorrected value, or by combinations of these and other alterations. An attempt to alter the characterization of an instance attribute can include a trackable user confirmation of the value with an electronic document, e.g., one available on the Internet. A record of a user correction can thus include one or more of a corrected value, an uncorrected value, and a notation of whether a confirmation was made. A record that includes multiple user corrections of one or more values can reflect the collective wisdom and work of a number of human users. The present inventors have recognized that such a record can be used to improve the usefulness of a search system for subsequent users.

System 100 includes a search engine 105, a user correction history 1 10, and a client 115. A current user can interact with client 115 to enter a search query, the response to which includes an attribute value for an instance. For example, the search query can inquire about the value of an attribute of an instance. Search engine 105 can respond to the search query by searching, e.g., electronic documents of a document collection such as the Internet, a store of information characterizing electronic documents, or a structured database organized and viewed through a database management system (DBMS). Search engine 105 can operate with an internal or external module to rank the results in a result set, e.g., according to the relevancy of the results to a search query. Search engine 105 can be implemented on one or more computers deployed at one or more geographical locations that are programmed with one or more sets of machine-readable instructions for searching in response to requests originating from multiple client devices. In certain circumstances, search engine 105 can conduct a search and return a result set of one or more values responsive to the search query. As described further below, the content of the result set, the arrangement of results in the result set, or both can reflect corrections that have previously been made by users and recorded in user correction history 110.

User correction history 110 stores information characterizing corrections that have previously been made by users. In some implementations, the corrections can be received from users who interact with a client in the context of a search. For example, as described further below, a user can interact with a structured presentation displayed at client 1 15, such as the structured presentations shown in FIGS. 3-5.

User correction history 110 can be stored on one or more data storage devices deployed at one or more geographical locations. The information in user correction history 110 is accessible either directly by search engine 105 or by one or more intermediate modules that can provide information characterizing the information content of user correction history 110 to search engine 105.

Client 115 is a device for interacting with a user and can be implemented on a computer programmed with machine-readable instructions. Client 115 can include one or more input/output devices, such as a display screen 120 for displaying information to the current user. For example, client 1 15 can display a presentation 125 on display screen 120. Presentation 125 indicates that an attribute of an instance is characterized by a value

130 (e.g., "THE ATTRIBUTE X OF INSTANCE Y IS: VALUE Z."). Other presentations indicating that an attribute of an instance is characterized by a value 130, namely, structured presentations, are described in further detail below.

In general, a presentation indicating that an attribute of an instance is characterized by a value will be displayed during a search session. For example, a user who is currently interacting with client 1 15 can enter a search query using an input device such as a mouse or a keyboard. The response to the search query can include an attribute value for an instance. In some implementations, the search query can identify both an instance and an attribute of the instance that is to be characterized. For example, the search query can be an instance: attribute pair (e.g., "France: capital" or "mayor:Birmingham"). As another example, the search query can be formed so that identifiers of the instance and the attribute are found in a linguistic pattern indicating that a value characterizing the attribute of the instance is desired. Examples of such patterns include "what is the <attribute> of <instance>," "who is < instance>'s <attribute>," and the like. As another example, a user can enter a search query by interacting with or referring to a structured presentation displayed on display screen 120. For example, as described further below, a user can click on a cell in a structured presentation or manually formulate a search query that refers to cells in a structured presentation as attribute and instance (e.g., "CELL_1 :CELL_2").

In some implementations, a search query need not identify both an instance and an attribute of the instance that is to be characterized. Rather, a search query can merely identify either an attribute or an instance, e.g., in a context that indicates that one or more attributes of one or more instances are to be characterized. For example, a query "mayors" can be taken as an inquiry requesting that values of the attribute "mayor" of city instances be identified. As another example, a query "richest women in the world" can be taken as an inquiry requesting that requesting that values of the attribute "name" of "richest women in the world" instances to be identified.

In response to receipt of the search query, client 115 transmits a representation of the search query, or the search query itself, to search engine 105 in a message 135. Message 135 can be transmitted over a data communications network. Search engine 105 can receive message 135 and use the content of message 135 to define parameters for searching. For example, the content of message 135 can be used to define terms used to search an indexed collection of electronic documents, to define a query in a DBMS query language, or combinations of these and other approaches.

Search engine 105 performs the search according to the parameters for searching defined by the content of message 135. The search can yield a result set of one or more values responsive to the search query described in message 135. The content of the result set, the arrangement of results in the result set, or both can reflect corrections that have previously been made by users and recorded in user correction history 1 10. For example, user corrections recorded in history 110 can be incorporated into a database or other body of data that is searched by search engine 105. The user corrections can thus themselves be the source of values included in the result set. As another example, user corrections recorded in history 110 can be used in ranking values in the result set. The values in the value result set are candidates for characterizing one or more attributes of one or more instances and are responsive to the search query. The content and arrangement of values in the value result set can reflect one or more changes in the confidence that particular values correctly characterize an attribute of an instance. For example, when a user correction is a source of a value included in the result set, that value may go from having a low confidence and hence being excluded from the result set to having a confidence that is high enough to justify inclusion in the result set. As another example, the ranking of values in the result set can reflect the confidence in the individual values. In particular, a value that is more likely to correctly characterize an attribute of an instance will generally be ranked above a value that is less likely to correctly characterize an attribute of an instance.

Search engine 105 transmits a representation of the result set that reflects user corrections to client 115 in a message 140. Message 140 can be transmitted, e.g., over the same data communications network that transmitted message 135. Client 115 can receive message 140 and use the content of message 140 to display a presentation 125 on display screen 120. Presentation 125 characterizes an attribute of an instance with a value 130 which is found in the value result set that reflects user corrections. In some implementations, presentation 125 can use text to indicate that an attribute of an instance is characterized by a value 130, as shown. In some implementations, presentation 125 can use the arrangement of identifiers of an attribute and an instance to indicate that the identified attribute of the identified instance is characterized by a value 130. For example, presentation 125 can be a structured presentation that displays values and identifier of instance attributes in an organized, systematic arrangement so that the characterization of an instance attribute by a value is apparent to a user, as described further below. In some implementations, systems such as system 100 can be used to supplement user correction history 110.

FIG. 2 is a schematic representation of the supplementation of user correction history 110 in system 100. As shown, a correction tracker 205 is coupled to client 115. Correction tracker 205 is a component for tracking corrections of characterizations of instance attributes made by a user at client 115. For example, correction tracker 205 can be implemented on one or more computers deployed at one or more geographical locations that are programmed with one or more sets of machine-readable instructions. Correction tracker 205 can be implemented using in client 115, for example, in a client side script, or it can be implemented search engine 105, or elements of correction tracker 205 can be implemented in both.

In the illustrated implementation, a user at client 1 15 has corrected presentation 125. In particular, the user has deleted an uncorrected value 130 and replaced it with a corrected value 205.

Correction tracker 205 can track the correction by recording a representation of the alteration(s) made by the user. Correction tracker 205 can also transmit data representing the user correction directly or indirectly in a message 210 to search engine 105 for storage in user correction history 110. Message 210 can be an XML document or other form of data package. The content of message 210 can be used to create a new record 215 of the user correction. New record 215 supplements the historical record of user corrections at user correction history 110. FIGS. 3-5 are examples of structured presentations that associate attributes of instances with values. FIG. 3 is a schematic representation of an example table structured presentation 300. Table 300 is an organized, systematic arrangement of one or more identifiers of instances, as well as the values of particular attributes of those instances. In some implementations, structured presentations such as table 300 can also include identifiers of attributes, as well as identifiers of the units in which values are expressed.

The grouping, segmentation, and arrangement of information in table 300 can be selected to facilitate understanding of the information by a user. In this regard, table 300 includes a collection of rows 302. Each row 302 includes an instance identifier 306 and a collection of associated attribute values 307. The arrangement and positioning of attribute values 307 and instance identifiers 306 in rows 302 thus graphically represents the associations between them. For example, a user can discern the association between attribute values 307 and the instance identifier 306 that is found in the same row 302.

Table 300 also includes a collection of columns 304. Each column 304 includes an attribute identifier 308 and a collection of associated attribute values 307. The arrangement and positioning of attribute values 307 and attribute identifier 308 in columns 304 thus graphically represent the associations between them. For example, a user can discern the association between attribute values 307 and the attribute identifier 308 that is found in the same column 304 based on their alignment.

Each row 302 is a structured record 310 in that each row 302 associates a single instance identifier 306 with a collection of associated attribute values 307. Further, the arrangement and positioning used to denote these associations in one structured record 310 is reproduced in other structured records 310 (i.e., in other rows 302). Indeed, in many cases, all of the structured records 310 in a structured presentation 106 are restricted to having the same arrangement and positioning of information. For example, values 307 of the attribute "ATTR_2" are restricted to appearing in the same column 304 in all rows 302. As another example, attribute identifiers 308 all bear the same spatial relationship to the values 307 appearing in the same column 304. Moreover, changes to the arrangement and positioning of information in one structured record 310 are generally propagated to other structured records 310 in the structured presentation 106. For example, if a new attribute value 307 that characterizes a new attribute (e.g., "ATTR_2³Λ") is added to one structured record 310, then a new column 304 is added to structured presentation 106 so that the values of attribute "ATTR_2³/4" of all instances can be added to structured presentation 106. In some implementations, values 307 in table 300 can be presented in certain units of measure. Examples of units of measure include feet, yards, inches, miles, seconds, gallons, liters, degrees Celsius, and the like. In some instances, the units of measure in which values 307 are presented are indicated by unit identifiers 309. Unit identifiers 309 can appear, e.g., beside values 307 and/or beside relevant attribute identifiers 308. The association between unit identifiers 309 and the values 307 whose units of measure are indicated is indicated to a viewer by such positioning. In many cases, all of the values 307 associated with a single attribute (e.g., all of the values 307 in a single column 304) are restricted to being presented in the same unit of measure.

The values in a value result set (such as the value result set described in message 140 (FIG. I)) can be used to populate table 300 or other structured presentation in a variety of different ways. For example, a structured presentation can be populated automatically (i.e., without human intervention) with a collection of values drawn from multiple search result sets that are each responsive to queries for instance attributes. For example, the individual values most likely to correctly characterize the instance attributes can be displayed in the structured presentation by default. A user can alter, or attempt to alter, those values by, e.g., interacting with or referring to the structured presentation. Other values in a value result set can be presented as candidates for replacing the value which the search engine has determined is most likely to correctly characterize the instance attributes.

FIG. 4 is a schematic representation of another implementation of a structured presentation, namely, a structured presentation table 400. In addition to including attribute identifiers 308, instance identifiers 306, values 307, unit identifiers 309 organized into rows 302 and columns 304, table 400 also includes a number of interactive elements for interacting with a user. In particular, table 400 includes a collection of instance selection widgets 405, a collection of action triggers 410, a collection of column action trigger widgets 415, and a notes column 420.

Instance selection widgets 405 are user interface components that allow a user to select structured records 310 in table 400. For example, instance selection widgets 405 can be a collection of one or more clickable checkboxes that are associated with a particular structured record 310 by virtue of arrangement and positioning relative to that structured record 310. Instance selection widgets 405 are "clickable" in that a user can interact with widgets 405 using a mouse (e.g., hovering over the component and clicking a particular mouse button), a stylus (e.g., pressing a user interface component displayed on a touch screen with the stylus), a keyboard, or other input device to invoke the functionality provided by that component.

Action triggers 410 are user interface components that allow a user to trigger the performance of an action on one or more structured records 310 in table 400 selected using instance selection widgets 405. For example, action triggers 410 can be clickable text phrases, each of which can be used by a user to trigger an action described in the phrase. For example, a "keep and remove others" action trigger 410 triggers the removal of structured records 310 that are not selected using instance selection widgets 405 from the display of table 400. As another example, a "remove selected" action trigger 410 triggers the removal of structured records 310 that are selected using instance selection widgets 405 from the display of table 400. As yet another example, a "show on map" action trigger 410 triggers display of the position of structured records 310 that are selected using instance selection widgets 405 on a geographic map. For example, if a selected instance is a car, locations of car dealerships that sell the selected car can be displayed on a map. As another example, if the selected instances are vacation destinations, these destinations can be displayed on a map. Column action trigger widgets 415 are user interface components that allow a user to apply an action to all of the cells within a single column 304. When a user interacts with the clickable '+' sign, a further user interface component is displayed which offers to the user a set of possible actions to be performed. The actions in this set can include, e.g., removing the entire column 304 from the structured presentation 400 or searching to find values for all the cells in column 304 which are currently blank. Notes column 420 is a user interface component that allows a user to associate information with an instance identifier 306. In particular, notes column 420 includes one or more notes 425 that are each associated with a structured record 310 by virtue of arrangement and positioning relative to that structured record 310. The information content of notes 425 is unrestricted in that, unlike columns 304, notes 425 are not required to be values of any particular attribute. Instead, the information in notes 425 can characterize unrelated aspects of the instance identified in structured record 310.

In some implementations, table 400 can include additional information other than values of any particular attribute. For example, table 400 can include a collection of images 430 that are associated with the instance identified in a structured record 310 by virtue of arrangement and positioning relative to that structured record 310. As another example, table 400 can include a collection of text snippets 435 extracted from electronic documents in collection 102. The sources of the snippets can be highly ranked results in searches conducted using instance identifiers 306 as a search string. Text snippets 435 are associated with the instance identified in a structured record 310 by virtue of arrangement and positioning relative to that structured record 310.

As another example, table 400 can include one or more hypertext links 440 to individual electronic documents in collection 102. For example, the linked documents can be highly ranked results in searches conducted using instance identifiers 306 as a search string. As another example, the linked documents can be source of a value 307 that was extracted to populate table 400. In some instances, interaction with hypertext link 440 can trigger navigation to the source electronic document based on information embedded in hypertext link 440 (e.g., a web site address).

FIG. 5 is a schematic representation of another implementation of a structured presentation, namely, a card collection 500. Card collection 500 is an organized, systematic arrangement of one or more identifiers of instances, as well as the values of particular attributes of those instances. The attributes of an instance can be specified by values. Moreover, card collection 500 generally includes identifiers of attributes, as well as identifiers of the units in which values are expressed, where appropriate. The grouping, segmentation, and arrangement of information in card collection 500 can be selected to facilitate an understanding of the information by a user. In this regard, card collection 500 includes a collection of cards 502. Each card 502 includes an instance identifier 306 and a collection of associated attribute values 307. The arrangement and positioning of attribute values 307 and instance identifiers 306 in cards 502 thus graphically represents the associations between them. For example, a user can discern the association between attribute values 307 and the instance identifier 306 that is found on the same card 502.

In the illustrated implementation, cards 502 in card collection 500 also include a collection of attribute identifiers 308. Attribute identifiers 308 are organized in a column 504 and attribute values 307 are organized in a column 506. Columns 504, 506 are positioned adjacent one another and aligned so that individual attribute identifiers 308 are positioned next to the attribute value 307 that characterizes that identified attribute. This positioning and arrangement allows a viewer to discern the association between attribute identifiers 308 and the attribute values 307 that characterize those attributes. Each card 502 is a structured record 310 in that each card 502 associates a single instance identifier 306 with a collection of associated attribute values 307. Further, the arrangement and positioning used to denote these associations in one card 502 is reproduced in other cards 502. Indeed, in many cases, all of the cards 502 are restricted to having the same arrangement and positioning of information. For example, the value 307 that characterizes the attribute "ATTR 1" is restricted to bearing the same spatial relationship to instance identifiers 306 in all cards 502. As another example, the order and positioning of attribute identifiers 308 in all of the cards 502 is the same. Moreover, changes to the arrangement and positioning of information in one card 502 are generally propagated to other cards 502 in card collection 500. For example, if a new attribute value 307 that characterizes a new attribute (e.g., "ATTR-PA") is inserted between the attribute values "value_l 1" and "value_2_l" in one card 502, then the positioning of the corresponding attribute values 307 in other cards 502 is likewise changed.

In some implementations, cards 502 in card collection 500 can include other features. For example, cards 502 can include interactive elements for interacting with a user, such as instance selection widgets, action triggers, attribute selection widgets, a notes entry, and the like. As another example, cards 502 in card collection 500 can include additional information other than values of any particular attribute, such as images and/or text snippets that are associated with an identified instance. As another example, cards 502 in card collection 500 can include one or more hypertext links to individual electronic documents in collection 102.

Such features can be associated with particular instances by virtue of appearing on a card 502 that includes an instance identifier 306 that identifies that instance.

During operation, a viewer can interact with the system presenting card collection 500 to change the display of one or more cards 502. For example, a viewer can trigger the side- by-side display of two or more of the cards 502 so that a comparison of the particular instances identified on those cards is facilitated. As another example, a viewer can trigger a reordering of card 502, an end to the display of a particular card 502, or the like. As another example, a viewer can trigger the selection, change, addition, and/or deletion of attributes and/or instances displayed in cards 502. As yet another example, a viewer can trigger a sorting of cards into multiple piles according to, e.g., the values of an attribute values 307 in the cards.

In some implementations, cards 502 will be displayed with two "sides." For example, a first side can include a graphic representation of the instance identified by instance identifier 306, while a second side can include instance identifier 306 and values 307. This can be useful, for example, if the user is searching for a particular card in the collection of cards 500, allowing the user to identify the particular card with a cursory review of the graphical representations on the first side of the cards 502.

FIG. 6 is a flow chart of a process 600 for improving search with user corrections. Process 600 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 600 can be performed by the search engine 105 in system 100 (FIG. 1). In some implementations, process 600 can be performed in response to the receipt of a trigger, such as a user request to use user corrections to improve search. Process 700 can be performed in isolation or in conjunction with other digital data processing operations.

The system performing process 600 can receive a description of a user correction of a value of an instance attribute (step 605). A user correction is an alteration or an attempted alteration of a value. A user correction may be submitted to prevent a mischaracterization of the attribute of the instance by a false value, to characterize the attribute of the instance correctly using an appropriate value, or to refine the characterization of the attribute of the instance. Example corrections of a value of an instance attribute can thus include, e.g., deleting a value, adding a new value, changing a value, or confirming the value with a source document. Example changes to a value include, e.g., correcting the spelling of the value, adding a time constraint to the value, increasing the accuracy of a value, and the like. The system performing process 600 can also change a confidence value that indicates a degree of confidence that the uncorrected value correctly characterizes the attribute of the instance (step 610). An uncorrected value is the value prior to correction by the present user. For example, as described further below, an uncorrected value can be a value returned after an initial search of a document collection or a database. The initial search — and the uncorrected value itself — can reflect corrections by other users.

Confidence is a characterization of the likelihood that a value correctly characterizes an attribute of an instance. For example, a value with a high confidence is one that has been determined to be likely to correctly characterize the attribute of the instance. On the other hand, it has been determined to be unlikely that a value with a low confidence correctly characterizes the attribute of the instance.

The confidence that a value correctly characterizes an attribute of an instance can be embodied in a confidence score or other parameter. The system can change or create a confidence parameter in response to the received user correction of a value of an attribute, as described further below. In some implementation, the confidence parameter can be a scaled rating of the confidence in the value of the attribute. For example, the confidence parameter can be percent certainty (e.g., "90% certain") that a value correctly characterizes the attribute of the instance. In other implementations, the confidence parameter can be an increment (i.e., a "delta") that can be applied to a scaled rating of the confidence in the value of the attribute. For example, the confidence parameter can be an increase or decrease in the percent certainty (e.g., "2% more certain" or "3% less certain") that the value correctly characterizes the attribute of the instance.

FIG. 7 is a flow chart of a process 700 for improving search with user corrections. Process 700 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 700 can be performed by the search engine 105 in system 100 (FIG. 1). In some implementations, process 700 can be performed in response to the receipt of a trigger, such as a user request to use user corrections to improve search. Process 700 can be performed in isolation or in conjunction with other digital data processing operations. The system performing process 700 can receive a description of a user correction of a value of an instance attribute (step 605) and change the confidence that the uncorrected value correctly characterizes the attribute of the instance (step 610).

The system performing process 700 can also change the confidence that the corrected value correctly characterizes the attribute of the instance (step 705). A corrected value is the value after correction by the present user. For example, as described further below, a corrected value can be a value selected from a list of candidate values, a changed version of the uncorrected value, or an entirely new value entered by the user. The change in confidence can be embodied in a confidence parameter, such as a scaled rating or a delta that can be applied to a scaled rating. FIG. 8 is a schematic representation of a structured presentation in which a user correction of a value of an instance attribute can be received, namely, a structured presentation 800. Structured presentation 800 can be used to receive a user correction of a value of an instance attribute, e.g., at step 605 in methods 600, 700 (FIGS. 6, 7).

Structured presentation 800 can be any form of structured presentation, including any of the structured presentations described above. For example, structured presentation 800 can be a data table displayed in a spreadsheet framework, as shown. The data table of structured presentation 800 includes a collection of rows 302 and columns 304. Each row 302 includes a respective instance identifier 306 and each column 304 includes a respective attribute identifier 308. The arrangement and positioning of instance identifiers 306 and attribute identifiers 308 in rows 302 and columns 304 associates each cell of the spreadsheet framework in which structured presentation 800 is displayed with an instance and an attribute. For example, a cell 805 in structured presentation 800 is associated with the instance identified as "Tesla Roadster" and the attribute identified as "mpg." A cell 810 in structured presentation 1000 is associated with the instance identified as "Chevy Volt" and the attribute identified as "range." A cell 815 in structured presentation 800 is associated with the instance identified as "Myers NmG" and the attribute identified as "top speed." A cell 1020 in structured presentation 800 is associated with the instance identified as "Myers NmG" and the attribute identified as "mpg." The associations between instance, attributes, and cells such as cells 805, 810, 815,

820 can be used to identify the attribute of the instance that is being corrected by a user. For example, receipt of user interaction selecting cell 820 can identify the attribute identified as "mpg" of the instance identified as "Myers NmG." User interaction selecting a cell can include, e.g., receipt of input positioning a cursor 825 over the cell, the user clicking on the cell, or the like. In some implementations, the selection of a cell can be denoted by positioning a visual indicia such a perimetrical highlight 830 in or around the cell.

In the illustrated implementation, selected cell 820 includes an uncorrected value 835 (i.e., "50 mpg") at the time of selection. For example, cell 820 in structured presentation 800 could have been populated with the results of a search performed, e.g., using an instance: attribute pair, in response to a user interacting with cell 820, or in response to a user referring to cell 820. Value 835 is an uncorrected value in that value 835 is the value of the attribute identified as "mpg" of the instance identified as "Myers NmG" displayed by the system.

FIG. 9 is a schematic representation of structured presentation 800 after a user correction of value 835 has been received. As shown, value 835 has thus been deleted from cell 820. The user may have deleted value 835 from cell 820 to correct what the user saw as a mischaracterization of the attribute identified as "mpg" of the instance identified as "Myers NmG" by value 835.

FIG. 10 is a schematic representation of structured presentation 800 after a corrected value 1005 has been received. As shown, the blank space left by deletion of value 835 from cell 820 has been filled with value 1005, which is provided by the user. Structured presentation 800 has thus been corrected to include value 1005 (i.e., "75 mpg") in cell 820. The user may have made this deletion and replacement to correct what the user sees as a mischaracterization of the attribute identified as "mpg" of the instance identified as "Myers NmG" by value 835 and to correctly characterize the attribute identified as "mpg" of the instance identified as "Myers NmG" with value 1005.

FIG. 1 1 is a schematic representation of a structured presentation in which a user correction of a value of an instance attribute can be received, namely, a structured presentation 1 100. Structured presentation 1 100 can be used to receive a user correction of a value of an instance attribute, e.g., at step 605 in methods 600, 700 (FIGS. 6, 7). In particular, user interaction selecting or referring to cell 820 can be used to trigger the presentation of a candidate window 1105. Candidate window 1105 presents candidate corrected values that are considered likely to be suitable for replacing an uncorrected value currently characterizing an instance attribute. In some implementations, the candidate values can be other values in a value result set, such as a value result set described in message 140 (FIG. 1). Thus, in some implementations, the nature and ranking of candidate corrected values can themselves reflect prior user corrections.

Candidate window 1105 includes a header 1 1 10, a collection of selection widgets 1115, a collection of identifiers 1120 of corrected candidate values, a collection of source identifiers 1 125, a collection of snippets 1 130, and a collection of search interactive elements 1135, a selection trigger 1 140, a full search trigger 1145, and a cancel trigger 1 150.

Header 11 10 can include text or other information that identifies the attribute of the instance which is characterized by a value which can be corrected. In the illustrated implementation, the attribute and instance (i.e., Myers NmG: mpg) that are characterized by the value 835 in cell 820 are identified.

Selection widgets 1 115 are interactive display devices that allow a user to select a value that is to be used to characterize the attribute and the instance identified in header 1110. In the illustrated implementation, the user can select from among the uncorrected value 835 and two candidate corrected values identified by value identifiers 1 120.

Value identifiers 1 120 include text or other information that identifies candidate corrected values for characterizing the attribute and the instance identified in header 1 110. The candidate corrected values identified by value identifiers 1 120 can be drawn, e.g., from electronic documents in an electronic document collection such as the Internet. Source identifiers 1 125 include text or other information that identifies one or more electronic documents in which value 835 and the candidate corrected values identified by value identifiers 1625 appear. In some implementations, source identifiers 1125 can also include hyperlinks to one or more electronic documents in which the value 835 and candidate corrected values identified by value identifiers 1125 appear. A user can follow such a hyperlink to confirm the respective of uncorrected value 835 and the corrected values identified by value identifiers 1120 directly with one or more source documents.

Each snippet 1130 is text or other information that describes the context of value 835 and the candidate corrected values identified by value identifiers 1120 in an electronic document. Snippets 1130 can allow a user to confirm the respective of uncorrected value 835 and the candidate corrected values identified by value identifiers 1120 indirectly, i.e., from candidate window 1 105 without linking to a source document.

Search interactive elements 1 135 are hyperlinks that allow a user to navigate to an electronic document in which the respective of value 835 or values identified by value identifiers 1 125 appears. A user can follow a search interactive element 1 135 to confirm the respective of uncorrected value 835 and the candidate corrected values identified by value identifiers 1 120 directly from the linked electronic document.

Selection trigger 1 140 is an interactive element that allows a user to consent to the use of a value to characterize the attribute and the instance identified in header 1110. In particular, selection trigger 1 140 allows the user to consent to the use of uncorrected value 835 or either of the candidate corrected values identified by value identifiers 1120. When a user consents to the use of either of the candidate corrected values, the selected value is substituted for value 835 in cell 820. The selected value is thus no longer a candidate corrected value but rather a corrected value. Search trigger 1 145 is an interactive element that triggers a search of an electronic document collection. Search trigger 1 145 can allow a user to confirm an uncorrected value 835, as well as both of the corrected values identified by value identifiers 1 120, directly from another source, such as an electronic document on the web. The search triggered by search trigger 1805 can be a "full search" in that it is conducted using a general purpose Internet search engine such as the GOOGLE™ search engine available at www. google. com. In some implementations, the search engine can be presented with a query that is automatically generated using the attribute of the instance identified in heading 1 1 10. The confirmation of a value by the user using a search can be recorded.

Cancel trigger 1150 is an interactive element that allows a user to cancel a correction of the value characterizing the attribute of the instance identified in heading 11 10. Cancel trigger 1150 can be used, e.g., when a user mistakenly identifies the wrong cell.

FIG. 12 is a flow chart of a process 1200 for improving search with user corrections. Process 1200 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 1200 can be performed by the search engine 105 using a historical record of user corrections 110 in system 100 (FIGS. 1, 2). In some implementations, process 1200 can be performed in response to the receipt of a trigger, such as a user request to use user corrections to improve search. Process 1200 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 1200 can be performed as either of processes 600, 700 (FIGS. 6, 7).

The system performing process 1200 can receive a description of a user correction of a value of an instance attribute (step 605). For example, the system performing process 1200 can receive a user correction made in interacting with displays such as structured presentations 800, 1100 (FIGS. 8-11).

The system performing process 1200 can also classify the user correction (step 1205). The user correction can be classified according to the activities performed by the user in correcting a value. For example, in some implementations, a user correction can be classified into one of the seven different classes shown in Table 1 below.

CORRECTION CLASSES

Class 1 : User selection of a candidate corrected value from a collection, without direct confirmation with a source.

Class 2: User selection of a candidate corrected value from a collection, after user directly confirming with a source.

Class 3 : User replacement of an uncorrected value with a corrected value, without user directly confirming with a source.

Class 4: User replacement of an uncorrected value with a corrected value, after user directly confirming with a source. Class 5 : User did not change an uncorrected value after user directly confirming with a source (i.e., a failed attempted alteration).

Class 6: User deletion of an uncorrected value without replacement by a corrected value, without user directly confirming with a source.

Class 7: User deletion of an uncorrected value without replacement by a corrected value, after user directly confirming with a source. TABLE 1

The activities used to classify a user correction (including any search for a confirmation) can be recorded during user interaction with displays such as structured presentations 800, 1100 (FIGS. 8-11), as described above.

The system performing process 1200 can log the user correction, e.g., by storing it in a digital data storage device (step 1210). The user correction can be logged as a collection of information that identifies the attribute of the instance that was corrected, the uncorrected value, and any corrected values. In general, the log of a user correction will also include an identification of the classification of the correction.

FIG. 13 is a schematic representation of a user correction log, namely, a data table 1300 that includes records 1305, 1310, 1315, 1320, 1325 of user corrections. Data table 1300 is a data structure stored in a digital data storage device for access by a computer program operating on a digital data processing system. Table 1300 includes a collection of columns 1330, 1335, 1340, 1345, 1350. Column 1330 includes instance identifiers that identify the instances in the logged corrections. Column 1335 includes attribute identifiers that identify the attributes of the instances in the logged corrections. Column 1340 includes correction classification identifiers that identify the classifications of the logged corrections. For example, column 1340 can include integers that corresponding to the numbering of the correction classes listed in Table 1. Column 1345 includes uncorrected value identifiers that identify the uncorrected values of the logged corrections. Column 1345 includes corrected value identifiers that identify the corrected values of the logged corrections. In situations where there is no corrected value (e.g., correction class 5: when a user did not change an uncorrected value after direct confirmation from a source), then the respective entry in column 1350 can remain empty or include a dummy value.

As shown in FIG. 12, the system performing process 1200 can repeatedly receive, classify, and log user corrections (steps 605, 1205, 1210). For example, the system can form a database of user corrections, such as historical record of user corrections 110 (FIG. 1). The system performing process 1200 can receive a search query, the response to which includes an attribute value for an instance (step 1215). For example, the received search query can identify both an instance and an attribute of the instance that is to be characterized in a linguistic pattern or as a consequence of interaction with or reference to a structured presentation. The system performing process 1200 can access the user correction log (step 1220). For example, the system can read the user correction log from one or more digital data storage devices. The system can also determine whether the contents of a result set responsive to the received search query match a correction of an instance attribute recorded in the user correction log (step 1225). For example, the system can compare the instance and an attribute of the instance that are the subject of the received search query with identifiers of instances and attributes in the user correction log. In the context of a user correction log such as data table 1300 (FIG. 13), the system can first compare the instance that is the subject of the search query with the contents of column 1330 to identify which of user corrections logs 1305, 1310, 1315, 1320, 1325 are relevant to the received search query. The system can then compare the attribute of the instance with the contents of column 1335 in the relevant user corrections logs 1305, 1310, 1315, 1320, 1325.

If the system determines that the received search query does not match a recorded user correction of an instance attribute, the system can return to receive additional descriptions of user corrections at step 605. If the system determines that the received search query matches a recorded user correction of an instance attribute, the system can change the confidence that one or both of the uncorrected value and the corrected value of the instance attribute correctly characterizes the instance attribute (step 1230). The change or changes in confidence can be embodied in one or more confidence parameters, such as scaled ratings or deltas that can be applied to scaled ratings.

FIG. 14 is a flow chart of a process 1400 for improving search with user corrections. Process 1400 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 1400 can be performed by the search engine 105 in system 100 (FIG. 1). In some implementations, process 1400 can be performed in response to the receipt of a trigger, such as a user request to use user corrections to improve search. Process 1400 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 1400 can be performed in conjunction with the activities of one or more of processes 600, 700, 1200 (FIGS. 6, 7, 12). The system performing process 1400 can receive a description of a user correction of a value of an instance attribute (step 605). The system can also verify the user correction (step 1405). In some implementations, the verification can establish the suitability of the format and syntax of a value. For example, capitalization, spelling, and the units (meters, feet, inches, etc.) of a value can be confirmed by corroborating the correction with other sources, e.g., one or more electronic documents available on the Internet. In some implementations, such verifications can be used as a preliminary threshold screening to determine whether subsequent activities — such as changing the confidence that a value correctly characterizes an instance attribute — are to be performed. For example, a user correction of characterization of the "height" attribute of the "Great Pyramid of Giza" instance from a value of "455 feet" to a value of "139 meters" need not result in a change in the confidence of either value. Instead, the system can automatically recognize and confirm unit conversions, e.g., feet to meters, mpg to liters-per-100 km, and so on.

In some implementations, a collection of user corrections are verified and assembled into an aggregated feedback data collection. An aggregated feedback data collection can include information describing attributes of instances, candidate values for those attributes of instances, and description information characterizing a collection of user corrections. Such an aggregation of user corrections can be used to determine the extent to which confidence in candidate values has been increased or decreased by user corrections, as described below. FIG. 15 is a schematic representation of an aggregated feedback data collection, namely, an aggregated feedback data table 1500. Data table 1500 is a data structure stored in a digital data storage device for access by a computer program operating on a digital data processing system. Data table 1500 includes a collection of records 1505, 1510, 1515, 1520, 1525, 1530 that each include description information characterizing one or more user corrections of a value that is potentially suitable for characterizing a particular attribute of a particular instance.

Table 1500 includes a collection of columns 1535, 1540, 1545, 1550. Column 1535 includes instance identifiers that identify the instances for which description information has been aggregated. Column 1540 includes attribute identifiers that identify the attributes of the instances for which signaling information derived from user corrections has been aggregated. Column 1545 includes value identifiers that identify the values for which description information has been aggregated. The values identified in column 1545 potentially characterize the attributes of the instances identified in columns 1535, 1540.

Column 1550 includes an inventory of correction information characterizing categories of user corrections involving the attribute of the instance identified in columns 1535, 1540 and the value identified in columnl545. In the illustrated implementation, the categories characterized in column 1550 are delineated on an individual, correction-by- correction, basis by both the class of the user corrections and whether the value identified in column 1545 was a corrected value or an uncorrected value. In the illustrated implementation, the categories of each individual user correction is categorized using a three unit code of the form "w#B," where:

- "w" is an identifier indicating that a user correction is being categorized;

- the number "#" identifies the classification of each individual user correction (here, an integer between one and seven, corresponding to the seven classes described in Table 1); and

- the value "B" is a value that identifies whether the value identified in column 1545 was a corrected value or an uncorrected value in the user correction (here, "U" indicating uncorrected and "C" indicating corrected). In other implementations, user corrections can also be categorized in an aggregated feedback data collection based on information such as the identity of the user making the correction, the date when corrections were made, weighting factors characterizing the correctness of other corrections made by certain users, the context in which corrections were made, and the like. As shown in FIG. 14, the system performing process 1400 can also change the confidence that one or both of the uncorrected value and the corrected value of the instance attribute correctly characterizes the instance attribute (step 1230). In implementations in which user corrections are individually categorized in an aggregated feedback data collection, confidence can be changed by weighting the individual correction categories. For example, the individual correction categories can be weighted using weighting parameters collected in a weighting parameter data collection.

FIG. 16 is a schematic representation of a weighting parameter data collection, namely, a weighting parameter data table 1600. Data table 1600 is a data structure stored in a digital data storage device for access by a computer program operating on a digital data processing system. Data table 1600 includes a collection of records 1605, 1610, 1615, 1620, 1625, 1630, 1635, 1640 that each include information characterizing the weight of certain categories of user corrections.

Table 1600 includes a collection of columns 1645, 1650. Column 1645 includes correction category identifiers that characterize categories of user corrections. For example, the correction category identifiers can identify categories of user corrections in the same manner that categories of user corrections are characterized in an aggregated feedback data collection, such as in column 1550 of aggregated feedback data able 1500 (FIG. 15).

Column 1650 includes weighting parameters that embody the magnitude of the change in confidence associated with user corrections of the corresponding category. For example, in the illustrated implementation, a weight of 0.9 in records 1615 can indicate that the corrected value selected by a user from a collection after review and direct confirmation from a source (i.e., class 2) has a larger impact on the confidence than when that same value is selected by a user (as the "corrected value") without review and direct confirmation from a source.

Since the weighting of different categories of user corrections is different, appropriate changes to the confidence that a value correctly characterizes an attribute of an instance can be made. For example, corrections that are made after searches can have a larger impact on confidence than corrections made without searches. As another example, attempts to alter a value by confirming the value directly with a source can have a larger impact on confidence than user deletions of an uncorrected value without direct confirmation from a source.

In other implementations, other characteristics of user corrections can be considered in categorizing and/or weighting user corrections. For example, user corrections made by individuals who have a history of making appropriate corrections can be weighed more heavily that user corrections made by other individuals. As another example, more recent user corrections can be weighed more heavily than older user corrections.

As shown in FIG. 14, the system performing process 1400 can also rank one or both of uncorrected value and corrected value of instance attribute in a result set responsive to the search query (step 1410). In this regard, a value that is more likely to correctly characterize an attribute of an instance is generally ranked above a value that is less likely to correctly characterize an attribute of an instance.

The ranking can reflect the change in confidence that the value correctly characterizes the instance attribute. For example, corrections of different categories can be weighed differently, e.g., using weighting parameters such as shown in weighting parameter data table 1600 (FIG. 16), to generate deltas that are applied to a scaled rating.

For example, in some implementations, a search for an attribute of a value can be conducted in a database or an electronic document collection. The database can include information characterizing, e.g., a collection of structured presentations displayed previously for other users. The search can yield candidate values, each having an individual initial confidence score that embodies the likelihood that the candidate value correctly characterizes the attribute of the instance. Such an initial confidence score can be based on measures such as, e.g., keyword matching, fonts, subdivisions, the precise location of each word, the content of neighboring web pages, and the like. The initial confidence score can be in the form of a scaled rating (e.g., a rating scaled between a lowest possible value (e.g., "0") and a highest possible value (e.g., a "1").

Deltas that embody the change in confidence that the value correctly characterizes the instance attribute can then be applied to the initial confidence scores. The application of the deltas to the initial confidence score can yield a changed confidence score that can be used, e.g., to change the content of a result set or re-rank the results in a result set. For example, if inclusion in a result set requires a certain minimum confidence level, the application of deltas to the initial confidence score in a value increase the confidence in that value above the minimum confidence level so that the content of a result set changes. As another example, the application of deltas to the initial confidence score of one value can increase confidence in that value above the confidence score of another value (or decrease confidence in that value below the confidence score of another value). If the results in a result set are ranked, then such changes in the level of confidence can change the ordering of results in the result set. If the results in a result set are constrained to a certain number (e.g., constrained to the four most likely results), then such changes in the level of confidence in results can change the content of the result set.

In some implementations, the application of deltas to initial confidence scores includes multiplying the number of occurrences of each category of user correction by weighting parameters that embody the magnitude (and possibly direction) of the change in confidence associated with that category. The products can then be added to respective initial confidence scores. In some implementations, the magnitude of the weighting parameters, as well as, e.g., the magnitude of a scalar value applied to ensure that the weighting is scaled in accordance with the scale of the initial confidence score, can be determined to maximize the total number of values that are correct after application the confidence scores. The results in a result set can be ranked based on the sum. The result set with the ranked value or values can be provided to a user, e.g., in a message transmitted over a data transmission network, e.g., message 140 (FIG. 1).

FIG. 17 is a flow chart of a process 1700 for improving search with user corrections. Process 1700 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 1700 can be performed by the search engine 105 in system 100 (FIG. 1). In some implementations, process 1700 can be performed in response to the receipt of a trigger, such as a user request to use user corrections to improve search. Process 1700 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 1700 can be performed in conjunction with the activities of one or more of processes 600, 700, 1200, 1400 (FIGS. 6, 7, 12, 14).

The system performing process 1700 can receive a description of a search query (the response to which includes an attribute value for an instance), a result set of candidate values for characterizing the instance attribute, and initial confidences that those values correctly characterize the instance attribute (step 1705). The system can also access a user correction log, such as user correction history 110 (FIG. 1), to search for user corrections of the candidate values in the result set (step 1710).

The system performing process 1700 can also determine whether corrections of the candidate values in the result set are found in the user correction log (step 1715). If the system determines that corrections of the candidate values in the result set are not found, then the system can leave the initial confidences that those values correctly characterize the instance attribute unchanged (step 1717). If the system determines that corrections of the candidate values in the result set are found, then the system can weight the different categories of user corrections (step 1720). For example, in some implementations, the system can weight the different categories of user corrections using the weighting parameters in weighting parameter data table 1600 (FIG. 16).

FIG. 18 is a schematic representation of another weighting parameter data table 1800. Data table 1800 is a data structure stored in a digital data storage device for access by a computer program operating on a digital data processing system. Data table 1800 includes a collection of records 1805, 1810, 1815, 1820, 1825, 1830, 1835, 1840, 1845, 1850, 1855, 1860, 1865, 1870 that each include information characterizing the weight of certain categories of user corrections.

Table 1800 includes a collection of columns 1875, 1880. Column 1875 includes correction category identifiers that characterize categories of user corrections. For example, the correction category identifiers can identify categories of user corrections in the same manner that categories of user corrections are characterized in an aggregated feedback data collection, such as in column 1550 of aggregated feedback data able 1500 (FIG. 15).

Column 1880 includes weighting parameters that embody both the magnitude and the direction of the change in confidence associated with user corrections of the corresponding category. For example, in the illustrated implementation, the negative weights in records 1805, 1810, 1815, 1820, 1830, 1835 indicate that the confidence in the values subject to user corrections of the corresponding categories has been decreased. As another example, in the illustrated implementation, the positive weights in records 1825, 1840, 1845, 1850, 1855 indicate that the confidence in the values subject to user corrections of the corresponding categories has been increased. The magnitude of the changes in confidence is indicated by the absolute value of the weights.

As shown in FIG. 17, the system performing process 1700 can aggregate the weights of the corrections of various candidate values (step 1725). In some implementations, the system can sum the weights in order to aggregate them. For example, in the context of the weighting parameters in data table 1800 (FIG. 18), the system can arrive at a sum of "10" when five user corrections of the category W5U have been made. As another example, the system can arrive at a sum of "-10" when five user corrections of the category W4U have been made.

The system performing process 1700 can also assign the impact that the aggregated weights are to have on the confidences in the values in the result set (step 1730). The assigned impact of the aggregated weights need not scale linearly with the magnitude of the aggregation of the weights. For example, in some implementations, the impact of the aggregated weights is a sigmoid function of the magnitude of the aggregation of the weights. For example, the impact of the aggregated weights can be assigned using Equation 1,

1

F(s) = 1 _i_ ,(-**) Equation 1 l + e^[ where F(s) is the impact of the aggregated weights "s" and k is a form parameter that helps determine the relationship between the impact of the aggregated weights and the magnitude of the aggregated weights. In implementations where weights such as those in column 1880 of data table 1800 (FIG. 18) are aggregated by summing, k can have a value of approximately two.

The system performing process 1700 can also change the confidence that one or more of the values in the result set correctly characterize the instance attribute (step 1735). For example, the system can multiply the individual confidences received at step 1705 by the respective impacts of the aggregated weights assigned at step 1730. The system can also rank the values in the result set according to their respective confidences (step 1740).

FIG. 19 is a schematic representation of a system 1900 in which a group of related instances is identified. Related instances are instances that share one or more common attributes. In system 1900, a group of related instances is identified in response to a search query. The search query specifies the attributes that are shared by the related instances. The attributes that are shared by a group of related instances can be specified by the search query explicitly, implicitly, or both explicitly and implicitly. For example, a search query "cities" implicitly specifies instances of discrete densely populated urban areas. As another example, a search query "cities located in North America" explicitly identifies that such urban areas are to be located in North America.

System 1900 includes a search engine 1905, a collection 1910 of groups of identifiers of instances, and a client 1915. Client 1915 is a device for interacting with a user and can be implemented as a computer programmed with machine-readable instructions. Client 1915 can include one or more input/output devices and can receive, from a user, a search query that specifies the attributes that are shared by a group of related instances. For example, a user who is currently interacting with client 1915 can enter a search query using an input device such as a mouse or a keyboard. The search query can include text. Examples of textual search queries include "US presidents" and "North American cities." As another example, a user can enter a search query by interacting with or referring to a graphical elements displayed on display screen 1920. For example, a user can click on a cell in a structured presentation or formulate a search query that refers to a feature that appears in a structured presentation (e.g., "ROW 1 "). Structured presentations are described in more detail below. Client 1915 can also present a group of identifiers of related instances that shares the attributes specified by the search query. In the illustrated example, client 1915 includes a display screen 1920 that displays a presentation 1925. Presentation 1925 indicates that a group (i.e., "CATEGORY X") includes a collection of related instances (i.e., identified by identifiers "INSTANCE A," "INSTANCE_B," and "INSTANCE_C"). In the illustrated implementation, presentation 1925 is text. Other presentations can indicate that a group includes a collection of related instances. For example, structured presentations can identify a group in a column header and a collection of related instances in the cells in the column under that header. In response to receipt of the search query, client 1915 transmits a representation of the search query, or the search query itself, to search engine 1905 in a message 1935. Message 1935 can be transmitted over a data communications network. Search engine 1905 can receive message 1935 and use the content of message 1935 to define parameters for searching. Search engine 1905 can be implemented on one or more computers deployed at one or more geographical locations that are programmed with one or more sets of machine-readable instructions for identifying a relevant group of related instances from the groups of instances in collection 1910. In some implementations, other functionality — i.e., functionality in addition to the functionality of search engine 1905 — can be implemented on the one or more computers. Search engine 1905 identifies a relevant group of related instances according to the parameters for searching defined by the content of message 1935. The search can yield a result set of relevant instances responsive to the search query described in message 1935. The content of the result set, the arrangement of instances in the result set, or both can reflect the likelihood that the constituent instances are relevant to the search query. In some implementations, the content or arrangement of instances in the result set can also reflect other factors, such as the relative importance of the instances or a confidence that the instances are indeed responsive to the search query.

The groups of identifiers of instances in collection 1910 can be found in or drawn from electronic documents of an unstructured electronic documents collection. For example, collection 1910 can be groups of identifiers of instances that are found in electronic documents available on the Internet. The source documents of the groups of identifiers of instances are thus not necessarily constrained to conform with a predetermined structure that can be exploited for the extraction of information. For this reason, one or more computers can execute one or more sets of machine-readable instructions that are tailored to identify and extract groups of identifiers of instances from an unstructured electronic documents collection. Machine-readable instructions that are tailored in this way can be referred to as "extractors."

Collection 1910 can include, e.g., lists of instance identifiers 1945, tables of instance identifiers 1950, and structured text 1955 that includes instance identifiers. A list of instance identifiers 1945 is an ordered series of words or numerals. A list of instance identifiers can be found in text and identifiable, e.g., by grammatical conventions or mark-up tags. For example, the instance identifiers in a list can be delineated by commas or semicolons in text. A table of instance identifiers 1950 is a systematic arrangement of instance identifiers. For example, instance identifiers can be arranged in rows or columns. In electronic documents, tables can be identified, e.g., by lines or spaces that delineate rows and columns or by markup tags. Structured text 1955 includes other structured arrangements of instance identifiers, such as instance identifiers ordered by bullet points or instances in a series of paragraph headings. In electronic documents, structured text 1955 can be identified, e.g., by the structural features of the arrangement of instances or by mark-up tags.

In some implementations, collection 1910 can also include groups of one or more instance identifiers that are formed using text extraction techniques. In particular, textual patterns that explicitly or implicitly indicate that an identified instance has certain attributes can be used to form groups of one or more instance identifiers. For example, text such as "New York, the largest city in North America, . .." and "Quebec was the first North American city to be designated a UNESCO World Heritage Site. .." can be identified using pattern identification techniques. For example, text extraction techniques using Hearst patterns or the approaches described in, e.g., "The Role of Documents vs. Queries in Extracting Class Attributes from Text," by M. Pasca, B. Van Durme, and N. Garera

(CIKM'07, November 24-8, 2007, Lisboa, Portugal) and "Weakly-Supervised Acquisition of Open-Domain Classes and Class Attributes fromWeb Documents and Query Logs" by M. Pasca, B. Van Durme (Proceedings of ACL-08: HLT, pages 19-27, Columbus, Ohio, USA, June 2008) can be used. Instance identifiers can be extracted from text and combined to form a group of instance identifiers having the attributes explicitly and implicitly associated with, e.g., North American cities. Extractors can use such characteristics to identify and extract groups of instances from an unstructured electronic documents collection.

Search engine 1905 can transmit a representation of the result set to client 1915 in a message 1940. Message 1940 can be transmitted, e.g., over the same data communications network that transmitted message 1935. Client 1915 can receive message 1940 and use the content of message 1940 to display a presentation 1925 on display screen 1920. Presentation 1925 indicates that one or more common attributes are shared by a group of instances, namely, at least some the instances in the result set described in message 1935. In some implementations, presentation 1925 can use text to identify the shared attributes and instance identifiers. For example, in the illustrated implementation presentation 1925 describes that instance identified as "INSTANCE_A," "INSTANCE_B," and "INSTANCE_C" share the attribute of belonging to a category "CATEGORY_X." Category "CATEGORY_X" can explicitly or implicitly specify the attributes shared by instance identified as "INSTANCE_A," "INSTANCE_B," and "INSTANCE_C." In some implementations, presentation 1925 can use the spatial arrangement and positioning of information to identify that a group of instance shares one or more common attributes. For example, presentation 1925 can be a structured presentation, as described further below.

FIG. 20 is a flow chart of a process 2000 for identifying a group of related instance identifiers. Process 2000 can be performed by one or more computers that perform operations by executing one or more sets of machine-readable instructions. For example, process 2000 can be performed by the search engine 1905 in system 1900. The system performing process 2000 receives a query (step 2005). For example, in the context of system 1900 (FIG. 19), the system can receive a representation of the search query or the search query itself in message 1935 over a data communications network.

The system performing process 2000 identifies that the query inquires about a group of related instances (step 2010). The query can be identified as inquiring about a group of related instances from the content of the query, the context of the query, or both. For example, the terms in a text search query "cities in California" can be identified as inquiring about a group of related instances such as "San Diego," "Los Angeles," and "Bakersfield" due to the plural term "cities" being characterized by a common attribute of those instances, namely, being "in California." As another example, the terms in a search query "Ivy League schools" can be identified as inquiring about a group of related instances (such as "Cornell," "Columbia," and "Brown") due to the plural term "schools" being characterized by a common attribute "Ivy League." The context of the receipt of the search query can also be used to identify that a query inquires about a group of related instances. For example, an express indication by a user or the history of previous queries can be used to identify that a search query inquires about a group of related instances.

The system performing process 2000 identifies electronic documents that are relevant to the search query (step 2015). The electronic documents can be identified by matching text, concepts, or both to entries in an indexed database of electronic documents. The match between the text or concepts in the electronic documents can be used to determine a page rank that embodies the relevance of the electronic documents to the search query, as well as other factors. Examples of these other factors include, e.g., the age of the electronic document, the number of links from other electronic documents to the electronic document, the likelihood that the electronic document is a "spam document," and the like. The system performing process 2000 identifies groups of instance identifiers in the relevant electronic documents (step 2020). For example, groups of instance identifiers can be identified by delineations, mark-up tags, or other characteristics of the arrangement of instance identifiers in the relevant electronic documents. In some implementations, the groups of instance identifiers can be extracted from their respective source electronic documents and assembled into a collection, e.g., collection 1910 in system 1900 (FIG. 19).

The system performing process 2000 determines the relevance of each group of instance identifiers to the search query (step 2025). In general, the relevance of a group of instance identifiers to a search query will differ from the relevance or page rank of its source electronic document to that same query. For example, at least some of the text and concepts that appear in a source electronic document will generally be omitted from a group of instance identifiers in that document. In some implementations, the relevance of a group of instance identifiers can be determined according to the relevance or page rank of its source electronic document and other factors, as described further below. The system performing process 2000 scores the relevance of the instances that appear in the groups individually (step 2030). The scores of the individual instance identifiers can embody the likelihood that each individual instance is relevant to the search query. In some implementations, individual instance identifiers are scored according to the relevance of the groups in which the instance identifiers appear, the overlap between the instance identifiers that appear in different groups, other features of the search that identified the groups in which the instance identifiers appeared, or combinations of these and other factors. A single group of instance identifiers from a single source electronic document can thus include a collection of instance identifiers that are scored differently. Examples of different approaches to scoring the relevance of groups are described further below. The system performing process 2000 ranks the individual instance identifiers according to their scores (step 2035). The ranking can characterize the likelihood that individual instances are relevant to the search query. For example, an instance with a high ranking is one that is likely to be an entity that has the attributes explicitly or implicitly identified in the search query. On the other hand, an instance with a low ranking is one that is unlikely to be an entity that has the attributes explicitly or implicitly identified in the search query. The ranked instance identifiers can be output in a result set provided to a user, e.g., in a message transmitted over a data transmission network, e.g., message 1940 (FIG. 19).

FIG. 21 is a schematic representation 2100 of a process for identifying a group of related instance identifiers. The process can be performed by one or more computers that perform operations by executing one or more sets of machine-readable instructions. For example, representation 2100 can represent the identification of related instance identifiers using process such as process 2000 (FIG. 20) in a system such as system 1900 (FIG. 19).

A collection of electronic documents 2105 can be searched to yield a collection 21 10 of groups of instance identifiers. Collection 2105 can be an unstructured collection of electronic documents 2105. The search can be performed in response to a search query that is used to define parameters for searching. The search can identify relevant documents that include groups of instance identifiers. These groups of instance identifiers can be extracted from their respective source documents to yield collection 2110. The individual instance identifiers within the groups of instances in collection 2110 can then be ranked according to relevance to the search query. The instances can thus be entities that share one or more attributes implicitly or explicitly identified in the search query. The ranked instance identifiers can be output in a result set provided to a user. In some implementations, the top-ranked instance identifiers can be found in different groups of instance identifiers in collection 21 10. For example, the highest-ranked instance identifier may be found in a first group of instance identifiers, whereas the second highest-ranked instance identifier may be absent from that first group of instance identifiers.

FIG. 22 is a flow chart of a process 2200 for identifying electronic documents relevant to a query. Process 2200 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 2200 can be performed by the search engine 1905 in system 1900 (FIG. 19). Process 2200 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 2200 can be performed in conjunction with the activities of process 2000, e.g., at step 2015 (FIG. 20).

The system performing process 2200 receives a search query (step 2205). For example, in the context of system 1900 (FIG. 19), the system can receive a representation of the search query or the search query itself in message 1935 over a data communications network. The system performing process 2200 forms one or more new search queries that are biased to identify groups of instance identifiers (step 2210). Such a biased query can be formed by combining text or concepts represented in the received search query with text or concepts that are biased toward the identification of groups of instance identifiers. For example, text drawn from the received search query (e.g., "rollercoasters" or "hybrid vehicles") can be combined with text biased toward the identification of groups (e.g., "list of [query text]," "this year's [query text]," "my favorite [query text]," "group of [query text]," "the best [query text]," "[query text] such as," "[query text] including," and the like). In some implementations, a biased query can include text or concepts that are intended to prevent certain groups of instance identifiers from being identified by the biased query. For example, in some implementations, a collection of biased queries can be formed, with each including text that specifies a subcategory of a broader category specified by the query text. Examples of such biased queries include "[subcategory_l] [query text] such as," "[subcategory_2] [query text] such as," and "[subcategory_3] [query text] such as." By way of example, suppose the search query "restaurants" is received. As discussed above, a query that is biased to identify groups of instance identifiers such as "[restaurants] including" can be formed. However, in addition to identifying individual restaurants (e.g., the instance identifiers "Bodo's Bagels," "Point Loma Seafood," and "Pat's Pizza"), this biased query could also identify identifiers of instances of culinary sub-categories of restaurants (e.g., "French restaurants," "Italian restaurants," "Thai restaurants," and "fast food restaurants"). In such instances, text that specifies such sub-categories of the broader category can be included in a collection of biased queries. For example, biased queries such as "[French] [restaurants] including," "[Italian] [restaurants] including," and "[Thai] [restaurants] including," "[fast food] [restaurants] including") can be formed.

The system performing process 2200 also forms one or more new search queries that are constrained to search certain sources (step 2215). In some implementations, searches can be constrained to one or more compendia, such as encyclopedia (e.g., www.wikipedia.org) or dictionaries. In some implementations, the searched sources are constrained according to the subject mater of the query. For example, a search for "hybrid vehicles" may be constrained to searching news media and consumer agencies that deal with motor vehicles.

The system performing process 2200 conducts a search using the received search query, the search queries that are biased to identify groups of instance identifiers, and the search queries that are constrained to search certain sources (step 2220). The searches can be run in series or in parallel. The searches using the received search query and the biased search queries can be conducted on the same unstructured collection of electronic documents (e.g., the electronic documents available on the Internet). Each of the searches can yield a separate search result set that identifies electronic documents relevant to the respective search query. The individual documents in each search result set can be scored and ranked, e.g., according to the relevance to the respective search query and other factors.

The system performing process 2200 combines the search result sets yielded by the different searches into a combined search result set (step 2225). The electronic documents identified in the combined search result set can be ranked, e.g., according to the relevancy score or page rank determined in the individual searches. In some implementations, the relevancy scores or page ranks determined in the individual searches are normalized to a standard, e.g., so that the highest ranked electronic documents in each search result set are the three highest ranked electronic documents in the combined search result set. In other implementations, the relevancy scores or page ranks are weighted to prefer electronic documents found in multiple search result sets or electronic documents found in search result sets yielded by a certain one of the searches. For example, the relevancy scores or page ranks of electronic documents in search result sets yielded by queries that are constrained to search certain sources can be preferentially weighted to appear higher in the rankings of the combined search result set. FIG. 23 is a schematic representation 2300 of a process for identifying electronic documents relevant to a query. The process can be performed by one or more computers that perform operations by executing one or more sets of machine-readable instructions. For example, representation 2300 can represent the identification of electronic documents using a process such as process 2200 (FIG. 22) in a system such as system 1900 (FIG. 19). An unstructured collection 2305 of electronic documents (e.g., the documents available on the Internet) can be searched multiple times to yield a source-constrained query result set 2310, a result set 2315 yielded by a query that is biased to identify groups, and a query result set 2320. Result sets 2310, 2315, 2320 can identify the same or different electronic documents in collection 2305. Result sets 2310, 2315, 2320 can be combined together to form a combined result set 2325. Combined result set 2325 identifies electronic documents which appear in unstructured collection 2305.

FIG. 24 is a flow chart of a process 2400 for determining the relevance of groups of instance identifiers to a search query. Process 2400 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 2400 can be performed by the search engine 1905 in system 1900 (FIG. 19). Process 2400 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 2400 can be performed in conjunction with the activities of process 2000, e.g., at step 2025 (FIG. 20). The system performing process 2400 receives a search query (step 2405). For example, in the context of system 1900 (FIG. 19), the system can receive a representation of the search query or the search query itself in message 1935 over a data communications network.

The system performing process 2400 computes the relevance of each of a collection of source documents to the query (step 2410). The relevance can be computed, e.g., by matching a query to text, concepts, or both in an electronic document. The match between the text or concepts in the electronic documents can be used to determine a page rank that embodies the relevance of the electronic documents to the search query and potentially other factors. The system performing process 2400 computes the likelihood that potential groups of instance identifiers in the source documents are actually groups of instance identifiers (step 2415). As described above, delineations, mark-up tags, or other characteristics of the arrangement of instance identifiers in the relevant electronic documents can be used to identify potential groups of instance identifiers. In some circumstances, it is not completely certain that a group of instance identifiers has in fact been identified. For example, although commas are generally used to delineate members of a list in text, commas may sometimes be omitted from a list inadvertently or otherwise. In such cases, the certainty that a series of instance identifiers is in fact a list is decreased. As another example, different textual patterns can be more or less likely to exclusively identify instance identifiers that have certain attributes. The likelihood that a potential group of instance identifiers assembled using such textual patterns actually includes correct instance identifiers can be computed according to the accuracy of the textual patterns used.

As another example, mark-up HTML tags such as <b>, <li>, <td>, <a> , and the like can be used to identify potential groups of instance identifiers. However, such HTML tags do not always delineate lists of items. Instead, HTML authors can use them for other purposes. For example, the HTML tag <li> — which is designed to define list items — can also be used for other formatting purposes or to contain ancillary text that does not identify a group of instance identifiers. Thus, it is not completely certain that even mark-up tags designed to define a group of instance identifiers can actually be used to identify a group of instance identifiers.

The likelihood that a group of instance identifiers has been identified can be computed and expressed as a normalized value between absolute certainty that a group of instance identifiers has been identified (e.g., a "1") and absolute certainty that a group of instance identifiers has not been identified (e.g., a "0").

The system performing process 2400 computes the relevance of each potential group of instance identifiers to the source document that includes the potential group (step 2420). In some circumstances, a group of instance identifiers is unrelated to other content of the electronic document that includes the group of instance identifiers. For example, the cover page of a company newsletter may include a table setting forth the addresses where the company has offices. Although the table is a group of instance identifiers, the content of this table (e.g., office addresses) may be unrelated to other content of the newsletter. The system can compute the relevance of each potential group of instance identifiers to the source document that includes the potential group by comparing the text, the concepts, or both in the potential group of instance identifiers to the text, the concepts, or both in the source document.

The system performing process 2400 ranks the potential groups according to the relevance of source document to query, the likelihood that potential group of instance identifiers is a group, and the relevance of potential group to source document (step 2420). For example, a merit score "SG" can be computed for each potential group of instance identifiers according to a formula that relies upon multiplication, addition, exponentiation, or other computation using the relevance of the source document of the potential group of instance identifiers to the query, the likelihood that the potential group of instance identifiers is in fact a group, and the relevance of the potential group of instance identifiers to the source document that includes the potential group of instance identifiers. For example, in some implementations, the merit score "SG" is computed for each potential group of instances according to the formula:

S_G = R_DQL_GR_GD Equation 1

where "RDG" is the relevance of the source document of the potential group of instance identifiers to the query, "LG" is the likelihood that the potential group of instance identifiers is in fact a group, and "RGD" is the relevance of the potential group of instance identifiers to the source document that includes it. The merit score SG of each potential group of instance identifiers can thus embody the relevance of those potential groups to a search query.

As another example, a merit score "SG" can be computed for each potential group of instance identifiers using machine-learning techniques. For example, the relevance of source document to query, the likelihood that potential group of instance identifiers is a group, and the relevance of potential group to source document can be input into as features into a predictive analytic tree-building algorithm that has been trained using a groups of known relevance to a search query. The merit score "SG" yielded by a predictive analytic tree- building algorithm can embody the percentage of decision trees that have voted for a group. This percentage can be expressed as a number between 0 and 1. In some implementations, the percentage of decision trees that have voted for a group can be adjusted to account for factors such as the number of times that a group appears, the extent to which the members of the group have been refined, and other factors.

FIG. 25 is a flow chart of a process 2500 for scoring instance identifiers according to relevance of groups in which instance identifiers appear. Process 2500 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 2500 can be performed by the search engine 1905 in system 1900 (FIG. 19). Process 2500 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 2500 can be performed in conjunction with the activities of process 200, e.g., at step 2030 (FIG. 20).

The system performing process 2500 receives description information describing potential groups (including the identity of the instance identifiers in the potential groups) and the relevance of these potential groups to a search query (step 2505). For example, the system can receive a listing of the instance identifiers in each potential group and a merit score SG for each potential group.

The system performing process 2500 estimates the likelihood that each instance identifier appears in a relevant group according to relevance of potential groups in which instance identifier appears (step 2510). A group of instance identifiers is relevant to a search query when the group includes instance identifiers that share the attributes that are implicitly or explicitly specified in the search query. The likelihood that each instance identifier appears in a relevant group can thus embody the relevance of the instance identifier to a search query.

In some implementations, the likelihood that each instance identifier appears in a relevant group is estimated according to a method that relies on an expectation maximization algorithm. An expectation maximization algorithm makes maximum-likelihood estimates of one or more parameters of a distribution from a set of data that is incomplete and missing variables. An expectation maximization algorithm can pick a set of parameters that best describes a set of data given a model. In the present context, the set of data are the potential groups. The model assumes that some potential group are relevant to the query (groups "R") whereas other potential groups are not relevant to the query (groups "N"). Further, a given item (i) has a probability of occurring in a relevant group "P(i|R)" and a probability of occurring in an irrelevant group "P(i|N)". The probabilities P(i|R), P(i|N) can initially be estimated based on, e.g., the relevance of the source document of the group to a search query, the likelihood that a group of instances is indeed a group, and the relevance of the group to its source document. The probabilities P(i|R), P(i|N) can then be maximized using the expectation maximization algorithm. The expectation maximization algorithm can be implemented as iterative processes which alternates between expectation steps and maximization steps. In expectation steps, missing variables are estimated from the observed data and current estimates of the parameters of the distribution. In maximization steps, estimates of the parameters of the distribution is maximized under the assumption that the missing variables are known, i.e., have the values estimated in the previous expectation step. As the steps are iteratively repeated, the estimates of the parameters of the distribution converge. Expectation maximization algorithms are described in more detail, e.g., in "Maximum Likelihood from Incomplete Data via the EM Algorithm" by A.P. Dempster, N. M. Laird, D.B. Rubin Journal of the Royal Statistical Society, Series B (Methodological) 39 (1) pp. 1-38 (1977).

FIG. 26 is a flow chart of a process 2600 for scoring instance identifiers according to the relevance of groups in which instance identifiers appear. Process 2600 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 2600 can be performed by the search engine 1905 in system 1900 (FIG. 19). Process 2600 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 2600 can be performed in conjunction with the activities of process 2000, e.g., at step 2030 (FIG. 20).

The system performing process 2600 receives description information describing potential groups (including the identity of the instance identifiers in the potential groups) and the relevance of these potential groups to a search query (step 2605). For example, the system can receive a listing of the instance identifiers in each potential group and a merit score SG for each potential group.

The system performing process 2600 represents features of the instance identifiers in the potential groups in one or more vertex-edge graphs (step 2610). A vertex-edge graph is a representation of a set of objects where some pairs of the objects are connected by links. The interconnected objects are represented by vertices and the links that connect some pairs of vertices are called edges.

FIG. 27 is a schematic representation of a vertex-edge graph 2700 that represents features of the instance identifiers in the potential groups. Vertex-edge graph 2700 includes vertices 2705, 2710, 2715, 2720, 2725, 2730 that are connected pairwise by groups of one or more edges 2735, 2740, 2745, 2750, 2755, 2760, 2765. Vertex-edge graph 2700 is a undirected graph. Each of vertices 2705, 2710, 2715, 2720, 2725, 2730 represents an instance identifier which is found in a potential group that was identified in one or more searches. For example, vertex 2720 represents the instance identifier "George Washington," vertex 2720 represents the instance identifier "Franklin D. Roosevelt," and vertex 2730 represents the instance identifier "Martha Washington." The potential groups from which vertices 2705, 2710, 2715, 2720, 2725, 2730 are drawn can be constrained to have at least some threshold level of relevance to the search query. The relevance of the potential groups to the search query can be determined, e.g., using process 2400 (FIG. 24).

Each group of edges 2735, 2740, 2745, 2750, 2755, 2760, 2765 represents co- occurrences of the vertices connected by the edges in a potential group. For example, the four different edges in edge group 2755 can represent that "George Washington" vertex 2720 was found in four potential groups that also included "Franklin D. Roosevelt." In some implementations, other features can be can be represented by edges. Table 1 is a list of examples of such features.

EXAMPLE FEATURES

-query that identified source document that includes vertex pair;

-class of query (e.g., biased query, source-constrained query) that identified source document(s) that include vertex pair;

-number of potential groups identified by the query that identified source document(s) that include vertex pair;

- relevance of the source document

-source document of the vertex pair; -extractor that identified vertex pair;

-other instances in potential groups where vertex pair is found;

TABLE 1

In some implementations, other features that can be represented by edges can be determined from the characteristics of neighboring items.

FIG. 28 is a schematic representation of another vertex-edge graph 2800 that represents features of the instance identifiers in the potential groups. Vertex-edge graph 2800 includes vertices 2805, 2810, 2815, 2820, 2825, 2830 that are connected pairwise by individual edges 2835, 2840, 2845, 2850, 2855, 2860, 2865. Each of edges 2835, 2840, 2845, 2850, 2855, 2860, 2865 is weighted by a respective weight 2870, 2875, 2880, 2885, 2890, 2895, 2899. Vertex-edge graph 2800 is thus a weighted undirected graph. Each of vertices 2805, 2810, 2815, 2820, 2825, 2830 represents a potential group of instance identifiers. For example, vertex 2815 represents a group of six instance identifiers, vertex 2820 represents a group of three instance identifiers, and vertex 2825 represents a group of three instance identifiers. The potential groups represented in vertices 2805, 2810, 2815, 2820, 2825, 2830 can be constrained to have at least some threshold level of relevance to the search query. The relevance of the potential groups to the search query can be determined, e.g., using process 2400 (FIG. 24).

Each of edges 2735, 2740, 2745, 2750, 2755, 2760, 2765 represents the "overlap" between the pair of vertices it connects. The "overlap" between two vertices is the number of instance identifiers common to the potential groups represented by those vertices. The overlap can be represented by the respective weight 2870, 2875, 2880, 2885, 2890, 2895, 2899 associated with each edge 2735, 2740, 2745, 2750, 2755, 2760, 2765. For example, weight 2880 represents that there are no instance identifiers which are common to the potential groups represented by vertices 2815, 2820 and weight 2885 represents that there are three instance identifiers which are common to the potential groups represented by vertices 2815, 2825. For the sake of clarity, other zero weight edges have been omitted from vertex- edge graph 2800. Vertex-edge graph 2800 thus represents the overlap between the potential groups in which instance identifiers are found.

The vertices and edges of graphs 2700, 2800 need not be displayed in pictorial form, as shown. Rather, graphs 2700, 2800 can remain abstract representations, e.g., in a computer that performs digital data processing operations.

Returning to FIG. 26, the system performing process 2600 scores the instance identifiers in the potential groups according to the features represented by the edges in the vertex-edge graph (step 2615). The nature of the scoring can depend on the features represented in the vertex-edge graph as well as the role of the instance identifiers themselves in the vertex-edge graph.

In some implementations, the instance identifiers in the potential groups can be scored using the result of a machine-learning technique performed by a computers that executes one or more sets of machine-readable instructions. A training set of data can first be used to allow a machine to establish a set of rules for scoring instance identifiers. This set of rules for scoring can then be applied to other sets of data.

For example, in the context of vertex-edge graph 2700 (FIG. 27), a predictive analytic tree-building algorithm such as classification and regression tree analysis can score the instances according to the likelihood that they belong in a relevant group, classify the instance identifiers as to whether they belong in a relevant group, or both. Tree-building algorithms determine a set of if-then logical rules for scoring instance identifiers that permit accurate prediction or classification of cases. Trees are built by a collection of rules based on values of variables in a modeling data set. The rules can be selected based on how well splits based on the values of different variables can differentiate observations. Examples of tree- building algorithms are described, e.g., in: "Classification and Regression Trees," Breiman et al., Chapman & Hall (Wadsworth, Inc.) New York (1984); "CART: Tree-structured Non- parametric Data Analysis," Steinberg et al., Salford Systems, San Diego, Calif, U.S.A. (1995); and "Random Forests," Breiman, Machine Learning, Vol. 45 : 1. (2001), pp. 5-32. Such a predictive analytic tree-building algorithm can be trained using a group of instance identifiers of confirmed accuracy that are relevant to a search query, a set of potential groups of instance identifiers that have been identified from an unstructured collection of electronic documents, and features of the instance identifiers in the potential groups. The decision trees can make their decisions based on features, e.g., the features listed in Table 1. For example, an exhaustive list of the Presidents of the United States of America, a set of potential groups of instance identifiers that have been identified in response to a search query inquiring about the Presidents of the United States of America, and features of the instance identifiers in these potential groups can be used by a machine to establish a classification and regression tree. The set of if-then logical rules for scoring in this classification and regression tree can then be applied to other sets of potential groups of instance identifiers that have been identified in response to other search queries, as well as the features of the instance identifiers in these other potential groups. The application of these logical conditions can score the instance identifiers in these other potential groups according to the likelihood that they belong in a relevant group, classify the instances as to whether they belong in a relevant group, or both.

In some implementations, the instance identifiers in the potential groups can be scored by identifying cliques in a vertex-edge graph. A clique is a set of pairwise adjacent vertices, or in other words, an induced subgraph which is a complete graph. The size of a clique is the number of vertices within the clique. In the context of vertex-edge graph 2800 (FIG. 28), vertices 2815, 2830 form a complete bipartite graph (or a "biclique") in which every instance identifier in vertex 2815 is also found in vertex 2830. This high degree of overlap is represented by the relatively high value of weight 2890 (i.e., a value of six). Vertices 2815, 2825 have a middling degree of overlap and share only three constituent instance identifiers. This middling degree of overlap is represented by the intermediate value of weight 2885 (i.e., a value of three). Vertices 2820, 2830 do not overlap at all and this lack of overlap is represented by the zero value of weight 2899.

The identification of cliques and the overlap between vertices can be used to score the instance identifiers in the potential groups represented by these vertices. For example, instance identifiers in large cliques and/or with a high degree of overlap can be treated as more likely to have the attributes specified by a search query, whereas instance identifiers in small cliques and/or with a low degree of overlap can be treated as less likely to have the attributes specified by the search query. In some implementations, the size of the clique can be weighted more heavily in scoring than the degree of overlap in smaller cliques. For example, vertices 2815, 2825, 2830 form a three-vertex clique edges having edges with a minimum weight of three, whereas vertices 2815, 2830 form a two-vertex clique edges having edges with a minimum weight of six. The larger three-vertex clique can be taken as a collection of independent sources confirming that three common instance identifiers are likely to have the attributes specified by a search query. In some implementations, a representation of the set of scored instance identifiers can then be transmitted to a client, e.g., client 1915 in system 1900 (FIG. 19).

FIG. 29 is a flow chart of a process 2900 for rescoring instance identifiers. Process 2900 can be performed by one or more computers that perform digital data processing operations by executing one or more sets of machine-readable instructions. For example, process 2900 can be performed by the search engine 1905 in system 1900 (FIG. 19). Process 2900 can be performed in isolation or in conjunction with other digital data processing operations. For example, process 2900 can be performed in conjunction with the activities of process 2500, e.g., after step 2510 (FIG. 25) or in conjunction with the activities of process 2600, e.g., after step 2615 (FIG. 26). The system performing process 2900 receives description information describing a search query and a collection of scored instance identifiers (step 2905). The instance identifiers can be scored according to the likelihood that they have the attributes specified by the received search query.

The system performing process 2900 can remove instance identifiers that match the text of the received search query, or permutations of the text of the received search query

(step 2910). For example, if a search queries that inquires about "U.S. Presidents," instance identifiers such as "presidents," "U.S. President," and the like can be removed from the set of scored instance identifiers. In some implementation, other instance identifiers such as vulgar words can be removed from the set of scored instance identifiers. The system performing process 2900 can change the score of like or related instance identifiers in a set of scored instance identifiers (step 2915). Examples of like or related instance identifiers include that which identify the same instance using words that originate from different orthographies (e.g., defense/defence, behavior/behaviour), words that are different transliterations of foreign words (e.g., tsar/czar/csar), words that are abbreviations or diminutives (Robert Kennedy/Bobby Kennedy/R. F. Kennedy), and words that are a substring of another instance identifier (e.g., George Washington/Biography of George Washington). In some implementations, like or related instance identifier can be combined into a single instance identifier. The system performing process 2900 can also weight the scores of instance identifiers according to the frequency at which the instance identifiers appear in the electronic documents of an unstructured electronic documents collection (step 2920). For example, as a group of electronic documents is being indexed, the number of occurrences of different terms (including the instance identifier terms) appearing in the electronic documents can be determined. The scores for different instance identifiers can then be scaled, e.g., by multiplying the scores by a value that is approximately the inverse of the number of occurrences. As a result, the scores of instance identifiers that appear often in the electronic documents can be decreased relative to the scores of instance identifiers that appear only rarely in the electronic documents. In some implementations, other activities can be used to rescore a collection of instances. For example, in some implementations, instance identifiers that match a fixed blacklist can be removed from the collection altogether, in effect, reducing their score to zero. The blacklist can include individual instance identifiers or identifier/search query pairs.

In some implementations, the score of an instance identifier can be changed to reflect the likelihood that the identifier characterizes a category of instances. In some implementations, the likelihood that the identifier characterizes a category of instances can be determined from a log of search queries submitted by different human users. For example, in response to a user switching between searching with a search query that identifies a scored instance (e.g., the search query "car") to searching with a search query that uses that identifier to identify a category (e.g., the search queries "types of car" and "list of cars"), the score of that instance identifier can be decreased. As another example, in response to a user switching between searching with a search query that identifies a scored instance (e.g., the search query "car") to searching with an identifier of a more specific instance within that category (e.g., the search query "prius" within the category "car"), the score of the more specific instance identifier can be increased.

In some implementations, a representation of the set of rescored instance identifier can then be transmitted to a client, e.g., client 1915 in system 1900 (FIG. 19). FIGS. 3-5 are examples of structured presentations 300, 400, 500 that present a group of related instance identifiers to user. Structured presentations 300, 400, 500 can be presented to a user, e.g., by client 1915 in presentation 1925 on display screen 1920 (FIG. 19). Structured presentations 300, 400, 500 use the spatial arrangement and positioning of information to identify that a group of instance shares one or more common attributes. Embodiments of the subject matter and the operations described in this specification can be implemented in digital electronic circuitry, or in computer software, firmware, or hardware, including the structures disclosed in this specification and their structural equivalents, or in combinations of one or more of them. Embodiments of the subject matter described in this specification can be implemented as one or more computer programs, i.e., one or more modules of computer program instructions, encoded on a computer storage medium for execution by, or to control the operation of, data processing apparatus. Alternatively or in addition, the program instructions can be encoded on an artificially-generated propagated signal, e.g., a machine-generated electrical, optical, or electromagnetic signal, that is generated to encode information for transmission to suitable receiver apparatus for execution by a data processing apparatus. A computer storage medium can be, or be included in, a computer-readable storage device, a computer-readable storage substrate, a random or serial access memory array or device, or a combination of one or more of them. Moreover, while a computer storage medium is not a propagated signal, a computer storage medium can be a source or destination of computer program instructions encoded in an artificially-generated propagated signal. The computer storage medium can also be, or be included in, one or more separate physical components or media (e.g., multiple CDs, disks, or other storage devices).

The operations described in this specification can be implemented as operations performed by a data processing apparatus on data stored on one or more computer-readable storage devices or received from other sources.

The term "data processing apparatus" encompasses all kinds of apparatus, devices, and machines for processing data, including by way of example a programmable processor, a computer, a system on a chip, or multiple ones, or combinations, of the foregoing The apparatus can include special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit). The apparatus can also include, in addition to hardware, code that creates an execution environment for the computer program in question, e.g., code that constitutes processor firmware, a protocol stack, a database management system, an operating system, a cross-platform runtime environment, a virtual machine, or a combination of one or more of them. The apparatus and execution environment can realize various different computing model infrastructures, such as web services, distributed computing and grid computing infrastructures.

A computer program (also known as a program, software, software application, script, or code) can be written in any form of programming language, including compiled or interpreted languages, declarative or procedural languages, and it can be deployed in any form, including as a stand-alone program or as a module, component, subroutine, object, or other unit suitable for use in a computing environment. A computer program may, but need not, correspond to a file in a file system. A program can be stored in a portion of a file that holds other programs or data (e.g., one or more scripts stored in a markup language document), in a single file dedicated to the program in question, or in multiple coordinated files (e.g., files that store one or more modules, sub-programs, or portions of code). A computer program can be deployed to be executed on one computer or on multiple computers that are located at one site or distributed across multiple sites and interconnected by a communication network. The processes and logic flows described in this specification can be performed by one or more programmable processors executing one or more computer programs to perform actions by operating on input data and generating output. The processes and logic flows can also be performed by, and apparatus can also be implemented as, special purpose logic circuitry, e.g., an FPGA (field programmable gate array) or an ASIC (application-specific integrated circuit).

Processors suitable for the execution of a computer program include, by way of example, both general and special purpose microprocessors, and any one or more processors of any kind of digital computer. Generally, a processor will receive instructions and data from a read-only memory or a random access memory or both. The essential elements of a computer are a processor for performing actions in accordance with instructions and one or more memory devices for storing instructions and data. Generally, a computer will also include, or be operatively coupled to receive data from or transfer data to, or both, one or more mass storage devices for storing data, e.g., magnetic, magneto-optical disks, or optical disks. However, a computer need not have such devices. Moreover, a computer can be embedded in another device, e.g., a mobile telephone, a personal digital assistant (PDA), a mobile audio or video player, a game console, a Global Positioning System (GPS) receiver, or a portable storage device (e.g., a universal serial bus (USB) flash drive), to name just a few. Devices suitable for storing computer program instructions and data include all forms of non- volatile memory, media and memory devices, including by way of example semiconductor memory devices, e.g., EPROM, EEPROM, and flash memory devices; magnetic disks, e.g., internal hard disks or removable disks; magneto-optical disks; and CD-ROM and DVD-ROM disks. The processor and the memory can be supplemented by, or incorporated in, special purpose logic circuitry. To provide for interaction with a user, embodiments of the subject matter described in this specification can be implemented on a computer having a display device, e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor, for displaying information to the user and a keyboard and a pointing device, e.g., a mouse or a trackball, by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback, e.g., visual feedback, auditory feedback, or tactile feedback; and input from the user can be received in any form, including acoustic, speech, or tactile input. In addition, a computer can interact with a user by sending documents to and receiving documents from a device that is used by the user; for example, by sending web pages to a web browser on a user's client device in response to requests received from the web browser.

Embodiments of the subject matter described in this specification can be implemented in a computing system that includes a back-end component, e.g., as a data server, or that includes a middleware component, e.g., an application server, or that includes a front-end component, e.g., a client computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the subject matter described in this specification, or any combination of one or more such back-end, middleware, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication, e.g., a communication network. Examples of communication networks include a local area network ("LAN") and a wide area network ("WAN"), an inter-network (e.g., the Internet), and peer-to-peer networks (e.g., ad hoc peer- to-peer networks).

The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. In some embodiments, a server transmits data (e.g., an HTML page) to a client device (e.g., for purposes of displaying data to and receiving user input from a user interacting with the client device). Data generated at the client device (e.g., a result of the user interaction) can be received from the client device at the server.

While this specification contains many specific implementation details, these should not be construed as limitations on the scope of any invention or of what may be claimed, but rather as descriptions of features specific to particular embodiments of the particular inventions. Certain features that are described in this specification in the context of separate embodiments can also be implemented in combination in a single embodiment. Conversely, various features that are described in the context of a single embodiment can also be implemented in multiple embodiments separately or in any suitable subcombination. Moreover, although features may be described above as acting in certain combinations and even initially claimed as such, one or more features from a claimed combination can in some cases be excised from the combination, and the claimed combination may be directed to a subcombination or variation of a subcombination.

Similarly, while operations are depicted in the drawings in a particular order, this should not be understood as requiring that such operations be performed in the particular order shown or in sequential order, or that all illustrated operations be performed, to achieve desirable results. In certain circumstances, multitasking and parallel processing may be advantageous. Moreover, the separation of various system components in the embodiments described above should not be understood as requiring such separation in all embodiments, and it should be understood that the described program components and systems can generally be integrated together in a single software product or packaged into multiple software products.

Thus, particular embodiments of the subject matter have been described. Other embodiments are within the scope of the following claims. For example, in some implementations, systems such as system 100 include mechanisms for excluding corrections made by non-human users from user correction history 1 10. In some cases, the actions recited in the claims can be performed in a different order and still achieve desirable results. In addition, the processes depicted in the accompanying figures do not necessarily require the particular order shown, or sequential order, to achieve desirable results. In certain implementations, multitasking and parallel processing may be advantageous. These technologies may also be implemented in one or more of the following embodiments.

Embodiment 1. A method performed by one or more data processing apparatus, the method comprising: the data processing apparatus receiving a search query at a data processing apparatus, the search query specifying attributes shared by a group of related instances; the data processing apparatus identifying groups of instance identifiers in an unstructured collection of electronic documents with the data processing apparatus; the data processing apparatus determining relevance of the groups of instance identifiers to the search query with the data processing apparatus; the data processing apparatus scoring at least some of the instance identifiers in the groups of instance identifiers individually with the data processing apparatus; and the data processing apparatus ranking the at least some instance identifiers according to the scores with the data processing apparatus.

Embodiment 2. The method of embodiment 1 , wherein determining the relevance of the groups of instance identifiers to the search query comprises: computing relevance of the groups of instance identifiers to source documents that include the groups of instance identifiers; computing likelihoods that the identified groups of instance identifiers are indeed groups of instance identifiers; and computing relevance of source documents which include the groups of instance identifiers to the search query.

Embodiment 3. The method of embodiment 1 , wherein identifying the groups of instance identifiers comprises: forming a first new query biased to identify groups; forming a second new query constrained to search compendia sources; and searching the unstructured collection of electronic documents with the received query, the first new query, and the second new query.

Embodiment 4. The method of embodiment 1 , further comprising the data processing apparatus rescoring the at least some instance identifiers before ranking.

Embodiment 5. The method of embodiment 1 , wherein scoring at least some of the instance identifiers in the groups of instance identifiers comprises: representing features of the instance identifiers in a vertex-edge graph; and scoring the instance identifiers according to the features represented in the vertex-edge graph. Embodiment 6. The method of embodiment 5, wherein: vertices in the vertex-edge graph represent groups of instance identifiers; and respective edges in the vertex-edge graph are weighted according to overlap between the vertices connected by the edge. Embodiment 7. The method of embodiment 5, wherein: vertices in the vertex-edge graph represent individual instance identifiers; and respective edges in the vertex-edge graph represent features shared by the instance identifiers.

Embodiment 8. The method of embodiment 6, wherein a first edge in the vertex-edge graph represents an extractor that identified a pair of vertices joined by the first edge.

Embodiment 9. The method of embodiment 6, wherein a first edge in the vertex-edge graph represents other instance identifiers in potential groups where vertices joined by the first edge are found.

Embodiment 10. The method of embodiment 6, wherein a first edge in the vertex- edge graph represents a class of the query that identified source document where vertices joined by the first edge are found.

Embodiment 11. The method of embodiment 5, wherein scoring the instance identifiers comprises identifying cliques in the vertex-edge graph.

Embodiment 12. The method of embodiment 1, wherein scoring the instances identifiers comprises scoring the instance identifiers using a predictive analytic tree-building algorithm.

Embodiment 13. The method of embodiment 1, wherein scoring the instance identifiers using the predictive analytic tree-building algorithm comprises: training the predictive analytic tree-building algorithm using a group of instance identifiers of confirmed accuracy that are relevant to a search query, a set of potential groups of instance identifiers that have been identified from an unstructured collection of electronic documents, and features of the instance identifiers in the potential groups; and generating a classification and regression tree.

Embodiment 14. One or more computer storage media encoded with a computer program, the program comprising instructions that when executed by one or more data processing apparatus cause the data processing apparatus to perform operations, the operations comprising: receiving a search query at a data processing apparatus, the search query specifying attributes shared by a group of related instances; searching an electronic document collection to identify identifiers of instance that are responsive to the search query; representing features of the instance identifiers in a vertex-edge graph; and scoring relevance of the instance identifiers to the search query according to the features represented in the vertex-edge graph.

Embodiment 15. The computer storage medium of embodiment 14, wherein the operations further comprise: identifying groups of instance identifiers in the electronic documents of the collection; and determining relevance of the groups of instance identifiers to the search query; and a first feature represented in the vertex-edge graph comprises the relevance of the groups that include respective instance identifiers to the search query.

Embodiment 16. The computer storage medium of embodiment 14, the operations further comprising: identifying electronic documents available on the Internet that are relevant to the search query; and extracting groups of instance identifiers from the electronic documents that are relevant to the search query.

Embodiment 17. The computer storage medium of embodiment 16, the operations further comprising: computing relevance of the electronic documents from which the groups of instance identifiers are extracted to the search query; computing relevance of the groups of instance identifiers to the electronic documents from which the groups of instance identifiers are extracted; and computing likelihoods that the groups of instance identifiers are groups of instance identifiers.

Embodiment 18. The computer storage medium of embodiment 15, wherein identifying the groups of instance identifiers comprises: forming a new query biased to identify groups; and searching the electronic document collection with the new query.

Embodiment 19. The computer storage medium of embodiment 14, wherein a first edge in the vertex-edge graph represents a class of the query that identified a pair of vertices joined by the first edge. Embodiment 20. The computer storage medium of embodiment 14, wherein a first edge in the vertex-edge graph represents other instance identifiers in potential groups where vertices joined by the first edge are found.

Embodiment 21. The computer storage medium of embodiment 14, wherein scoring relevance of the instance identifiers to the search query comprises identifying cliques in the vertex-edge graph.

Embodiment 22. A system comprising: a client device; and one or more computers programmed to interact with the client device and the data storage device, the computers programmed to perform operations comprising: receiving a search query from the client device, the search query explicitly or implicitly specifying attributes of instances; searching an electronic document collection to identify identifiers of instances that may have the attributes specified by the search query; representing features of the search of the electronic document collection in a vertex-edge graph; scoring the instance identifiers that may have the attributes specified by the search query according to the features represented in the vertex- edge graph; and outputting, to the client device, instructions for visually presenting at least some of the instance identifiers.

Embodiment 23. The system of embodiment 22, wherein: outputting the instructions comprises outputting instructions for visually presenting a structured presentation at the client device; and the client device is configured to receive the instructions and cause the structured presentation to be visually presented.

Embodiment 24. The system of embodiment 22, further comprising a data storage device storing a data describing multiple groups of instances.

Embodiment 25. The system of embodiment 22, further comprising a data storage device storing machine-readable instructions tailored to identify and extract groups of instance identifiers from electronic documents in an unstructured collection.

Embodiment 26. The system of embodiment 22, wherein:representing features comprises representing the relevance of the groups in which the instance identifiers appear in the vertex-edge graph; and scoring the instance identifiers comprises scoring the instance identifiers individually according to the relevance of the groups in which the instance identifiers appear to the search query.

Embodiment 27. The system of embodiment 22, wherein scoring the instance identifiers comprises identifying cliques in the vertex-edge graph.

Embodiment 28. The system of embodiment 22, wherein scoring the instance identifiers comprises scoring the instance identifiers according to an extractor represented in the vertex-edge graph.

Embodiment 29. The system of embodiment 22, wherein scoring the instance identifiers comprises scoring the instance identifiers according to a class of a query represented in the vertex-edge graph.

What is claimed is:

Claims

1. A method performed by one or more data processing apparatus, the method comprising: receiving a value result set at the data processing apparatus, the value result set comprising a collection of one or more values, the values being candidates for characterizing an attribute of an instance; accessing historical records of user corrections stored at one or more data storage devices, the historical records describing user corrections of the characterization of instance attributes by values; determining that the historical records of user corrections describe a first user correction involving a first value in the value result set, wherein the first value is involved in the correction as either a corrected value or an uncorrected value; and changing a confidence parameter embodying a confidence that the first value correctly characterizes the attribute of the instance.

2. The method of claim 1 , further comprising: ranking the values in the value result set according to the changed confidence parameter; and visually displaying at least a portion of the value result set on a display screen according to the ranking.

3. The method of claim 2, wherein: visually displaying at least the portion of the value result set comprises presenting a structured presentation to a user; the structured presentation is populated with a first value included in the value result set; and the first value has a confidence parameter indicating that the first value is the value in the value result set that is most likely to correctly characterize the instance attribute.

4. The method of claim 2, wherein visually displaying at least a portion of the value result set comprises displaying a candidate window that includes candidate values for characterizing an instance attribute.

5. The method of claim 1 , wherein changing the confidence parameter comprises applying a delta value suitable to a scaled confidence rating, the scaled confidence rating embodying the confidence that the involved value correctly characterizes the attribute of the instance.

6. The method of claim 5, wherein generating the delta value comprises weighting a category of a user correction of the involved value.

7. The method of claim 5, wherein generating the delta value comprises categorizing the user correction.

8. A computer storage medium encoded with a computer program, the program comprising instructions that when executed by data processing apparatus cause the data processing apparatus to perform operations, the operations comprising: receiving a description of a user correction involving a value characterizing an instance attribute, wherein the value is involved in the correction as either a corrected value or an uncorrected value; changing a confidence parameter reflecting the likelihood that the value correctly characterizes the instance attribute; and ranking a collection of candidate values that includes the value according to respective confidence parameters, including the changed a confidence parameter.

9. The computer storage medium of claim 8, wherein the operations further comprise transmitting a description of the ranked collection of candidate values over a data communication network in response to receipt of a search query, the response to which includes an attribute value for an instance.

10. The computer storage medium of claim 8, wherein receiving the description of the user correction comprises receiving a description of whether that the user confirmed the correction with a source.

11. The computer storage medium of claim 8, wherein receiving the description of the user correction comprises receiving a description that the user did not change an uncorrected value after reviewing an electronic document.

12. The computer storage medium of claim 8, wherein receiving the description of the user correction comprises receiving a description of the uncorrected value prior to the user correction and the corrected value after the user correction.

13. The computer storage medium of claim 8, wherein changing the confidence parameter comprises: categorizing the user correction; and weighting the impact of the user correction on the confidence parameter according to the categorization of the user correction.

14. The computer storage medium of claim 13, wherein weighting the impact of the user correction comprises weighting user corrections made after confirmation from a source more heavily than user corrections made without confirmation from a source.

15. The computer storage medium of claim 13, wherein weighting the impact of the user correction comprises weighting more recent user corrections more heavily than earlier user corrections.

16. The computer storage medium of claim 8, wherein changing the confidence parameter comprises changing the confidence parameter reflecting the likelihood that an corrected value correctly characterizes the instance attribute.

17. A system comprising : a client device comprising an input device, a display screen, and a digital data processing device operable to display, on the display screen, characterizations of instance attributes by values and to receive, over the input device, user input correcting characterizations of instance attributes; a correction tracker operable to interact with the client to track the user input correcting the characterizations of the instance attributes and to store descriptions of the user input in records of the user correction history; one or more data storage devices storing the records of the user correction history; and a search engine operable to interact with the one or more data storage devices to access the records of the user correction history and to change a confidence that a first value correctly characterizes a first instance attribute in response to identifying a record describing a user correction correctin *-&g the characterization of the first instance attribute.

18. The system of claim 17, wherein the display screen displays a structured presentation under the direction of the digital data processing device, the structured presentation associating instance attributes with values.

19. The system of claim 18, wherein the structured presentation comprises interactive elements selectable by a user to identify an instance attribute whose characterization by a value is to be corrected.

20. The system of claim 19, wherein the interactive elements comprise cells of the structured presentation.

21. The system of claim 18, wherein the structured presentation comprises a deck of cards.

22. The system of claim 17, wherein the display screen displays a candidate window under the direction of the digital data processing device, the candidate window presenting candidate corrected values for replacing an uncorrected value characterizing an instance attribute.

23. A method performed by one or more data processing apparatus, the method comprising: the data processing apparatus receiving a search query at a data processing apparatus, the search query specifying attributes shared by a group of related instances; the data processing apparatus identifying groups of instance identifiers in an unstructured collection of electronic documents with the data processing apparatus; the data processing apparatus determining relevance of the groups of instance identifiers to the search query with the data processing apparatus; the data processing apparatus scoring at least some of the instance identifiers in the groups of instance identifiers individually with the data processing apparatus; and the data processing apparatus ranking the at least some instance identifiers according to the scores with the data processing apparatus.

24. A system comprising: a client device; and one or more computers programmed to interact with the client device and the data storage device, the computers programmed to perform operations comprising: receiving a search query from the client device, the search query explicitly or implicitly specifying attributes of instances; searching an electronic document collection to identify identifiers of instances that may have the attributes specified by the search query; representing features of the search of the electronic document collection in a vertex-edge graph; scoring the instance identifiers that may have the attributes specified by the search query according to the features represented in the vertex-edge graph; and outputting, to the client device, instructions for visually presenting at least some of the instance identifiers.