Predicting Interestingness of Questions in Community Question Answering

Abstract

Exemplary methods, computer-readable media, and systems are presented for learning to recommend questions and other user-generated submissions to community sites based on user ratings. The size of available training data is enlarged by taking into consideration questions without user ratings, which in turn benefits the learned model. Question or other user-generated submissions are obtained by crawling Internet-accessible Web sites including community sites. Questions and other submissions, even when not tagged, voted or indicated as “popular” or “interesting” by users are quantitatively indentified as “interesting.”

Description

BACKGROUND

Prior to making purchases, consumers and others often conduct research, read reviews and search for best prices for products and services. Information about products and services can be found at a variety of types of Internet-accessible Web sites including community sites. Such information is abundant. Product developers, vendors, users and reviewers, among others, submit information to a variety of such sites. Some sites allow users to post opinions about products and services. Some sites also allow users to interact with each other by posting questions and receiving answers to their questions from other users.

Ordinary search services yield thousands and even millions of results for any given product or service. A search of a community site often yields far too many hits with little filtering. Results of a search of a community site are typically presented one at a time and in reverse chronological order merely based on the presence of search terms.

A search of typical question and answer community sites typically results in a listing of questions. For example, a search for a product such as a “Mokia L99” cellular telephone could yield hundreds of results. Only a few results would be viewed by a typical user from such a search. Each entry on a user interface to a search result could be made up of part or all of a question, all or part of an answer to the corresponding question and other miscellaneous information such as a user name of each user who submitted each respective question or answer. Other information presented would include when the question was presented and how many answers were received for a particular question. Each entry listed as a result of a search could be presented as a link so that a user could access a full set of information about a particular question or answer matching a search query. A user would have to follow each hyperlink to view the entire entry to attempt to find useful information.

Such searching of products and services is time-consuming and is often not productive because search queries yield either too much information, not enough information, or just too much random information. Such searching also typically fails to lead a user to the most useful entries on community and other sites because there is little or no automatic parsing or filtering of the information—just a dump of entries matching one or more of desired search terms. Users would have to click through page after page and link after link with the result of spending excessive amounts of time looking for the most useful information responsive to a relatively simple inquiry. To further compound the problem, product and service information is spread over a myriad of sites and is presented in many different formats.

Some community sites offer a means for voting or recommending certain content. In particular, on certain community sites, users can vote for or recommend certain questions and corresponding answers that may contain information of interest to users. However due to such large volumes of submitted questions, many questions (and corresponding answers) do not receive enough hits and thus any voting associated with such questions does not adequately reflect their likely interest to the users of the community site. Voting or recommending can also be skewed by when a particular question or answer is submitted. For example, timing may be important such as what time of day or night the question is submitted or what day of the week the question is submitted. Further, some sites do not offer the ability for users to vote or recommend questions and answers. These and other conditions of voting or recommending of content on community sites present challenges for users to find content which is most valuable or useful.

SUMMARY

This Summary is provided to introduce a selection of concepts in a simplified form that are further described below in the Detailed Description. This Summary is not intended to identify key or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

Information from question-answer community and other Internet-accessible Web sites are crawled and information such as questions and answers are extracted from these sites.

A plurality of questions from the information is identified such that each of a subset of the questions has an indication of preference such as a vote or indication of “interestingness.” For each user, instance pairs are identified for a majority of users whose input reflects question interestingness for all users. Training data from a minority of users is screened out to avoid the use of input that does not reflect question interestingness for all users.

Then, a user weight for each user is determined. The closer a user's indication(s) of “interestingness” matches that of the majority, the more weight is given to that particular user's questions for training purposes. A statistical model is trained by emphasizing training data from instance pairs from the majority of users whose input reflects question interestingness for all users. The training uses the user weights. The questions are then sorted by a value of reflective of “interestingness.”

BRIEF DESCRIPTION OF THE DRAWINGS

The Detailed Description is set forth and the teachings are described with reference to the accompanying figures. In the figures, the left-most digit(s) of a reference number identifies the figure in which the reference number first appears. The use of the same reference numbers in different figures indicates similar or identical items.

FIG. 1 is an exemplary user interface showing an exemplary use of predicting question “interestingness.”

FIG. 2 shows an overview of the topology of an exemplary system used to predict “interestingness” of questions from community question and answer sites.

FIG. 3 is a diagram showing parts of a product or service information indexing and search.

FIG. 4 is flow chart showing a process for a product or service information indexing and search.

FIG. 5 is a bar chart showing a distribution of questions which were voted as “interesting” out of a large sample of questions obtained from a community answer website.

FIG. 6 is a graph showing an accumulated count of users grouped by cosine similarity of users' preferences compared to a learned preference as described herein.

DETAILED DESCRIPTION

This disclosure is directed to predicting or estimating “interestingness” of content (such as questions) on community Internet sites. Herein, while reference may be made to a product, a service, information, data or something else may just as easily be the subject of the features described herein. For the sake of brevity and clarity, not limitation, reference is made to a question about a product.

Further, community sites as understood herein include community-based question submission and question answering sites, and various forum sites, among others. Community sites as used herein include at least community question and answer (community QnA) sites, blogs, forums, email threads of conversation and the like. In short, the techniques described herein can be applied to any set of information connected to members of a group or community Thus, the techniques can be generally applied to information chunks selected from community sites or other sources.

One problem associated with community sites is that what is considered interesting or useful to one user is not necessarily interesting to another user. Yet another problem is that newly submitted information may not get enough exposure for user interaction and thus information that would have been considered very interesting by many users is not identified at the time a particular community user seeks information.

As described herein, in a particular illustrative implementation, instead of a conventional search result, a user receives an enhanced and aggregated search result upon entering a query. The result 100 of such illustrative query is shown in FIG. 1 using “Mokia L99,” an exemplary product. Such search result includes the use of a method of predicting “interestingness” or popularity of a question or other user-generated content.

Exemplary User Interface and Search Results

With reference to FIG. 1, a product summary 102 is provided to a user as part of the result 100. Such a summary 102 includes by way of example, without limitation, a title 140, a picture 142, a range of prices 152 at which the product is being offered for sale, a link to a list of sites containing prices 154, a composite average of ratings made by users 144, a link to a list of Web pages of user reviews 148, a composite average of ratings made by experts or commercial entities 146, a link to a list of Web pages of expert or commercial reviews 150, and an exemplary description of the product 156.

In one implementation, a product feature summary 104 is also provided to a user. This product feature summary 104 includes, by way of example, an overall summary of questions from community sites, some of which are flagged or tagged by users as “interesting” 106 and questions grouped according to product feature 108. For example, in FIG. 1, about five percent of 1442 questions have been marked as “interesting.” In one implementation, questions flagged as “interesting” also include those questions which have programmatically been predicted as likely to be flagged as interesting according to a method described in more detail below. If a user desires more information about “all questions,” the “all questions” is presented as a link leading to a Web page which includes a listing of all questions, preferably where the questions tagged as “interesting” by users are presented first, grouped together, or otherwise set off from the others.

Product features 108 may be generated by users, automatically generated by a computer process, or identified by some other method or means. These product features 108 may be presented as links to respective product feature Web pages which each contain listing of questions addressed to a single feature or group of related features. For example, in FIG. 1, a user is presented with a link to “sound” as a feature of the Mokia L99 cellular telephone. If a user selects the link to sound, questions addressing sound of the Mokia L99 would be listed on a separate Web page where one of the seven questions would be identified as “interesting” (about 14 percent of the seven questions as shown in FIG. 1).

Product feature Web pages preferably list questions marked as “interesting” ahead of, or differently from, other questions addressing the same product feature. A user would then be directed in a hierarchal fashion to specific product features and then to questions or answers or both questions and answers that have been marked by community site users as “interesting” or programmatically identified as likely to be “interesting.” Another designation other than “interesting” may be used and correlated or combined with those items flagged as “interesting.”

In the lower left portion of FIG. 1, a user is also presented with a tag cloud 110 or listing of keywords or “hot topics” found in the 1442 indexed questions. The size or presentation of each keyword or phrase is in proportion to its relative frequency in the set of indexed questions. For example, the word “provider” 112 is smaller than the word “Microsoft” 114 because the word “Microsoft” 114 appears more frequently then provider 112 as to those results which pertain to “Mokia L99.” The number and sizes of words and phrases in the tag cloud vary depending on the set of indexed questions.

With reference to FIG. 1, a sample of questions from the set of indexed questions is presented in a questions listing section 160. Questions may be presented in a variety of ways in this section including most recent 116, comparative 118, interesting 120 and most popular 122. In one implementation, a user is presented with a link for accessing information that is sorted in one of these ways. A set of sample comparative questions 118 is shown in FIG. 1; the word “interesting” 120 is bolded to indicate this type of question. Each question in the comparative listing of questions addresses two or more products of the same type as that identified by the query or search terms. For example, the first sample question addresses “Mokia L99” 132 and “Samsun Q44” cellular telephone telephones. Questions, answers and other types of information may be identified and to a user interface or other destination in response to selecting a comparative 118 option.

In one implementation, a summary of information about each question is presented in the questions listing section 160. For example, such a question summary includes a user rating 130 for a particular question, a bolding of a search term in the question 132 or in an answer 134 to a question. A user rating 130 may take the form of a number of stars (as shown in FIG. 1) along a scale such as from 1 to 5, or as a vote of “interesting” or some other designator such as a thumbs up.

The site from which the question appears 136 is also shown. A short summary of each answer and links or other navigation to see other answers 138 to a particular question are also provided. In FIG. 1, three comparative questions are shown. However, any number of questions may be shown on a single page of a user interface.

In summary as to the user interface 100, a user is simultaneously presented with a variety of features with which to check product details, compare prices provided by a plurality of sites, and gain access to opinions from many other users from one or more sites having questions or from users who have provided answers to questions about a particular product.

Illustrative Network Topology

FIG. 2 shows an exemplary network topology 200 of one implementation of an improved product service including the use of a method of predicting question interestingness as described herein. A single server 210 is shown, but many servers may be used. The server 210 houses memory 212 on which operates a crawler and extractor application 214 and an indexer application 216. The crawler and extractor application 214 interoperates with the indexer application 216. The crawler and extractor application 214 and indexer application 216 acquire, read and store data in one or more databases. FIG. 2 shows a single database 220 for convenience. This database receives data from at least a plurality of community sites and community QnA sites 202, as obtained by the crawler and extractor application 214, and from the indexer application 216. A processing unit 218 is shown and represents one or more processors as part of the one or more servers 210. The server 210 connects to community sites 202 and to user machines 204 through a network 206 such as the Internet.

An exemplary implementation of a process to generate the user interface shown in FIG. 1 is shown in FIG. 3 and FIG. 4.

With reference to FIG. 3, one implementation of the process involves crawling and extracting information from community sites 202 and other sites including forum sites 302. Crawling and extracting are done by a crawler and extractor appliance, application or process 214 operating on one or more servers 210. For convenience, a single server is shown in FIG. 3. Crawling and extracting also takes information from forum site wrappers 304 and posts or threads of users' discussions 306 of forum sites 302. The crawling and extracting further takes information from community site wrappers 308 of community sites 202. Questions and answers 326 are taken from the extracted information.

Using a taxonomy of product names 310, questions (and answers) are grouped by product names 328. Metadata is prepared for each question (and answer) 330 from the extracted information. A metadata extractor 350 prepares such metadata through several functions. The metadata extractor 350 identifies comparative questions 312, predicts question “interestingness” 314 (as explained more fully below), predicts question popularity 316, extracts topics within questions 318, and labels questions by product feature 320.

Metadata is then indexed by question ID 322 and answers are indexed by question ID 324. Using the metadata, questions are grouped by product names 332 and questions are ranked by lexical relevance and using metadata 334.

Predicting question interestingness 314 includes flagging a question or other information as “interesting” when it has not been tagged as “interesting” or with some other user-generated label. Indexing also comprises labeling questions by feature 308 such as by product feature. While question or questions are referenced, the process described herein equally applies to answers to questions and to all varieties of information.

When a search for information about a product or service is desired, a query is submitted 338 through a user device 204. For example, a user submits a query for a “Mokia L99” in search of information about a particular cellular telephone. In response, the server 210 ranks questions, answers and other information by lexical relevance and by using metadata 334 and then generates search results 336 which are then delivered to the user device 204 or other destination. In one implementation, questions are sorted by a relevance score. A user can then interact 340 with the search results which may involve a re-ranking of questions 334.

FIG. 4 shows one implementation of a method to provide questions, answers and other product or service information sorted by relevance or other means. Community and other sites are crawled and certain information is extracted therefrom 402. If any questions (or answers or other information) have not been tagged as interesting, a prediction 404 is done to identify which of these questions would likely have been tagged, voted or labeled as preferred, “interesting” or “popular.” Prediction may be done by determining the number of answers provided in response to a question, similarity to other questions or answers that were tagged as interesting, or by other method such as one described more fully below.

With reference to FIG. 4, questions, answers and other information are indexed, labeled or both indexed and labeled by feature 406. Topics about products or services are extracted 408 from the information extracted from the community and other sites. Comparative questions, answers and other information are identified 410. Questions, answers and other information are indexed 412. In one implementation, these actions or steps are performed prior to receiving a query 414. Indexing may use a relevance value to rank query results.

Next, a query may be entered by a user or may be received programmatically from any source. Based on the query, questions and other information are ranked by lexical relevance or interestingness, or relevance and interestingness 416. Then, questions, answers and other information are provided in a sorted or parsed format. In a preferred implementation, such information is provided sorted by relevance or a combined score 418.

In one implementation, through a user interface, after indexing and ranking are completed, a user is able to browse relevant questions, answers and other information addressing a particular product or service sorted by feature. Questions can also be browsed by topic since questions that address the same or similar topic are grouped together so as to provide a user-friendly and user-accessible interface. Further, search results from question and answer community sites and other types of sites are sorted and grouped by similar comparative questions. Product search is enhanced by providing an improved search of questions, answers and other information from community sites. The new search can save effort by users in browsing or searching community sites when users conduct a survey on certain products.

An improved search of questions and answers helps users not only to make decisions when users want to purchase a product or service but also to get instructions after users have already purchased a product or service. Further implementation details for one embodiment are now presented.

Product or Service Features

Each type of product or service is associated with a respective set of features. For example, for digital cameras, product features are zoom, picture quality, size, and price. Other features can be added at any time (or dynamically) and the indexing and other processing can then be re-performed so as to incorporate any newly added feature. Features can be generated by one or more users, user community, or programmatically through one or more computer algorithms and processing.

In one implementation, a feature indexing algorithm is implemented as part of a server operating crawling and indexing of community sites. The feature indexing algorithm uses an algorithm similar to an opinion indexing algorithm. This feature indexing algorithm is used to identify the features for each product or type of product from gathered data and metadata. Features are identified by using probability and identifying nouns and other parts of speech used in questions and answers submitted to community sites and, through probability, identifying the relationships between these parts of speech and the corresponding products or services.

In particular, when provided with sentences from community sites, the feature algorithm or system identifies possible sequences of parts of speech of the sentence that are commonly used to express a feature and the probability that the sequence is the correct sequence for the sentence. For each sequence, the feature identifying system then retrieves a probability derived from training data that the sequence contains a word that expresses a feature. The feature identification system then retrieves a probability from the training data that the feature words of the sentence are used to express a feature. The feature identification system then combines the probabilities to generate an overall probability that a particular sentence with that sequence expresses a feature. Potential features are then identified. Potential features across a plurality of products of a given category of product are then gathered and compared. A set of features is then identified and used. A restricted set if features may be selected by ranking based on a probability score.

In another embodiment, product or service features are determined using two kids of evidence within the gathered data and metadata. One is “surface string” evidence, and the other is “contextual evidence.” An edit distance can be used t compare the similarity between the surface strings of two product feature mentions in the text of questions and answers. Contextual similarity is used to reflect the semantic similarity between two identifiable product features. Surface string evidence or contextual evidence are used to determine the equivalence of a product or service feature in different forms (e.g. battery life and power).

When using contextual similarity, all questions and answers are split into sentences. For each mention of a product feature, the feature “mention,” or term which may be a product feature, is taken as a query and search for all relevant sentences. Then, a vector is constructed for the product feature mention by taking each unique term in the relevant sentences as a dimension of the vector. The cosine similarity between two vectors of product feature mentions can then be present to measure the contextual similarity between the two feature mentions.

Product or Service Topics

Usually, a topic around which users ask questions cannot be predicted or fall within a fixed set of topics for a product or service. While some user questions may be about features, most questions are not. For example, a user may submit “How do I add songs to my Zoon music player?” Thus, the process described herein provides users with a mechanism to browse questions around topics that are automatically extracted from a corpus of questions. To extract the topics automatically, questions are grouped around types of question, and then sequential pattern mining and part-of-speech (POS) tags-based filtering are applied to each group of questions.

POS tagging is also called grammatical tagging or word-category disambiguation. POS tagging is the process of marking up or finding words in a text as corresponding to a particular part of speech. The process is based on both its definition as well as its context—i.e., relationship with adjacent and related words in a phrase, sentence, or paragraph. A simplified form of POS tagging is commonly taught to school-age children, in the identification of words as nouns, verbs, adjectives and adverbs. Once performed by hand, POS tagging is now done in the context of computational linguistics, using algorithms which associate discrete terms, as well as hidden parts of speech, in accordance with a set of descriptive tags. Questions, answers and other information extracted from sites are treated in this manner.

Comparative Questions

Sometimes, users not only care about the product or service that they want to purchase, but also want to compare two or more products or services. As shown in FIG. 1, comparative questions are found and presented on a user interface. Further, such batch of questions can be filtered or sorted according to “interestingness” making it easier for a user to find desired or usable information.

User Labeling

Some sites such as community sites allow users to label, tag, star or vote certain questions, answers or other information as “interesting.” Product search and product comparisons are merely examples of where a prediction of “interestingness” can be used.

In one particular implementation, “interestingness” is defined as a quadruple (u, x, v, t) such that a user u (is an element of all users U) provides a vote v (interesting or not) for a question x which is posted at a specific time t (within R+). It is noted that v is contained within the set {1, 0} where 1 means that a user provides an “interesting” vote and 0 denotes no vote given. The set of questions with a positive “interestingness” label can be expressed as Q+={x: (u, x, v, t), v=1}.

In the implementation described herein, such a designation of “interesting” is a user-dependent property such that different users may have different preferences as to whether a question is interesting. It is assumed that the identity of users is not available. It is also assumed for purposes of the described implementation that there is a commonality of “interestingness” over all users and this is referred to as “question interestingness,” an indication of whether a question is worthy for recommendation. This term “interestingness” is formally defined in this implementation as the likelihood that a question is considered “interesting” by most users. “Interestingness” is characterized by a measure called question popularity. The higher the popularity of a question, the more likely the question is recommended by most users. For any given question that is labeled as “interesting” by many users, it is probable that it is “interesting” for any individual user in U. A description follows of one implementation to estimate question popularity and then use question popularity to recommend questions.

Data Construction

It is somewhat difficult to make a judgment as to how likely a question is to be recommended. It is easier (computationally) to determine which is more likely to be recommended given a pair of questions. A preference relationship is defined between any two questions such that x⁽¹⁾x⁽²⁾ if and only if the popularity of question x⁽¹⁾ is greater than that of x⁽²⁾. The preference relationship is defined on the basis of user ratings. Two definitions of are provided.

Definition 1: a preference order

x⁽¹⁾¹x⁽²⁾  (1)

exists if and only if {u: (u,x⁽¹⁾, v,t)εQ+}|−{u: (u,x⁽²⁾, v,t)εQ+}|≧Δv, where ΔvεN+ and where the operation |{}| represents the size of a set. The more votes a question receives, the more popular or likely to be recommended it is. Thus, ¹ is defined on the basis of the number of votes that a question receives. The parameter Δv is introduced to control the margin of separation in terms of votes between x⁽¹⁾ and x⁽²⁾.

The preference relationships derived according to ¹ can be reliable when Δv is set to a relatively large value (e.g. 5). Definition 1 is used to build a test set.

One disadvantage with ¹ is that it can only be used to judge the preference order between questions already having votes by users. In an exemplary collection of data in an experimental use of Definition 1, not all of the questions were voted upon. For example, in a category of “travel,” only 13% of questions were voted or identified as “interesting.” Thus, the use of such sparse data makes the training data less reliable and less desirable when used to learn a “question recommendation” model.

One method for addressing data sparsity is just to include all questions without user ratings or votes into Definition 1 directly, which can be done simply by replacing Q+ with Q. The method implicitly assumes that all questions without user ratings are “not recommended.” However, the questions without votes could be worth being recommended as well. As users are not obligated to rate questions in community sites or services, users may not rate a question even if the users feel the question is interesting or recommendable. Thus, to better use questions without votes, definition 2 is introduced.

Definition 2: a preference order

x⁽¹⁾_u²x⁽²⁾  (2)

exists if and only if

• • ∃(u, x⁽¹⁾, v₁, t₁) and (u, x⁽²⁾, v₂, t₂)
  • such that v₁>v₂, |t₁−t₂|<Δt, and ΔtεR+.

Questions at community sites are usually sorted by posting time when they are presented to users as a list of ranked items. That is, the latest posted question is ranked highest, and then older questions are presented in reverse chronological order. The result is that questions with close posting times tend to be viewed by a particular user within a single page which means that they have about the same chance of being seen by user and about the same chance of being labeled as “interesting” by the user. With the assumption that a user u sees x⁽¹⁾ and x⁽²⁾ at about the same time within a single page, it can lead to the result that x⁽¹⁾ can be tagged as “interesting” and x⁽²⁾ left as not “interesting” by a user. Therefore, it is relatively safe to accept that for any given user, x⁽¹⁾ is more “interesting” or popular than x⁽²⁾.

By using definition 2, more caution is used in identifying whether questions without user votes are “not recommended” or “not interesting.” Particularly, only questions which do not have users' votes and share similar user browsing contexts with questions having user votes are considered “not recommended.”

According to definition 2 (Equation 2), it is possible to build a set of ordered (question) instance pairs for any given user as follows:

$\begin{array}{cc}{S}_u={\left\{x_i^{\left(1\right)},x_i^{\left(2\right)},z_i\right\}}_{i=1}^{l_u}& \left(3\right)\end{array}$

where z_i equals 1 for x⁽¹⁾_u² x⁽²⁾ and −1 otherwise, where i runs from 1 to l_u where l_u is the number of instance pairs given by a user u. The number of sets is the size of all users U (denoted |U|). S is the union ∪S_u.

One assumption is that a majority of users share a common preference about “question interestingness.”

Problem Statement

It is assumed that question x comes from an input space X which is a subset of Rⁿ, where n denotes a number of features of a product (e.g. x⊂Rⁿ). A set of ranking functions f exists where each f is an element of all functions F (e.g. fεF). Each function f can determine the preference relations between instances as follows:

x _i ² x _j if and only if f(x_i)>f(x_j)  (4)

The best function f* is selected from F that respects the given set of ranked instances S. It is assumed that f is a linear function such that

f _w(x)=w,x  (5)

where w denotes a vector of weights and •,• denotes an inner product. Combining Equation 4 and Equation 3 yields

x _i ² x _j if and only ifw,x_i−x_j>0  (6)

Note that the relation x_iu²x_j between instance pairs x_i and x_j is expressed by a new vector x_i−x_j. A new vector is created from any instance pair and the relationship between the elements of the instance pair. From the given training data set S, a new training data set S′ is created that contains 1 (lower-case letter “L”) (=Σ_ul_u) labeled vectors.

$\begin{array}{cc}{S}^{\prime }={\left\{x_i^{\left(1\right)}-x_i^{\left(2\right)},z_i\right\}}_{i=1}^l>0& \left(7\right)\end{array}$

Similarly, S′_u is created for each user u.

S′ is taken as classification data and a classification model is constructed that assigns either a positive label z=+1 or a negative label z=−1 to any vector x_i⁽¹⁾−x_i⁽²⁾.

A weight vector w* is learned by the classification model. The weight vector w* is used to form a scoring function f_w* for evaluating “interestingness” or popularity of a question x. A popularity score determines the likelihood that the question is recommended by many users.

f _w*(x)=w,x  (8)

In one implementation, the Perceptron algorithm is adapted for the above presented learning problem by guiding the learned function by a majority of users. The Perceptron algorithm is a learning algorithm for linear classifiers. A particular variant of the Perceptron algorithm is used and is called the Perceptron algorithm with margins (PAM). The adaptation as disclosed herein is referred to as Perceptron algorithm for preference learning (PAPL). A pseudocode listing for PAPL is as follows.

Listing 1
Input: training examples {xi(1) − xi(2), zi}i=1m,
       training rate η is an element in R+,
       margin parameter τ is an element in R+
1      w0 = 0; t = 0;
2      repeat
3      for i = 1 to m do
4      if zi <wt,xi(1) − xi(2)> ≦ τ then
5      Wt+1= wt + ηzi(xi(1) − xi(2));
6      bt+1 = bt + ηzi maxj ∥xi(1) − xi(2)∥2; // this step
commented out
7      t ←t + 1;
8      end if
9      end for
10     until no updates made within the for loop
11     return Wt;

In this implementation, PAPL makes at least two changes when compared to PAM. First, transformed instances (instead of raw instances) as given in Equation 8 are used as input. Second, an estimation of an intercept is no longer necessary (as in line 6 of Listing 1). The changes do not influence the convergence of the PAPL algorithm.

For each user u, Listing 1 can learn a model (denoted by weight vector w_u) on the basis of S′_u. However, none of the users can be used for predicting question “interestingness” or popularity because such indications are personal to a particular user, not to all users.

An alternative implementation is to use the model (denoted by w₀) learned on the basis of S′. The insufficiency of the model w₀ originates from an inability to avoid influences of a minority of users which diverges from the majority of users in terms of preferences about “interesting,” popularity or whether a question is recommended. This influence can be mitigated and w₀ can be enhanced or boosted as explained further below.

It is noted that different users might provide different preference labels for a same set of instance pairs. In one implementation, instance pairs from a majority of users are used and instance pairs from an identified minority of users are ignored as noise or weighed less important. In such implementation, this process is done automatically by identifying the majority from the minority.

One solution for mitigating the problem associated with the minority is to give a different weight to each instance of pairs where a bigger weight means the particular instance pair is more important. In this implementation, it is assumed that all instance pairs from a user u share the same weight α_u. The next step is to determine a weight for each user.

Every w obtained by PAPL (from Listing 1) is treated as a directional vector. Predicting a preference order between two questions x_i⁽¹⁾ and x_i⁽²⁾ is achieved by projecting x_i⁽¹⁾ and x_i⁽²⁾ onto the direction denoted by w and then sorting them on a line. Thus, the directional vector w_u denoting a user u agreeing with a majority should be close to the directional vector w₀ denoting the majority. Furthermore, the closer a user vector is to w₀, the more important the user data is.

As one implementation, cosine similarity is used to measure how close two directional vectors are to each other. A set of user weights {α_u} is found as follows:

$\begin{array}{cc}{\alpha }_u={〈w_o,w_u〉}_N=\frac{〈w_o,w_u〉}{w_o·w_u}& \left(9\right)\end{array}$

This implementation is termed majority-based perceptron algorithm (MBPA) and emphasizes its training on the instance pairs from a majority of users such as by using Equation 9. Listing 2 provides pseudo code for one implementation of this method.

Listing 2
Input: training examples {xi(1) − xi(2), zi}i=1m,
       users' weight vectors {wu}u=1k,
       training rate η is an element in R+,
       margin parameter τ is an element in R+,
       lower bound of correlation δ is an element in R+,
       initial weight vector w0 satisfying ∥ w0 ∥ = 1;
1      t = 0;
2      repeat
3      for i = 1 to m do
4      if <wt,wu(i)>N ≧ δ then
5      if zi <wt,xi(1) − xi(2)> ≦ τ<wt,wu(i)>N then
6      wt+1 = wt + ηzi(xi(1) − xi(2))/<wt,wu(i)>N;
7      t ←t + 1;
8      end if
9      end if
9      end for
10     until no updates made within the for loop
11     return Wt;

In MBPA, at iteration 0 (t=0), the condition at line 4 of Listing 2 prevents the minority from participating in the training process. Note that u(i) represents a user who is involved in generating the preference pair x_i⁽¹⁾ and x_i⁽²⁾ (such as found in definition 2). Further, at line 5 of Listing 2, training is emphasized over important instance pairs according to Equation 9. At iteration 1, w₀ is replaced with w₁ and the procedure is iterated where it is expected that w_t+1 represents the majority better than w_t.

As MBPA is an iterative algorithm, it is helpful to discuss its convergence. Theorem 1 guarantees the convergence of MBPA. First, Definition 3 is defined:

The margin of γ(w, S′) of a score function fw is minimal real-valued output on the training set S′. Specifically,

$\begin{array}{cc}\gamma (w,S’)={\min }_{x_i^1-x_i^2\in S^{\prime }}\frac{z_i〈w,x_i^1-x_i^2〉}{w}& \left(10\right)\end{array}$

Theorem 1:

$\mathrm{Let}S’={\left\{x_i^{\left(1\right)}-x_i^{\left(2\right)},z_i\right\}}_{i=1}^l$

be a set of training examples, and let r:=max∥x_i⁽¹⁾−x_i⁽²⁾∥. Suppose there exists w_optεRⁿ such that ∥w_opt∥=1 and

γ(w_opt,S′)≧Γ  (11)

Then, if w_opt, w₀>0 the number of updates made by the algorithm MBPA on S′ is bounded by

$2\left({\left(\frac{r}{\mathrm{\delta \Gamma }}\right)}^2+\frac{1}{{\eta }^2{\Gamma }^2}\right)+\frac{1}{{\eta }^2{\Gamma }^2}$

. Theorem 1 is an extension of Novikoff's theorem.

Learning Features

At community sites, a question is usually associated with three kinds of entities: (a) an asker who posts a question; (b) answers who provide answers to the question; and (c) answers to the question. Using the exemplary method described above, popularity is predicted for not only questions with answers, but also questions without answers. Thus, when modeling question popularity, two aspects are features are explored: features about questions and features about askers of the questions. Table 1 provides a list of features about questions (QU) and Table 2 provides a list of features about askers of questions (AS).

TABLE 1
Features about Questions (QU)
Feature Alias Description
Title Length  Number of words in title of the question.
Description   Number of words in description of the question.
Length
KL-Divergence Ratio between KL-divergence of a question to
Score         “interesting” questions and KL-divergence of
              the question to “not interesting” questions,
              both within a particular training set.
WH-Type       WH-word leading the title of a question; WH-words
              include why, what, where, when, who, whose and
              how. “None” is used to indicate that none of
              the WH-words occurs.
Posting Time  Time when a question is posted.

TABLE 2
Features about Askers (AS)
Feature Alias    Description
Total Questions  Total number of questions that an asker posted
Posted           in the past.
Total Stars      Total number of stars (or other indicator) that
Received         an asker received in the past.
Ratio of Starred Total questions with stars/total questions
Questions        posted.
Stars per        Average star that one question posted by the
Question         asker receives.
Total Answer     Total number of all the answers that an asker
                 obtained for his questions.
Answers per      Average number of answers that one question
Question         posted by an asker receives.

Features about questions (as shown in Table 1) come only from metadata of questions. In one implementation, a question comprises a title, description and posting time. In one implementation, a “bag-of-words” feature in reference to a question title and question description is not used. Features about askers are extracted from historical behaviors or askers. An asker's historical information can indicate if he is a skilled question asker or has a history of asking “interesting” questions.

Experimental Results

Using the above-described technology, experimental results were obtained. As to the dataset—297,919 questions were crawled from the Yahoo! Answers website under the top-level category of “travel.” The questions were posted within nine months between Aug. 1, 2007 and Apr. 30, 2008. Each question comprised two fields, title and description. Each question was also identified by asker of the question. Users of Yahoo! Answers rate or recommend questions by the label of “interesting.”

The following procedure was used to build training sets, a development set, and a test set.

1—Randomly separated all questions into two sets denoted Set-A and Set-B.2—With Set-A, built two training sets:

TR-1—Extracted from Set-A all questions voted as “interesting” by more than four users and then applied Definition 1 onto the extracted questions (Δv=5). The resulting preference pairs comprise TR-1.

TR-2—Applied Definition 2 to all questions in Set-A which resulted in a data set as Equation 3. The data set is denoted by TR-2.

3—Questions voted as “interesting” by more than four users were extracted. Definition 1 was then applied to the extracted questions (Δv=5). The result was then split into two subsets: a development set DEV and a test set TST.

Among the crawled questions, only about 13% of questions were voted by users as “interesting.” TR-1 was then considered sparse. FIG. 5 shows a distribution of questions 500 which were voted as “interesting.” The number of users which voted a question “interesting” (by, for example, giving a question a “star”) is represented along the horizontal axis 502, and the number of questions represented on the vertical axis 504. The horizontal axis 502 is labeled as number of stars or “# Stars.”

The number of preference pairs in the resulting data sets is as follows: TR-1 (188,638 pairs), TR-2 (1,090,694 pairs), DEV (49,766 pairs), and TST (49,148 pairs). TR-2 was larger than TR-1. TR-1 was obtained by setting Δv=5.

Three of the most “interesting” or “popular” questions in the data set which according to users' votes were: “Where in the world would you love to visit?” “Any suggestions for preventing seasickness?” and “Often do hotels have the comforters and pillows washed?”

Error rates of preference pairs were determined using a formula of the form ER=|mistakenly predicted preference pairs|/| all preference pairs in TST|. The use of different features of questions (and answers) as to “interestingness” or “popularity” were evaluated in two ways: (a) calculated information gain of each feature; and (b) evaluated the contribution of each feature in terms of predicting capability.

Table 3 shows the information gain (IG) for each of a list of learning features sorted or ranked by IG as calculated on the training set TR-2. Features about asker (AS) play a major role in predicting question “interestingness” or “popularity.” From the data, the history of an asker posting starred questions is the most important (AS: Ratio of Starred, AS: Stars per Question, and AS: Total Stars Received). In comparison, the WH-words features are weak features in terms of predicting “interestingness” or “popularity.”

Also shown in Table 3 shows the error rates for a series of models trained with PAPL and used on the data set DEV. The error rates do not decrease monotonically, meaning that the features are not independent from each other. The error rates also show that WH-word features do not help (much) in terms of “error rate of preference pairs.”

TABLE 3
IG	Feature	ER
0.127476	AS: Ratio of Starred	0.456
0.077378	AS: Stars per Question	0.313
0.058141	AS: Total Stars Received	0.322
0.012919	QU: KL-Divergence Score	0.319
0.007207	AS: Total Answers	0.304
0.005480	AS: Answers per Question	0.307
0.004009	AS: Total Questions Posted	0.348
0.00596	QU: WH-Type-Why	0.349
0.00418	QU: Title Length	0.351
0.000389	QU: WH-Type-Where	0.354
0.000355	QU: WH-Type-What	0.352
0.000319	QU: WH-Type-None	0.355
0.00218	QU: Description Length	0.352
0.00159	QU: WH-Type-How	0.347
8.63E−05	QU: WH-Type-Who	0.351
5.99E0−05	QU: WH-Type-When	0.350
8.41E−06	QU: WH-Type-Whose	0.352
6.80E−06	QU: Posting Time	0.345

Effectiveness

The following shows the evaluation according to two aspects: (a) how does the training set TR-2 help boost performance? and (2) how well does the method MBPA perform when compared with PAPL? In the experiments, all features of Table 1 were used. The parameters for PAPL and MBPA were tuned with the development set DEV.

	TABLE 4
	Algorithm	Training Set	ER
	PAPL	TR-1	0.362
	PAPL	TR-2	0.345
	MBPA	TR-2	0.283

Table 4 shows the results of the evaluation of effectiveness. With reference to Table 4, the training set TR-1 was obtained by setting Δv to 5 which is the same as that in TST. From Table 4, the algorithm MBPA trained with the training set TR-2 outperformed both the PAPL trained with TR-1 and the PAPL trained with TR-2 significantly (e.g. sign-test, p-value<0.01). This result shows that (1) taking into consideration questions without user ratings (or votes) incorporates more evidence than the training set given by Definition 1 (by noting that PAPL trained with TR-2 performs better than the PAPL trained with TR-1); and (2) the majority-based perceptron algorithm (MBPA) is effective in filtering noisy training data.

It is noted that the size of TR-2 is much larger than the size of TR-1. It could be argued that the size of TR-1 could be increased by setting Δv smaller (e.g. <5) to achieve a possibly better performance. Table 5 shows the results of setting Δv smaller than 5. The test set is TST and the model is PAPL. With reference to Table 5, the size of TR-1 becomes larger but the error rate of the corresponding PAPL increases as Δv gets smaller. When Δv=1, the size of TR-1 is even comparable with TR-2, but the model learned with TR-1 still performs significantly worse than that learned with TR-2. This further confirms the use of TR-2 built with the data construction method.

	TABLE 5
	Δv	Number Preference Pairs	ER
	6	132,868	0.396
	5	188,638	0.362
	4	273,316	0.371
	3	399,550	0.398
	2	583,463	0.398
	1	844,802	0.387

Prediction is easier when finer categories of questions are considered. Users tend to converge in their preference about “interesting” or “popularity” when topics of questions are constrained within a sub-category. For example, it is relatively easy for users to find the same preference when only topics of Asian as a travel area are considered. Table 6 shows the results of predicting “interestingness” or “popularity” for “Asia Pacific” and “Europe” sub-categories of travel questions.

	TABLE 6
	ER
	Sub-Category	PAPL (TR-1)	PAPL (TR-2)	MBPA
	Asia Pacific	0.286	0.280	0.239
	Europe	0.270	0.267	0.217

There were 46,541 questions under “Asia Pacific” and 23,080 questions under “Europe.” By comparing Table 6 with Table 4, it can be seen that question “popularity” is predicted more accurately constrained within categories of question topics.

Insights

There is a relationship between a learned preference and users' preferences (represented by

${\left\{w_u\right\}}_{u=1}^{U}$

). FIG. 6 is a plot 600 that shows the accumulated count of users 604 (vertical axis) grouped by cosine similarity 602 of users' preferences (horizontal axis) compared to the learned preference. The values shown in FIG. 6 were generated as follows: (1) let ŵ denote the weight vector learned by MBPA (606) and then calculate the cosine similarities (w₀, w_u)_N and (ŵ, w_u)_N for each user u (note that w₀ denotes the weight vector learned by PAPL (608)); (2) for each type of similarity, count the number of users whose similarities are less than −0.9, then −0.8, . . . , and 1.0.

FIG. 6 shows that most users have larger cosine similarities whenever compared to MBPA or PAPL and only a small portion of users have smaller cosine similarities suggesting that there exists certain commonality in users' preferences.

Users also tend to have larger cosine similarities compared to ŵ than compared to w₀. In one implementation, for ŵ, the algorithm only uses data from users whose similarities are larger than 0 (line 4 of Listing 2 ensures this). FIG. 6 also confirms or shows that the preference learned by MBPA (606) agrees with most users more than PAPL (608) does and implies that MBPA (606) can automatically lower the influence of the noisy data from the minority users.

The subject matter described above can be implemented in hardware, or software, or in both hardware and software. Although the subject matter has been described in language specific to structural features or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts are disclosed as exemplary forms of implementing the claimed subject matter. For example, the methodological acts need not be performed in the order or combinations described herein, and may be performed in any combination of one or more acts.

Claims

1. A system for sorting information extracted from one or more community sites, the system comprising: a memory and a processor;a crawler stored in the memory, and configured when executed on the processor, to crawl and extract information from one or more community sites; andan indexer stored in the memory and configured, when executed on the processor, to perform acts comprising: identifying a plurality of information chunks from the information, wherein each of a subset of the plurality of the information chunks have an indication of preference;identifying a user identifier for each of the plurality of information chunks;identifying instance pairs for each user from a majority of users whose input reflects interestingness for all users;screening out training data from a minority of users whose input does not reflect interestingness for all users;determining a user weight for each user;training a statistical model by emphasizing training data from instance pairs from the majority of users whose input reflects interestingness for all users according to the user weights giving proportionately more weight to data from users according to a degree to which each user agrees with the majority of users; andproviding the information chunks sorted by a value of interestingness.

2. The system of claim 1 wherein the indexer is further configured to predict interestingness using a factor common to each of the users.

3. The system of claim 2 wherein the factor common to each of the users is a feature selected from a list of features including: total information chunks posted by the user,total preference indications that the user received prior to posting a given information chunk,ratio of total information chunks with preference indications where the information chunks were submitted by a particular user relative to total information chunks posted by the particular user,an average of a number of preference indications that an information chunk posted by a particular user received,a total number of all responses that a particular user obtained for his information chunks, andan average number of responses that a information chunk posted by a particular user received.

4. The system of claim 1 wherein the indexer is further configured to predict interestingness using a factor based on a feature common to each information chunk of the plurality of information chunks.

5. The system of claim 4, wherein an information chunk is a question, and wherein the feature common to each information chunk of the plurality of information chunks is a feature selected from a list of features comprising: question title length in number of words, question description length in number of words, and word leading each question title.

6. The system of claim 1 wherein the indexer is further configured to: identify any information chunks which have been tagged with a user-generated label as tagged information chunks; andidentify any information chunks which have not been tagged with a user-generated label as untagged information chunks.

7. The system of claim 1 wherein the information extracted from one or more community sites is a user-generated submission.

8. A method of ranking information submitted from users to one or more community sites, the method comprising: crawling one or more community sites to extract information;identifying a plurality of portions of information submitted by users to the one or more community sites, wherein each of a subset of the plurality of the portions of information have an indication of preference;identifying a user identifier for each of the plurality of portions of information;identifying instance pairs of portions of information for each user from a majority of users whose input reflects interestingness for all users;screening out training data from a minority of users whose input does not reflect interestingness for all users;determining a user weight for each user;training a statistical model by emphasizing training data from instance pairs from the majority of users whose input reflects interestingness for all users according to the user weights giving proportionately more weight to portions of information from users according to a degree to which each user agrees with the majority of users; andproviding the portions of information sorted by a value of interestingness.

9. The method of claim 8 wherein the portions of information are either a question or an answer, and wherein the one or more community sites are sites that accept user generated questions and answers.

10. The method of claim 8 wherein the user weight is a user weight αu and is determined according to a formula of the form:

11. The method of claim 8 wherein the method further comprises: extracting a plurality of topics from the plurality of portions of information;identifying portions of information which are related to any of the plurality of topics;grouping into a topic group, one topic group for each topic, any portions of information which are identified as related to a particular topic of the plurality of topics; andproviding the portions of information related to any of the topics sorted by topic group.

12. The method of claim 8 further comprising predicting interestingness using a factor common to each of the users (question askers).

13. The method of claim 11, wherein the factor common to each of the users is a feature selected from a list of features including: total questions posted by the user,total preference indications that the user received prior to posting a given question,ratio of total questions with preference indications where the questions were submitted by a particular user relative to total questions posted by the particular user,an average of a number of preference indications that a question posted by a particular user received,a total number of all answers that a particular user obtained for his questions, andan average number of answers that a question posted by a particular user received.

14. The method of claim 8 wherein the method further comprises: identifying any questions or answers which compare two or more products or two or more services as respectively comparative questions and comparative answers; andrespectively grouping into comparative question groups or comparative answer groups the respective comparative questions and comparative answers which compare a same two or more products or two or more services.

15. One or more computer-readable storage media comprising computer-readable instructions that, when executed by a computing device, cause the computing device to perform a method, the method comprising: crawling one or more community sites to extract information;identifying a plurality of portions of information submitted by users to the one or more community sites, wherein each of a subset of the plurality of the portions of information have an indication of preference;identifying a user identifier for each of the plurality of portions of information;identifying instance pairs of portions of information for each user from a majority of users whose input reflects interestingness for all users;screening out training data from a minority of users whose input does not reflect interestingness for all users;determining a user weight for each user;training a statistical model by emphasizing training data from instance pairs from the majority of users whose input reflects interestingness for all users according to the user weights giving proportionately more weight to portions of information from users according to a degree to which each user agrees with the majority of users; andproviding the portions of information sorted by a value of interestingness.

16. The computer-readable storage media of claim 15 wherein the portions of information are either a question or an answer, and wherein the one or more community sites are sites that accept user generated questions and answers.

17. The computer-readable storage media of claim 15 wherein the user weight is a user weight αu and is determined according to a formula of the form:

18. The computer-readable storage media of claim 17 wherein training the statistical model additionally comprises identifying a margin γ for a scoring function that is a minimal real-valued output on a training set and that satisfies a formula of the form:

19. The computer-readable storage media of claim 16 wherein the method further comprises: identifying all portions of information which are a question;determining for each question a lexical relevance to a subject of a search query;identifying any questions which have been tagged with a user-generated label as tagged questions;identifying any questions which have not been tagged with a user-generated label as untagged questions;predicting, for each untagged question, whether the untagged question would likely have been tagged and identifying each such question as a likely tagged question;grouping likely tagged questions, if any, with tagged questions, if any, into a tagged question group;ranking each question by a relevance score, wherein the relevance score is a combination of lexical relevance and label; andproviding the questions of the tagged question group sorted by feature and then by ranking.

20. The computer-readable storage media of claim 16 wherein the method further comprises: determining for each portion of information a lexical relevance to a subject of a search query; andafter identifying the plurality of portions of information related to a particular product or service from each of the one or more community sites, ranking each portions of information by lexical relevance.