Customers often provide feedback, in the form of reviews, regarding offerings (products or services) of different enterprises. Reviews can be submitted online at third party sites, or alternatively, reviews can be received directly by an enterprise. There can be potentially a very large number of received reviews, which can make meaningful analysis of such reviews difficult and time-consuming.
The patent or application file contains at least one drawing executed in color. Copies of this patent or patent application publication with color drawing(s) will be provided by the Office upon request and payment of the necessary fee.
Some embodiments are described with respect to the following figures:
An enterprise (e.g. a company, educational organization, government agency, etc.) may collect feedback from customers (or more generally “users”) to better understand user sentiment regarding an offering of the enterprise. An offering can include a product or a service provided by the enterprise. A “sentiment” refers to an attitude, opinion, or judgment of a human with respect to the offering.
If there are a relatively large number of users, then there can be relatively large amounts of user feedback. An enterprise can provide an online website to collect feedback from users. Alternatively or additionally, the enterprise can also collect feedback through telephone calls or on paper survey forms. Furthermore, feedback can be collected at third party sites, such as travel review websites, product review websites, and so forth. Some third party websites provide professional reviews of offerings from enterprises, as well as provide mechanisms for users to submit their individual reviews.
The feedback data collected by an enterprise can include a relatively large number of attributes for which users can express sentiments. For example, attributes relating to a hotel include room, carpet, food, noise, luggage, housekeeping, and so forth. Users who have stayed at the hotel may offer opinions on any, some, or all of the foregoing attributes. There can potentially be hundreds if not thousands of attributes that can be found in feedback data. A challenge faced by an enterprise in attempting to understand user sentiments regarding an offering (or offerings) of the enterprise is in selecting an appropriate subset of the candidate attributes, where the selected subset of attributes includes the attributes that are most likely to be useful for sentiment analysis.
In accordance with some implementations, visualization analytic techniques or mechanisms are provided to allow for a systematic way of selecting subsets of candidate attributes in feedback data for understanding user sentiments with respect to at least one offering of an enterprise. Selecting a “subset” of attributes found in feedback data refers to selecting a portion less than all of the attributes that are present in the feedback data with respect to which users may offer feedback.
In some implementations, visualization analytics (including selection of a subset of attributes, sentiment analysis with respect to the selected attributes, and visualization of results of the sentiment analysis regarding the selected subset of attributes) can be performed in real-time. Performing visualization analytics in “real-time” refers to performing such visualization analytics as feedback data is continually received. Although reference is made to real-time visualization analytics, note that in alternative implementations, the visualization analytics can be performed off-line, where feedback data can be collected into a log or database and studied at a later time.
From among candidate attributes in the received data records, the process selects (at 104) a subset of the attributes that relate to user sentiment regarding the at least one offering. An attribute relates to user sentiment regarding the at least one offering if users have provided feedback regarding the attribute—the feedback can be used to ascertain user sentiment regarding the attribute. The selecting at 104 can be according to various selection criteria, including frequencies of occurrence of the candidate attributes, relative amounts of negative and positive feedback associated with the candidate attributes, and time density of feedback associated with the candidate attributes.
The selection of a subset of attributes based on various criteria allows the selection of attributes based on characteristics found in the received feedback data. In this manner, visualization analytics does not rely on a static list of attributes, but the set of attributes can vary depending on the characteristics of the received feedback data.
After selecting the subset of attributes from among the candidate attributes, the process presents (at 106) for display a visualization of at least one sentiment characteristic of the selected subset of attributes that relate to user sentiment. The sentiment characteristics of each attribute can be represented in any of a number of ways. Generally, a “sentiment characteristic” refers to an indication of a sentiment expressed with respect to an attribute—the sentiment characteristic can be a sentiment score assigned to the attribute, or the sentiment characteristic can be a percentage value (or other aggregate value) indicating a collective sentiment expressed with respect to the attribute.
In some examples, the displayed visualization can include graphs to depict percentages (e.g. percentage of positive reviews and percentage of negative reviews) representing user sentiment expressed with respect to the respective attributes. Alternatively or additionally, the displayed visualization can include visual indicators associated with the attributes in the subset to depict user sentiments expressed in feedback regarding the attributes. The visual indicators can be in the form of different colors assigned to cells in the visualization to represent user sentiment expressed with respect to each attribute in the subset (e.g. a red color can indicate a negative user sentiment, a green color can indicate a positive user sentiment, and a gray color can indicate a neutral user sentiment). In other examples, different colors can be used to express different user sentiments with respect to each attribute. In yet further examples, other types of indicators can be used to represent different user sentiment, such as different shadings or hatching.
As noted above, one of the selection criteria for selecting a subset of attributes from among candidate attributes is a criterion relating to frequency of occurrence (referred to as a “high-frequency criterion”). The frequency of occurrence of a candidate attribute refers to how frequently the candidate attribute occurs in the received data records. A candidate attribute is considered to occur frequently in received data records if an amount of occurrence of the candidate attribute in the received data records exceeds some predefined threshold. For example, the amount of occurrence of the candidate attribute can be expressed as a percentage of the data records in which the candidate attribute is present—thus, the candidate attribute can be considered to be a high-frequency candidate attribute if the percentage of data records in which the candidate attribute appears is greater than some predefined percentage threshold.
As listed above, another selection criterion considered in selecting the subset of attributes is relative amounts of negative and positive feedback associated with each candidate attribute (referred to as a “large difference criterion”). A given candidate attribute can have a first number of positive feedback and a second number of negative feedback. The number of positive or negative feedback can refer to the number of data records in which the attribute appears (such as within a particular time interval being studied) and for which users have expressed positive or negative feedback, respectively. The number of positive or negative feedback can also refer to the number of times that users have expressed positive or negative review, respectively, of the given candidate attribute in the received data records.
The first number is compared to the second number, and if the first number differs from the second number by some predefined threshold, then that candidate attribute would be assigned greater importance than another candidate attribute associated with numbers of positive and negative feedback that are relatively closer together. As an example, a first candidate attribute can appear in 1,000 data records. Within the 1,000 data records, the first candidate attribute is associated with positive feedback 100 times, but is associated with negative feedback 800 times. In the remaining 100 occurrences of the given attribute, the feedback received is neutral. In the foregoing, it can be seen that the difference between the number of positive feedback (100) and the number of negative feedback (800) is relatively large.
In contrast, a second candidate attribute also appeared in 1,000 data records, but for the second candidate attribute, 300 reviews are positive, 300 reviews are negative, and 400 reviews are neutral. In this case, the difference between the positive feedback and negative feedback is small (in fact, it is zero). In the foregoing example, the first candidate attribute is assigned greater importance for understanding user sentiment than the second candidate attribute, and thus the first candidate attribute would more likely be selected to be in the subset (selected at 104 in
In some implementations, a ratio can be taken between the first number of positive feedback and second number of negative feedback (for each candidate attribute), or vice versa. The ratios of respective candidate attributes can then be compared, with the candidate attributes having higher ratios assigned greater importance than candidate attributes having lower ratios. In alternative examples, instead of calculating a ratio between numbers of positive and negative feedback, differences can be taken, with such differences compared to determine which candidate attributes should be assigned greater importance than other candidate attributes.
As noted above, another selection criterion considered in selecting a subset of attributes from among candidate attributes is the time density of feedback associated with each candidate attribute (referred to as a “time density criterion”). “Time density” refers to the average time gap between successive data records that contain feedback for a given candidate attribute. For example, the average time gap between feedback for a first candidate attribute may be X, while the average time gap between feedback data records for a second candidate attribute is Y, where X<Y. In such an example, the first candidate attribute is considered more important that the second candidate attribute, since the average time gap between reviews for the first candidate attribute is smaller than the average time gap for the second candidate attribute. This is an indication that users are more interested in the first candidate attribute (and thus have expressed more feedback on the first candidate attribute). Alternatively, techniques can consider the change in time density—a sudden increase in time density of an attribute indicates that the attribute would be more important.
In addition to the three selection criteria listed above for selecting a subset of attributes from candidate attributes, other selection criteria can also be considered in further implementations. For example, an attribute associated with greater negativity is assigned greater importance than another attribute with less negativity. The “negativity” of an attribute refers to the amount of negative feedback received for the given attribute. For example, within 1,000 data records, a first candidate attribute may be associated with 500 negative reviews, whereas a second candidate attribute may be associated with 100 negative reviews. In this case, this selection criterion (referred to as the “negativity criterion”) would indicate that the first candidate attribute would be more important than the second candidate attribute. Alternatively, instead of a negativity criterion, a positive criterion can be more relevant in certain application—an attribute with a greater number of positive reviews would be more important.
Another selection criterion that can be considered for selecting a subset of attributes from among candidate attributes is an application-specific criterion. An “application-specific criterion” refers to a criterion indicating that, for a given application, certain attributes are more important than other attributes. For example, if the application is sentiment analysis for hotels, then the application-specific criterion can specify a certain group of attributes that should be considered for analysis, such as room, price, and location. The application-specific criterion allows for selection of candidate attributes that are known or understood to be important factors in user sentiment regarding a particular application.
The multiple selection criteria (high-frequency criterion, large difference criterion, time density criterion, negativity criterion, and application-specific criterion) can be combined to assign an importance score to each candidate attribute. Each of the selection criteria causes an individual score to be produced for the candidate attribute. Such individual scores can be combined (such as by use of a weighted sum) to produce an overall score indicating the relative importance of the candidate attribute in user sentiment understanding. A weighted sum approach involves assigning respective weights to the individual scores (the weights are multiplied to the individual scores, and the weighted individual scores are summed to produce the overall score). The weights assigned can be equal weights, or alternatively, the weights assigned can be different.
In some implementations, the selection criteria can also be dynamically discovered based on visualization analytics techniques. Based on visualizations of sentiment characteristics of various attributes, analysts or other users can determine which selection criteria would be more useful for selecting attributes for sentiment visualization. The analyst or user can select a subset of the selection criteria listed above, for example, or alternatively, the analyst or user can modify weights assigned to the selection criteria as noted above.
An attribute selection process is performed (at 210) to select a subset of candidate attributes, using selection criteria including a high-frequency criterion 212, a large difference criterion 214, a time density criterion 216, a negativity criterion 218, and an application-specific criterion 220. These five selection criteria are discussed above.
The output of the attribute selection process (210) is a subset of attributes, which can be subject to sentiment analysis (222) to produce various visualizations (e.g. 229 and 226 in
A pattern can be that there are a large number of negative reviews but a small number of positive reviews with respect to a particular attribute. A correlation can be that a large number of negative reviews of a first attribute coincides with a large number of negative reviews of a second attribute—thus, an analyst can determine that the root cause of negative feedback for the first attribute may be due to the second attribute. An exception can be that negative feedback for a particular attribute is an anomaly, since the feedback received for other attributes is vastly positive.
The sentiment analysis performed at 222 can be based on use of any of various different sentiment analysis techniques, which involve identifying words or phrases in the data records that relate to sentiment expressed by users with respect to each attribute. A sentiment score can be generated based on the identified words or phrases. The sentiment score provides an indication of whether the expressed sentiment is positive, negative, or neutral. The sentiment score can be a numeric score, or alternatively, the sentiment score can have one of several discrete values (e.g. Positive, Negative, Neutral).
In
The visualization 226 of
The color assigned to each cell can be according to the sentiment score assigned to an attribute in the sentiment analysis 222. The cells within each bar 232 are arranged in an order according to the sentiment scores, such that cells having like colors are grouped adjacent each other to form the portions of different colors contained in each bar 232.
The “room” attribute is associated with various information in parentheticals, which in the example includes “(8592, +3677, −1829)”. The information in parentheticals is in the following format, according to some examples: (total number of reviews, number of positive reviews, number of negative reviews). The “+” symbol and “−” symbol in the parenthetical information indicates that the number is associated with positive or negative reviews, respectively.
In the calendar-view visualization 302, the cells corresponding to different data records are arranged in columns 306 and rows 308. Each column 306 represents a respective first time interval (e.g. month). Each row represents a respective second time interval (e.g. day) having a time length different from the first time interval. Each cell (represented by a small rectangular box in
The difference between
A processor can include a microprocessor, microcontroller, processor module or subsystem, programmable integrated circuit, programmable gate array, or another control or computing device.
The storage media 706 can be implemented as one or more computer-readable or machine-readable storage media. The storage media include different forms of memory including semiconductor memory devices such as dynamic or static random access memories (DRAMs or SRAMs), erasable and programmable read-only memories (EPROMs), electrically erasable and programmable read-only memories (EEPROMs) and flash memories; magnetic disks such as fixed, floppy and removable disks; other magnetic media including tape; optical media such as compact disks (CDs) or digital video disks (DVDs); or other types of storage devices. Note that the instructions discussed above can be provided on one computer-readable or machine-readable storage medium, or alternatively, can be provided on multiple computer-readable or machine-readable storage media distributed in a large system having possibly plural nodes. Such computer-readable or machine-readable storage medium or media is (are) considered to be part of an article (or article of manufacture). An article or article of manufacture can refer to any manufactured single component or multiple components. The storage medium or media can be located either in the machine running the machine-readable instructions, or located at a remote site from which machine-readable instructions can be downloaded over a network for execution.
In the foregoing description, numerous details are set forth to provide an understanding of the subject disclosed herein. However, implementations may be practiced without some or all of these details. Other implementations may include modifications and variations from the details discussed above. It is intended that the appended claims cover such modifications and variations.
Number | Name | Date | Kind |
---|---|---|---|
20080250326 | Mather | Oct 2008 | A1 |
20090048823 | Liu | Feb 2009 | A1 |
20090193328 | Reis | Jul 2009 | A1 |
20090265307 | Reisman | Oct 2009 | A1 |
20100023311 | Subrahmanian | Jan 2010 | A1 |
20100121707 | Goeldi | May 2010 | A1 |
20100306192 | Kapur | Dec 2010 | A1 |
Number | Date | Country |
---|---|---|
2007018285 | Jan 2007 | JP |
2009087636 | Jul 2009 | WO |
Entry |
---|
Gawalt et al., ACM, Discovering Word Associations in News Media via Feature Selection and Sparse Classification, Mar. 2010 (pp. 211-220). |
Wanner et al., Large-scale Comparative Sentiment Analysis of News Articles, 2009 (2 pages). |
Hao, “Visual Opinion Analysis of Customer Feedback Data,” IEEE Symposium on Visual Analytics Science and Technology. Oct. 12-13, 2009. http://ieeexplore.ieee.org/xpls/abs—all.jsp?arnumber=5333919&tag=1. |
Miao et al., 2008 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology, “A Unified Framework for Opinion Retrieval”, 2008 (pp. 739-742). |
Duan et al., Computer Science, Advances in Knowledge Discovery and Data Mining, SpringerLink—Abstract, CCRM: An Effective Algorithm for Mining Commodity Information from Threaded Chinese Customer Reviews, 2007 (1 page). |
Shi et al., Research Article—Science Alert—An Attribute-based Sentiment Analysis System, Aug. 2010 (9 pages). |
Zhang et al., IEEE Computer Society 2009, Improving Movie Gross Prediction Through News Analysis, 2009 (pp. 301-304). |
Balahur et al., IEEE Computer Society 2009, Opinion Mining on Newspaper Quotations, 2009 (4 pages). |
Abbasi et al., ACM Transactions on Information Systems, vol. 26, No. 3, Article 12, Sentiment Analysis in Multiple Languages: Feature Selection for Opinion Classification in Web Forums, Jun. 2008 (34 pages). |
Zhang et al., ACM Digital Library—Abstract, Sentiment Analysis of Chinese Documents: From Sentence to Document Level, Dec. 2009 (1 page). |
Crimson Hexagon Products (www.crimsonhexagon.com), Major Enhancement to Crimson Hexagon Platform Delivers Groundbreaking New Social Media Monitoring and Analysis Capabilities, Oct. 2010 (2 pages). |
SYSOMOS (www.sysomos.com) Best Global Brands Online, Oct. 2009 (4 pages). |
Goorha et al., ACM Digital Library—Abstract, Discovery of Significant Emerging Trends, Jul. 2010 (1 page). |
Hao et al., Chinese Receiving Office, International Appl. No. PCT/CN2011/000957entitled “Sentiment Trend Visualization Relating to an Event Occurring in a Particular Geographic Region” filed Jun. 8, 2011 (31 pages). |
Castellanos et al., U.S. Receiving Office, International Appl. No. PCT/US11/39566 entitled “Determining and Visualizing Social Media Expressed Sentiment” filed Jun. 8, 2011 (25 pages). |
Yingcai Wu, OpinionSeer: Interactive Visualization of Hotel Customer Feedback, IEEE, Oct. 2010 (10 pages). |
Hao et al., U.S. Appl. No. 13/209,739 entitled “Visualizing Sentiment Results with Visual Indicators Representing User Sentiment and Level of Uncertainty”, filed Aug. 15, 2011 (34 pages). |
Number | Date | Country | |
---|---|---|---|
20120317049 A1 | Dec 2012 | US |