Opinion Summaries in Google Maps Show Sentiment
You buy a new phone, and it doesn’t work as advertised, and customer service is even worse. Many people in your shoes would go online and write a negative review somewhere.
You go on a vacation and stay at an inexpensive and charming bed and breakfast. You have a wonderful time, in no small part to the thoughtfulness and suggestions of your hosts, and their incredible hospitality. Chances are, you write a glowing review about the experience on the Web.
The number of reviews and review sites on the Web has been growing over the past few years. Google’s recent “review” search option is one attempt to help people find both positive and negative reviews.
Google also presents reviews in Google Maps results. If you search for businesses and organizations in Google Maps, you’ll see under each listing a link to “write a review” for each business listed. If you click upon the “more info” link for a listed business, you’ll see a “review” tab in the box that appears in the middle of the map for that business. The results that show reviews are summaries, which often contain some level of sentiment about the businesses listed.
How does Google come up with those opinion summaries from reviews?
Sentiment and Reviews
In providing reviews, Google had to try to find a way to bring meaningful and helpful reviews to their search results. This means going past ratings that might be limited to a certain number of stars out of five, or that provide ratings on aspects of a product or service that might be immaterial to someone looking for more information.
For example, someone looking for a digital music player might be more concerned about the quality of sound and the battery life of the player than its weight or the number of colors it is available in.
A trio of patent applications from Google provides some insights into their approach for finding opinions from people who write reviews online about music, books, films, restaurants, hotels, electronics, and many other products and services.
I wrote about one of the patent filings in a recent post, titled GoogleΓ’β¬β’s New Review Search Option and Sentiment Analysis. It explored how the language that expresses different sentiments in reviews of different types. For example, “very small” might be a positive phrase when it comes to many electronic components, but the same phrase might be seen as negative when it comes to the size of hotel rooms.
Google has followed up with two more patent applications on reviews and the sentiments contained within them. One of them discusses how Google might summarize sentiments in reviews. The other delves into how Google might create snippets for reviews that appear search results. Both the summaries and the snippets may be displayed to someone who may be interested in seeing reviews and may click through a search result to read more.
Today’s post is about the creation and use of sentiment summaries. In a future post, I’ll write about the creation of sentiment snippets.
Summarizing Sentiment in Reviews
This patent application attempts to find different aspects of products or services that are being reviewed, and find phrases in those reviews that express some kind of opinion or sentiment about those aspects, or features.
Aspect-Based Sentiment Summarization
Invented by George Reis, Sasha Blair-Goldensohn, Ryan T. McDonald
US Patent Application 20090193328
Published July 30, 2009
Filed: March 19, 2008
Abstract
Reviews express sentiment about one or more entities. Phrases in the reviews that express sentiment about a particular aspect are identified. Reviewable aspects of the entity are also identified.
The reviewable aspects include static aspects that are specific to particular types of entities and dynamic aspects that are extracted from the reviews of a specific entity instance. The sentiment phrases are associated with the reviewable aspects to which the phrases pertain.
The sentiment expressed by the phrases associated with each aspect is summarized, thereby producing a summary of sentiment associated with each reviewable aspect of the entity. The summarized sentiment and associated phrases can be stored and displayed to a user as a summary description of the entity.
We’re told in the patent filing that the process by which reviews are summarized and displayed requires the use of a sentiment summarizing engine, which consists of three parts:
A sentiment summarizer – provides summaries of sentiment about aspects of reviewable objects or services.
Aspects are properties that can be evaluated by someone. A restaurant might be provided with summaries of sentiment about the food served there, and about the service itself. The summary could include a rating, such as three out of five stars or a letter grade. The sentiment summaries can come from reviews found on web sites, and in other locations.
Aspects that may be summarized may be statically generated for certain types of objects or services. Static aspects to be summarized are predefined ones – for example, reviews of hotels would probably always include location or service, so those aspects will be included in all summaries for hotels.
Aspects may also be dynamically determined for different objects and services. That means that a reviewer may express an opinion or sentiment about some aspect of what they are reviewing that isn’t predefined. In a review of a pizzeria, the reviewer may have expressed opinions about the “cheesesteaks,” the “salads,” and the “chicken fingers” they had while they dined there.
The sentiment summarizer looks for phrases about different static and dynamic aspects of objects and services in reviews and creates a summary of the review that includes those sentiments.
A data repository – The actual source reviews, and the summaries of the reviews may be stored in a data repository.
These can include both professional reviews and user-provided reviews from web sites on the internet. A wide range of reviews can be included in the data repository, beyond just restaurants, hotels, and electronics. The patent application points out many examples of other entities that might be included, such as “hair salons, schools, museums, retailers, auto shops, golf courses, etc.” It’s possible that in addition to including the actual reviews in this database, links pointing back to the sources of the reviews may be included.
The summaries stored in the database include sentiment phrases from the source reviews. For example, if a restaurant is reviewed, and the aspect is “service,” sentiment phrases that might be included could be things such as “service was quite good” or “truly awful service.”
A sentiment display engine – In addition to summarizing opinions and storing them, this process requires a way to show the summaries to searchers.
While we might see sentiment summaries in Google’s search options review tab, the patent filing points out that summaries might be shown in a local search.
In one embodiment, the sentiment display engine is associated with a search engine that receives queries about entities local to geographic regions. For example, the search engine can receive a query seeking information about Japanese restaurants in New York, N.Y. or about hotels in San Francisco, Calif.
The search engine provides the query and/or related information (such as a list of entities satisfying the query) to the sentiment display engine, and the sentiment display engine provides summaries of aspects of matching entities in return. Thus, if the query is for Japanese restaurants in New York, the sentiment display engine returns summaries of aspects of Japanese restaurants in the New York area.
The summaries can include a star rating for each aspect, as well as relevant snippets of review text on which the summaries are based.
How Sentiment Summarization Works
The sentiment classification approach requires the search engine to collect a large body of reviews in text form and to go through them and break them down to a word by word level, where each word is tagged with a “part of speech” token that classifies it. The Ultimate goal is to identify phrases like the following:
- Very good sound quality
- This is my favorite pizzeria ever!!
- Print quality was good even on ordinary paper.
Before we can get to phrases like that, the classifier program needs to understand what kinds of words are included in the reviews and to see if they fit into certain patterns or regular expressions.
The “part of speech” tagging can identify punctuation, adjectives, verbs, nouns, pronouns. It may use natural language processing to stem words to their roots, understand different senses or meanings of words used, and recognize compound words.
Regular expression identification can then be used to extract phrases from the text that has been “part of speech” tagged. Here’s the example used in the patent filing:
The following regular expressions are given in standard regular expression notation. In this notation, the second set of parentheses represents an example of the text that is extracted.
1. Adjective+Noun: “(.*?)(A+N+)( )” (e.g. great pizza)
2. Adverb+Adjective+Noun: “(.*?)(R+A+N+)( )” (e.g. really great pizza)
3. Model Verb+Verb+Adjective+Noun: “(.*?)(MV ?A+N+)( )” (e.g. can make a great pizza)
4. Pronoun+Verb+Adverb (optional)+Adjective+Noun: “(.*?)(PV ?R*A+N+)( )” (e.g. I love the really great pizza)
5. Punctuation+Verb+Adverb (optional)+Adjective+Noun, if preceded by punctuation: “( |.*?Q)(V+?R*A+N+)( )” (e.g. Love the great pizza)
6. Noun/Pronoun+Verb+Adverb (optional)+Adjective: “(.*?)((?: N+|P)+V+R*A+)(Q|$)” (e.g. the pizza is really great)
In alternate embodiments, other methods of identifying sentiment phrases are used, such as syntax trees or semantic grammars.
Once sentiment phrases are identified by the identification of regular expressions in the reviews, the phrases are scored based upon how strong the opinions or sentiments expressed within them might be. This might be done by looking at the words used in the phrases, to see if they tend to indicate fairly strong opinions. For example, the word “quality” tends to be fairly positive, while the word “disease” tends to be generally negative.
I wrote in my earlier post involving sentiment analysis about how some words may express a positive sentiment for some types of objects or services and negative sentiment for others. The phrase “lightweight” might be positive when it comes to a smartphone, and not so positive when it comes to a book about politics.
The next step involves the association of sentiment phrases with different aspects of things and places and services being reviewed.
Aspects and Sentiment
If you replace the word “aspects” with the word “features,” it’s much easier to get a grasp on why this patent filing focuses upon providing summaries that cover different aspects of whatever is being reviewed.
When someone looks for information about a restaurant, they may want to know a number of different things about that restaurant:
- How good is the food?
- Do they have a good wine list?
- How is the service?
- Is the atmosphere formal, casual, comfortable, cramped, noisy?
- Where are they located?
- Do they have good seafood, good steaks, great desserts?
These are all features that can be reviewed by a reviewer, and a sentiment summarizer would be much more valuable to people looking for reviews if the reviews covered different aspects or features of the thing being reviewed.
The sentiment summarizer specifically looks for different aspects of the sentiment phrases identified. Some of these may be predefined, such as “location” and “service” for a hotel. Those are such important features of a hotel, that they should be covered in a summary. They are “static” aspects or hotel reviews. Static aspects of electronic devices or DVDs would be different.
Other things may come up in a review outside of predefined static aspects, and these are referred to in the patent filing as “dynamic” aspects. For example, a restaurant that serves “fish tacos” may be reviewed on those fish tacos. That isn’t one of the static aspects that might be covered in all reviews of restaurants, such as price, decor, and location. It’s a feature that is unique to that restaurant.
Here’s part of an image from the patent application that shows sentiment phrases organized by different aspects:
Phrases are matched up with aspects identified in a review, and associated together, like in the image above.
Once that is done, the phrases may be ranked to see which contains the strongest sentiments, whether positive or negative. The ones that do may be combined to create a sentiment summary.
The phrases taken from reviews about a specific object or service may be analyzed to see what percentage of the phrases tend to be positive and which percentage are negative. If 90 percent are positive and 10 percent are negative, and there are slots for 10 phrases in the summary, 9 positive phrases may be shown, and 1 negative phrase.
The sources that the phrases are taken from would also be looked at, and an attempt to present sentiments from as wide a variety of reviews as possible would be made.
Different aspects of the objects or services being reviewed would also be a focus in determining which phrases to include in the sentiment summary.
Conclusion
Google will try to cover different features, or aspects, of objects or services reviewed in their summaries, and will try to match the percentage of positive or negative sentiments in the summary phrases they see those reviews within the summary that they present. Here’s the summary for one restaurant that I found in Google Maps:
Great place to have dinner and a good time with friends. The service was perfect. The server seemed to be able to read my mind. Everything I needed was delivered before I even asked for it. Great decor and delicious food. …
Notice that it includes five different sentiment phrases, covering a range of different features or aspects of the restaurant. In this particular instance, those phrases were all taken from the same review, which covered a range of different aspects of the restaurant. The phrases could have easily been taken from many reviews about the same restaurant.
In the first patent application we looked at from Google about sentiment analysis, we learned that Google might see the same word as expressing a positive or negative sentiment or opinion based upon what was being reviewed.
In this patent application, we see that Google attempts to construct sentiment summaries that cover a range of features from reviews.
In my next post on Google’s sentiment analysis, we’ll look more closely at how sentiment phrases are created, and how they are chosen to be used as snippets in search results.
Brilliant article… and a brilliant idea. This may be a dumb question, but where is this Google “review” search option??
Hi Bill. Thought provoking as usual! We have a lot of hotel clients so the reviews shown in Google Maps are of great interest to us. Understanding some of the concepts involved here is very useful. Many thanks…
I would love to hear one of the engineers talk about how they deal with slang, esp. the language of youth who often say things like “That pizza is sick!” and mean it as the highest compliment possible. π
I wonder how this would translate to non-En content? Language is always a fabulous barrier!
The other major issue is the source of the reviews. It’s obvious that Google has to handpick the review sites they seed the reviews from. Sadly spam likely precludes them from doing otherwise.
Bill – Have you seen these sentiment-based review listings in the wild ever?
Hi Kim,
Thank you. You can get to the “review” search option by clicking on the “Show options…” link just above search results after performing a search, and then clicking upon the “Reviews” link in the left hand column. Again, there are also reviews that show up in Google Maps, and I’ve seen some reviews show up in regular search results as well.
Hi Bullaman,
You’re welcome. I’m also looking at some of the older papers and patent filings from Google regarding reviews. It seemed like it was worth doing, to see what this sentiment analysis adds to it (or takes away from it). Hopefully, I’ll get to write about those sometime in the near future.
Hi Richard,
Good questions. I know Google has been actively engaged in creating statistical language models. We can see that in their Google Translate service, and it’s mentioned in a number of patents that attempt to understand the meanings of words based upon their context. I think it might translate well into non-English content, though it’s quite possible that English would be the first language to focus upon.
I don’t know when Google might expand the sources of their reviews outside those “handpicked” sources, but I imagine they might at some point in time. It’s a topic that I am going to discuss in more detail in a post on sentiment snippets, since that patent bushes upon that topic.
We were told in Google’s searchology presentation this year that they are using sentiment analysis in their reviews. After looking at a number of reviews, I believe that they are making decisions about what to present based upon some level of sentiment analysis. As for the local search results, I haven’t seen anything that looks quite like the screen shot that I included in my post, but again, I think there’s some level of sentiment analysis going on.
Hi Matt,
π
I imagine that slang is one challenge that is to be expected. Regardless of whether a phrase like “That pizza is sick” is used in a positive manner or a negative manner, it’s clear that it’s a strong sentiment either way.
The first of the patent filings on sentiment analysis and reviews that I wrote about previously described a method that has to have some level of flexibility to it when it comes to the senses and meanings of words, and how they might apply to different kinds of things being reviewed. It requires the creation of a language model that can make good guesses as to whether a particular term or phrase might be positive or negative based upon context. I would hope that language model is flexible enough to get some grasp of slang if it is seen enough. If many reviews of pizzerias have 5 out of 5 stars with statements in them like “that pizza is sick,” that should hopefully be a sign that’s a good thing in a pizza. π
As always Bill – awesome analysis.
I agree that slang is an issue – heck, have the time I don’t know what the youth are saying let alone systems. π
Thanks as always !
Chris
Hi Chris,
Thanks. The evolution of language and the use of slang has to make attempts to do sentiment analysis interesting.
Hi Bill. I’m wondering – has location in google maps some influence on serps (think about regional keywords). Probably not, but i’m not sure
Hi Adi,
I’m not sure that location in Google Maps has that much influence on search results at this point. I’ve seen manybusinesses appear ranked highly in Google Maps that don’t have much of a presence in organic search results for regional queries. I think that’s one of the benefits of blending in local search results into organic search results – it creates the possibility that the local search results might point to places that the organic results might not show.
Congrats Bill, outstanding article and great blog.
Adi, I don’t think location has an influence on Google web search results. It has, however, a clear influence on Google Maps search results. There are many factors believed to play into Google Maps’ ranking algorithm, among them:
– Location. If searched for “pizza Chicago” on Google Maps, for example, proximity to the center of the city of Chicago will be evaluated
– Optimization of the listing, including keywords in the business name, description, categories, etc
– Age of listing
– Local phone number on listing
– Number of reviews
And many more.
It takes time and work to achieve good results in local search, and it’s even more challenging to get a local listing to show in the universal search results, but I personally consider local search to be an exciting and changing environment.
Cheers,
Antonio
Hi Bill,
Thanks for your answer. Actually after re-reading my comment I realized that I didn’t communicate my thoughts properly. I completely agree with you in that location of the searcher has a significant influence in Google web search results, specially if the keywords being searched for trigger local results. We can see this by searching for “pizza” from two different locations, and seeing how not only the Google Maps but also web search results change.
What I meant to say is that the location of the business itself on Google Maps has not much influence on web search results.
Cheers,
Antonio
Hi Antonio,
Thank you. Location can have an impact on Google Web search results in a few different ways.
For instance, Google may bias the results you see by what it believes your “preferred” country might be. Google has also customized search results based upon where it believes your location might be.
I do agree with you on local search – it’s a pretty exciting area and one that’s also changing fairly quickly.
Hi Antonio,
Good points – that’s definitely one of the challenges that the search engines face – the location that a web site operates from doesn’t necessarily reflect the geographic area of the people that it serves or the geographic focus of the content that it provides. A site operating a few miles away from me when I lived in Delaware was one of the most informative sites on the Web about all things British. It’s focus was on the British Isles, it’s audience was global, and its Google Maps location was in Delaware, USA.
Another thing that makes local search interesting and sometimes challenging are sites with organizations behind them that have multiple locations.