Lecture on Nov 9, 2010. (Slides)
Information Visualization for Search Interfaces, Marti Hearst, Search User Interfaces, Chapter 10
Information Visualization for Text Analysis, Marti Hearst, Search User Interfaces, Chapter 11
Mapping Text with Phrase Nets. Frank van Ham, Martin Wattenberg, Fernanda B. Viégas. IEEE InfoVis 2009.
Viewing query hits based on position in the document I thought was neat, especially when combined with a small multiples approach as in the State of the Union visualization in Chapter 11 of the reading. Just telling the number of results in the document is ambiguous: is the query talked about in the entire page or is just a few paragraphs of dense usage? That said I'm not sure how much it helps besides in a "huh, that's cool" way.
I was also caught by surprise when I went to use google after doing the reading and finding out that they've added the ability to preview websites from the main page (click the magnifying glass). Again, I'm not sure how useful this is though, since it's almost as much work as just clicking the page and going back if the page wasn't what you thought it was.
Reading through the chapters from the Search User Interfaces book I got a distinct impression of how fundamentally hard text visualization is in the general case. Chapter 10 was mostly an overview of a large spectrum of visualization schemes applied to searching text, each one successful only in particular application cases. I completely agree with Hearst's argument that the very high dimensionality nominal nature of text makes its reduction to lower dimensionality spaces extremely challenging. In fact, in the absence of a well defined mapping from nominal text to at least an ordinal space (which usually comes along with a particular application domain, such as for example when analyzing word frequencies within a given corpus) all attempts at visualization seem to be relatively powerless. The popular saying that "a picture is worth a thousand words" seems to succinctly portray the fact that the direction of text-to-picture is fundamentally a compression (typically lossy) and as such likely to be successful only for particular domains where the information lost is not required in achieving the given task. (Apparently compression research has dabbled in lossy text compression only "semi-seriously" so far, at least according to the Wikipedia article on Lossy Compression)
@estrat Are you talking about "Wonder Wheel"?
I must admit that's a pretty odd visualization interface, but I haven't fully digested it yet. In terms of examples to cite, I think Visual Thesaurus tops the list (Wordle is also pretty nice - Marti Hearst also refers to them in Chapter 11). I think what's exciting about text-based visualization is that it poses such a challenge (in terms of effective space use contrasted with visibility). In particular, since I love languages with a passion, the idea of using semantic structures like "X begat Y" seems almost too intuitively simple of an idea.
I wonder if anyone has done lyrics visualizations? # Oh but someone from Stanford already has: https://ccrma.stanford.edu/~jieun5/cs448b/final/Oh_final.pdf
@yanzhudu - I agree, but there aren't a lot of great alternatives that don't integrate interactivity. Innovation is much needed in this space because I agree the tag cloud isn't right, but it's at least a compromise. I think adding some semantic meaning to the text is a great place to start so certain terms can be aggregated together. This obviously requires more tweaking by the person making the visualization, but it would contribute to reducing the chaotic form of text in the tag cloud.
I really enjoyed Heer's description of the Wikipedia edit history visualization of the stacked graphs in the context of text wars.
@rakaska No, it's a new feature and I guess it hasn't been rolled out to everyone yet. Here's what it looks like:
Tag cloud has been over-used. It looks cluttered and confused. This might be suitable for random twits. But for normal blogs, Tag cloud might mislead its readers.
I wonder if anyone has done serious usability test on tag cloud and how user perceive and use tag cloud.
The discussion on source code visualization reminds me of gource, for example this video.
Source code visualization is, in my opinion, still very inadequate. But then again I think the same thing of programming language expressiveness.
In lecture, Jeff repeatedly made the point that text visualizations often (almost always?) convey models of text rather than raw data. This did indeed seem true for many of the visualizations we saw (e.g. using frequency counts, G^2 measures, etc). However, the NYTimes example of ? and the Darwin example were two visualizations that actually showed the "raw data", namely by highlighting where certain words appeared in the text. This also held true for much of the "diff" revision visualizations which showed the words themselves rather than abstractions of the words. These types of visualizations seemed like noteworthy exceptions to Jeff's general rule.
One of the most valuable insights I took away from today's lecture and the phrase net article was how the usefulness of many of these text visualization tools could vary alot depending on what type of analytical task one was interested in. When I initially saw the phrase net article, I was confused about the usefulness of the tool. However, reading it closer, I found the use cases illustrate that this tool provides quite a bit of insight into the details of how language is being used. For many analytical tasks, this is not of interest; for the discourse analysis tasks in questions, the tool seems well-suited.
I think it would be cool to combine the sketch based query system of TimeSearcher with a dataset like the Enron emails. You could ask questions like: who else sends emails that are similar to these people's emails? I select a link between people or a specific email and any communications that have similar text qualities are highlighted. You could also restrain the search so that only communications are highlighted that are connected to the original edge thought highlighted edges, so you could highlight the flow of a specific email's information, similar to how TimeSearcher let's you highlight the evolution of a certain selected part of the data.
I fell completely in love with Ben Fry's origin of species visualization. Some features that weren't highlighted in class: If you hover of the visualization you can read that segment of text, and as the edits go through, the on hover will show the edited words in the corresponding edition color.
I like this visualization not only because it's showing something that is interesting, but because of the attention to detail. The chapter representation is intuitive and beautiful, and the chosen fonts makes me feel like I'm actually looking at the text that is some 150 years old, rather than a digital representation of it. The color coding is subtle but very efficient both in the animation and the text zooming. The design is clean, but has everything needed.
All in all, one of my favorite visualizations so far. If I ever get close to creating something this elegant I'll be super excited.
When I was reading chapter 10, the following line struck me: "When reading text, one is focused on that task; it is not possible to read and visually perceive something else at the same time. Furthermore, the nature of text makes it difficult to convert it to a visual analogue." I had to take a moment and think about whether or not I agreed with the statement. Mainly it was because I had just come to the realization that so much of my interaction is routine nowadays that I probably have habits that I just dont notice anymore. For whatever reason, my thinking immediately went to games--online multiplayer games which display tremendous amounts of data simultaneously. Many design decisions which I had previously not considered very strongly now were backed by this principle. For example, something as simple as where the damage amount is displayed in text. Long ago in old games, the text used to appear in a scrolling text box, but at some point games realized that it made more sense to display the numbers on the character that was being hit. When I thought about the difference in this kind of interaction, I immediately saw a connection to the above statement...it certainly helps when the thing you are reading and supposed to be visually perceiving (here, the fact that the character was damaged) are in the same place.
I thought further to another strong counterexample: subtitles. If the statement above were true, then it would be very difficult to watch shows in other languages. Therefore, I think the statement is altogether too strong and absolute. Instead I think that we are capable of doing so, just perhaps there are more efficient ways. Comic books, for example, take this a bit further and do a very good job of putting text in places that make it easy to understand their relevance and context where they are. I'm pretty sure I look at the images while I am reading the text.
I think the statement mainly pertains to textual data in mid-to-large sized volumes. But we're not actually creating visual analogues necessarily when we read text. Instead, as suggested in lecture today, we're creating a visual way to comprehend and explore the text. We're looking to visualize a model for the text that we can understand and interact with. Using "visual analogue" to describe this is a bit awkward to me.
I noticed someone mentioned twitter above. It think that is an interesting example since its domain in purely based on short texts, yet there are so many creative ways people have created tools to visualize it. For example, this link provides a brief overview on 17 different ways to visualize twitter, and includes some examples we actually saw on today's lecture http://flowingdata.com/2008/03/12/17-ways-to-visualize-the-twitter-universe/. This is also another interesting interactive visualization of twitter's text: http://twistori.com/ Oh, and I agree with @yanzhudu in that tag-clouds are overused.
I think that there is a big issue with creating visualizations that are composed of written works which has not been addressed and is, in my opinion, a pitfall that a lot of the visualizations we've considered fall into. That problem is simply Readability. In the effort to encode a ton of information into a small interactive viz, the words become either too small, too transparent, too close together or too badly oriented to be readable. I find that on any one visualization with a lot of text in it, I spend a LOT of time trying to read relatively few of the words and often have a hard time making them out. If the goal is to highlight the few giant bold solid words that I can read then it's not an issue. but often I want a little more then a glimpse and getting that extra 10% of information is very hard.
So, the additional constraint i would add to making text visualizations is: Make it easier to read!!! If a word takes a long time to decipher then chances are that rather then add any value to the visualization it actually subtracts from the overall value by cluttering.
Reading the "INFORMATION VISUALIZATION FOR SEARCH INTERFACES" article reminded me of a social game i used to play with others where the 1st person would write down a sentence, pass it to the #2 person who would draw a picture of it and pass it to #3 who would write a sentence of what they thought was conveyed in the picture, etc all the way to #10. Essentially a different form of the phone game. Whats interesting about this game is that a visulization or picture that focuses on fewer points (meaning the simpler the drawing the better the meaning stays unchanged) will have less data/meaning loss from person to person. I believe this also to be the case as drawn from the article which showed that search was more effective when there was less noise or other distracting patterns.
It would have been neat to see some coverage of statistically improbable phrases as a visualization of its own.
The concept of phrases that are more or less likely to occur than the average English text was brushed on a few times during lecture--we talked about unigram/bigram tradeoffs, we looked at relative unigram frequency in the Democratic/Republican abortion debate comparison. It would be interesting to elaborate on this further, since looking for phrases that are unique to a text, and visualizing it in the same order as the text, would make for a quick summary of the text suitable for search results, viewing a large list of documents for comparison of other traits, etc.
More on this here: http://en.wikipedia.org/wiki/Statistically_Improbable_Phrases . The idea is used by Amazon.com to browse the content of texts.
My favorite visualization from this lecture was the one representing the complete text of Alice in Wonderland. While it wasn't the prettiest visualization ever, and I wouldn't expect the word cloud to be too useful on it's own, I think the package as a whole is an elegant way to explore a large text document. In particular, I like how the full text is wrapped around the word cloud so it's easy to link any word to all it's locations in the document. Also, as a result, some interesting data can be encoded in the position of words in the cloud. The thing that really tops it off, though, is the ability to view the raw source text. This was emphasized in lecture, and I think allowing the user to drill into the data like this really provides a lot of valuable context. One potential improvement could be to expand this model to simple phrases (maybe just two words), as shown in the text-description example.
On a separate note, one thing I wish we covered a bit more in the lecture is grammatical analysis. We went into a lot of depth on single words or simple relationships between words, but we didn't talk much about visualizing more complex grammatical information. I'm not sure if this space is just less well explored so far (since it does seem more challenging), but I'd be curious to know what has been done so far.
@wulabs: I like that idea of a new phone game using the pictures, and I think you provided some good insight into visual compression in relation to verbosity (if that is a word).
I was impressed by how much better digrams did over the G2's unigram method. I suppose this makes sense in terms of the significant (quadratic) increase in options we get by essentially squaring the number of choices we have in the language we are dealing with. Of course that comes at a computation hit, but I think as we move forward it should be more feasible to move to trigrams and on. I imagine that there is or will be a stopping point where we have some sort of normalized curve that suggests the number of phrases we as humans are comfortable using to describe things. I think it would be around 5 or 6 phrases, however I think that number may be dropping as I see kids today talking in text. That is, they actually say the 'lol' now or other shorthand letters that they would use to type. They do this out of a constraint to both bandwidth (well perhaps initially anyway) and typing speed. Yet, these constraints on one medium are not shaping our spoken communication to some extent.
I believe that the simpler, inanimate "CVS annotate" or git/SVN versions if you prefer, provides a much better visualization (different versions given different colors) than the animated origin of species visualization we saw in class. This viz let us see how different editions of the book had evolved. The animation does clearly show multiple edits performed to the same section over time but makes it harder for a person to stitch together information about how a particular section changed over time.
I think after class i realized how hard it is to get text visualization right. I think the most challenging aspect of it is how do you reduce reading by alot and at the same time also increase the amount of information you take out. There were some compelling examples in class like the word frequency application or the "origin of species" visualization of changes in the book. However, one of my favorites more practical examples is the version comparison that the eclipse IDE does. I think the reader really abstracts away the majority of the source text and only focuses at a very small portion of that. This reduces the amount of reading significantly.
@arievans While I think you are definitely right in saying that it is possible to read and visually perceive other things at the same time, I think that the point being made is that reading takes a lot of focus cognitively, leaving less focus for other simultaneous tasks. For example, I know that when watching a movie with subtitles, I am often much less able to notice some of the visual cinematic qualities of the film. Which is not to say that it's too difficult or not enjoyable to watch, but that my eyes are constantly being drawn to one particular part of the screen and I don't notice the others as much. I think your gaming example is an excellent point on how text can be used to supplement a visualization, and in particular I am a big fan of numbers in visualization since they are highly descriptive, quick to read, and don't take up much space (but I'm not sure that I would really consider them to be "text" in a visualization in quite the same sense).
Some of the critical points here for me are that it's hard for people to a) process large amounts of text quickly as would be desired in a visualization and b) to notice patterns in large volumes of text. I think that text visualizations can be quite helpful to address these problems by directing us to key patterns or insights, but of course we need a good model for representing the aspects of the text that we care about when trying to answer questions with a visualization.
I thought that graph visualization was tough but text is even more difficult. I think the main problem (drawing form Tufte early lessons) is the fact that you use a lot of ink to communicate minimal information and that usually leads to cluttered and confusing visualizations. Images are so more powerful to convey information whereas text is much more important for precision. Another important issue is that text is better understood in a specific, hence the simple rule that Jeff mentioned in class that you always need to add the actual text to the vis. However some of the techniques presented in class did provide useful insights and trends. I especially enjoyed the evolution of Darwins work and the Wikipedia articles.
In case people are interested, there was a write-up about Jonathan Harris and WordCount in the Montreal Gazette this week. It's not particularly deep, but he shares some of his thoughts on WordCout, QueryCount, and We Feel Fine:
I was interested in the details of Jason's visualization of the different department theses. Why was the layout in a circle? Was that due to the natural structure of the data or just a decision of the layout. I find it hard to believe completely that laying out the distances from different categories of documents results in a perfect circle with only a specific few middle points. However, it looks appealing.
I wish we saw an interactive version of that application as it had a very interesting topic and showed what we might have assumed and augmented it with new discoveries describing the relationships between departments.
I think someone should make a site of all of the cool visualizations. At least from Stanford!
I think the phrase net article is very interesting in the sense that it grants users the freedom to visualize important text phases and their relationships through a set of well defined specifications. For example, users can choose the "X of Y" option to visualize character relationships in the text or the "X at Y" option to visualize location relationships. This is a great tool that let users effectively extract semantic information out of a text of interest.
@yanzhudu - I agree with you that Tag cloud is over-used. But I do think that it is still very effective in visually representing certain features of words in a text. Upon reading Ch 11 of text analysis visualization, I think that by comparing Figure 11.10 and 11.11, its very obvious that Word cloud is way more cluttered and confusing than Tag cloud.
Perhaps this is blasphemous to say, but I'm not sure I'm entirely convinced that visualizations would actually be beneficial to your basic text-based search. Basic changes in things like typography do make a big difference, but a number of the visualizations (such as the various graph=based ones in the Ch. 10 reading) seem more exploratory than practical. To make a sweeping and probably erroneous generalization, it seems as though the farther you get from the basic text, the less useful the visualization ultimately becomes.
If you intend your search queries to be deliberately explorative, however, then perhaps this is preferable. But otherwise, I am not entirely convinced that visualizations are the way to go. Surely this means that there is much more innovative research to be done in the field.
I really liked the visualization examples that showed the changes in text over time like Ben Fry's visualization of the text changes between different editions of Darwin's On the Origin of Species . Ben Fry has done some other interesting text visualization work related to text, including some projects on the human genome (like this one http://benfry.com/isometricblocks/ ) and his visualization of the evolution of the Processing source code (http://benfry.com/revisionist/ ).
I've also noticed that the technique of highlighting query term locations in a document is really widely used by Google; like Hearst, I've noticed it being used in the search results in Google Chrome, as well as in Google Books. I find this highlighting to be really effective at helping the user retain context when navigating to each search term location. The TileBars technique also seems interesting for the same reason, but the layout of having the results for each term stacked on top of the other results seems a bit confusing. Woodruff's textually-enhanced thumbnail summary also seems like a really effective method for showing webpage previews since it places the textual content within its graphical context.
It would be interesting to analyze scientifically how typographical features convey meaning in text. We know that position is great for communicating a quantitative variable; could we rank font weight or size? How does it stack against area, for example? Displaying different types could be more effective than displaying different colors when it comes to discriminating a nominal variable.
I can see from the class and the readings that text visualizations are inherently very difficult to make and make effectively. From the text search article, they make a good point that one of the difficulties in creating a visualization for textual search is that it often makes the user take more time to process the search results then the plain text alone would. It also seems to not really aid in their understanding of the search results over just the plain text, and can even distract and make searching more difficult. I found it interesting that despite this, the studies also suggested that people still claimed to prefer image based visualizations to textual search, even if it did not help them to search more effectively. On a side not, the concept of 'faceted navigation' that they talk about visualizing in the search chapter seems very relevant to my final project idea...So I will definitely read more about that in their chapter 8!
I think the issue of the "double gulf of evaluation" makes text visualization an interesting challenge. One of the important principles given in class, "always give access to the source text," seems to be aimed at mitigating the double gulf problem to some extent. However, I'm not sure that I agree that providing access to the source text is always important. Sometimes, the "model" constructed to visualize the text is simple enough that source access may not be important. For example, the "who mentioned who" visualization of primary candidates seems like it really would not benefit from access to the source text, at least for the particular purpose of that graphic. Perhaps another way to think about this problem is how we can construct models simple enough that they are easy for viewers to understand and access to the source text is not as critical.
@andreaz - thanks for sharing those, those are neat! More generally: I'm always skeptical of automatic text summarization (non-visual) because I fear that crucial concepts in the text will be omitted. Visual summaries present a nice alternative: not only do they give a global sense of the text, but they can also allow the user to drill down to fine detail. I was especially interested in the phrase nets as a mechanism for exploring and comparing author styles; in many ways I feel that this is the most useful text visualization I have seen. For example, I can imaging using those in a poetry or literature class; I find it harder to conceive using other text visualizations I've seen in real-life scenarios, other than for artistic appeal and confirmation of what I already know about the text. That said, artistic appeal is important, too, and I feel more and more than some of the more beautiful text visualizations will make their way into advertising in the future.
@estrat - Bing has been supporting that for a long time now. I think that this is a neat feature since it reduces the number of clicks by half (you don't have hit the back button), and you can still see other search results on the same page as the preview.
The text visualization lecture was very useful, considering that my final project involves a lot of text vis. We feel fine and I want you to want me both display the original text of the data, and the double gulf of evaluation discussion was very relevant in this context. It was interesting to note that text should be used to represent text as well.
@nchen11 – I partially agree with your point. I think one major reason that text visualization is difficult is that text is, in and of itself, a visualization of ideas and information. When we read, context (even to lengths of pages, chapters, or entire works) is important to meaning, and this context is often lost when text visualizations chunk text into one or two word parts. Additionally, the same ideas might be conveyed textually in many different ways.
Given this, one dimensions along which text visualizations might be placed is that with (1) focus on meaning at one extreme and (2) focus on the text itself at the other. Perhaps raw text is a visualization that is entirely focused on meaning where as something like Ben Fry's Animated Traces might be more focused on the text itself (in this case the "evolution" of a document). The result of running Prof. Heer's regression model for extracting the most descriptive words from a document might be used in visualizations that fall somewhere in the middle.
So, @nchen11, if you care most about conveying the ideas presented in the original text in the greatest detail, the raw text seems likely to be among the best solution candidates. Other visualizations have value as well though, since they can reveal things about the structure, evolution, or social phenomena surrounding documents – information that is usually not made visually salient by the text itself.
I wasn't the biggest fan of the PhraseNet visualization. I feel network style visualizations in general aren't meant to be followed along on a micro basis and are more effective at showing large scale structure such as groupings. For networks with a medium to large number of nodes there just isn't enough space to fit the text of a node at a comfortable size. I think the PhraseNet visualization may be more effective if it worked in a highly interactive manner in that instead of showing the whole graph at once it showed only a minimum number of nodes and would update when moving the focus to a new node.
cs448b_data_visualization: Text Visualization (last edited 2010-11-10 02:53:17 by jheer)