Intent Tag Clouds and 6th Sense Devices

25 02 2009

Recently, my research group has been doing some work on Intent Tag Clouds or – in other words – Intent Annotations. The idea of Intent Annotations is not to focus on what a given resource is about, but what goals it refers to/ what goals it helps to achieve / what goals it aspires to.

Here’s an example from a small project a student of mine is working on right now, which shows a comparison of Intent vs. Traditional Tag Clouds generated from speeches given by US presidential candidates in 2008:

Intent Tag Clouds vs. Traditional Tag Clouds, Based on transcripts of political speeches given by the two US presidential candidates in 2008.

Intent Tag Clouds vs. Traditional Tag Clouds, Based on transcripts of political speeches given by the two US presidential candidates in 2008.

On the left, you can see an Intent Tag Cloud describing what goals are referenced in the political speeches given by the candidates, while on the right you can see a traditional tag cloud of the same text, based on word frequency and a simple stop word list. The difference lies mainly in the observation that intent tags reference future states (often in combination with a verb) focusing on aspirations, while traditional tags focus on reflecting content. In cooperation with A. Gordon from USC, we have developed an algorithm for intent annotation based on an existing social-psychological theoretical framework [1] that can produce intent annotations with useful results. However, our approach is a prototype, and some flaws can easily be identified from the intent tag clouds above.

Which brings me to a recent demonstration of Pattie Maes (MIT Media Lab) presenting some of their research on Ambient Intelligence (Now Fluid Interfaces) at TED 2009, particulary their research on portable “sixth sense” devices:

Media Lab Sixth Sense Device

Media Lab Sixth Sense Device

Here’s a link to the full video (it seems that you have to endure a sponsor message in order to watch). Make sure to have a look at min 01:40 – 02:00. Pattie’s group demos a vision of a device that can project relevant information on arbitrary objects in a person’s immediate surroundings. An example that was shown is what I would call a “Human Tag Cloud”: Projecting a tag cloud onto a person’s body while talking to him/her. The idea here is that the Tag Cloud provides some clues about the person’s interest or character during interaction.

While having a traditional tag cloud projected on a person’s t-shirt could definitely spark interesting conversations, such an approach is limited in terms of what it can say about a person (an example from the video would be the projection of the tag “Canon EOS 650” on the person’s T-shirt, it’s not quite clear what that would say about a person). Of course this is by no means a limitation of the 6th sense interface, but imagine how interesting it would be to have an idea about a persons goals, his aspirations and motivations, before entering into a conversation – Knowing that a person cares deeply about the environment or is enthusiastic about charity! Apart from privacy concerns that would certainly have to be dealt with, such technology could prove useful in the formation of real-world groups or the initiation of conversations.

A paper describing the technical details of our approach is currently under review, and we are working on a web interface to produce intent annotations from different textual resources, so stay tuned!

[1] Chulef, A. S., Read, S. J. and Walsh, D. A. (2001), A Hierarchical Taxonomy of Human Goals. In Motivation and Emotion, 25 (3), 191–232, 2001.




8 responses

25 02 2009

Hi Markus,
how do intent clouds handle goals that may not be directly mentioned in the text? Do you have a structured taxonomy that relates hierarchies of goals?

I’m working on a paper that uses software quality taxonomies to probe developer conversations to see how the terms occur over time.

25 02 2009

Hi Neil,

Great question! Our work assumes that goals are not directly mentioned in text (in fact, we looked at several textual resources and found it difficult to find explicit references of goals), but can be inferred from actions that are indicative of intent. Now the question becomes: How can we 1) identify indicative actions and 2) relate them to goals?

For that, we are using a knowledge base that maps ~170.000 indicative actions to one of 135 categories of human intent (such as “Looking Young”). For each sentence in an input document (e.g. Obama’s speech) we calculate its similarity with each intent category, which produces a characteristic intent vector for each input text.

So how did we build the knowledge base? Each category in the knowledge base is described via descriptive phrases (e.g. “Looking Young”, in this case the same as the category title) and combined with causal phrases (e.g. “in order to”). We then used Yahoo to obtain websites that contain these phrases, and we assumed that in the textual neighborhood of these phrases (e.g. “in order to look young you need to work out”), indicative actions occur. We stored these indicative actions in the knowledge base and can now compute similarity with an input text. This approach works to some extent but also suffers from a series of problems, which are detailed in our paper – I’ll send you a current version!

1 03 2009
Pranam Kolari


This is an interesting research agenda. Tag clouds haven’t see much use beyond basic textual re-representation.

I am very curious to learn the details of your approach, and if you are using a vocabulary based, or a machine-learning based approach. Also interested in how disparate social media corpora render themselves to such analysis.

Please post on your blog once available.

1 03 2009


Thanks for your comments! At the moment, our approach is vocabulary-based, using simple similarity metrics. We are interested in translating the problem into an ML problem though.
Regarding your second comment: I suspect that the performance of our method depends on certain characteristics of the corpora. In fact we have chosen transcripts of political speeches as a first dataset because of the lack of noise and the suspected expression of political/social intentions within them. The applicability of our approach to other corpora (blogs, tweets, etc) is something we are looking at right now.
I’ll post more information on the blog when it is available, so stay tuned!

23 04 2009
Annotating Textual Resources on the Web « Intentialicious

[…] Textual Resources on the Web 23 04 2009 I have recently written about current research in our group attempting to annotate textual resources with i…. A preliminary report from a study that focused on annotating transcripts from speeches given by […]

5 03 2011

Hi Markus,
I’m interested in your intent tag cloud… You mentioned the following in 2009….”A paper describing the technical details of our approach is currently under review, and we are working on a web interface to produce intent annotations from different textual resources, so stay tuned!”

Has the paper been completed and is there a working web interface available?

Many thanks.


8 03 2011
Markus Strohmaier


Thanks for your interest, the paper can be found at:

The web interface can be found here: (currently offline)

3 03 2014
Comparing the Clouds | Koltrast

[…] and girls, while the other is two identical set of words picked from two tag clouds  (flickr & obama’s presidential speech) with only difference in font size. To put it simply, either we choose two sets of words sharing […]

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )


Connecting to %s

%d bloggers like this: