Saturday, September 17, 2011

Cool Dozen

At TechDisrupt 2011 Vinod Khosla enthusiastically spoke about what he calls his “cool dozen” areas that excite him in the technology space. A few of them are: data reduction, big data+machine learning to focus on user preferences, apps (context-aware and mobile) that tap emotion, eduction, health, utilities, the democratization of publishing, interest graphs.

Points to ponder in some of the broad directions that VCs (Venture Assistant as Vinod calls) are headed:
Social Advertising: We have seen digital advertising evolve from banner ads to search ads and now social ads. But what is the definition of social and how can we turn user engagement into business? More importantly, how can we encourage people to do targeted advertising for us via their social channels (facebook,twitter, foursquare etc.)?

Clean Technology: How can we make existing mass production processes more not only better, faster and cheaper but also CLEANER !

Tuesday, June 14, 2011

Voicify : Crowdsourcing Voice Aquisition

1. Upload a photo of text, document, URL pointing to the data you want speechified
2. Select from a list of existing users, judging by samples or open an audition for a sample !
3. If satisfied by sample continue to completion !
4. Website provides quality assistance: cleanup of voice samples, voting mechanism for the same.

Acquiring userbase:
1. List of people who post on elance, mturk for voice overs
2. Signup people from Mturk / Crowdflower and ask them to post sample of voices
3. Spread the word !

Research and Technology:
1. Storage on E2 cluster
2. Background noise cleanup (filters etc)
3. Quick samples can be obtained on the web or on devices similar to -,

Who may be interested:
1. Animation companies (Project description must be more detailed to accomodate such)
2. Speech Researchers looking for voice overs
3. Every startup wants to spread its wings into new domains possibly with a video
4. Audio books
5. Voicify your Websites to reach people with hearing disabilities

Existing: (A site for crowdsouring music) (A site for crowdsourcing speech transcription - other way round) (Record and send voice emails)

Wednesday, June 8, 2011

CollabSum: Collaborative Summarization of Webpages

Every page on the internet is read by at least more than two users. We can observe a power-law distribution over number of users that read a webpage. I am convinced at this point that most useful pages on the internet are read by a few thousands of users everyday. Imagine some of those users clicking on at least one sentence from the article to let us know its importance to the meaning of the article. Now imagine accumulating that information from a few hundreds of such users and producing a synopsis of the article for the quick consumption of everyone else.

I think technically this is a easy to build application which has a far-reaching impact and changes the way we consume content on Internet.

So what do we need to achieve this:
1. A client-side plugin that records clicks of users on a sentence of the news article.
2. A server-side that records these clicks and accumulates them
3. A method to push the accumulated summary back to the Client.
4. A way to overlay the summary on the article (highlight the sentences, heatmap, pop-up)

Who are involved:
1. User: He provides a single click, but benefits from the summaries of others. One way of addressing information overload.
2. Newspapers: Can understand their users better and create synopses and daily digests much faster.
3. Third-party: APIs can be provided to anyone that needs these summaries
4. Researchers: Automatic summarization systems can benefit from such data

What do we do with such data:
1. Improve existing summarization systems for the web
2. Build iphone apps that can provide summarized versions of the existing blogs
3. Create better RSS feeds with summaries and smarter digests for paid subscription users
4. Provide SEO support for user-generated social media by guiding Adsense programs towards relevant text
5. Kindle can now use the data to provide "highlights" for newspaper subscriptions

And all it takes is a single click of the user. At the risk of sounding a cliche - "A single click a day, can keep the information overload away".

Tuesday, June 7, 2011

Crowdsourcing Translation and Commercial Models

Existing Market:
MyGengo translates thousands of texts daily in 14 languages for clients such as ShapeUp, Evernote, Youversion and Producteev.

Potential business:
- Typing scanned documents (ULIB style)
- Localization of Websites (Can we directly substitute it into the website and make it available for download?)
- Translating Advertisements
- Sub-text translation of DVDs for Movies
- Translating Polls and Surveys

Sunday, June 5, 2011

Language Learning Games

Commercial Websites:
Combined current environment into the learning of a language.
(Generates reports of which words went wrong .... should be useful in
the word alignment game)
(A cool set of iphone apps associated with the site)

Non-profit Outfits: More details and interesting log of activity in the project at the author's blog - : Donate rice by playing games, some of which are language learning flash-card style games. Interesting categorization and flash games.

Research Projects:
Lingua Mechanica: Eric Horovitz's group at Microsoft. Sample games: Word Tetris

Duolingo: A project from Luis Von Ahn, CMU.

Friday, May 6, 2011

What will you work for? Incentive system in Crowdsourcing

A few days ago, I was pitching an idea at the Entrepreneurship Club at CMU to move away from micro-payments in crowdsourcing to something what people would care for. While micro-payments like $0.01 are meaningful in developing countries, it is not attractive for most others. And the point here is that the reward mechanism has severely narrowed down the kind of tasks that can be performed online in a crowdsourcing manner. If we can find an incentive system that is attractive to a wider range of audience, it will automatically drive more users to participate and in-turn more companies will start moving towards crowdsourcing.

Here is a potential list of things that I think people would work for and surprisingly most of them are still unexplored incentive systems. Unless current platforms expand and provide flexible incentive systems, the adoption of crowdsourcing as mainstream will be limited.
1. Micro-payments (Mturk, Crowdflower etc)
2. Free mobile minutes (Txteagle)
3. Frequent flyer miles (?)
4. Coupons (?)
5. Movie tickets (?)
6. Discount points that can be used at other sites
7. Facebook points, Zynga points (Gambit+Crowdflower)
8 .....

Think about it. Amazon's Mechanical Turk is where most of the crowdsourcing happens today, and at any given instance the number of tasks on the site are still less than 200,000 ! So little for such a widespread concept of crowdsourcing. For more details of studies realated to MTurk, I will refer you to Panos's blog.

Wednesday, May 4, 2011

News Games: Annotating information nuggets in news via Games

Most of us read news every morning, some more often than others. While reading can we have a bit of fun with others who are reading the same article? And by doing so can we in-turn help enhance the reading experience of everyone else as well?

The idea is to device a bunch of simple 2-player games around the content of the newspaper which when annotated gives information that can be used in improving the presentation of the newspaper as well. Imagine the following scenario -

Scenario 1: You are reading a newspaper in your local newspaper (I speak Telugu) Does YSR, and Y.S. Rajasekar Reddy mean the same? How does a computer know? Can a single player play a "Spot the Person" game, where you click on a word thereby highligting it. You get points for identifying the maximum occurences

Scenario 2: The same scenario above can now be played as a 2-player game. Consider a "sentiment analysis" task. Given a newsarticle, ask a question two both participants about "What is Jagan doing in Hyderabad?" and you select a sentence that highlights the activity, lets say "campaigning". Both users get points when they agree on the click !

What does all this mean to your local news reading experience?
- We can now cluster news around people causing them (Named Entity Tagging)
- We can highlight events and relationships between people (Relation Extraction)
- We can better analyze news to create time-lines , location tracking and also understand general pulse of people (Sentiment Analysis)

Tuesday, May 3, 2011

Donate a word


The idea of the site is to collect an English word and its translation into a favorite language of anyone willing to contribute. This summer, I wish to enhance the collection process using a web interface with ample feedback and visualization.

Motivation factor for people to contribute to such a site -
1. Work towards a nobel cause (donate rice -
2. Fun to play (A 2-player game, or a scrabble style game) with a point system
3. Learning experience (123teachmespanish style)

Our motivation here is to enable "Assistive reading" technologies. An accompanying application/plugin will be built to enhance an average non-english speaker's reading experience on wikipedia , by suggesting translations at word-level in their native language.
The source vocabulary will cover all words from English wikipedia and a user can select his target language for providing a translation. Existing dictionaries free online ill be used to seed the application/plugin so that user's get a feel for what their contributions can achieve.

Visualizations will be the motivation:
1.How much of Wikipedia has been covered for the user's language
2. How many articles will the user help by translating a particular word

Visualizing twitter streams from different languages

Twitter has become a great hub that gives a sneak peek at what the world is talking about. However we are only equipped to analyze the English tweets (e.g recent work on dialect identification on twitter, part of speech analysis, sentiment analysis etc.). Can we go beyond in analyzing other languages ?

Why do we need to do this?
More than 50 percent of the messages on twitter are non-English. Infact as of today the world speaks about 4000 languages, and restricting ourselves to English tweets does not give us a complete picture of the world's communication.

For instance, Egypt problem as viewed from united states is different from what people in Egypt sees it. What about the opinion of people who are on twitter but blog/tweet in a different language? Has this to do with the language twitter speaks? What if they don't tweet in English and don't follow the conventions of #hash tags

1. Visualize the twitter/facebook streams for a particular query "egypt revolution" and overlay it on the world-map
2. Use a different lens (a dictionary for Arabic-English) and translate more of the twitter streams (word for word) and overlay it on the map. Does this look radically different?
3. Zoom-in and out on different countries to see what they think about an issue.

Well, what about the rest of the people who don't blog, tweet, facebook? Thats a problem for another day.

Monday, May 2, 2011

Language Translation in the Crowd: Part 1

In the past I have tried involving crowds for translating text to be used in seeding an automatic translation systems.(refer: 2010, Ambati and Vogel 2010)

A few problems with crowd:
1. Too many spammers, how do you know who is doing the right thing, when you don't know what's right.
2. Too few bilingual speakers for any language-pair you pick. There are 50 major language pairs, and 3950 other languages in the world. Think of creating translation systems for translating between 4000X4000 !!!
3. How do you make it interesting for the users to contribute the site and not feel that they are being stolen of their Intellectual property ! (Give them money. Not feasible when you think of the few thousands of language-pairs you are considering)

Now there are some projects out there which have looked at a sub-set of the problems I mention above, although I am not convinced yet we have a silver bullet yet.
1. Monotrans: Effort from Maryland, which is by now well published now and the results have been applied to translation of Children books from the ICDL
2. Duolinguo has been making some noise for about a year now, but its not yet seen by the world outside. I hope and wish its really good, coz the success of such projects give focus to similar efforts !

My take on this is the success of translation in the crowd is going to need the following: (Some of which I am working on and will be publishing in my research!):
1. Translation task needs to become verifiable: Task-breakdown
2. Involve two vs. one person: Collaboration
3. Make it fun or a learning experience for all: Challenging Innovative Games

In a continuing post, I will talk about some of the designs we have come up with to build, collaborative, constructive and verifiable methods for involving the crowd and hopefully its fun and motivating enough for people to contribute without making them feel they are robbed of their time or knowledge. After all, knowledge can only be shared and it rightly should be !

Sunday, May 1, 2011

PollSense: Its the ad-sense of Polls

Polls when, where and how?: A task-integrated context-sensitive polls

Ask them when they are willing to provide?

We need the opinion of people who are well educated in the issue and so identifying the context gives us an understanding of our user. If he is reading about the "Health bill", he perhaps is a right user to vote on it. Reduces noise when compared to the traditional polling methods.

Engage users . Similar to context sensitive ads.
Have widgets and provide services

language independent , domain independent
Time-sensitive, based on social media, current-affairs ?

- As more sites adopt it, you have an understanding of Page-ranks , what pages people visit etc


- Create thousands of automatica polls from news, and provide a toolbar for users
- Share data or just results ?
- Can we learn more about the users themselves (tie up with the web)

Bloomberg is getting into it.

Whats out there:

Create content for your site:

Tuesday, April 5, 2011

SearchGuru: Assisting Computer Newbies in Search

SearchGuru: (firefox plugin, chrome plugin that connects to a service)
- When a non-native user types a search query, if we can identify the sense or meaning of the word/phrase typed based on the results returned, then we show the “phishguru” style help online for the illiterate population, old population that are not comfortable with Google (VillageGoogle)
- Provides data for Crosslingual IR (queries that users affirm for a set of results can be cross-indexed)
- Provides data for sense-disambiguation
- A cool mascot is needed to cater to Kids and Elderly !

Monday, April 4, 2011

Meta-Search Engine for Micro-tasks

Crowdflower, odesk, crowdcloud, castingwords, mechanical turk and the list goes on.
Lots of websites and the users have no clue of where to contribute how and make best use of the few minutes of time they have per day.

We propose a meta-search engine for harvesting tasks in real-time from these micro-task markets and act as a thin-layer of abstraction.

Now, apart from having the luxury of a single-point of entry, what else can we offer so that users embrace such a crowdsourcing platform.
1. Larger volume of tasks becomes quickly attractive for users who are looking for scale
2. Personalized search catering to the constraints of the user
3. Recommendation based on the skillset of the user will be benficial not only to user but also the requester who has to deal with low-quality otherwise.

As researchers embrace micro-task markets for eliciting human input, the nature of the posted tasks moves from those requiring simple mechanical labor to requiring specific cognitive skills. On the other hand, increase is seen in the number of such tasks and the user population in micro-task market places requiring better search interfaces for productive user participation. We posit that understanding user skill sets and presenting them with suitable tasks not only maximizes the over quality of the output, but also attempts to maximize the benefit to the user in terms of more successfully completed tasks.

Summarizing local news for Global Communication

mNews -

Wednesday, March 16, 2011

Flash games with movie stars

I was wondering if there are any websites out there that create online flash games featuring movie stars. India is a land for movies where people idolize heros and heroines, and I figure flashgames may cater to kids and adults equally. More users on the site, more advertising revenue. Sites like Mochi Media ( and even Google now provides in-game advertisements.

Imagine a fruit ninja like Rajnikanth slicing fruits. Or imagine Emran Hashmi in the role of this smoocher here. :)

- Speech synthesis systems that can mimic the voices of public figures
- Computer simulations of gait and mannerisms

- Do IP problems come into play. Do we need to pay studios for images/voices of film personalities?
- Is the research there yet?

Thursday, March 3, 2011

Video Ad Annotation

Ads in all languages, local brands, but inaccessible as they are not rich in content

We provide exposure.
1. You upload a video
2. We tag , enhance the content for SEO purpose
3. We also annotate existing content for Google

1. Users
2. Local brands:
3. Ad and Marketing Giants:
4. Search Engines

More details...TBD