Archives

A job posting that really doesn’t suck

I just got a pile of money to build a piece of state-of-the-art open-source visualization software, to allow journalists and curious people everywhere to make sense of enormous document dumps, leaked or otherwise.

Huzzah!

Now I am looking for a pair of professional developers to make it a reality. It won’t be hard for the calibre of person I’m trying to find to get some job, but I’m going to try to convince you that this is the best job.

The project is called Overview. You can read about it at overview.ap.org. It’s going to be a system for the exploration of large to very large collections of unstructured text documents. We’re building it in New York in the main newsroom of The Associated Press, the original all-formats global news network. The AP has to deal with document dumps constantly. We download them from government sites. We file over 1000 freedom of information requests each year. We look at every single leak from Wikileaks, Anonymous, Lulzsec. We’re drowning in this stuff. We need better tools. So does everyone else.

So we’re going make the killer app for document set analysis. Overview will start with a visual programming language for computational linguistics algorithms. Like Max/MSP for text. The output of that will be connected to some large-scale visualization. All of this will be backed by a distributed file store and computed through map-reduce. Our target document set size is 10 million. The goal is to design a sort of visualization sketching system for large unstructured text document sets. Kinda like Processing, maybe, but data-flow instead of procedural.

We’ve already got a prototype working, which we pointed at the Wikileaks Iraq and Afghanistan data sets and learned some interesting things. Now we have to engineer an industrial-strength open-source product. It’s a challenging project, because it requires production implementation of state-of-the-art, research-level algorithms for distributed computing, statistical natural language processing, and high-throughput visualization. And, oh yeah, a web interface. So people can use it anywhere, to understand their world.

Because that’s what this is about: a step in the direction of applied transparency. Journalists badly need this tool. But everyone else needs it too. Transparency is not an end in itself — it’s what you can do with the data that counts. And right now, we suck at making sense of piles of documents. Have you ever looked at what comes back from a FOIA request? It’s not pretty. Governments have to give you the documents, but they don’t have to organize them. What you typically get is a 10,000 page PDF. Emails mixed in with meeting minutes and financial statements and god-knows what else. It’s like being let into a decrepit warehouse with paper stacked floor to ceiling. No boxes. No files. Good luck, kiddo.

Intelligence agencies have the necessary technology, but you can’t have it. The legal profession has some pretty good “e-discovery” software, but it’s wildly expensive. Law enforcement won’t share either. There are a few cheapish commercial products but they all choke above 10,000 documents because they’re not written with scalable, distributed algorithms. (Ask me how I know.) There simply isn’t an open, extensible tool for making sense of huge quantities of unstructured text. Not searching it, but finding the patterns you didn’t know you were looking for. The big picture. The Overview.

So we’re making one. Here are the buzzwords we are looking for in potential hires:

  • We’re writing this in Java or maybe Scala. Plus JavaScript/WebGL on the client side.
  • Be a genuine computer scientist, or at least be able to act like one. Know the technologies above, and know your math.
  • But it’s not just research. We have to ship production software. So be someone who has done that, on a big project.
  • This stuff is complicated! The UX has to make it simple for the user. Design, design, design!
  • We’re open-source. I know you’re cool with that, but are you good at leading a distributed development community?

And that’s pretty much it. We’re hiring immediately. We need two. It’s a two-year contract to start. We’ve got a pair of desks in the newsroom in New York, with really nice views of the Hudson river. Yeah, you could write high-frequency trading software for a hedge fund. Or you could spend your time analyzing consumer data and trying to get people to click on ads. You could code any of a thousand other sophisticated projects. But I bet you’d rather work on Overview, because what we’re making has never been done before. And it will make the world a better place.

For more information, see :

Thanks for your time. Please contact jstray@ap.org if you’d like to work on this.

The editorial search engine

It’s impossible to build a computer system that helps people find or filter information without at some point making editorial judgements. That’s because search and collaborative filtering algorithms embody human judgement about what is important to know. I’ve been pointing this out for years, and it seems particularly relevant to the journalism profession today as it grapples with the digital medium. It’s this observation which is the bridge between the front page and the search results page, and it suggests a new generation of digital news products that are far more useful than just online translations of a newspaper.

It’s easy to understand where human judgement enters into information filtering algorithms, if you think about how such things are built. At some point a programmer writes some code for, say, a search engine, and tests it by looking at the output on a variety of different queries. Are the results good? In what way do they fall short of the social goals of the software? How should the code be changed? It’s not possible to write a search engine without a strong concept of what “good” results are, and that is an editorial judgement.

I bring this up now for two reasons. One is an ongoing, active debate over “news applications” — small programs designed with journalistic intent — and their role in journalism. Meanwhile, for several years Google’s public language has been slowly shifting from “our search results are objective” to “our search results represent our opinion.” The transition seems to have been completed a few weeks ago, when Matt Cutts spoke to Wired about Google’s new page ranking algorithm:

In some sense when people come to Google, that’s exactly what they’re asking for — our editorial judgment. They’re expressed via algorithms. When someone comes to Google, the only way to be neutral is either to randomize the links or to do it alphabetically.

There it is, from the mouth of the bot. “Our editorial judgment” is “expressed via algorithms.” Google is saying that they have and employ editorial judgement, and that they write algorithms to embody it. They use algorithms instead of hand-curated lists of links, which was Yahoo’s failed web navigation strategy of the late 1990s, because manual curation doesn’t scale to whole-web sizes and can’t be personalized. Yet hand selection of articles is what human editors do every day in assembling the front page. It is valuable, but can’t fulfill every need.

Informing people takes more than reporting
Like a web search engine, journalism is about getting people the accurate information they need or want. But professional journalism is built upon pre-digital institutions and economic models, and newsrooms are geared around content creation, not getting people information. The distinction is important, and journalism’s lack of attention to information filtering and organization seems like a big omission, an omission that explains why technology companies have become powerful players in news.

I don’t mean to suggest that going out and getting the story — aka “reporting” — isn’t important. Obviously, someone has to provide the original report that then ricochets through the web via social media, links, and endless reblogging. Further, there is evidence that very few people do original reporting. Last year I counted the percentage of news outlets did their own reporting on one big story, and found that only 13 of 121 stories listed on Google News did not simply copy information found elsewhere. A contemporaneous Pew study of the news ecosystem of Baltimore found that most reporting was still done by print newspapers, with very little contributed by “new media,” though this study has been criticized for a number of potentially serious category problems. I’ve also repeatedly experienced the power that a single original report can have, as when I made a few phone calls to discover that Jurgen Habermas is not on Twitter, or worked with AP colleagues to get the first confirmation from network operators that Egypt had dropped off the internet. Working in a newsroom, obsessively watching the news propagate through the web, I see this every day: it’s amazing how few people actually pump original reports into the ecosystem.

But reporting isn’t everything. It’s not nearly enough. Reporting is just one part of ensuring that important public information is available, findable, and known. This is where journalism can learn something from search engines, because I suspect what we really want is a hybrid of human and algorithmic judgement.

As conceived in the pre-digital era, news is a non-personalized, non-interactive stream of updates about a small number of local or global stories. The first and most obvious departure from this model would be the ability to search within a news product for particular stories of interest. But the search function on most news websites is terrible, and mostly fails at the core task of helping people find the best stories about a topic of interest. If you doubt this, try going to your favorite news site and searching for that good story that you read there last month. Partially this is technical neglect. But at root this problem is about newsroom culture: the primary product is seen to be getting the news out, not helping people find what is there. (Also, professional journalism is really bad at linking between stories, and most news orgs don’t do fine-grained tracking of social sharing of their content, which are two of primary signals that search engines use to determine which articles are the most relevant.)

Story-specific news applications
We are seeing signs of a new kind of hybrid journalism that is as much about software as it is about about reporting. It’s still difficult to put names to what is happening, but terms like “news application” are emerging. There has been much recent discussion of the news app, including a session at the National Institute of Computer-Assisted Reporting conference in February, and landmark posts on the topic at Poynter and NiemanLab. Good examples of the genre include ProPublica’s dialysis facility locator, which combines investigative reporting with a search engine built on top of government data, and the Los Angeles Time’s real-time crime map, which plots LAPD data across multiple precincts and automatically detects statistically significant spikes. Both can be thought of as story-specific search engines, optimized for particular editorial purposes.

Yet the news apps of today are just toes in the water. It is no disrespect to all of the talented people currently working in the field say this, because we are at the beginning of something very big. One common thread in recent discussion of news apps has been a certain disappointment at the slow rate of adoption of the journalist-programmer paradigm throughout the industry. Indeed, with Matt Waite’s layoff from Politifact, despite a Pulitzer Prize for his work, some people are wondering if there’s any future at all in the form. My response is that we haven’t even begun to see the full potential of software combined with journalism. We are under-selling the news app because we are under-imagining it.

I want to apply search engine technology to tell stories. “Story” might not even be the right metaphor, because the experience I envision is interactive and non-linear, adapting to the user’s level of knowledge and interest, worth return visits and handy in varied circumstances. I don’t want a topic page, I want a topic app. Suppose I’m interested in — or I have been directed via headline to — the subject of refugees and internal migration. A text story about refugees due to war and other catastrophes is an obvious introduction, especially if it includes maps and other multimedia. And that would typically be the end of  the story by today’s conventions. But we can do deeper. The International Organization for Migration maintains detailed statistics on the topic. We could plot that data, make it searchable and linkable. Now we’re at about the level of a good news app today. Let’s go further by making it live, not a visualization of a data set but a visualization of a data feed, an automatically updating information resource that is by definition evergreen. And then let’s pull in all of the good stories concerning migration, whether or not our own newsroom wrote them. (As a consumer, the reporting supply chain is not my problem, and I’ve argued before that news organizations need to do much more content syndication and sharing.) Let’s build a search engine on top of every last scrap of refugee-related content we can find. We could start with classic keyword search techniques, augment them by link analysis weighted toward sources we trust, and ingest and analyze the social streams of whichever communities deal with the issue. Then we can tune the whole system using our editorial-judgment-expressed-as-algorithms to serve up the most accurate and relevant content not only today, but every day in the future. Licensed content we can show within our product, and all else we can simply link to, but the search engine needs to be a complete index.

Rather than (always, only) writing stories, we should be trying to solve the problem of comprehensively informing the user on a particular topic. Web search is great, and we certainly need top-level “index everything” systems, but I’m thinking of more narrowly focussed projects. Choose a topic and start with traditional reporting, content creation, in-house explainers and multimedia stories. Then integrate a story-specific search engine that gathers together absolutely everything else that can be gathered on that topic, and applies whatever niche filtering, social curation, visualization, interaction and communication techniques are most appropriate. We can shape the algorithms to suit the subject. To really pull this off, such editorially-driven search engines need to be both live in the sense of automatically incorporating new material from external feeds, and comprehensive in the sense of being an interface to as much information on the topic as possible. Comprehensiveness will keep users coming back to your product and not someone else’s, and the idea of covering 100% of a story is itself powerful.

Other people’s content is content too
The brutal economics of online publishing dictate that we meet the needs of our users with as little paid staff time as possible. That drives the production process toward algorithms and outsourced content. This might mean indexing and linking to other people’s work, syndication deals that let a news site run content created by other people, or a blog network that bright people like to contribute to. It’s very hard for the culture of professional journalism to accept this idea, the idea that they should leverage other people’s work as far as they possibly can for as cheap as they can possibly get it, because many journalists and publishers feel burned by aggregation. But aggregation is incredibly useful, while the feelings and job descriptions of newsroom personnel are irrelevant to the consumer. As Sun Microsystems founder Bill Joy put it, “no matter who you are, most of the smartest people work for someone else,” and the idea that a single newsroom can produce the world’s best content on every topic is a damaging myth. That’s the fundamental value proposition of aggregation — all of the best stuff in one place. The word “best” represents editorial judgement in the classic sense, still a key part of a news organization’s brand, and that judgement can be embodied in whatever algorithms and social software are designed to do the aggregation. I realize that there are economic issues around getting paid for producing content, but that’s the sort of thing that needs to be solved by better content marketplaces, not lawsuits and walled gardens.

None of this means that reporters shouldn’t produce regular stories on their beats, or that there aren’t plenty of topics which require lots of original reporting and original content. But asking who did the reporting or made the content misses the point. A really good news application/interactive story/editorial search engine should be able to teach us as much as we care to learn about the topic, regardless of the state of our previous knowledge, and no matter who originally created the most relevant material.

What I am suggesting comes down to this: maybe a digital news product isn’t a collection of stories, but a system for learning about the world. For that to happen, news applications are going to need to do a lot of algorithmically-enhanced organization of content originally created by other people. This idea is antithetical to current newsroom culture and the traditional structure of the journalism industry. But it also points the way to more useful digital news products: more integration of outside sources, better search and personalization, and story-specific news applications that embody whatever combination of original content, human curation, and editorial algorithms will best help the user to learn.

[Updated 27 March with more material on social signals in search, Bill Joy’s maxim, and other good bits.]
[Updated 1 April with section titles.]

Investigating thousands (or millions) of documents by visualizing clusters

This is a recording of my talk at the NICAR (National Institute of Computer-Assisted Reporting) conference last week, where I discuss some of our recent work at the AP with the Iraq and Afghanistan war logs.

References cited in the talk:

  • “A full-text visualization of the Iraq war logs”, a detailed writeup of the technique used to generate the first set of maps presented in the talk.
  • The Glimmer high-performance, parallel multi-dimensional scaling algorithm, which is the software I presented in the live demo portion. It will be the basis of our clustering work going forward. (We are also working on other large-scale visualizations which may be more appropriate for e.g. email dumps.)
  • “Quantitative Discovery from Qualitative Information: A General-Purpose Document Clustering Methodology.” Justin Grimmer, Gary King, 2009. A paper that everyone working in document clustering needs to read. It clearly makes the point that there is no “best” clustering, just different algorithms that correspond to different pre-conceived frames on the story — and gives a method to compare clusterings (though I don’t think it will scale well to millions of docs.)
  • Wikipedia pages for bag of words model, tf-idf, and cosine similarity, the basic text processing techniques we’re using.
  • Gephi, a free graph visualization system, which we used for the one-month Iraq map. It will work up to a few tens of thousands of nodes.
  • Knight News Challenge application for “Overview,” the open-source system we’d like to build for doing this and other kinds of visual explorations of large document sets. If you like our work, why not leave a comment on our proposal?

A computational journalism reading list

[Last updated: 18 April 2011 — added statistical NLP book link]

There is something extraordinarily rich in the intersection of computer science and journalism. It feels like there’s a nascent field in the making, tied to the rise of the internet. The last few years have seen calls for a new class of  “programmer journalist” and the birth of a community of hacks and hackers. Meanwhile, several schools are now offering joint degrees. But we’ll need more than competent programmers in newsrooms. What are the key problems of computational journalism? What other fields can we draw upon for ideas and theory? For that matter, what is it?

I’d like to propose a working definition of computational journalism as the application of computer science to the problems of public information, knowledge, and belief, by practitioners who see their mission as outside of both commerce and government. This includes the journalistic mainstay of “reporting” — because information not published is information not known — but my definition is intentionally much broader than that. To succeed, this young discipline will need to draw heavily from social science, computer science, public communications, cognitive psychology and other fields, as well as the traditional values and practices of the journalism profession.

“Computational journalism” has no textbooks yet. In fact the term barely is barely recognized. The phrase seems to have emerged at Georgia Tech in 2006 or 2007. Nonetheless I feel like there are already important topics and key references.

Data journalism
Data journalism is obtaining, reporting on, curating and publishing data in the public interest. The practice is often more about spreadsheets than algorithms, so I’ll suggest that not all data journalism is “computational,” in the same way that a novel written on a word processor isn’t “computational.” But data journalism is interesting and important and dovetails with computational journalism in many ways.

Visualization
Big data requires powerful exploration and storytelling tools, and increasingly that means visualization. But there’s good visualization and bad visualization, and the field has advanced tremendously since Tufte wrote The Visual Display of Quantitative Information. There is lots of good science that is too little known, and many open problems here.

  • Tamara Munzner’s chapter on visualization is the essential primer. She puts visualization on rigorous perceptual footing, and discusses all the major categories of practice. Absolutely required reading for anyone who works with pictures of data.
  • Ben Fry invented the Processing language and wrote his PhD thesis on “computational information design,” which is his powerful conception of the iterative, interactive practice of designing useful visualizations.
  • How do we make visualization statistically rigorous? How do we know we’re not just fooling ourselves when we see patterns in the pixels? This amazing paper by Wickham et. al. has some answers.
  • Is a visualization a story? Segal and Heer explore this question in “Narrative Visualization: Telling Stories with Data.”

Computational linguistics
Data is more than numbers. Given that the web is designed to be read by humans, it makes heavy use of human language. And then there are all the world’s books, and the archival recordings of millions of speeches and interviews. Computers are slowly getting better at dealing with language.

Communications technology and free speech
Code is law. Because our communications systems use software, the underlying mathematics of communication lead to staggering political consequences — including whether or not it is possible for governments to verify online identity or remove things from the internet. The key topics here are networks, cryptography, and information theory.

  • The Handbook of Applied Cryptography is a classic, and free online. But despite the title it doesn’t really explain how crypto is used in the real world, like Wikipedia does.
  • It’s important to know how the internet routes information, using TCP/IP and BGP, or at a somewhat higher level, things like the BitTorrent protocol. The technical details determine how hard it is to do things like block websites, suppress the dissemination of a file, or remove entire countries from the internet.
  • Anonymity is deeply important to online free speech, and very hard. The Tor project is the outstanding leader in anonymity-related research.
  • Information theory is stunningly useful across almost every technical discipline. Pierce’s short textbook is the classic introduction, while Tom Schneider’s Information Theory Primer seems to be the best free online reference.

Tracking the spread of information (and misinformation)
What do we know about how information spreads through society? Very little. But one nice side effect of our increasingly digital public sphere is the ability to track such things, at least in principle.

  • Memetracker was (AFAIK) the first credible demonstration of whole-web information tracking, following quoted soundbites through blogs and mainstream news sites and everything in between. Zach Seward has cogent reflections on their findings.
  • The Truthy Project aims for automated detection of astro-turfing on Twitter. They specialize in covert political messaging, or as I like to call it, computational propaganda.
  • We badly need tools to help us determine the source of any given online “fact.” There are many existing techniques that could be applied to the problem, as I discussed in a previous post.
  • If we had information provenance tools that worked across a spectrum of media outlets and feed types (web, social media, etc.) it would be much cheaper to do the sort of information ecosystem studies that Pew and others occasionally undertake. This would lead to a much better understanding of who does original reporting.

Filtering and recommendation
With vastly more information than ever before available to us, attention becomes the scarcest resource. Algorithms are an essential tool in filtering the flood of information that reaches each person. (Social media networks also act as filters.)

  • The paper on preference networks by Turyen et. al. is probably as good an introduction as anything to the state of the art in recommendation engines, those algorithms that tell you what articles you might like to read or what movies you might like to watch.
  • Before Google News there was Columbia News Blaster, which incorporated a number of interesting algorithms such as multi-lingual article clustering, automatic summarization, and more as described in this paper by McKeown et. al.
  • Anyone playing with clustering algorithms needs to have a deep appreciation of the ugly duckling theorem, which says that there is no categorization without preconceptions. King and Grimmer explore this with their technique for visualizing the space of clusterings.
  • Any digital journalism product which involves the audience to any degree — that should be all digital journalism products — is a piece of social software, well defined by Clay Shirky in his classic essay, “A Group Is Its Own Worst Enemy.” It’s also a “collective knowledge system” as articulated by Chris Dixon.

Measuring public knowledge
If journalism is about “informing the public” then we must consider what happens to stories after publication — this is the “last mile” problem in journalism. There is almost none of this happening in professional journalism today, aside from basic traffic analytics. The key question here is, how does journalism change ideas and action? Can we apply computers to help answer this question empirically?

  • World Public Opinion’s recent survey of misinformation among American voters solves this problem in the classic way, by doing a randomly sampled opinion poll. I discuss their bleak results here.
  • Blogosphere maps and other kinds of visualizations can help us understand the public information ecosystem, such as this interactive visualization of Iranian blogs. I have previously suggested using such maps as a navigation tool that might broaden our information horizons.
  • UN Global Pulse is a serious attempt to create a real-time global monitoring system to detect humanitarian threats in crisis situations. They plan to do this by mining the “data exhaust” of entire societies — social media postings, online records, news reports, and whatever else they can get their hands on. Sounds like key technology for journalism.
  • Vox Civitas is an ambitious social media mining tool designed for journalists. Computational linguistics, visualization, and more.

Research agenda
I know of only one work which proposes a research agenda for computational journalism.

This paper presents a broad vision and is really a must-read. However, it deals almost exclusively with reporting, that is, finding new knowledge and making it public. I’d like to suggest that the following unsolved problems are also important:

  • Tracing the source of any particular “fact” found online, and generally tracking the spread and mutation of information.
  • Cheap metrics for the state of the public information ecosystem. How accurate is the web? How accurate is a particular source?
  • Techniques for mapping public knowledge. What is it that people actually know and believe? How polarized is a population? What is under-reported? What is well reported but poorly appreciated?
  • Information routing and timing: how can we route each story to the set of people who might be most concerned about it, or best in a position to act, at the moment when it will be most relevant to them?

This sort of attention to the health of the public information ecosystem as a whole, beyond just the traditional surfacing of new stories, seems essential to the project of making journalism work.