reading: Readings in Machine Translation

Post before the end of the year!

I'm really enjoying Readings in Machine Translation -- it's got all of these great MT papers from past decades, going from the Warren Weaver memo from 1949 to the Brown et al. paper where they make stat-mt fashionable again in the early 90s. Apparently, a lot of the papers in the volume were somewhat hard to find online in 2003.

Really interesting: so far, the early papers have had some very detailed descriptions of the low-level particulars. "Well, we're going to need this many memory drums...", "oh, and the words will be stored in memory in alphabetical order" (which seems very archaic), and a fixation on picking the right word in the target language, in sort of a word-sense disambiguation sense (which is slightly fashionable again!).

So for people into MT who want a sense of history, these are papers that it seems like one should read -- I mean, Sergei Nirenburg and friends picked them out, so they've got to be good, right?

If you haven't read the 1949 Warren Weaver memo, though -- even if you're not an NLP person -- do yourself a favor and go ahead and read it!

securing your MoinMoin wiki

I really like MoinMoin; it's straightforward, it's Pythonic, it's got WikiWords. Great! I've been keeping a bunch of notes on one.

The IU Computational Linguistics group had an aging MoinMoin too, but most of the edits and new accounts were spam. I suspect the edits were being done by humans too, because they were pretty good at fitting into the markup of existing pages.

We replaced it with a new install (here it is) and make it a bit more secure. We disallowed anonymous edits, and made it so that you can't just arbitrarily create accounts, which took a code change. I added the "if not request.user.isSuperUser() ..." block (suggested here) to MoinMoin/action/newaccount.py. The rest of the changes described on that page aren't necessary -- just make it refuse new account requests.

Then it occurred to me: spammers have probably been trying to spam my personal wiki too! I checked: there were about a hundred spam accounts; a new one every day or two. My Moin was allowing arbitrary account creation, but they were all useless because only I could edit pages!

So in the interest of discouraging future webcrap, let me issue a warning: CyrilleVincent, SabaFaulkner, CasinoBonus, life insurance quotes, and ChickyBowen -- I'm coming for you. And you'd best sleep with one eye open, paydayloansuk214. If that's your real name.

reading: Religious Literacy by Stephen Prothero

Not too long ago, I read the very thought-provoking Religious Literacy: What Every American Needs to Know and Doesn't, by Stephen Prothero, likely because it came recommended by Dale McGowan. The discussion of the history of religious education in the US is fantastic.

The main argument of the book is that, until recent decades, we knew quite a lot about Protestantism,  through instruction at home, in churches, and even in the school system. People apparently used the verb "catechize", as something you did to children. But this knowledge of what's actually in the Bible, and actual church doctrines is, these days, largely lost on us in the US, even though we're very caught up in our religious identities and church attendance is huge.

Political discourse is full of religious allusions, but we often don't get the references. I would have appreciated more examples of problems that this causes in practice -- is it really an issue for being a citizen in a democracy?

Prothero notes, early on, that Europeans are much less religious than Americans but more familiar with religious content. Dale McGowan says "faith is most easily sustained in ignorance"; knowing a thing or three about a few different religions makes it easier to not get caught up in any of them. Is the major problem caused by religious ignorance that it makes it easier for preachers and politicians to jerk people around by telling them that God says thus-and-such?

While Prothero doesn't address the question of why the more broadly-educated Europeans don't tend to be churchgoers, he does put forth a policy suggestion, that our curricula should have more information about the world's religions. And while I agree that it's probably a good idea, he doesn't say much about the sorts of changes we might see, with better religious studies education. I must admit, I have a hard time thinking of education-about-religion as anything except a strategy against the influence of seemingly-devout people.

Perhaps I'll pick up his more recent book, God Is Not One, about the fundamental disagreements between different religions, contrasting with the framing you'd get from Huston Smith or Karen Armstrong, who argue that different religions are grasping towards the same fundamental truth. I'm really curious about his personal position, because Prothero identifies himself as an Episcopalian, but hasn't thus far talked about any particular benefits of people believing any particular thing.

cross-lingual word sense disambiguation

Have I mentioned what I've been working on recently? Maybe I haven't.

In general, I'm working on cross-lingual word- and phrase-sense disambiguation. WSD/PSD is the problem of deciding, for a given word or phrase, which meaning was intended, for some pre-defined sets of meanings. You might get the possible senses out of a dictionary, where they're nicely enumerated, or perhaps from WordNet. The stock example is "bank" -- is it the side of a river, or is it a building where they do financial services? Or, is it the abstract financial institution?

There's a brilliant bit from the prescient Warren Weaver, from 1955 (via):

If one examines the words in a book, one at a time as through an opaque mask with a hole in it one word wide, then it is obviously impossible to determine, one at a time, the meaning of the words . . . . But if one lengthens the slit in the opaque mask, until one can see not only the central word in question but also say N words on either side, then if N is large enough one can unambiguously decide the meaning of the central word . . . . The practical question is: "What minimum value of N will, at least in a tolerable fraction of cases, lead to the correct choice of meaning for the central word?

The "cross-lingual" kind of WSD means that we care about exactly the distinctions that cause you to pick a different word in a given target language, typically because the CLWSD system is meant to be integrated into an MT system; that's becoming fashionable (Carpuat and Wu, 2007). So in this setting, say if you're translating "bank" from English into Spanish, your system doesn't have to decide if it's the building or the institution that owns it -- it's still "banco". Now a riverbank is an "orilla".

In the general case, your system might end up learning how to make distinctions that you as a human didn't know you had to make -- for example, I'm given to understand that Japanese doesn't have just one word for "brother", but "older brother" and "younger brother", which are different enough concepts that they get totally separate words.

Making these choices is typically treated as a classification problem: you get some features for a bunch of instances of usage of a source word, and do supervised learning to get a classifier with (hopefully) good accuracy on the problem of predicting whether this is a "banco" usage or an "orilla" usage. The features are typically things like "which words are in the surrounding context?", or perhaps something fancier based on a parse of the sentence or knowledge about the document as a whole -- whatever you think will be predictive of what the target-language word should be. Hopefully your learning algorithm has some good way of filtering out irrelevant features.

And then, once that's all put together, hopefully you have some extra signal to feed into your translation system, and it makes better word choices, and everybody's happy.

And that's cross-lingual word/phrase-sense disambiguation!

Mr. Verb on metaphor

Oh, also: there was this great Mr. Verb post, where they talk about government funding to do research on metaphor. Link to an article in The Atlantic: Why Are Spy Researchers Building a 'Metaphor Program'?

Here's the job posting, which sounds awesome, except that it will probably ultimately lead to people getting exploded:

The Metaphor Program will exploit the fact that metaphors are pervasive in everyday talk and reveal the underlying beliefs and worldviews of members of a culture. In the first phase of the two-phase program, performers will develop automated tools and techniques for recognizing, defining and categorizing linguistic metaphors associated with target concepts and found in large amounts of native-language text. The resulting conceptual metaphors will be validated using empirical social science methods. In the second phase, the program will characterize differing cultural perspectives associated with case studies of the types of interest to the Intelligence Community. Performers will apply the methodology established in the first phase and will identify the conceptual metaphors used by the various protagonists, organizing and structuring them to reveal the contrastive stances.

Metaphors We Live By

I just finished reading Lakoff and Johnson's Metaphors We Live By. The most central claim of the book is that metaphor isn't just about language, but that it's a core part of cognition, and that a great deal of our experience is shaped by the metaphors that we use to understand the world.

The authors talk a few times about how we understand love: is it more like a journey, or more like a collaborative work of art? Is a marriage like a partnership, or like a haven from the outside world? One's way of thinking about either of these things probably has a direct effect on experience and behavior.

I'm left with some questions, though, and maybe you've got insights that you'd like to share. In the later chapters, the authors go on to say that understanding metaphor this way requires rethinking a lot of philosophy, particularly what it means to say that a sentence is true. I don't know enough about philosophy-of-language to say whether it's actually been shaken to its core (philosophy is a physical structure), but fairly common sentences like "she's a cool drink of water" (attractive person is refreshing), "he threw in the towel" (this situation is a boxing match and he gave up) or even "I see what you mean" (knowing/understanding is seeing) seem pretty hard to represent in propositional logic.

I think I buy the argument that establishing their truth or falsity is possible only if you understand the metaphor that the speaker is using -- but it's a very typical case, that the speaker is using some metaphor or another. It's nearly impossible to speak or reason without using metaphors (similarity is physical closeness), so maybe in general a good theory of meaning really does have to take this into account.

Is there a point at which a metaphor becomes an honestly frozen form, though? What happens if you learn an expression without knowing its etymology, and just say it because it's that kind of situation? You learn a classifier for "throw in the towel" situations, and say it appropriately, but maybe you don't know anything about boxing. I'll start watching for this sort of situation: maybe it happens a lot.

Also, do Lakoff and Johnson essentially agree or disagree with Doug Hofstadter, who says that analogy is the core of cognition? (the video of Doug's talk is totally worth watching).

I'll probably read some of the more recent work on this: More than Cool Reason (metaphor in poetry), and Moral Politics (metaphor in reasoning about politics) both sound really interesting, possibly even relevant to the real world.

Python tip: 3.2 adds a memoization decorator

This is a really handy thing to have around, in case you've got expensive computations that you want done more than once. I've written a clumsier version of this myself, once or twice -- but it makes sense to have in the standard library!

http://docs.python.org/py3k/library/functools.html#functools.lru_cache

There's some other neat stuff in the functools package too: currying and reducing. Pretty slick!

CMU's Avenue Project

Before there was HLTDI, there was Carnegie Mellon's Avenue Project, which seems to have had basically the same goal as us -- produce good machine translation systems for under-resourced languages, especially those spoken by under-resourced indigenous people.

Avenue itself doesn't seem to have been under-resourced, though -- they sent people to South America (Chile, Peru, Bolivia...) to collect training data, and seemed to have a lot of contacts with local educators and language experts. They got quite a few papers out of this line of research, and apparently wrote a lot of good software. They had a much deeper pool of money (and arguably talent) than we do.

And now... the website is dormant, the PhD students involved seem to have graduated, the data and software are not publicly available, and the researchers seem to have moved on to other things. (one of the resulting doctors is the illimitable Kathrin Probst, who hipped me to Avenue when we were both at Google Atlanta, although I didn't really grasp how serious it was at the time -- darn her for being so humble!)

They were pretty gracious in giving us the Quechua data that they collected (and said we could redistribute it), and I've been reading a bunch of their papers, but I'm left some sadness about the whole enterprise -- they surely already went through a lot of the problems that HLTDI is going to have to address. Why can't we just check out and fork their code?

... maybe I should ask for their software too. Science is supposed to be easily replicable, isn't it?

did you know: F-Spot can export to picasa, flickr, etc!

If you're using Picasa Web Albums, and you're on Linux (maybe a small audience for this post), you don't have to manually upload your files on the web interface or use the Picasa client. The Picasa client is OK, I guess, but it's actually a thinly veiled Windows app bundled with a version of Wine. It's not all that snappy, and it doesn't feel like the rest of the apps on the Linux desktop, and it's not open source.

But! F-Spot is native, and well-integrated with Gnome, and Free Software... and it can very easily upload your photos to picasaweb, or other online photo-hosting places. That's pretty rad.

Scripts, Plans, Goals and Understanding

So, years ago, Kurt Eiselt, who was then a professor at Georgia Tech, did an independent study with me on NLP/NLU. It was pretty great, though mostly because it got me excited about the field. We wrote really simple parsers for really tiny subsets of English, in Common Lisp, and talked about how language might work cognitively. He had me get a copy of Schank and Abelson's classic Scripts, Plans, Goals and Understanding. I read some of it, but for the most part, it's just sat on my shelf for the better part of a decade.

Out of a sense of "geez, this is something I should have read, I'm an NLP researcher now", I've been working through it. It's kind of slow going: I find Schank kind of light on the details and heavy on the intuition, fairly vague. Maybe the later chapters have more detail about the story understanding system they purportedly built in the 70s.

But if there's been a chunk that's worth reading, it's this, at the very end of chapter 5.

John couldn't get a taxi, so he rode his horse downtown and into a restaurant. He beat up another customer and took a menu from him. He decided to have a steak. The waiter came along and John offered him a bottle of scotch if he listened to John tell him what he wanted to eat. John went to the kitchen and told the cook to give him a steak, because the cook could always deduct the gift from his income tax. When the cook refused, John offered to give him guitar lessons, and that worked. While John was eating the steak, the waiter came back and stole $10 from John's wallet. Then John got on his horse and rode out.

Python tip: dir() with no arguments

If you do Python, you probably knew that you can ask an object what it has on it by saying dir(theobject).

I just found this out, though: you can also say dir() with no arguments to find out everything that's in your current namespace. No more losing temporary variables that you forgot about earlier in the repl. It works for finding out which modules you have loaded, too. Holy cow.

>>> dir()
['__builtins__', '__doc__', '__name__', '__package__']
>>> x = "this string is amazing"
>>> dir()
['__builtins__', '__doc__', '__name__', '__package__', 'x']

Machine of Death Remix Corpus!

Machine of Death is fantastic. And it turns out that most of the stories in Machine of Death allow derivative works...

And when I hear "derivative works", I think the word "remix". But what's a remix of a text? Cut-up poetry? n-gram models to generate ominous-sounding emails to your family? Counting how many times the word "of" appears? The original text, but with randomized capitalization? This is for you to decide. Well, it is now. Here's all the derivs-allowed stories, conveniently pulled out and converted to ASCII.

Machine of Death Remix Corpus.

Use this to train your poetry bot, or your story understanding system, or your unexpected other thing! Share and enjoy :)

empiricism, faith, computational linguistics

Mike sent me a fantastic piece by Ted Pederson, calling for NLP/CL researchers to care more about having reproducible results and maintainable software.

Empiricism Is Not a Matter of Faith.

It's sad that this is a problem; it should be easy to get other researchers' software up and running, reproduce the results reported in papers, and plug things into other things -- but I think we're moving in that direction. At least one CL conference, CICLING, explicitly calls for open software and reproducible results. Which is pretty cool.

keeping notes on a wiki

Stepping firmly into 2005, I decided to keep notes and links and miscellanea on my own personal wiki! I'm trying to be better about remembering things, and also about keeping my stuff in a somewhat public, web-accessible place.

And I really like MoinMoin! It was easy to set up on our Ubuntu server. Up until now, I've mostly just kept a bunch of text files in version control, organized chronologically. But a wiki can be nonlinear, which is cool. Maybe sooner or later, my labmates will wiki with me on a shared wiki space...

How do you manage notes to yourself? Do you use a physical notebook? Some other kind of software? (where does it run?)