28 November 2016

Workshops and mini-conferences

I've attended and organized two types of workshops in my time, one of which I'll call the ACL-style workshop (or "mini-conference"), the other of which I'll call the NIPS-style workshop (or "actual workshop"). Of course this is a continuum, and some workshops at NIPS are ACL-style and vice versa. As I've already given away with phrasing, I much prefer the NIPS style. Since many NLPers may never have been to NIPS or to a NIPS workshop, I'm going to try to describe the differences and explain my preference (and also highlight some difficulties).

(Note: when I say ACL-style workshop, I'm not thinking of things like WMT that have effectively evolved into full-on co-located conferences.)

To me, the key difference is whether the workshop is structured around invited talks, panels, discussion (NIPS-style) or around contributed, reviewed submissions (ACL-style).

For example, Paul Mineiro, Amanda Stent, Jason Weston and I are organizing a workshop at NIPS this year on ML for dialogue systems, Let's Discuss. We have seven amazing invited speakers: Marco Baroni, Antoine Bordes, Nina Dethlefs, Raquel Fernández, Milica Gasic, Helen Hastie and Jason Williams. If you look at our schedule, we have allocated 280 minutes to invited talks, 60 minutes to panel discussion, and 80 minutes to contributed papers. This is a quintessential NIPS-style workshop.

For contrast, a standard ACL-style workshop might have one or two invited talks with the majority of the time spent on contributed (submitted/reviewed) papers.

The difference in structure between NIPS-style and ACL-style workshops has some consequences:

  • Reviewing in the NIPS-style tends to be very light, often without a PC, and often just by the organizers.
  • NIPS-style contributed workshop papers tend to be shorter.
  • NIPS-style workshop papers are almost always non-archival.
My personal experience is that NIPS-style workshops are a lot more fun. Contributed papers at ACL-style workshops are often those that might not cut it at the main conference. (Note: this isn't always the case, but it's common. It's also less often the case at workshops that represent topics that are not well represented at the main conference.) On the other hand, when you have seven invited speakers who are all experts in their field and were chosen by hand to represent a diversity of ideas, you get a much more intellectually invigorating experience.

(Side note: my experience is that many many NIPS attendees only attend workshops and skip the main conferences; I've rarely heard of this happening at ACL. Yes, I could go get the statistics, but they'd be incomparable anyway because of time of year.)

There are a few other effects that matter.

The first is the archival-ness of ACL workshops which have proceedings that appear in the anthology:
(This is from EACL, but it's the same rules across the board.) I personally believe it's absurd that workshop papers are considered archival but papers on arxiv are not. By forcing workshop papers to be archival, you run the significant risk of guaranteeing that many submissions are things that authors have given up on getting into the main conference, which can lead to a weaker program.

A second issue has to do with reviewing. Unfortunately as of about three years ago, the ACL organizing committee almost guaranteed that ACL workshops have to be ACL-style and not NIPS-style (personally I believe this is massive bureaucratic overreaching and micromanaging):
By forcing a program committee and reviewing, we're largely locked into the ACL-style workshop. Of course, some workshops ignore this and do more of a NIPS-style anyway, but IMO this should never have been a rule.

One tricky issue with NIPS-style workshops is that, as I understand it, some students (and perhaps faculty/researchers) might be unable to secure travel funding to present at a non-archival workshop. I honestly have little idea how widespread this factor is, but if it's a big deal (e.g., perhaps in certain parts of the world) then it needs to be factored in as a cost.

A second concern I have about NIPS-style workshops is making sure that they're inclusive. A significant failure mode is that of "I'll just invite my friends." In order to prevent this outcome, the workshop organizers have to make sure that they work hard to find invited speakers who are not necessarily in their narrow social networks. Having a broader set of workshop organizers can help. I think that when NIPS-style workshops are proposed, they should be required to list potential invited speakers (even if these people have not yet been contacted) and a significant part of the review process should be to make sure that these lists represent a diversity of ideas and a diversity of backgrounds. In the best case, this can lead to a more inclusive program than ACL-style workshops (where basically you get whatever you get as submissions) but in the worst case it can be pretty horrible. There are lots of examples of pretty horrible at NIPS in the past few years.

At any rate, these aren't easy choices, but my preference is strongly for the NIPS-style workshop. At the very least, I don't think that ACL should predetermine which type is allowed at its conferences.

08 November 2016

Bias in ML, and Teaching AI

Yesterday I gave a super duper high level 12 minutes presentation about some issues of bias in AI. I should emphasize (if it's not clear) that this is something I am not an expert in; most of what I know is by reading great papers by other people (there is a completely non-academic sample at the end of this post). This blog post is a variant of that presentation.

Structure: most of the images below are prompts for talking points, which are generally written below the corresponding image. I think I managed to link all the images to the original source (let me know if I missed one!).

Automated Decision Making is Part of Our Lives

To me, AI is largely the study of automated decision making, and the investment therein has been growing at a dramatic rate.

I'm currently teaching undergraduate artificial intelligence. The last time I taught this class was in 2012. The amount that's changed since there is incredible. Automated decision making is now a part of basically everyone's life, and will only be more so over time. The investment is in the billions of dollars per year.

Things Can Go Really Badly

If you've been paying attention to headlines even just over the past year, the number of high stakes settings in which automated decisions are being made is growing, and growing into areas that dramatically affect real people's real life, their well being, their safety, and their rights.

This includes:
This is obviously just a sample of some of the higher profile work in this area, and while all of this is work in progress, even if there's no impact today (hard to believe for me) it's hard to imagine that this isn't going to be a major societal issue in the very near future.

Three (out of many) Source of Bias

For the remainder, I want to focus on three specific ways that bias creeps in. The first I'll talk about more because we understand it more, and it's closely related to work that I've done over the past ten years or so, albeit in a different setting. These three are:
  1. data collection
  2. objective function
  3. feedback loops

Sample Selection Bias

The standard way that machine learning works is to take some samples from a population you care about, run it through a machine learning algorithm, to produce a predictor.

The magic of statistics is that if you then take new samples from that same population, then, with high probability, the predictor will do a good job. This is true for basically all models of machine learning.

The problem that arises is when your population samples are from a subpopulation (or different population) for those on which you're going to apply your predictor.

Both of my parents work in marketing research and have spent a lot of their respective careers doing focuses groups and surveys. A few years ago, my dad had a project working for a European company that made skin care products. They wanted to break into the US market, and hired him to conduct studies of what the US population is looking for in skin care. He told them that he would need to conduct four or five different studies to do this, which they gawked at. They wanted one study, perhaps in the midwest (Cleveland or Chicago). The problem is that skin care needs are very different in the southwest (moisturizer matters) and the northwest (not so much), versus the northeast and southeast. Doing one study in Chicago and hoping it would generalize to Arizona and Georgia is unrealistic.

This problem is often known as sample selection bias in the statistics community. It also has other names, like covariate shift and domain adaptation depending on who you talk to.

One of the most influential pieces of work in this area is the 1979 Econometrica paper by James Heckman, for which he won the 2000 Nobel Prize in economics. He's pretty happy about that! If you haven't read this paper, you should: it's only about 7 pages long, it's not that difficult, and you won't believe the footnote in the last section. (Sorry for the clickbait, but you really should read the paper.)
There's been a ton of work in machine learning land over the past twenty years, much of which builds on Heckman's original work. To highlight one specific paper: Corinna Cortes is the Head of Google Research New York and has had a number of excellent papers on this topic over the past ten years. One in particular is her 2013 paper in Theoretical Computer Science (with Mohri) which provides an amazingly in depth overview and new algorithms. Also a highly recommended read.

It's Not Just that Error Rate Goes Up

When you move from one sample space (like the southwest) to another (like the northeast), you should first expect error rates to go up.

Because I wanted to run some experiments for this talk, here are some simple adaptation numbers for predicting sentiment on Amazon reviews (data due to Mark Dredze and colleagues). Here we have four domains (books, DVDs, electronics and kitchen appliances) which you should think as standins for the different regions of the US, or different demographic qualifiers.

The figure shows error rates when you train on one domain (columns) and test on another (rows). The error rates are normalized so that we have ones on the diagonal (actual error rates are about 10%). The off-diagonal shows how much additional error you suffer due to sample selection bias. In particular, if you're making predictions about kitchen appliances and don't train on kitchen appliances, your error rate can be more than two times what it would have been.

But that's not all.

These data sets are balanced: 50% positive, 50% negative. If you train on electronics and make predictions on other domains, however, you get different false positive/false negative rates. This shows the number of test items predicted positively; you should expect it to be 50%, which basically is what happens in electronics and DVDs. However, if you predict on books, you underpredict positives; while if you predict on kitchen, you overpredict positives.

So not only do the error rates go up, but the way they are exhibited chances, too. This is closely related to issues of disparate impact, which have been studied recently by many people, for instance by Feldman, Friedler, Moeller, Scheidegger and Venkatasubramanian.

What Are We Optimizing For

One thing I've been trying to get undergrads in my AI class to think about is what are we optimizing for, and whether the thing that's being optimized for is what is best for us.

One of the first things you learn in a data structures class is how to do graph search, using simple techniques like breadth first search. In intro AI, you often learn more complex things like A* search. A standard motivating example is how to find routes on a map, like the planning shown above for me to drive from home to work (which I never do because I don't have a car and it's slower than metro anyway!).

We spend a lot of time proving optimality of algorithms in terms of shortest path costs, for fixed costs that have been given to us by who-knows-where. I challenged my AI class to come up with features that one might use to construct these costs. They started with relatively obvious things: length of that segment of road, wait time at lights, average speed along that road, whether the road is one-way, etc. After more pushing, they came up with other ideas, like how much gas mileage one gets on that road (either to save the environment or to save money), whether the area is “dangerous” (which itself is fraught with bias), what is the quality of the road (ill-repaired, new, etc.).

You can tell that my students are all quite well behaved. I then asked them to be evil. Suppose you were an evil company, how might you come up with path costs. Then you get things like: maybe businesses have paid me to route more customers past their stores. Maybe if you're driving the brand of car that my company owns or has invested it, I route you along better (or worse) roads. Maybe I route you so as to avoid billboards from competitors.

The point is: we don't know, and there's no strong reason to a priori assume that what the underlying system is optimizing for is my personal best interest. (I should note that I'm definitely not saying that Google or any other company is doing any of these things: just that we should not assume that they're not.)

A more nuanced example is that of a dating application for, e.g., multi-colored robots. You can think of the color as representing any sort of demographic information you like: political leaning (as suggested by the red/blue choice here), sexual orientation, gender, race, religion, etc. For simplicity, let's assume there are way more blue robots than others, and let's assume that robots are at least somewhat homophilous: they tend to associate with other similar robots.

If my objective function is something like “maximize number of swipe rights,” then I'm going to want to disproportionately show blue robots because, on average, this is going to increase my objective function. This is especially true when I'm predicting complex behaviors like robot attraction and love, and I don't have nearly enough features to do anywhere near a perfect matching. Because red robots, and robots of other colors, are more rare in my data, my bottom line is not affected greatly by whether I do a good job making predictions for them or not.

I highly recommend reading Version Control, a recent novel by Dexter Palmer. I especially recommend it if you have, or will, teach AI. It's fantastic.
There is an interesting vignette that Palmer describes (don't worry, no plot spoilers) in which a couple engineers build a dating service, like Crrrcuit, but for people. In this thought exercise, the system's objective function is to help people find true love, and they are wildly successful. They get investors. The investors realize that when their product succeeds, they lose business. This leads to a more nuanced objective in which you want to match most people (to maintain trust), but not perfectly (to maintain clientèle). But then, to make money, the company starts selling its data to advertisers. And different individuals' data may be more valuable: in particular, advertisers might be willing to pay a lot for data from members of underrepresented groups. This provides incentive to actively do a worse job than usual on such clients. In the book, this thought exercise proceeds by human reasoning, but it's pretty easy to see that if one set up, say, a reinforcement learning algorithm for predicting matches that had long term company profit as its objective function, it could learn something similar and we'd have no idea that that's what the system was doing.

Feedback Loops

Ravi Shroff recently visited the CLIP lab and talked about his work (with Justin Rao and Shared Goel) related to stop and frisk policies in New York. The setup here is that the “stop and frisk” rule (in 2011, over 685k people were stopped; this has subsequently been declared unconstitutional in New York) gave police officers the right to stop people with much lower thresholds than probable cause, to try to find contraband weapons or drugs. Shroff and colleagues focused on weapons.

They considered the following model: a police officer sees someone behaving strangely, and decide that they want to stop and frisk that person. Before doing so, they enter a few values into their computer, and the computer either gives a thumbs up (go ahead and stop) or a thumbs down (let them live their life). One question was: can we cut down on the number of stops (good for individuals) while still finding most contraband weapons (good for society)?

In this figure, we can see that if the system thumbs downed 90% of stops (and therefore only 10% of people that police would have stopped get stopped), they are still able to recover about 50% of the weapons. With stopping only about 1/3 of individuals, they are able to recover 75% of weapons. This is a massive reduction in privacy violations while still successfully keeping the majority of weapons off the streets.

(Side note: you might worry about sample selection bias here, because the models are trained on people that the policy did actually stop. Shroff and colleagues get around this by the assumption I stated before: the model is only run on people who policy have already decided are suspicious and would have stopped and frisked anyway.)

The question is: what happens if and when such a system is deployed in practice?

The issue is that policy officers, like humans in general, are not stationary entities. Their behavior changes over time, and it's reasonable to assume that their behavior would change when they get this new system. They might feed more people into the system (in “hopes” of thumbs up) or feed fewer people into the system (having learned that the system is going to thumbs down them anyway). This is similar to how the sorts of queries people issue against web search engines change over time, partially because we learn to use the systems more effectively, and learn what to not consider asking a search engine to do for us because we know it will fail.

Now, once we've (hypothetically) deployed this system, it's collecting its own data, which is going to be fundamentally different from the data is was originally trained one. It can continually adapt, but we need good technology for doing this that takes into account the human behavior of the officers.

Wrap Up and Discussion

There are many things that I didn't touch on above that I think are nonetheless really important. Some examples:
  1. All the example “failure” cases I showed above have to do with race or (binary) gender. There are other things to consider, like sexual orientation, religion, political views, disabilities, family and child status, first language, etc. I tried and failed to find examples of such things, and would appreciate pointers. For instance, I can easily imagine that speech recognition error rates skyrocket when working for users with speech impairments, or with non-standard accents, or who speak a dialect of English that not the “status quo academic English.” I can also imagine that visual tracking of people might fail badly on people with motor impairments or who use a wheelchair.
  2. I am particularly concerned about less “visible” issues because we might not even know. The standard example here is: could a social media platform sway an election by reminding people who (it believes) belong to a particular political party to vote? How would we even know?
  3. We need to start thinking about qualifying our research better with respect to the populations we expect it to work on. When we pick a problem to work on, who is being served? When we pick a dataset to work on, who is being left out? A silly example is the curation of older datasets for object detection in computer vision, which (I understand) decided on which objects to focus on by asking five year old relatives of the researchers constructing the datasets to name all the objects they could see. As a result of socio-economic status (among other things), mouse means the thing that attaches to your computer, not the cute furry animal. More generally, when we say we've “solved” task X, does this really mean task X or does this mean task X for some specific population that we haven't even thought to identify (i.e., “people like me” aka the white guys problem)? And does “getting more data” really solve the problem---is more data always good data?
  4. I'm at least as concerned with machine-in-the-loop decision making as fully automated decision making. Just because a human makes the final decision doesn't mean that the system cannot bias that human. For complex decisions, a system (think even just web search!) has to provide you with information that helps you decide, but what guarantees do we have that that information isn't going to be biased, either unintentionally or even intentionally. (I've also heard that, e.g., in predicting recidivism, machine-in-the-loop predictions are worse than fully automated decisions, presumably because of some human bias we don't understand.)
If you've read this far, I hope you've found some things to think about. If you want more to read, here are some people whose work I like, who tweet about these topics, and for whom you can citation chase to find other cool work. It's a highly biased list.
  • Joanna Bryson (@j2bryson), who has been doing great work in ethics/AI for a long time and whose work on bias in language has given me tons of food for thought.
  • Kate Crawford (@katecrawford) studies the intersection between society and data, and has written excellent pieces on fairness.
  • Nick Diakopoulos (@ndiakopoulos), a colleague here at UMD, studies computational journalism and algorithmic transparency.
  • Sorelle Friedler (@kdphd), a former PhD student here at UMD!, has done some of the initial work on learning without disparate impact.
  • Suresh Venkatasubramanian (@geomblog) has co-authored many of the papers with Friedler, including work on lower bounds and impossibility results for fairness.
  • Hanna Wallach (@hannawallach) is the first name I think of for machine learning and computational social science, and has recently been working in the area of fairness.
I'll also point to less biased sources. The Fairness, Accountability and Transparency in Machine Learning workshop takes place in New York City in a week and a half; check out the speakers and papers there. I also highly recommend the very long reading list on Critical Algorithm Studies, which covers more than just machine learning.

02 November 2016

ACL 2017 PC Chairs Blog

In case you missed it, Regina Barzilay and Min-Yen Kan, who are program chairs for ACL 2017, are running an open blog describing what they're doing:


Check it out!

24 August 2016

Debugging machine learning

I've been thinking, mostly in the context of teaching, about how to specifically teach debugging of machine learning. Personally I find it very helpful to break things down in terms of the usual error terms: Bayes error (how much error is there in the best possible classifier), approximation error (how much do you pay for restricting to some hypothesis class), estimation error (how much do you pay because you only have finite samples), optimization error (how much do you pay because you didn't find a global optimum to your optimization problem). I've generally found that trying to isolate errors to one of these pieces, and then debugging that piece in particular (eg., pick a better optimizer versus pick a better hypothesis class) has been useful.

For instance, my general debugging strategy involves steps like the following:

  1. First, ensure that your optimizer isn't the problem. You can do this by adding "cheating" features -- a feature that correlates perfectly with the label. Make sure you can successfully overfit the training data. If not, this is probably either an optimizer problem or a too-small-sample problem.
  2. Remove all the features except the cheating feature and make sure you can overfit then. Assuming that works, add feature back in incrementally (usually at an exponential rate). If at some point, things stop working, then probably you have too many features or too little data.
  3. Remove the cheating features and make your hypothesis class much bigger; e.g., by adding lots of quadratic features. Make sure you can overfit. If you can't overfit, maybe you need a better hypothesis class.
  4. Cut the amount of training data in half. We usually see test accuracy asymptote as the training data size increases, so if cutting the training data in half has a huge effect, you're not yet asymptoted and you might do better to get some more data.
The problem is that this normal breakdown of error terms comes from theory land, and, well, sometimes theory misses out on some stuff because of a particular abstraction that has been taken. Typically this abstraction has to do with the fact that the overall goal has already been broken down into an iid/PAC style learning problem, and so you end up unable to see some types of error because the abstraction hides them.

In an effort to try to understand this better, I tried to make a flow chart of sorts that encompasses all the various types of error I could think of that can sneak into a machine learning system. This is shown below:
I've tried to give some reasonable names to the steps (the left part of the box) and then give a grounded example in the context of ad placement (because it's easy to think about). I'll walk through the steps (1-11) and try to say something about what sort of error can arise at that step.
  1. In the first step, we take our real world goal of increasing revenue for our company and decide to solve it by improving our ad displays. This immediately upper bounds how much increased revenue we can hope for because, well, maybe ads are the wrong thing to target. Maybe I would do better by building a better product. This is sort of a "business" decision, but it's perhaps the most important question you can ask: am I even going after the right things?
  2. Once you have a real world mechanism (better ad placement) you need to turn it into a learning problem (or not). In this case, we've decided that the way we're going to do this is by trying to predict clickthrough, and then use those predictions to place better ads. Is clickthrough a good thing to use to predict increased revenue? This itself is an active research area. But once you decide that you're going to predict clickthrough, you suffer some loss because of a mismatch between that prediction task and the goal of better ad placement.
  3. Now you have to collect some data. You might do this by logging interactions with a currently deployed system. This introduces all sorts of biases because the data you're collecting is not from the final system you want to deploy (the one you're building now), and you will pay for this in terms of distribution drift.
  4. You cannot possibly log everything that the current system is doing, so you have to only log a subset of things. Perhaps you log queries, ads, and clicks. This now hides any information that you didn't log, for instance time of day or day of week might be relevant, user information might be relevant, etc. Again, this upper bounds your best possible revenue.
  5. You then usually pick a data representation, for instance quadratic terms between a bag of words on the query side and a bag of words on the ad side, paired with a +/- on whether the user clicked or not. We're now getting into the position where we can start using theory words, but this is basically limited the best possible Bayes error. If you included more information, or represented it better, you might be able to get a lower Bayes error.
  6. You also have to choose a hypothesis class. I might choose decision trees. This is where my approximation error comes from.
  7. We have to pick some training data. The real world is basically never i.i.d., so any data we select is going to have some bias. It might not be identically distributed with the test data (because things change month to month, for instance). It might not be independent (because things don't change much second to second). You will pay for this.
  8. You now train your model on this data, probably tuning hyperparameters too. This is your usual estimation error.
  9. We now pick some test data on which to measure performance. Of course, this test data is only going to be representative of how well your system will do in the future if this data is so representative. In practice, it won't be, typically at least because of concept drift over time.
  10. After we make predictions on this test data, we have to choose some method for evaluating success. We might use accuracy, f measure, area under the ROC curve, etc. The degree to which these measures correlate with what we really care about (ad revenue) is going to affect how well we're able to capture the overall task. If the measure anti-correlates, for instance, we'll head downhill rather than uphill.
(Minor note: although I put these in a specific order, that's not a prescriptive order, and many can be swapped. Also, of course there are lots of cycles and dependencies here as one continues to improve systems.)

Some of these things are active research areas. Things like sample selection bias/domain adaptation/covariate shift have to do with mismatch of train/test data. For instance, if I can overfit train but generalization is horrible, I'll often randomly shuffle train/test into a new split and see if generalization is better. If it is, there's probably an adaptation problem.

When people develop new evaluation metrics (like Bleu for machine translation), they try to look at things like #10 (correlation with some goal, perhaps not exactly the end goal). And standard theory and debugging (per above) covers some of this too.

I'm very curious if y'all have topics/tricks that you like that aren't mentioned here.

Related reading:

16 August 2016

Feature (or architecture) ablation

I wrote my first (and only) coreference paper back in 2005. At the time, my goals were to (a) do well on coref, (b) integrate background knowledge (like "Bush" is "president") using simple techniques, and (c) try to figure out how important different (types of) features were for making coreference decisions.

For the last, there is a reasonably extensive feature-type ablation experiment using backward selection (which I trust far more than forward selection). After writing the paper, I had many internal dialogues about why experiments like that are interesting. I have had, over the years, a couple of answers:

  1. The obvious answer is "it tells us something interesting about language." It would be nice if this were true, but I'm not totally sure it is, and it's definitely not true if one doesn't put a bunch more effort into it than I put into that 2005 paper. What can we say? Yeah, spelling is important. Knowledge is important. Syntax is hard to actualize. I don't know that we didn't already know these things before.
  2. Engineering. Suppose someone wanted to build a similar system. They want to put their effort where it's most valuable, and so feature ablation experiments tell you where you're likely to get the most bang for the buck. In a sense, you can see these as a type of negative result. Which features actually aren't that important. In the 2005 paper, you could remove syntactic, semantic, and class-based features with zero performance degradation; and also get rid of pattern-based features with minor performance degradation. This saves a lot of effort because some of these are actually quite a pain to implement and/or are slow and/or require lots of external resources.
Today, I mostly lean toward the engineering answer, or at least that's what I want to use as a jumping off point here.

Now that we're partially allergic to feature engineering and prefer to replace it with architecture engineering, I think the charge is stronger, not weaker, to do ablation experiments. Does that thing really need to be a biLSTM? Would an RNN suffice? What about just averaged bag of word embeddings? Do you need two layers of attention there or would one suffice? Do you need attention at all? Does that layer really need to be that wide?

These are all easy questions to ablate and answer.

There's never going to be a crisp answer like "yes, if I cut my hidden state from 493 units to 492 units performance goes down the drain." Many things will be gradual, but not all.

Why do I think this is important? Precisely for reason #2 above, but about a bajillion times more so. Training these really complicated models with wide hidden units, bidirectional stuff, etc., is really slow. Really really slow. If you tell me I can be within 1% accuracy but can train 100 times faster, I'm going to do it. Sure, for a final test run I might crank everything up again (and then report that!) but for development, it's super useful to have a system you can train and evaluate efficiently.

Does this tell us anything interesting about language? Almost certainly not (or at least not without a huge amount of extra work). But it does make everyone's life better.

14 August 2016

Some papers I liked at ACL 2016

A conference just ended, so it's that time of year! Here are some papers I liked with the usual caveats about recall.

Before I go to the list, let me say that I really really enjoyed ACL this year. I was completely on the fence about going, and basically decided to go only because of giving a talk at Repl4NLP, and wanted to attend the business meeting for the discussion of diversity in the ACL community, led by Joakim Nivre with an amazing report that he, Lyn Walker, Yejin Choi and Min-Yen Kan put together. (Likely I'll post more, separately, about both of these; for the latter, I tried to transcribe much of Joakim's presentation.)

All in all, I'm supremely glad I decided to go: it was probably my favorite conference in recent memory. This was not just because there were lots of great papers (there were!) but also because somehow it felt more like a large community conference than others I've attended recently. I'm not sure what made it like this, but I noticed it felt a lot less clique-y than NAACL, a lot more broad and interesting than ICML/NIPS (though that's probably because of my personal taste in research) and in general a lot friendlier. I don't know what the organizers did that managed this great combination, but it was great!

Okay, so on to the papers, sorted by ACL id.

P16-1009: Rico Sennrich; Barry Haddow; Alexandra Birch
Improving Neural Machine Translation Models with Monolingual Data

I like this paper because it has a nice solution to a problem I spent a year thinking about on-and-off and never came up with. The problem is: suppose that you're training a discriminative MT system (they're doing neural; that's essentially irrelevant). You usually have far more monolingual data than parallel data, which typically gets thrown away in neural systems because we have no idea how to incorporate it (other than as a feature, but that's blech). What they do here is, assuming you have translation systems in both directions, back translate your monolingual target-side data, and then use that faux-parallel-data to train your MT system on. Obvious question is: how much of the improvement in performance is due to language modeling versus due to some weird kind of reverse-self-training, but regardless the answer, this is a really cool (if somewhat computationally expensive) answer to a question that's been around for at least five years. Oh and it also works really well.

P16-1018: E.Dario Gutierrez; Ekaterina Shutova; Tyler Marghetis; Benjamin Bergen
Literal and Metaphorical Senses in Compositional Distributional Semantic Models
I didn't see this paper presented, but it was suggested to me at Monday's poster session. Suppose we're trying to learn representations of adjective/noun pairs, by modeling nouns as vectors and adjectives as matrices, evaluating on unseen pairs only. (Personally I don't love this style, but that's incidental to the main ideas in this paper.) This paper adjusts the adjective matrices depending on whether they're being used literally ("sweet candy") or metaphorically ("sweet dreams"). But then you can go further and posit that there's another matrix that can transform literal metaphors into metaphorical metaphors automatically, essentially implementing the Lakoff-style notion that there is great consistency in how metaphors are created.

P16-1030 [dataset]: Hannah Rashkin; Sameer Singh; Yejin Choi
Connotation Frames: A Data-Driven Investigation
This paper should win some sort of award for thoroughness. The idea is that in many frames ("The walrus pummelled the sea squirt") there is implied connotation/polarity/etc. on not only the agent (walrus) and theme (sea squirt) of the frame but also tells us something about the relationship between the writer/speaker and the agent/theme (the writer might be closer to the sea squirt in this example, versus s/pummelled/fought/). The connotation frame for pummelled collects all this information. This paper also describes an approach to prediction of these complex frames using nice structured models. Totally want to try this stuff on our old plotunits data, where we had a hard time getting even a much simpler type of representation (patient polarity verbs) to work!

P16-1152: Artem Sokolov; Julia Kreutzer; Christopher Lo; Stefan Riezler
Learning Structured Predictors from Bandit Feedback for Interactive NLP
This was perhaps my favorite paper of the conference because it's trying to do something new and hard and takes a nice approach. At a high level, suppose you're Facebook and you're trying to improve your translation system so you ask users to give 1 star to 5 star ratings. How can you use this to do better translation? This is basically the (structured) contextual bandit feedback learning problem. This paper approaches this from a dueling bandits perspective where I show you two translations and ask which is better. (Some of the authors had an earlier MT-Summit paper on the non-dueling problem which I imagine many people didn't see, but you should read it anyway.) The technical approach is basically probabilitic latent-variable models, optimized with gradient descent, with promising results. (I also like this because I've been thinking about similar structured bandit problems recently too :P.)

P16-1231: Daniel Andor; Chris Alberti; David Weiss; Aliaksei Severyn; Alessandro Presta; Kuzman Ganchev; Slav Petrov; Michael Collins
Globally Normalized Transition-Based Neural Networks
[EDIT 14 Aug 2:40p: I misunderstood from the talk and therefore the following is basically inaccurate. I'm leaving this description and paper here on the list because Yoav's comment will make no sense otherwise, but please understand that it's wrong and, I hate to say this, it does make the paper less exciting to me. The part that's wrong is struck-out below.] There's a theme in the past two years of basically repeating all the structured prediction stuff we did ten years ago on our new neural network technology. This paper is about using Collins & Roark-style incremental perceptron for transition-based dependency parsing on top of neural networks. The idea is that label-bias is perhaps still a problem for neural network dependency parsers, like their linear grandparents. Why do I like this? Because I think a lot of neural nets people would argue that this shouldn't be necessary: the network can do arbitrarily far lookahead into the future and therefore should be able to learn to avoid the label-bias problem. This paper shows that current techniques don't achieve that: there's a consistent win to be had by doing global normalization.

P16-2013: Marina Fomicheva; Lucia Specia
Reference Bias in Monolingual Machine Translation Evaluation
This paper shows pretty definitively that human evaluations against a reference translation are super biased toward the particular reference used (probably because evaluators are lazy and are basically doing ngram matching anyway -- a story I heard from MSR friends a while back). The paper also shows that this gets worse over time, presumably as evaluators get tireder.
P16-2096: Dirk Hovy; Shannon L. Spruit
The Social Impact of Natural Language Processing
This is a nice paper summarizing four issues that come up in ethics that also come up in NLP. I mostly liked this paper because it gave names to things I've thought about off and on, but didn't have a name for. In particular, they consider exclusion (hey my ASR system doesn't work on people with an accent, I guess they don't get a voice), overgeneralization (to what degree are our models effectively stereotyping more than they should), over- and under-exposure (hey lets all work on parsing because that's what everyone else is working on, which then makes parsing seem more important...just to pick a random example :P), and dual-use (I made something for good but XYZ organization used it for evil!). This is a position/discussion-starting paper, and I thought quite engaging.

Feel free to post papers you like in the comments!


05 August 2016

Fast & easy baseline text categorization with vw

About a month ago, the paper Bag of Tricks for Efficient Text Categorization was posted to arxiv. I found it thanks to Yoav Goldberg's rather incisive tweet:

Yoav is basically referring to the fact that the paper is all about (a) hashing features and (b) bigrams and (c) a projection that doesn't totally make sense to me, which (a) vw does by default (b) requires "--ngrams 2" and (c) I don't totally understand I don't think is necessary. (See this tutorial for more on how to do NLP in VW.)

At the time, I said if they gave me the data, I'd run vw on it and report results. They were nice enough to share the data but I never got around to running it. The code for their technique ("fastText") was just released, which goaded me into finally doing something.

So my goal here was to try to tell, without tuning any parameters, how competitive a baseline vw is to the results from fastText with minimal effort.

Here are the results:

ag news1
ag news23s92.55s92.3
amazon full1
amazon full233s60.269s56.6
amazon polarity1
amazon polarity252s94.668s94.2
sogou news1
sogou news236s96.830s96.9
yahoo answers1
yahoo answers227s72.348s71.0
yelp full1
yelp full218s63.937s60.0
yelp polarity1
yelp polarity215s95.720s95.5

(Average accuracy for fastText is 83.2; for vw is 82.2.)

In terms of accuracy, the two are roughly on par. vw occasionally wins; when it does, it's usually by 0.1% to 0.5%. fastText wins a bit more often, and on one dataset it wins significantly (yelp full: winning by 3%-4%) and on one a bit less (yahoo answers, up by about 1.3%). But the numbers are pretty much in line, and could almost certainly be brought up for vw with a wee bit of hyperparameter tuning (namely the learning rate, which is tuned in fastText).

In terms of training time, fastText is maybe 30% faster on average, though these are such small datasets (eg 500k examples) that a difference of 52s versus 68s is not too significant. I also noticed that for most of the datasets, simply writing the model to disk for vw took a nontrivial amount of time. But wait, there's more. That 30% faster for fastText was run on 20 cores in parallel whereas the vw run did not use parallelized learning (vw runs two threads, one for I/O and one for learning).
That said, a major caveat on comparing the training times. They're run on different machines. I don't know what type of machine the fastText results were achieved on, but it was a parallel 20-core run. The vw experiments were run on a single core, one pass over the data, on a 3.1Ghz Core i5-2400. Yes, I could have hogwild-ed vw and gotten it faster but it really didn't seem worth it for datasets this small. And yes, I could've rerun fastText on my machine, but... what can I say? I'm lazy.

What did I do to get these vw numbers? Here's the entire training script:
% cat run.sh 
for ngram in 1 2 ; do
  cat $d/train.csv | ./csv2vw.pl | \
    time vowpal_wabbit/vowpalwabbit/vw --oaa `cat $d/classes.txt | wc -l` \
                                  -b25 --ngram $ngram -f $d/model.$ngram
  cat $d/test.csv  | ./csv2vw.pl | \
    time vowpal_wabbit/vowpalwabbit/vw -t -i $d/model.$ngram
Basically the only flags to vw are (1) telling it to do multiclass classification with one-against-all, (2) telling it to use 25 bits (not tuned), and telling it to either use unigrams or bigrams. [Comparison note: this means vw is using 33m hash bins; fastText used 10m for unigram models and 100m for bigram models.]

The only(*) data munging that occurs is in csv2vw.pl, which is a lightweight script for converting the data, lowercasing, and doing very minor tokenization:
% cat csv2vw.pl
#!/usr/bin/perl -w
use strict;
while (<>) {
    if (/^"*([0-9]+)"*,"(.+)"*$/) {
        print $1 . ' | ';
        $_ = lc($2);
        s/","/ /g;
        s/([^a-z0-9 -\\]+)/ $1 /g;
        print $_ . "\n";
    } else { 
        die "malformed line '$_'";
There are two exceptions where I did slightly more data munging. The datasets released for dbpedia and Soguo were not properly shuffled, which makes online learning hard. I preprocessed the training data by randomly shuffling it. This took 2.4s for dbpedia and 12s for Soguo.

[[[EDIT 2:20p 5 Aug 2016: Out of curiosity, I upped the number of bits that vw uses for the experiments to 27 (so that it's on par with the 100m used by fastText). This makes it take about 5 seconds longer to run (writing the model to disk is slower). Performance stays the same on: ag news, amazon polarity, dbpedia, sogou, and yelp polarity; and it goes up from from 53.6/56.6 to 55.0/58.8 on amazon full, from 70.6/71.0 to 71.1/71.6 on yahoo answers, from 56.9/60.0 to 58.5/61.6 on yelp full. This puts the vw average with more bits at 82.6, which is 0.6% behind the fastText average.]]]

Long story short... am I switching from vw to fastText? Probably not any time soon.

29 July 2016

A quick comment on structured input vs structured output learning

When I think of structured input models, I typically think of things like kernels over discrete input spaces. For instance, the famous all-substrings kernel for which K(d1,d2) effectively counts the number of common substrings in two documents, without spending exponential time enumerating them all. Of course there are many more ways of thinking about structured inputs: tree-to-string machine translation has a tree structured input. RNNs (on the input side) are essentially structured input models for sequence structures.

When I think of structured output models, I typically think of things like CRFs, structured SVMs/M3Ns, multilabel predictors (those are borderline), various transition-based methods (eg., shift/reduce parsers), etc. Here, my internal model for the structure is essentially at prediction time: find a high scoring structure from this complicated discrete output space.

Perhaps this has been obvious to everyone-but-me for a decade, but I only recently came to the recognition that these are essentially the same, at least if you restrict the sort of models you're willing to consider. (In particular, if you ignore things like imitation learning/learning to search for a moment.)

In a pure structured input setting, you have some simple label space Y (let's assume it's the real numbers) and some complex input space X. Typically you want to learn a function f : X ➝ Y, which has low loss. In particular you want to minimize the expectation of loss(y, f(x)) over random draws of x,y. And the "interesting" thing is that x isn't just a vector, so you have to be clever.

In the pure structure output setting, in, for instance, the structured SVM/CRF setup, you have some input space X (which may or may not be structured) and some complex output space Y. As before, you want to learn a function f : X ➝ Y, which has low loss. However, in the most common setups, the way you accomplish this is that instead of directly learning f, you instead learn a scoring function s that scores x,y pairs based on how "good" that y is for the corresponding x. For a fixed scoring function s, you derive f according to the argmax rule: fs(x) := argmaxy s(x,y). In this way, you have effectively separated the learning problem (get a good s) from the structured problem (solve the argmax). [Whether this is good or not is up to debate; I'm personally on the "nay" side.] You then want to minimize something like the expectation of loss(y, argmaxy' s(x,y')) over random draws x,y.

The observation is that these two problems are essentially the same thing. That is, if you know how to do the structured input problem, then the structured output problem is essentially the same thing, as far as the learning problem goes. That is, if you can put structure in f(x) for structured input, you can just as well put structure in s(x,y) for structured output. Or, by example, if you can predict the fluency of an English sentence x as a structured input problem, you can predict the translation quality of a French/English sentence pair x,y in a structured output problem. This doesn't solve the argmax problem -- you have to do that separately -- but the underlying learning problem is essentially identical.

You see similar ideas being reborn these days with papers like David Belanger's ICML paper this year on energy networks. With this framework of think-of-structured-input-and-structured-output-as-the-same, basically what they're doing is building a structured score function that uses both the input and output simultaneously, and throwing these through a deep network. (Ok it's a bit more than that, but that's the cartoon.)

At any rate, maybe obvious to everyone but me, but I thought I'd write it down anyway :).

26 July 2016

Decoding (neural?) representations

I remember back in grad school days some subset of the field was thinking about the following question. I train an unsupervised HMM on some language data to get something-like-part-of-speech tags out. And naturally the question arises: these tags that come out... what are they actually encoding?

At the time, there were essentially three ways of approaching this question that I knew about:

  1. Do a head-to-head comparison, in which you build an offline matching between induced tags and "true" tags, and then evaluate the accuracy of that matching. This was the standard evaluation strategy for unsupervised POS tagging, but is really just trying to get at the question of: how correlated are the induced tags with what we hope comes out.
  2. Take a system that expects true POS tags and give it induced POS tags instead (at both training and test time). See how much it suffers (if at all). Joshua Goodman told me a few times (though I can't find his paper on this) that word clusters were just as good as POS tags if your task was NER.
  3. Do something like #2, but also give the system both POS tags and induced tags, and see if the POS tags give you anything above and beyond the induced tags.
Now, ten years later since we're in "everything old is new again" phase, we're going through the same exercises, but with word embeddings instead of induced tags. This makes things slightly more complicated because it means that we have to have mechanisms that deal with continuous representations rather than discrete representations, but you basically see the same ideas floating around.

In fact, of the above approaches, the only one that requires any modification is #1 because there's not an obvious way to do the matching. The alternative is to let a classifier do the matching, rather than an offline process. In particular, you take your embeddings, and then try to train a classifier that predicts POS tags from the embeddings directly. (Note: I claim this generalizes #1 because if you did this with discrete tags, the classifier would simply learn to do the matching that we used to compute "by hand" offline.) If your classifier can do a good job, then you're happy.

This approach naturally has flaws (all do), but I think it's worth thinking about this seriously. To do so, we have to take a step back and ask ourselves: what are we trying to do? Typically, it seems we want to make an argument that a system that was not (obviously) designed to encode some phenomenon (like POS tags) and was not trained (specifically) to predict that phenomenon has nonetheless managed to infer that structure. (We then typically go on to say something like "who needs no POS tags" even though we just demonstrated our belief that they're meaningful by evaluating them... but okay.)

As a first observation, there is an entire field of study dedicated to answering questions like this: (psycho)linguists. Admittedly they only answer questions like this in humans and not in machines, but if you've ever posed to yourself the question "do humans encode/represented phrase structures in their brains" and don't know the answer (or if you've never thought about this question!) then you should go talk to some linguists. More classical linguists would answer these questions with tests like, for instance, constituency tests or scoping tests. I like Colin Phillips' encyclopedia article on syntax for a gentle introduction (and is what I start with for syntax in intro NLP).

So, as a starting point for "has my system learned X" we might ask our linguist friends how they determine if a human has learned X. Some techniques are difficult to replicate in machine (e.g., eye movement experiments, though of course models that have something akin to alignment---or "attention" if you must---could be thought of as having something like eye movements, though I would be hesitant to take this analogy too far). But many are not, for instance behavioral experiments, analyzing errors, and, I hesitate somewhat to say it, grammaticality judgements.

My second comment has to do with the notion of "can these encodings be used to predict POS tags." Suppose the answer is "yes." What does that mean? Suppose the answer is "no."

In order to interpret the answer to these questions, we have to get a bit more formal. We're going to train a classifier to do something like "predict POS given embedding." Okay, so what hypothesis space does that classifier have access to? Perhaps you say it gets a linear hypothesis space, in which case I ask: if it fails, why is that useful? It just means that POS cannot be decoded linearly from this encoding. Perhaps you make the hypothesis space outrageously complicated, in which case I ask: if it succeeds, what does that tell us?

The reason I ask these questions is because I think it's useful to think about two extreme cases.
  1. We know that we can embed 200k words in about 300 dimensions with nearly orthogonal vectors. This means that for all intents and purposes, if we wanted, we could consider ourselves to be working with a one-hot word representation. We know that, to some degree, POS tags are predictable from words, especially if we allow for complex hypothesis spaces. But this is uninteresting because by any reasonable account, this representation has not encoded anything interesting: it's just the output classifier that's doing something interesting. That is to say: if your test can do well on the raw words as input, then it's dubious as a test.
  2. We also know that some things are just unpredictable. Suppose I had a representation that perfectly encoded everything I could possibly want. But then in the "last layer" it got run through some encryption protocol. All of the information is still there, so the representation in some sense "contains" the POS tags, but no classifier is going to be able to extract it. That is to say, just because the encoded isn't on the "surface" doesn't mean it's not there. Now, one could reasonably argue something like "well if the information is there in an impossible-to-decode format then it might as well not be there" but this slope gets slippery very quickly.
Currently, I much prefer to think about essentially the equivalent of "behavioral" experiments. For instance, if you're machine translating and want to know if your system can handle scoping, then give it a minimal pair to translate that only differs in the scoping of some negation. Or if you're interesting in knowing whether it knows about POS tags, perhaps look at errors in its output and see if they fall along POS categories.

EDIT 26 Jul 2016, 8:24p Eastern: It's unclear to a few people so clarification. I'm mostly not talking about type-level word embeddings above, but embeddings in context. At a type-level, you could imagine evaluating (1) on out of vocabular terms, which would be totally reasonable. I'm think more something like: the state of your biLSTM in a neural MT system. The issue is that if, for instance, this biLSTM can repredict the input (as in an autoencoder), then it could be that the POS tagger is doing all the work. See this conversation thread with Yoav Goldberg for some discussion.

12 July 2016

Some picks from NAACL 2016

Usual caveats: didn't see all talks, didn't read all papers, there's lot of good stuff at NAACL that isn't listed here! That said, here are some papers I particularly liked at NAACL, with some comments. Please add comments with papers you liked!

Automatic Summarization of Student Course Feedback by Luo, Liu, Liu and Litman.

Anyone who has taught has suffered the following dilemma. You ask students for feedback throughout the course, and you have to provide free text because if you could anticipate their problems, you'd have addressed them already. But now you have hundreds of responses that you can't quickly read through. This paper provides a dataset of such responses, together with summaries that you actually have time to read through. The approach is a reasonably standard ILP formulation, with some additional machinery to deal with the fact that the data is very sparse. The idea is to essentially induce a "bigram similarity" as part of the summarization problem. I like this paper because the problem is great (I think NLP should really push in the direction of helping learners learn and teachers teach!), the dataset is nice (if a bit small) and the approach makes sense. And they actually do a human evaluation, even for a short paper! Hooray!

A Latent Variable Recurrent Neural Network for Discourse Relation Language Models by Ji, Haffari and Eisenstein.

This paper combines the nice rich hypothesis classes you get with the "anonymous" latent variables in neural networks with the modeling power that one gets from the ability to marginalize "structured" latent variables from classic graphical models land. This is applied to document language modeling, in which the latent variables are discourse relations (in the PDTB sense). The model works well both for language modeling and for predicting implicit discourse relations and dialogue acts. (And, as one should expect from a paper with Eisenstein as an author, there are statistical significance tests!)

Structured Prediction with Output Embeddings for Semantic Image Annotation by Quattoni, Ramisa, Madhyastha, Simo-Serra and Moreno-Noguer.

If you want to label images with complex structured outputs, like play(dog,grass), you quickly run in to sparsity problems on the output. The proposal here is to decompose the outputs into embeddings (kind of like Vivek's work and Jags' work) and learning a bilinear model of inputs/outputs in that space. In general, I think there's a lot to be done in interesting modeling of structured output spaces, and this paper gives a new set of techniques in this space.

Deconstructing Complex Search Tasks: A Bayesian Nonparametric Approach for Extracting Sub-tasks by Mehrotra, Bhattacharya and Yilmaz.

The problem considered here is the following: if I want to go to ACL, I need to register, book a flight, book a hotel, find some attractions to go to while skipping sessions, look up all the good vegan restaurants in Berlin (hah!), etc. My overall task is going to ACL, but there are a number of highly related but different subtasks. The challenge is to infer these subtasks from search logs, so that you can provide better search support. The model is a Chinese Restaurant Process with word embeddings used to measuring similarities. And look, another short paper with a human evaluation!

A Corpus and Cloze Evaluation for Deeper Understanding of Commonsense Stories by Mostafazadeh, Chambers, He, Parikh, Batra, Vanderwende, Kohli and Allen.

This paper introduces the task of story cloze: given a story prefix, predict which of two endings is the "correct" (natural) one. Eg: "Jim got his first credit card in college. He didn't have a job so he bought everything on his card. After he graduated he amounted a $10,000 debt. Jim realized that he was foolish to spend so much money." Then you have to decide on the right ending: "Jim decided to device a plan for repayment." (correct) versus "Jim decided to open another credit card." (incorrect). The data set was quite well constructed (lots of checks), and a large collection of baseline models are run for comparison. The data is available. I like this paper for the task and the very very careful data set collection. Since this was establishing baselines, I would really liked to have seen error bars on the results so we can have more reasonable future comparisons. I'm really tempted to try to annotate some of these with plot units, but that was really really painful the first time around; but I feel like that's a good explanatory theory for a lot of the examples shown in the paper.

Learning Global Features for Coreference Resolution by Wiseman, Rush and Shieber.

The basic idea here is to take a "local" coreference model, that makes decisions about assigning mentions to entities, and augment it with a "global" model that models the entire cluster of mentions. In past work, cluster-level features have been hard to define (e.g., what are the right features to extract over sets of arbitrary size?). What's the solution? Essentially, embed the clusters. I like this paper because I wanted to try to do something like this and Wiseman did it better than I would have. I think there's also an interesting interpretation of this work in terms of things like neural Turing machines/memory networks, in which we now have technology for learning to update memory (where "memory" here is the "clusters"). My only gripe is that, like most papers in the new wave of coreference, the older ML approaches have been somewhat forgotten (except for Vincent Ng who is unforgettable); I'm thinking in particular about the work on ACE, like that of Luo, Ittycheriah, Jing, Kambhatla and Roukos on Bell-tree coreference, which I don't think gets enough notice these days for introducing a lot of the ideas that make up modern coref.

Visual Storytelling by Huang, Ferraro, Mostafazadeh, Misra, Agrawal, Devlin, Girshick, He, Kohli, Batra, Zitnick, Parikh, Vanderwende, Galley and Mitchell.

The main observation in this paper is that how you caption images in a sequence is different from how you caption them in isolation. For instance, if you have a temporal sequence of images from, say, a wedding, the captions for them should tell a story. This paper primarily introduces a dataset and baseline approaches for addressing the problem in this dataset. I like this paper, even if you remove the image stuff, because it emphasizes the fact that stories have a different structure than just sequences of sentences. The image stuff gives a grounding that's interesting beyond that. One problem pointed out in the paper is that automatic metrics aren't great here, which is problematic, but not particularly surprising.

05 July 2016

Rating the quality of reviews, after the fact

Groan groan groan reviewers are horrible people. Not you and me. Those other reviewers over there!

tldr: In general we actually don't think our reviews are that bad, though of course it's easy to remember the bad ones. Author perception of review quality is colored by, but not determined by, the overall accept/reject decision and/or the overall score that review gave to the paper.

NIPS did an experiment a bunch of years ago (can't dig it up any more, now it's an urban legend) where they asked reviewers at, I think, the time of author feedback, to rate the reviews. The anecdotal story was that there was almost perfect correlation between "this is a good review" and "this review gave my paper a high score." Of course this is not super surprising, even if you get rid of "emotions," because presumably I like my paper and so any review that doesn't like it is flawed.

For NAACL 2013, we did a similar experiment, but we asked authors for their responses several months after the fact (actually, even after the conference had taken place), at which point hopefully emotions had cooled a bit and they could look back at their reviews with a sort of fond recollection. We presented the contact author of each paper with the original review text for each of their reviews, but did not show them the original scores. We asked them on a standard Likert scale how helpful this review was, and how informative this review was.

Because this was long after the fact, response rate was of course not 100%, and it was also biased toward authors of papers that were accepted. We got responses from 128 authors on a total of 138 papers (some papers had same contact authors), covering a total of about 397 reviews (roughly one per paper, but some short papers only had two, and some papers had four).

All the plots below are restricted to this set of 138 papers, not to the full set of about 500.

First, let's get a sense of the data. Here are the overall results for this entire set of 138 papers:

(Note that the numbers add up to 397, not 138, because this is counting per-review not per-paper.) The first row shows the accept/reject ratio. Since NAACL 2013 had an acceptance rate between 25% and 30%, obviously survey results are biased toward accepted papers, but we still have a healthy response rate from rejected papers.

Overall, the vast majority (~80%) of reviews were considered both informative and helpful (score 4 or 5) according to the authors. So yes, we need to do something about the 20% of reviews that got a 1, 2 or 3 on the Likert scale, but we're actually not doing that horribly. (Modulo sample selection bias.) The papers themselves were considered overwhelmingly appropriate and clear. The overall score distribution matches (roughly) the overall score distribution for the entire conference.

Let's look at what happens if we look only at accepted or rejected paper:

Comparing these, we definitely see a bit of the NIPS effect. For accepted papers, the reviews were considered overwhelmingly informative and helpful (scores of 4 or 5 in 85% or more cases). However, for rejected papers, the reviews were still considered largely informative and helpful (~73% of cases were 4s and 5s). Not surprisingly, accepted papers fare quite well on the individual score metrics, in particular overall score (duh!).

We can alternatively condition the analysis on the overall paper score rather than the final accept/reject decision. Here's how that looks:

That's not substantially different.

So what makes the difference between good (informativeness 4 or 5) and bad (informativeness 1 or 2) reviews?

On average, the "good" reviews were about 15% longer than the bad reviews (on average 320 characters versus 280 characters).

Somewhat surprisingly, a linear classifier on bag of words data and distinguish with 90% accuracy "good" from "bad" reviews, but the features it gives high weight to are basically features that look like positive versus negative reviews, rather essentially exploiting of the correlation between informativeness and acceptance, rather than informativeness on its own.

30 June 2016

Subgradient descent, or something....

Last semester from grad ML, I totally revamped stuff and started using the awesome book Understanding Machine Learning by Shai Ben-David and Shai Shalev-Shwartz (I'll try to review this in a later post). Naturally, one of the things we talked about was subgradients and subgradient descent.

I imagine that for many of you, if I asked you to define a subgradient of a convex function f at x informally (let's stick in one dimension for now), you would say something like "it's any line that that makes contact with f at x and everywhere else lies below f." This is the definition given in UML, and the definition we always see in pictures, like the one in ciml:

Okay, so this is all great, and one of the fun things you can do is derive algorithms like (stochastic) subgradient descent, which involve picking any subgradient and then using that as if it were a gradient in a standard gradient descent procedure. Yeah, you run into some speed limits for optimization rates, but it basically works.

So then on the midterm I asked a question that was intended to be a freebie: give me the subderivatives of the ramp loss function. Ramp loss is like hinge loss (shown above), but where the negative part gets clamped at 1. Formally, it's ramp(x) = min(1, hinge(x)), where hinge(x) = max(0, 1-x).

Turns out this wasn't a freebie. The problem, obviously in retrospect, is that ramp is not convex, and therefore doesn't have subderivatives! Or at least it doesn't have subderivatives for any x less than zero, and it's (only) subderivative for x≥0 is zero.

Several students point this out, and several students just went through the motions and computed what-we-might-normally-call subderivatives anyway. And by "what-we-might-normally-call" I of course mean what I had originally intended, and what I would have done myself if I had been a student in this course, and also what pretty much any auto-diff toolkit would do.

And yet, it's clearly wrong according to the definition of subgradients that we all know and use everyday in our very non-convex neural networks, when we do things like relu or hardtanh units.

It turns out (not surprisingly) that subdifferentials of non-convex functions has been studied (extensively) since the 1970s. Murdukhovich and Shao give a brief history in their paper on Banach spaces. Unfortunately, I don't actually understand most of that paper (or its citations).

I did manage to find one set of slides that I could understand, though, by Adil Bagirov for a talk on Subgradient methods in nonsmooth nonconvexoptimization! Basically the idea that Adil proposes is that we can use a gradient-ified version of a quasisecant and most/some subgradient-like methods still go through and make sense with this generalized notion.

Why am I posting this? Because it caused my brain to reconfigure itself when I was forced to think about this by the very smart students in my class! Am I going to teach quasisecants in the future? Probably not. But I am going to explicitly point out that the standard definition of subgradient doesn't work for nonconvex functions (or, more specifically, it works, but you get an empty set in a lot of cases) and that there are generalizations but that I don't think we've really figured this all out (as a community).

If anyone has other pointers that I can use in the future, I'd love to see them!

24 June 2016

Language bias and black sheep

Tolga Bolukbasi and colleagues recently posted an article about bias in what is learned with word2vec, on the standard Google News crawl (h/t Jack Clark). Essentially what they found is that word embeddings reflect stereotypes regarding gender (for instance, "nurse" is closer to "she" than "he" and "hero" is the reverse) and race ("black male" is closest to "assaulted" and "white male" to "entitled"). This is not hugely surprising, and it's nice to see it confirmed. The authors additionally present a method for removing those stereotypes with no cost (as measured with analogy tasks) to accuracy of the embeddings. This also shows up on twitter embeddings related to hate speech.

There have been a handful of reactions to this work, some questioning the core motivation, essentially variants of "if there are biases in the data, they're there for a reason, and removing them is removing important information." The authors give a nice example in the paper (web search; two identical web pages about CS; one mentions "John" and the other "Mary"; query for "computer science" ranks the "John" one higher because of embeddings; appeal to a not-universally-held-belief that this is bad).

I'd like to take a step back and argue the the problem is much deeper than this. The problem is that even though we all know that strong Sapir-Whorf is false, we seem to want it to be true for computational stuff.

At a narrow level, the issue here is the question of what does a word "mean." I don't think anyone would argue that "nurse" means "female" or that "computer scientist" means "male." And yet, these word embeddings, which claim to be capturing meaning, are clearly capturing this non-meaning-effect. So then the argument becomes one of "well ok, nurse doesn't mean female, but it is correlated in the real world."

Which leads us to the "black sheep problem." We like to think that language is a reflection of underlying truth, and so if a word embedding (or whatever) is extracted from language, then it reflects some underlying truth about the world. The problem is that even in the simplest cases, this is super false.

The "black sheep problem" is that if you were to try to guess what color most sheep were by looking and language data, it would be very difficult for you to conclude that they weren't almost all black. In English, "black sheep" outnumbers "white sheep" about 25:1 (many "black sheep"s are movie references); in French it's 3:1; in German it's 12:1. Some languages get it right; in Korean it's 1:1.5 in favor of white sheep. This happens with other pairs, too; for example "white cloud" versus "red cloud." In English, red cloud wins 1.1:1 (there's a famous Sioux named "Red Cloud"); in Korean, white cloud wins 1.2:1, but four-leaf clover wins 2:1 over three-leaf clover. [Thanks to Karl Stratos and Kota Yamaguchi for helping with the multilingual examples.]

This is all to say that co-occurance frequencies of words definitely do not reflect co-occurance frequencies of things in the real world. And the fact that the correlation can both both ways means that just trying to model a "default" as something that doesn't appear won't work. (Also, computer vision doesn't really help: there are many many pictures of black sheep out there because of photographer bias.)

We observed a related phenomena when working on plot units. We were trying to extract "patient polarity verbs" (this idea has now been expanded and renamed "implicit sentiment": a much better name). The idea is that we want to know what polarity verbs inflict on their arguments. If I "feed" you, is this good or bad for you? For me? If I "punch" you, likewise. We focused on patients because action verbs are almost always good for the agent.

In order to accomplish this, we started with a seed list of "do-good-ers" and "wrong-do-ers." For instance, "the devil" was a wrong do-er, and so we can extract things that the devil does, and assume that these are (on average) bad for their patients. The problem was the "do-good-ers" don't do good, or at least they don't do good in the news. One of our do-good-ers was "firefighter". Firefighters are awesome. Even stereotyped, this is arguably a very positive social good, heroic profession. But in the news, what do firefighters do? Bad things. Is this because most firefighters do bad things in the world? Of course not. It's because news is especially poignant when stereotypically good people do bad things.

This comes up in translation too, especially when looking at looking at domain adaptation effects. For instance, our usual example for French to English translation is that in Hansards, "enceinte" transates as "room" but in EMEA (medical domain), it translates as "pregnant." What does this have to do with things like gender bias? In Canadian Hansards, "merde" translates mostly as "shit" and sometimes as "crap." In movie subtitles, it's very frequently "fuck." (I suspect translation direction is a confounder here.) This is essentially a form of intensification (or detensification, depending on direction). It is not hard to imagine similar intensifications happening between racial descriptions and racial slurs, or between gender descriptions and sexist slurs, depending on where the data came from.