Winograd Schema Challenge 2016

This wasn’t quite the Winograd Schema Challenge that I had set out on. Originally this language comprehension contest for A.I. was announced in July 2014, to be run in October 2015, but was postponed to February 2016, and then again to July 2016. I was just about to ship my program overseas, three weeks before the last-accepted arrival date of postal entries, when the contest announced changes to the rules and technical format.

Some universities had been training with ambiguous pronouns like this:

The birds ate the seeds because they were hungry.

I had been practising on the official Winograd schemas like this:

The foxes are getting in at night and attacking the chickens. I shall have to guard them.

Whereas the final test featured this:

Mark became absorbed in Blaze, the white horse. He was afraid the stable boys at the Burlington Stables struck at him and bullied him because he was timid, so he took upon himself the feeding and care of the animal.

The programs were now faced with any number of consecutively ambiguous pronouns in passages from 1940’s children’s novels, which made quite a difference. It turns out the organisers had already decided on this last year, as appears from their sensible enough explanation in a members-only AI magazine (Winograd schemas are too hard to compose). Unfortunately they somehow did not see fit to share these changes on the contest website until too late. While the benchmark of 65% had previously been feasible, it now quickly became unlikely that anyone would win anything this year. A number of would-been participants backed out.

The contest finally took place at the IJCAI conference in New York with four contestants: the Open University of Cyprus, the University of Science and Technology of China, the independent Denis Robert from France, and myself from the Netherlands. Curiously absent were a number of American universities who had previously reported successes of over 70% for solving Winograd schemas. The absence of Google, IBM, and other commercial powerhouses was less strange, if you consider that the winner was obligated to publish their methods so that others could reproduce them. And that anything below human level would be considered a failure in the media.

The glass is half full
The programs were asked to figure out 60 multiple choice pronouns, with such ambiguity that they were to be solved through an understanding of the context. With two to five potential answers per pronoun, the baseline score for guesswork was 45%. $1000 would be awarded for a 65% score, and $25000 for a 90% score, human level.
(Note: these are the scores after recount. There was some confusion as my program had omitted two answers)

Contestant Correct answers out of 60 Method
Quan Liu 35 / 35 / 29 (58% – 48%) deep neural network & ConceptNet
Nikos Isaak 29 (48%) probabilistic engine & knowledge extraction
Patrick Dhondt 29 (48%) logical axioms
Denis Robert 19 (32%) logical inferences

Quan Liu’s group entered three programs, which is a little unorthodox for contests. But if you see this as a scientific test then it makes sense to test which configuration of a neural network works best. Their machine learning approach gathered pairs of events (mainly verbs) that are commonly associated, e.g. “rob -> be arrested”, and then applied their probability of co-occurring. Two of their versions scored the highest, 58%, which is consistent with the track record of similar approaches.

There were nevertheless no winners that reached the 65% threshold. On the one hand one could say that technology is literally halfway human ability, on the other hand the programs did only a little better than one might by chance. Any conclusion drawn from just the scores is premature. If this test is to be a meaningful measure of progress, we should look at which areas the programs were better or worse in. To this I can at least answer about my own program.

Winograd schemas vs prose
The ambiguity in the new prose form was actually not so bad compared to previously published Winograd schemas. But the phrasing was often excessively long-threaded with all sorts of interjected tangents. Although I built my program for reading articles and dialogue alike, I had not covered the grammar of interrupting phrases that break up the main thread of a sentence. Such sentence structures are abundant in story novels but do not occur in Winograd schemas, and I wasn’t planning on having my A.I. read novels any time soon. The inclusion of some 1940’s vocabulary also complicated matters: “cook-shanty”, “red-letter days”, “a pallid young dandy”? Maybe it’s because I’m Dutch, but I can only guess what these are.

Compared to the wide variety of common sense axioms that I had programmed (see How to teach a computer common sense*), many solutions to the pronouns were ordinary cases of continuity. E.g. a pronoun with an active role typically refers to the last active-role noun (You won’t find this rule in a grammar book, because ambiguous pronouns are grammatically “incorrect” to begin with).

Always before, Larry had helped Dad with his work. But he could not help him now […]
The donkey wished a wart on its hind leg would disappear, and it did.
Mark was close to Mr. Singer’s heels. He heard him calling for the captain […]

This makes sense when you’re testing on novels: No storyteller wants to write in such a counter-intuitive way that the reader has to stop and think about it, contrary to Winograd schemas which are designed for exactly that purpose.
Where no particular common sense axiom applied, rules of continuity and grammar chose 21 of my 29 correct answers. Thus the majority of my success seemed not due to the application of common sense, but due to conventional writing. Curious, I ran the test again with all axioms disabled except continuity. The result was an equal amount of correct answers, but much more randomly distributed and obviously chosen for the wrong reasons. The common sense axioms clearly contributed by fencing off the exceptions to continuity, so the cause of the mistakes lay elsewhere.

A closer look at the results
The results below show which of the 60 pronouns my program got correct, which axioms were applicable, and/or which problems hindered their conclusion. Where no axiom applied or a problem occurred, the program defaulted to the grammatically correct choice: The candidate closest to the pronoun. Only 1/3rd of all pronouns actually conformed with this grammar rule, which explains why whenever a problem occurred, the answer was typically wrong.

I will highlight the most prominent mistakes:

2 & 3. Always before, Larry had helped Dad with his work. But he could not help him now […]

Logic could expect Dad to return the favour, were it not that “always” and “now” suggest a continuity, which the program did not pick up on. Consequently, the answers to both “he” and “him” were switched around. This also highlights why this test was more difficult than chance: The more ambiguous pronouns a passage contained, the more likely a mistake in one would carry over to the others.

9. What about the time you cut up tulip bulbs in the hamburgers because you thought they were onions?

For this the program compared the similarities of bulbs, hamburgers and onions, but of course knowledge of onions was lacking in the database, so the inference fell flat. Retrieving knowledge from the internet would slow things down, and though speed is no issue in a contest, in daily practice I want my program to read one page per second, not one sentence per second.

13. […] Antonio, takes Larry under his wing.

People aren’t known to have wings, otherwise the bodypart location paradox would have excluded Larry. Alternatively one would have to know figurative meanings of English idioms, an added layer of difficulty.

18. [Maude…] had left poor little Dora to do the best she could, alone.

The program considers “to…” to indicate Maude’s reason for leaving “in order to” do something.

30. […] Mr. Moncrieff has decided to cancel Edward’s allowance on the ground that he no longer requires his financial support.

“Backward” = “back”, “Southward” = “south”, therefore “Edward” = “Ed”. Although the pronoun was interpreted correctly, “Ed” was of course not found among the multiple choice answers.

40. Every day after dinner Mr. Schmidt took a long nap. Mark would let him sleep for an hour, then wake him up, scold him, and get him to work. He needed to get him to finish his work, because his work was beautiful.

As I mentioned in my previous post, the “what goes around comes around” axiom was the least reliable, causing five misinterpretations in this test. Sometimes it triggered on trivial events, other times the events did not make sense this way (scolding to get someone to do good). It had better be limited to events that are direct cause and result, as they had been in most Winograd schemas.

49. Of one thing Mark was sure. Harry knew much less than he did.

Consecutive mental activity is typically by the same person, but of course not when it’s a comparison. Though the context system does distinguish comparisons, the axioms did not.

56. Tatyana managed two guitars and a bag, and still could point out the Freemans: “Isn’t it nice that they have come, Mama!”

While the pronoun was interpreted correctly, there was a technical hitch with selecting “freemans” from the multiple choice answers.

59. Grant worked hard to harvest his beans so he and his family would have enough to eat that winter, His friend Henry let him stack them […]

“enough” was translated to “enough beans” but lost its plural status in the translation, after which the beans were no longer considered a candidate for plural “them”.

Most of these problems are easily fixed and are not inherent to the common sense axioms, apart from #40 and its like. The majority of problems were instead linguistic: Small flaws in the grammar rules, difficulty with long-threaded phrasing, limited coverage of the context system, and problems with the contest’s XML-format interface. It just goes to show how perfect every part of the system has to be before it pays off, and how little you can tell about a program’s abilities from the surface.

The language barrier
As a test of common sense I found this setup less suitable than the original plan with Winograd schemas, who were more concise and profound in which areas of common sense they tested (e.g. spatial relations, physics, social interactions). Had I known from the start that the qualifying round would mainly feature novel prose, I would probably not have embarked on this challenge. Now the prose passages contained too many variables to tell whether results were due to language or common sense, and it never got to the Winograd schema round. This puts us back at the Turing Test where it’s either everything or nothing, and that is not a useful measure of progress. Swapping the rounds would be a good idea for next time.

It was nice to see serious competitors with a wide variety of technology tackling the problem, and although the overall results are unimpressive, I am pleased that my partial solution did as well as some academic efforts, with a minimum of resources at that. I am not disappointed in my common sense axioms as many of them were well applicable in this test, including all the pronouns that weren’t graded. I will broaden their application to ambiguous locations and indirect object relations, where I have greater need for them.

However, my main interest is the development of intelligent processes and I do not intend to linger on this aspect of language processing more than necessary. It is worth remembering that much can be said without ambiguity, and software like Stanford’s Coreference Resolver already achieve 90% precision on average texts. Though common sense has widespread application, it ultimately serves to filter and limit possibilities, while the possibilities in areas like problem solving and planning have yet to expand. For that reason I do not expect human levels of common sense to be reached within ten years either, but we can certainly make strides towards.

The Winograd Schema Challenge

The Winograd Schema Challenge, a $25000 contest sponsored by the aptly named company Nuance Communications, has been put forth as a better test of intelligence than Turing Tests*. Although the scientific paper tiptoes around its claims, the organisers describe the contest as requiring “common sense reasoning”. This introductory article explains the test’s strengths and weaknesses in that regard.

Example of a Winograd Schema

I used a tissue to clean the key, and then I put it in the drawer.
I used a tissue to clean the key, and then I put it in the trash.

A Winograd Schema is a sentence with an ambiguous pronoun (“it”), that, depending on one variable word (“trash/drawer”), refers to either the first or the second noun of the sentence (“tissue/key”). The Challenge is to program a computer to figure out which of the two is being referred to, when this isn’t apparent from the syntax. So what did I put in the trash? The tissue or the key? To a computer who has never cleaned anything, it could be either. A little common sense would sure come in handy, and the contest organisers suggest that this takes intelligent reasoning.
common sense computers

Common sense, not Google sense

The hare beat the tortoise because it was faster.
The hare beat the tortoise because it was too slow.

Contrary to this example, good Winograd Schemas are supposed to be non-Googleable. In this case Googling “fast hare” would return 20x more search results than “fast tortoise”, so the hare is 20x more likely to be the one who “was faster”. Although statistical probability is useful, this would make the contest won simply by the company with the largest set of statistics. It takes no reasoning to count how many times word A happens to coincide with word B in a large volume of text. Therefore this example would preferably be written with neutral nouns like “John beat Jack”, subjects of who we have no predetermined knowledge, but can still figure out which was faster.

Having said that, some example schemas involving “crop dusters” and “bassinets” still suggest that a broad range of knowledge will be required. Although one could consult online dictionaries and databases, the contest will have restrictions on internet access to rule out remote control. So failure can also be due to insufficient knowledge rather than a lack of intelligence, though I suppose that is part of the problem to solve.

Early indications

If a bed doesn’t fit in a room because it’s too big, what is too big?
If Alex lent money to Joe because they were broke, who needed the money?

With the above two questions the 2015 Loebner Prize Turing Test gave a tiny glimpse of Winograd Schemas in practice, and the answers suggest that chatbots – the majority of participants – are not cut out to handle them. Only 2 of 15 programs even answered what was asked. One was my personal A.I. Arckon*, the other was Lisa. Chatbot systems are of course designed for chat, not logic puzzles, and typically rely on their creators to anticipate the exact words that a question will contain. The problem there is that the understanding of Winograd Schemas isn’t found in which words are used, but in the implicit relations between them. Or so we presume.

The mermaid swam toward Sue and waved her tail. (Googleable)
The mermaid swam toward Sue and made her gasp. (More than a single change)

A more noteworthy experiment was done by the University of Texas, tested on Winograd Schemas composed by students. To solve the schemas they used a mixed bag of methods based on human logic, such as memorising sequences of events (i.e. verb A -> verb B), common knowledge, sentiment analysis and the aforementioned Googling. All of this data was cleverly extracted from text by A.I. software, or retrieved from online databases. However, many of the schemas did not accord with the official guidelines, and though they usefully solved 73% in total, only 65% was solved without the use of Google.

According to the same paper, the industry standard “Stanford Coreference Resolver” only correctly solved 55% of the same Winograd Schemas. The Stanford Resolver restricts the possible answers by syntax, gender(“he/she”) and amount(“it/they”), but does not examine them by knowledge or reasoning. The reason is that this level of ambiguity is rare. In my experience with the same methods however, it is still a considerable problem that causes 1/10th of text-extracted knowledge to be mistaken, with the pronoun “it” being the worst offender. So it appears (see what I mean?) that any addition of common sense would already advance the state of the art.

How to hack Winograd Schemas
Guesswork: Since the answers are a simple choice of two nouns, a machine could of course randomly guess its way to a score of 50% or more. So I did the math: With 60 schemas to solve, pure guesswork has a 5% chance to score over 60%, and a 0.5% chance to score over 65%. With the odds growing exponentially unlikely, this is not a winning tactic.
That said, the participating A.I. still have to make a guess or default choice at those schemas that they fail to solve otherwise. If an A.I. can solve 30% of the schemas and guesses half of the rest right, its total score amounts to 65%, equaling Texas’ score. It wouldn’t be until it can solve around 80% of all schemas genuinely that it could reach the winning 90% score by guessing the final stretch. That’s a steep slope.

Reverse psychology: Since Winograd Schemas are deliberately made to not match Google search results, it seems that one can apply reverse psychology and deliberately choose the opposite. While I did notice such a tendency in Winograd Schemas composed by professors, others have noticed that Winograd Schemas composed by students simply did match Google search results. So the success of using reverse psychology heavily depends on the cleverness of the composers. A countermeasure would be to use only neutral names for answers, but this may also cut off some areas of genuine reasoning. Alternatively one could include an equal amount of schemas that match and mismatch Google search results, so that neither method is reliable.

Pairing: One dirty trick that could double one’s success lies in the fact that Winograd Schemas come in pairs, where the answer to the second version is always the alternate noun. So if the A.I. can solve the first version but not the second, it suffices to choose the remaining alternate answer. Vice versa when it can solve the second version but not the first. This rather undermines the reason for having pairs: To ascertain that the first answer wasn’t just a lucky guess. Although this hack only increases the success of guesswork by a few percent, it can certainly be used to make a weak contestant into a strong contender undeservedly.

I call these hacks because not only are they against intent, they are also entirely useless in real life application. No serious researcher should use them or they will end up with an inept product.

How you can’t hack Winograd Schemas
No nonsense: The judgement of the answers is clear and objective. There is only one correct answer to each schema. The A.I. are not allowed to dodge the question, make further inquiries or give interpretable answers: It’s either answer A or B.

No humans: Erratic human performance of the judges and control subjects does not influence the results. The schemas and answers have been carefully predetermined, and schemas with debatable answers do not make the cut.

No invisible goal: While the Turing Test is strictly a win or lose game with the goalposts at fields unknown, the WSC can reward gradual increase of the number of schemas answered correctly. Partial progress in one area of common sense like spatial reasoning can already show improved results, and some areas are already proving feasible. This encourages and rewards short-term efforts.
I must admit that the organisers could still decide to move the goalposts out of reach every year by omitting particular areas of common sense once solved, e.g. all schemas to do with spatial reasoning. I think this is even likely to happen, but at the same time I expect the solutions to cover such a broad range that it will become hard to still find new problems after 6 contests.

Mostly, the WSC trims off a lot of subjective variables from the Turing Test, making for a controlled test with clear results.

The Winograd Schema Challenge beats the Turing Test
From personal experience, Turing Tests that I have participated in* have at best forced me to polish my A.I.’s output to sound less robotic. That is because in Turing Tests, appearance is a first priority if one doesn’t want to be outed immediately at the first question, regardless how intelligent the answer is. Since keeping up appearances is an enormous task already, one barely gets around to programming intelligence. I’ve had to develop spell correction algorithms, gibberish detection, calculator functions, letter-counting games and a fictional background story before encountering the first intelligent question in a Turing Test. It stalls progress with unintelligent aspects and is discouragingly unrewarding.

Solving Winograd Schemas on the other hand forced me to program common sense axioms, which can do more than just figure out what our pronouns refer to. Indirect objects and locations commonly suffer from even worse ambiguity that can be solved by the same means, and common sense can be used to distinguish figurative speech and improve problem-solving. But I’ll leave that as a story for next time.
We should be careful to draw conclusions from yet another behavioural test, but whatever the Winograd Schema Challenge is supposed to prove, it offers a practical test of understanding language with a focus on common sense. As this has always been a major obstacle for computers, the resulting solutions are bound to be useful regardless how “intelligent” they may be found.

Read more in my report on the first Winograd Schema Challenge held in 2016.