Madly Ambiguous

A fill-in-the-blank game that teaches you about structural ambiguity
while you try to trick a computer!

Hi, I'm Mr. Computer Head!

More and more these days, people are interacting with computers through human language.
For example, my friends Alexa, Cortana and Siri have become quite popular.
But did you know that computers can have a very hard time understanding things that are easy for humans?

Well, it's true!

Computers don't have much common sense,
and that can make it tough to understand what people really mean!

What is particularly tough for computers is ambiguity.

If you swipe left or click on the arrow to the right,
I'll teach you about ambiguous sentences.

After that, we can play a game where I try to guess what you mean by some ambiguous sentences!

A sentence is ambiguous if it has more than one possible meaning.
For a person, figuring out which meaning of an ambiguous sentence is intended is kid stuff.
Not so for computers though!

A sentence can be ambiguous when it has a word with multiple meanings,
like the word “bank.”

If I say,

“Jane had a picnic by the bank,”

which one of these “banks” do you picture?
(click one)

Financial bank?
. . . or river bank?
In sentences like this, both meanings of “bank” are plausible.
This is called lexical ambiguity.

Jane might have had a picnic near the place she keeps her money,
or she might have had a picnic by a river.

If you want to know which type of bank it was, you need more information.

financial bank
river bank
Another kind of ambiguity is called structural ambiguity.

The confusion doesn't come from words with multiple meanings,
but instead from different ways the parts of the sentence can be put together.

Same pieces . . .
. . . different shapes!
For example, if I say,

“Jane ate spaghetti with a fork,”

would Jane likely be using the fork or eating the fork?
(click one)

Utensil . . .
. . . or food?
You would probably know that Jane used the fork to eat the spaghetti.

That's the way!
Ouch — that's a little TOO al dente!
Now, if I say,

“Jane ate spaghetti with meatballs,”

would she likely be eating meatballs with the spaghetti dish or using meatballs to eat the spaghetti?
(click one)

Food . . .
. . . or utensil?
I'm guessing you probably like to eat with a fork more than you like to eat with a meatball,
and that meatballs taste quite good in your dish.

Nice and tasty!
What a mess!
These two sentences have different syntactic structures.
One way to show the structure of a sentence is to use arrows to indicate how words combine into phrases:

As we’ve noted, “with a fork” can combine with “ate spaghetti” to tell us more about the action,
while “with meatballs” can combine with “spaghetti” to tell us more about the dish.
But think about it —
how do you really know that Jane didn't use the meatballs as utensils to eat the spaghetti . . .
What a mess . . .

. . . or that the fork wasn't a part of the spaghetti dish that Jane ate?
Ouch — that's a little TOO al dente!
After all, both sentences are of the same form:

“Jane ate spaghetti with        .”

And yet, without needing any other information,
you can easily tell how the words combine into sensible phrases,
so that that Jane ate meatballs and used a fork.

If it were the other way around, it would just be ridiculous!

Seems reasonable to me . . .
For a person like you, it's easy to see what these sentences really mean.

It couldn't be more obvious that a fork is a utensil and a meatball is a food,
so you know which way of combining the sentences makes sense!

Well, aren't you smart!

As smart as computers are getting,
most of us aren't very good at figuring out which interpretations make sense
. . . and which don't.

It takes a lot of work for a computer to figure out
why some interpretations are just too ridiculous to be possible.

But I'm no ordinary computer.

I know that you use forks and eat meatballs, not the other way around!
What's more, I know you can differentiate even more meanings of the same kind of sentence!

In the sentence,

“Jane ate spaghetti with gusto,”

you know that Jane exhibited gusto while she ate:
with gusto” again combines with “ate spaghetti” to tell us more about the action.

This time, however, the semantic role is to describe the manner of the action
and not the instrument used.

What? This is my “gusto” face!
And if I say,

“Jane ate spaghetti with Mary,”

you know that Mary was present as Jane's companion during the meal . . .

. . . and not part of the dish!

Or do you?
(Click one)

Companion . . .
. . . or food?
I'm pretty sure you know
it makes more sense for Mary to be Jane's companion here,
rather than part of the dish.

One way is a nice outing . . .
. . . while the other is a very bad day for Mary.

(Much less a utensil that Jane used to eat the spaghetti, or the way Jane acted while eating.)
Not to mention, it probably didn't even occur to you that
the sentences about a fork or meatballs could mean
those items were Jane's mealtime companions!

Nothing like a nice meal with my dear friends, Fork . . .
. . . and Meatball.
So, to determine the intended meaning out of all these possible ones,
a computer would have to figure out the right syntactic structure and the right semantic role:

Now that you know a bit about structural (and semantic) ambiguity,
see if you can beat me in a round of Madly Ambiguous!

All you have to do is come up with a word or phrase to complete the ambiguous sentence

“Jane ate spaghetti with        .”

picture what it means to you,
and then I'll take my best guess at your intended meaning of the sentence!

Do you think you can come up with a sentence that will stump me?

I'd like to see you try!

Basic Mode
Complete the following sentence
with a Noun Phrase
like “meatballs” or “a silver fork” or “my friend Joe”,
making sure to picture what it means to you.
Then I'll take my best guess at your intended meaning!

Jane ate spaghetti with
Basic Mode
I think that

“Jane ate spaghetti with a fork,”


Jane used a fork to eat spaghetti.

Am I right?

Jane used a fork to eat spaghetti.
Jane had spaghetti and a fork.
Jane exhibited a fork while eating spaghetti.
Jane ate spaghetti in the presence of a fork.
none of these options are correct.
(scroll left or right for more options)

In order to figure out what your sentence means,
I have to take a few steps.

First, if you give me a noun phrase with multiple words,
I have to figure out which words are important.

If you give me a phrase like “a silver fork,"
I just need to look at the noun “fork."

But I have to be careful with phrases like “a bunch of meatballs,”
where both “bunch” and “meatballs” are nouns.

In this case, I know that “meatballs” is the most important noun.

Once I pick out the most important noun,
I look up the word in WordNet,
which is like a very detailed thesaurus.

WordNet tells me what kinds of words I'm looking at.

In WordNet, not only are synonyms grouped together (e.g. “zest” and “gusto”),
but words are also linked with more general and
more specific words (e.g. “meatballs” are a particular kind of “dish”).

If my search in WordNet tells me that the main noun
is a kind of artifact, food, or feeling, I have my answer:
the semantic role is instrument, part-of, or manner, respectively.

If the noun isn't related to any of those things,
I assume that the sentence must be talking about Jane's company as she eats.

While my strategy works pretty well,
it sometimes yields unexpected results.

One reason is that WordNet includes some infrequent
or even archaic word meanings.

For example, it has an archaic meaning of “trump”
that leads me to think “with Trump” must specify a utensil.

You might also have noticed that my strategy is pretty specific
to this one kind of sentence and situation.

Well, it's true — I would have to make a completely new strategy
if I wanted to be able to understand ambiguity in a new sentence!

But that's just using my basic mode!

In advanced mode, I use a data-driven strategy
that can learn from observing how people use language.

In advanced mode, I take advantage of word embeddings
trained on about 100 billion words of Google News text
using a tool called word2vec.

Word embeddings are a way of representing the meaning of words
as points in a multi-dimensional space that aims to locate words that frequently
appear in similar contexts close to each other in this space.

For example, both “spaghetti” and “noodles” frequently follow the
verb “eat”, so they should be near each other.

To come up with a meaning for a phrase,
I simply average the embeddings for each word in the phrase,
which finds the mid-point of the points for each word.
(Words that are less frequent, and thus more informative,
are given more emphasis in this step.)

Next, I take examples of phrases that people have
told me have utensil, food, manner or
company meanings when following “Jane ate
spaghetti with”, and use them to come up with expected
meanings for each of these categories.

These meanings can be visualized in 2-D using
a tool called t-SNE,
with different colors for each category.

Here, I've grouped the examples into three sub-clusters each,
showing the phrase closest to the mid-point of each cluster in parentheses.

And here is a zoomed-in view, showing phrases that are outliers
in each category, and thus potentially difficult for me.

For example, the meaning for “an Italian” ends up close to
the one for “her Italian passion”, which means that it will be
hard for me to tell that the former should be interpreted
as company while the latter should be manner.

You might have noticed this time that even in advanced mode,
my strategy for finding the meaning of phrases is awfully simple,
and doesn't really take context into account.

True again — even a smart computer like me has a pretty long way to go
before I'm as good at figuring out ambiguity as humans are.

(Thankfully, researchers are working hard on these issues!)

But that's how I play Madly Ambiguous!

Maybe now that you know my secrets,
you have some new ideas on how to stump me.

Do you want to play another round?


Copyright 2017

Madly Ambiguous was created through the combined efforts of
Ajda Gokcen, Ethan Hill, and Michael White
(with David King, Matt Metzger, and Kaleb White)
and funded through NSF Grant 1319318.

Any opinions, findings, and conclusions or recommendations expressed in this material are those of the author(s)
and do not necessarily reflect the views of the National Science Foundation.