Title

Why Can't Computers
Use Language As Well
As People Do?
Victoria Muehleisen
Associate Professor
Tasks which computers do better than
humans



Calculating numbers: Computers are faster
and more accurate than people.
Searching and retrieving information:
Computers don‘t “forget” where information
is.
Recognizing patterns (such as “enhancing”
a photograph): Computers can compare
many pieces of information very quickly.

However, when it comes to language,
computers leave much to be desired.
Three programs in this Powerbook



Spell Checking (in TextEdit)
Speech Recognition (with Apple Speakable
Items)
Translation (with Sherlock)
You’ll see that each program does well with
things, but poorly with others.
Spell Checking

It‘s a kind of pattern matching. It checks the
words you type against words in its dictionary.
Problems with Spell Checking

It works fine for words in the dictionary, but a
lot of words aren‘t in the dictionary!

It usually doesn‘t catch a mistake when the
mistake is an actual word (just the wrong
word).
 Demonstration
Speech Recognition

A computer first must recognize the individual
sounds, then the words, then the meanings.
Problems with Speech Recognition

The program can only recognize a certain set
of words, but in natural languages, there are
many different ways to express the same
message.
 Demonstration
More problems with Speech
Recognition

Different speakers have different voices and
different accents.

Even one person sounds different at different
times (for example, talking quickly, talking
while sick with a cold.)
Translation

This is one of the most complicated things we
ask computers to do with language.

To translate, a person or a computer must
understand individual words, understand the
sentences, and understand the intended
meaning in one language.
Translation

Then the translator must find equivalent
words and sentences in the other language.
 Demonstration
Problems with Translation

There are many things which make
translation difficult, both for people and for
computers.

We’ll see several more examples in the rest
of this lecture.
Why can’t computers do better?

Of course, computers don‘t do anything with
language by themselves. They are
programmed by people!
So the question really is…

Why can‘t people program computers to deal
with language as well as people can?
In order to program computers to use
language …
1.
We have to understand the “rules” of
language.
2.
We have to express those rules in a way
that computers can use.
People working on this problem

Linguists: people who study the rules and
patterns of language

Computer scientists and applied linguists:
people creating programs based on the rules
and patterns
Understanding the Rules:
One Simple Rule of English

To form the past tense of most verbs,
add -ed to the present tense form.
live → lived
need → needed
We need to list the conditions and
exceptions.

Some verbs are irregular (wrote, saw), and these
must be listed individually in the program.

Speech production and speech recognition
programs would have to know that -ed has different
pronunciations: /d/ in lived, /id/ in needed, and /t/ in
slipped.
Computers can easily “learn” this kind
of simple rule, but...
…even with a rule so simple there may be variations
that are hard to include in the program.
 She
dived into the lake. (Used in many English
speaking countries)
 She
dove into the lake. (Used in some parts of
North America)
Many rules are more complex.

English speakers learn the difference
between “the” and “a” at an early age, but we
never learn explicit rules.

Linguists have tried to describe the rules for
using “the” and “a”, but there is still no
complete description.
A similar example in Japanese

You learned how to use “wa” and “ga” at an
early age, but there is still no complete set of
rules which can be used by adults learning
Japanese.
Many patterns in language are still not
completely understood…
…and if we don’t understand the patterns, we
can’t write computer programs.
An example:
Input: お湯がありますか?お茶を飲みたいです。
Sherlock’s translation: Is there a hot water? We
would like to drink the tea, is.
My translation: Is there any hot water? I want
some
tea.
Another example of rules
When translating from Japanese to English or
English to Japanese, we need rules to
describe the differences in basic word
order:
 Japanese: Subject - Object-Verb


私は、本を買った。
English: Subject - Verb - Object

I bought a book.
Word order rules in a computer
program
The difference in word order is something we
can easily make a rule for and put into a
computer program, as long as the sentence
is simple.
Sherlock’s translation
Input: Cats like mice.
Output: 猫がマウスを好む。
But not all sentences are this easy…
Consider a sentence with more than one
clause:
 Input: I love skiing, so whenever I get a
chance, I do it.
 Sherlock: 私はするそれをスキー、従って私がチ
ャンスを得る時はいつでも、私愛する。
A human translation (my husband’s)

私は、スキが好きです。チャンスがあれば、いつ
でもスキをやります。
There are many more difficulties for
Sherlock
For example, English sentences must always
have a subject, but in Japanese, the subject is
often not mentioned.
What is the subject of this sentence?
雨がすきで、雨が降ると雨をみる。

If I tell you that it is from a autobiographical
(自叙)essay, you can easily guess.
Human translation(mine):
Input:雨がすきで、雨が降ると雨をみる。
Output: I like rain, so when it rains, I watch it.
I know from the context that the person who
likes rain and the person who watches the
rain are both “I”.
Sherlock’s translation
Input:雨がすきで、雨が降ると雨をみる。
Output: Liking the rain, when the rain falls, you
see the rain.

My translation and Sherlock’s translation are
different because I understand the context of
the sentence.
Another example of the importance of
context
(Students of Japanese learn this one in class!)
In a restaurant:
 「私は、うなぎです。」
「私は、うなぎです。」

My translation: “I’ll have eel.”

Sherlock’s: “I am the eel.”
Possibility for improvement

It is possible to put more information about
grammar structure and contexts into
computer programs.
A better spell checker

TextEdit is free software, but a more
expensive program might catch mistakes like
this one:
“Can you sea this mistake?”
“See” is a verb, but “sea” is a noun.
A better translation program

Sherlock is a free program, but a more
expensive would probably know that we don’t
say “a hot water.”
お湯がありますか?⇒ Is there any hot water?
A better translation program

It might even be possible to make a program
that would know that in an autobiographical
essay, the subject is often “I”.
雨がすきで、雨が降ると雨をみる。
I like rain, so when it rains, I watch it.
But there’s a more serious difficulty…

In the widest sense, context includes the
whole world, including knowledge of social
relationships.

Humans can easily use this knowledge, but
can computers?
An example of social context
優美子:宝くじ一等当たったのよ。
貴子:うそ!
Sherlock’s translation
Yumiko: Lottery first class it hit, is.
Takako: Lie!
「うそ!」≠
“lie!”
We know that if Yumiko and Takako are
friends, Takako is not calling her friend a liar.
We can tell this both by the context and by
the pronunciation:「うっそ〜」.
A better translation

Yumiko: I won first prize in the lottery.
Takako: No! I don’t believe it!
“No! I don’t believe it!”

In this case too, the real meaning is not the
same as the literal meaning.

Takako does believe what Yumiko said, but
she is very surprised. This expression shows
her feelings.
Real meaning vs. literal meaning

It’s easy for humans to learn that the real
meaning is not always the literal meaning.
Although I never learned this use of うそ in
Japanese class in the U.S., I quickly learned
it when I got to Japan.

However, it’s hard to imagine how a computer
program could “learn” this kind of thing.
Summary

If linguists can write a language rule that
does not depend on the context, it’s easy for
computer scientists to write a computer
program using the rule.

But if the rule is not understood well, or if the
rule depends on the context, it’s very difficult,
or maybe impossible, to write a program
using it.

In the future, we can expect that computers
will get better and better at translating
“straight forward” texts (e.g., texts that just
give information).

However, there are some things that
computers might never be able to do.
If you want to learn more…

Go to the web sites on the handout
 Japanese
web sites about linguistics
 The Linguistic Society of America (check out
the videos!)
 The Linguist Network
Classes at the International College of
Waseda
 CO
201 Introduction to Language Studies
 CO 202 Introduction to Communication
 CO 203 Introduction to Application of
Language Studies
 GE 111 Introduction to Computer Literacy(情
報処理入門)

You can find the web links and this
PowerPoint lecture on my web page:
<http://www.f.waseda.jp/vicky/comp-lang.html>
Any questions?