Computational Linguistics: Crash Course Linguistics #15 - Free Educational videos for Students in K-12

Computational Linguistics: Crash Course Linguistics #15 - By Math and Science

Transcript


00:0-1	Hi , I'm taylor and welcome to crash chris linguistics
00:02	. Computers are pretty great , but they can only
00:04	do stuff that humans tell them to do counter intuitively
00:07	. This means that the more automatic a human skill
00:10	is , the more difficult it is for us to
00:12	teach to computers . It's easy for us to teach
00:15	a computer to calculate millions of digits of pi or
00:18	play chess but get a computer to recognize which image
00:21	contains a traffic light , surprisingly difficult . The same
00:25	thing goes for language . The parts that are difficult
00:27	for humans like learning . Lots of new words are
00:29	easy for computers and the parts that are easy for
00:32	humans like understanding across typos and accents or knowing of
00:36	someone sad or angry or joking are really , really
00:40	difficult for machines . Plus language isn't just one task
00:43	to teach . It's all the different things we've talked
00:45	about throughout the series and more programming computers to process
00:49	. Human language is called natural language processing or computational
00:53	linguistics . We rely on NLP for a whole range
00:56	of tasks , search engines , voice activated home systems
01:00	, spam filters , spell checkers , predictive text and
01:03	more . Today we'll look at what NLP is and
01:06	what makes language a difficult challenge for computers . Yeah
01:20	, getting a computer to work with something as complex
01:22	as language requires a lot of steps . First ,
01:24	we need to give the computer text to work with
01:27	. We can input it directly or get the computer
01:29	to transform speech sounds , handwriting or other physical text
01:33	into digital text . We do that with speech to
01:35	text , handwriting recognition or optical character recognition processes .
01:39	This step involves figuring out where the break between words
01:42	and sentences go , such as the difference between a
01:45	moist towelette versus a moist outlet or whether a small
01:50	speck is the daughter of an I . A .
01:52	Period or a flock of dirt . Once it has
01:55	the digital text , we then need the computer to
01:57	figure out a the meanings of the words and be
02:00	the relationship between them . It might use context to
02:03	disambiguate between things like bank and blank . A river
02:06	bank and a financial bank or common grounds and proper
02:09	towns . In this step , the machine figures out
02:12	approximately what is being said . The next step is
02:15	to get it to do something useful with that information
02:17	such as answer a question translated into another language or
02:21	find directions between two places . Each of these tasks
02:25	also requires a different system . All of this data
02:27	gets produced in some abstract form that the computer understands
02:30	like a mathematical equation or some lines of code .
02:33	The last step is to re encode that data into
02:35	natural human language , which can involve text generation depending
02:39	on what the user wants . The computer might need
02:41	to produce the answer as speech , in which case
02:44	it would use text to speech and speech synthesis .
02:47	That's a lot of steps . The nice thing about
02:49	splitting up natural language processing into different steps is that
02:52	we can reuse parts of it for other tasks .
02:55	For example , if we make one system that's good
02:57	at text to speech for english , it can read
02:59	aloud answers to questions , translations into english and directions
03:03	to go to places . We can also distinguish between
03:06	what needs to be customized for each human language and
03:08	what can always stay in computer code . That saves
03:11	programmers and computers . Sometime tools that perform just one
03:15	or two of these sub tasks can also be useful
03:17	by themselves . Automatic captions may just do the speech
03:20	to tax part screen readers may just do text to
03:23	speech and search or translation may start with text and
03:26	skip processing speech entirely . A similar set of steps
03:29	could work for signed languages too . Although this technology
03:32	is very underdeveloped compared to what's been created for a
03:35	few big spoken languages that could be something like signed
03:38	text , parsing signs , processing the results for a
03:41	computer to work with and rendering the output back into
03:44	signs . We could then also create systems that inter
03:47	operated between signed and spoken languages . For example ,
03:51	a computer could take input in english and translated to
03:53	a sl or vice versa . Just like with the
03:56	thousands of spoken languages , though each of the hundreds
03:59	of sign languages would still need to be supported separately
04:02	. One thing that won't really help is gloves .
04:05	Let's head to the thought bubble to pop that bubble
04:08	. You might have seen hyperbolic headlines about sign language
04:11	translation gloves in the news throughout the years . They
04:14	claim that these gloves can translate american sign language into
04:17	english speech by recognizing the wears hand shapes . Unfortunately
04:21	, these glove makers have made several fundamental misunderstandings about
04:25	how sign languages work . One is that the grammar
04:28	of signs languages isn't expressed just in the shape of
04:31	the hand . Signed languages also include facial expressions and
04:35	movements of the hands and arms in relation to the
04:37	rest of the body . To is that signed languages
04:40	use far more signs than the 26 letters of the
04:42	manual alphabet , which is all the gloves can detect
04:45	. Plus , signed languages tend to use the manual
04:48	alphabet to borrow technical words from spoken language is not
04:51	for core vocabulary , That's like making a translation system
04:54	for english that only recognizes the words that come from
04:57	greek three is that translation should enable two way communication
05:01	between hearing and deaf people . But gloves can only
05:04	translate from science to speech , never from speech to
05:07	a format accessible for deaf and hard of hearing people
05:09	. Which is ironic because the technology to produce written
05:12	captions of speech already exists , computational tools involving signed
05:16	languages could one day exist using other input sources that
05:19	can actually access full signs . But they're never going
05:22	to be any good if deaf people aren't consulted in
05:25	creating them . And many deaf researchers have already pointed
05:28	out that gloves are just never going to accomplish that
05:31	. Thanks . Thought bubble . So let's say we
05:32	have created a system that's pretty good at each of
05:34	the steps involved in natural language processing At least for
05:38	one or 2 languages . Does the system understand language
05:41	the way human does to answer that ? Let's pretend
05:44	we've trained a rabbit to press buttons A , B
05:46	and C . In order to get a treat .
05:48	We could relabel those buttons . I want food ,
05:51	but that wouldn't mean that the rabbit understands english .
05:54	The rabbit would press the same buttons if they were
05:56	labeled something entirely unrelated . The same goes for a
05:59	computer . If we tell a computer a few basic
06:02	instructions , it can give the appearance of understanding language
06:05	, but it might fall apart spectacularly when we ask
06:08	it to do something more complicated . That's part of
06:10	what makes teaching a computer to do language so tricky
06:13	. Originally , people taught computers to do language tasks
06:16	with long lists of more and more specific rules ,
06:19	such as make a word plural by adding s wait
06:22	unless the word is child , in which case add
06:25	ren instead and so on . For other exceptions ,
06:28	more modern approaches to machine learning involves showing computers a
06:31	whole bunch of data to train them on statistical patterns
06:34	and then testing how well they figured out these patterns
06:37	using a different set of data . A lot of
06:39	recent leaps and natural language processing have come from a
06:41	kind of statistical machine learning known as neural networks ,
06:45	neural nets are based on a very simplified model of
06:47	how neurons work in the brain , allowing them to
06:50	figure out for themselves which factors are the most relevant
06:54	in the training data . But because they work out
06:56	these factors for themselves , it's hard for humans to
06:59	know exactly what patterns they're picking up on early in
07:01	the neural nets . Training it will make really silly
07:04	non human like errors like returning a text E .
07:09	Because it's worked out that E is the most common
07:11	letter in english . Writing the machine will keep adjusting
07:14	itself based on the training data though , and eventually
07:16	it starts returning things that look more like words .
07:19	Well , almost in any kind of machine learning training
07:22	data is really important and there are two kinds of
07:24	data we can use . The first is data with
07:26	two corresponding parts that have been matched by humans ,
07:29	such as text with audio , words with definitions ,
07:32	questions with answers , sentences with translations or images with
07:36	captions using parallel data like this is known as supervised
07:39	learning and it's great , but it can be hard
07:42	to find enough data that has both parts . After
07:45	all , some humans have to create all of these
07:47	pairs . The second kind of data has only one
07:49	component , like a bunch of text or audio or
07:52	video . In one language using this kind of non
07:55	parallel data is known as unsupervised learning . It's much
07:59	easier to find , but it's harder to use to
08:01	train a computer since it has to learn only from
08:03	half of the pair . So researchers often use a
08:06	mix of both , a smaller amount of parallel data
08:08	to get things started and then a larger amount of
08:11	non parallel data . This combination is called semi supervised
08:15	learning , but none of this data just magically appears
08:18	. It gets created or gathered by humans and humans
08:21	have all sorts of bias . Is computer science researcher
08:24	Horeni Suresh created a framework to evaluate bias in machine
08:27	learning . We can use this framework to see how
08:30	bias affects the language tools we've discussed in this episode
08:33	. First historical bias is when a bias in the
08:36	world gets reflected in the output the computer produces .
08:39	For example , Turkish doesn't make a gender distinction in
08:41	any of its pronouns , whereas english does in the
08:44	third person singular between he she it and singular day
08:48	. So a translation system might pick agenda for pronouns
08:50	when translating them from Turkish to english making he is
08:53	a doctor but she is a nurse from the same
08:56	Turkish pronoun this might reflecting overall tendency in the world
08:59	, but our computer is still producing a gender bias
09:02	. Next representation bias is when some groups aren't as
09:05	well represented as others in the training data . For
09:08	instance , while researchers estimate that at least 2000 languages
09:11	are actively being used on social media , only a
09:14	few large languages are well represented in language tech tools
09:18	. The rest are barely represented or left out ,
09:20	including all signed languages . When the features and labels
09:24	in the training data don't accurately reflect what we're looking
09:26	for . That's measurement bias . The text that has
09:29	been translated into the most languages is the bible .
09:32	So it's often used as training data . But the
09:34	style of language and religious texts can be very different
09:37	from day to day conversation and can produce strange results
09:40	in google translate . Aggregation bias is when several groups
09:43	of data with different characteristics are combined and a single
09:46	system isn't likely to work well for all of them
09:48	at once . If we smushed all the varieties of
09:51	english into training data for an english speech to text
09:54	program , It could end up working better for standardized
09:56	english than say . African american english evaluation bias occurs
10:01	when researchers measure a program success based on something users
10:04	won't find useful . Researchers with an english first mentality
10:08	might focus on whether predictive text program predicts the next
10:11	word , whereas the program that predicts the next morphine
10:14	would work better for languages with longer words and more
10:17	morphine seems when a system was originally created for reasonable
10:20	purposes but then gets misused after its release . That's
10:23	deployment bias style analysis tools can be used to determine
10:27	whether a historic figure wrote an anonymous book , but
10:30	they can also be misused to identify anonymous whistleblowers .
10:33	Being aware of these sources of bias is the first
10:36	step in figuring out how to correct for them .
10:38	Like the whole field of computational linguistics , addressing these
10:42	biases is an active area of research . We have
10:44	a responsibility to use our increased understanding of language through
10:48	linguistics Too deeply consider the effects we have on each
10:51	other and the world we live in . This ethical
10:54	consideration is especially important in computational linguistics because we interact
10:59	with technology so much in our daily lives . Next
11:01	time we'll talk about a much older kind of language
11:04	technology , which is so common , we might not
11:06	even think of it as a technology writing system .
11:09	Thanks for watching this episode of crash course linguistics .
11:11	If you want to help keep all crash course free
11:14	for everybody forever , you can join our community on
11:17	Patreon .

Summarizer

DESCRIPTION:

Quality Math And Science Videos that feature step-by-step example problems!

OVERVIEW:

Computational Linguistics: Crash Course Linguistics #15 is a free educational video by Math and Science.

This page not only allows students and teachers view Computational Linguistics: Crash Course Linguistics #15 videos but also find engaging Sample Questions, Apps, Pins, Worksheets, Books related to the following topics.