PRATHYUSHA ENGINEERING COLLEGE

Department of Computer Science and Engineering

VI Semester (EVEN 2019-20)
CS8691ARTIFICIAL INTELLIGENCE

UNIT V APPLICATIONS 9
AI applications –Language Models –Information Retrieval-Information Extraction –Natural Language
Processing -Machine Translation –Speech Recognition –Robot –Hardware –Perception –Planning –Moving

1. AI APPLICATIONS

AI programs are developed to perform specific tasks, that is being utilized for a wide range of
activities including medical diagnosis, electronic trading platforms, robot control, and remote sensing.
AI has been used to develop and advance numerous fields and industries, including finance,
healthcare, education, transportation, and more.
 Information Retrieval  Agriculture
 Natural Language Processing  HealthCare
 Machine Translation  Gaming
 Speech Recognition  Space Exploration
 Robot Marketing  Autonomous Vehicles
 Banking  Chatbots
 Finance  Artificial Creativity

Marketing
 we search for an item on any e-commerce store, we get all possible results related to the item.
It’s like these search engines read our minds! In a matter of seconds, we get a list of all relevant
items.
With the growing advancement in AI, in the near future, it may be possible for consumers on the web
to buy products by snapping a photo of it
Banking
 A lot of banks have already adopted AI-based systems to provide customer support, detect
anomalies and credit card frauds. An example of this is HDFC Bank.
 HDFC Bank has developed an AI-based chatbot called EVA (Electronic Virtual Assistant),
built by Bengaluru-based Senseforth AI Research.
 Since its launch, Eva has addressed over 3 million customer queries, interacted with over half a
million unique users, and held over a million conversations.
 Eva can collect knowledge from thousands of sources and provide simple answers in less than
0.4 seconds.
 AI solutions can be used to enhance security across a number of business sectors, including
retail and finance.
 By tracing card usage and endpoint access, security specialists are more effectively preventing
fraud. Organizations rely on AI to trace those steps by analyzing the behaviors of transactions.
Companies such as MasterCard and RBS WorldPay have relied on AI and Deep Learning to detect
fraudulent transaction patterns and prevent card fraud for years now. This has saved millions of
dollars.
Finance
1
 Ventures have been relying on computers and data scientists to determine future patterns in the
market. Trading mainly depends on the ability to predict the future accurately.
 Machines are great at this because they can crunch a huge amount of data in a short span.
Machines can also learn to observe patterns in past data and predict how these patterns might
repeat in the future.
 Financial organizations are turning to AI to improve their stock trading performance and boost
profit.
Agriculture
 AI can help farmers get more from the land while using resources more sustainably.
 Issues such as climate change, population growth, and food security concerns have pushed the
industry into seeking more innovative approaches to improve crop yield.
 Organizations are using automation and robotics to help farmers find more efficient ways to
protect their crops from weeds.
 Blue River Technology has developed a robot called See & Spray which uses computer vision
technologies like object detection to monitor and precisely spray weedicide on cotton plants.
Precision spraying can help prevent herbicide resistance
 The image recognition app identifies possible defects through images captured by the user’s
smartphone camera. Users are then provided with soil restoration techniques, tips, and other
possible solutions. The company claims that its software can achieve pattern detection with an
estimated accuracy of up to 95%.
Health Care
 When it comes to saving our lives, a lot of organizations and medical care centers are relying on
AI.
 clinical decision support system for stroke prevention that can give the physician a warning when
there’s a patient at risk of having a heart stroke.
 Preventive care
 Personalized medicine
Gaming
 Artificial Intelligence has become an integral part of the gaming industry. In fact, one of the
biggest accomplishments of AI is in the gaming industry.
 DeepMind’s AI-based AlphaGo software, which is known for defeating Lee Sedol, the world
champion in the game of GO, is considered to be one of the most significant accomplishment in the
field of AI
 The actions taken by the opponent AI are unpredictable because the game is designed in such a
way that the opponents are trained throughout the game and never repeat the same mistakes.
 They get better as the game gets harder. This makes the game very challenging and prompts the
players to constantly switch strategies and never sit in the same position.

Space Exploration
 Space expeditions and discoveries always require analyzing vast amounts of data.
 Artificial Intelligence and Machine learning is the best way to handle and process data on this
scale.
 After rigorous research, astronomers used Artificial Intelligence to sift through years of data
obtained by the Kepler telescope in order to identify a distant eight-planet solar system.

2
 Artificial Intelligence is also being used for NASA’s next rover mission to Mars, the Mars 2020
Rover. The AEGIS, which is an AI-based Mars rover is already on the red planet.
 The rover is responsible for autonomous targeting of cameras in order to perform investigations on
Mars.
Autonomous Vehicles
 For the longest time, self-driving cars have been a buzzword in the AI industry. The development
of autonomous vehicles will definitely revolutionaries the transport system.
 Companies like Waymo conducted several test drives in Phoenix before deploying their first AI-
based public ride-hailing service.
 The AI system collects data from the vehicles radar, cameras, GPS, and cloud services to produce
control signals that operate the vehicle.
 Advanced Deep Learning algorithms can accurately predict what objects in the vehicle’s vicinity
are likely to do. This makes Waymo cars more effective and safer.
 Another famous example of an autonomous vehicle is Tesla’s self-driving car. Artificial
Intelligence implements computer vision, image detection and deep learning to build cars that can
automatically detect objects and drive around without human intervention

2. WUMPUS WORLD

 The Wumpus world is a simple world example to illustrate the worth of a knowledge-based agent and to
represent knowledge representation. It was inspired by a video game Hunt the Wumpus by Gregory Yob
in 1973.
 The Wumpus world is a cave which has 4/4 rooms connected with passageways. So there are total 16
rooms which are connected with each other.
 We have a knowledge-based agent who will go forward in this world. The cave has a room with a beast
which is called Wumpus, who eats anyone who enters the room.
 The Wumpus can be shot by the agent, but the agent has a single arrow.
 In the Wumpus world, there are some Pits rooms which are bottomless, and if agent falls in Pits, then
he will be stuck there forever.
 The exciting thing with this cave is that in one room there is a possibility of finding a heap of gold. So
the agent goal is to find the gold and climb out the cave without fallen into Pits or eaten by Wumpus.
 The agent will get a reward if he comes out with gold, and he will get a penalty if eaten by Wumpus or
falls in the pit.

3
There are also some components which can help the agent to navigate the cave. These
components are given as follows:
a. The rooms adjacent to the Wumpus room are smelly, so that it would have some stench.
b. The room adjacent to PITs has a breeze, so if the agent reaches near to PIT, then he will
perceive the breeze.
c. There will be glitter in the room if and only if the room has gold.
d. The Wumpus can be killed by the agent if the agent is facing to it, and Wumpus will emit a
horrible scream which can be heard anywhere in the cave.

The Wumpus world Properties:

o Partially observable: The Wumpus world is partially observable because the agent can only
perceive the close environment such as an adjacent room.
o Deterministic: It is deterministic, as the result and outcome of the world are already known.
o Sequential: The order is important, so it is sequential.
o Static: It is static as Wumpus and Pits are not moving.
o Discrete: The environment is discrete.
o One agent: The environment is a single agent as we have one agent only and Wumpus is not
considered as an agent.
Exploring the Wumpus world:
o Now we will explore the Wumpus world and will determine how the agent will find its goal by
applying logical reasoning.
Agent's First step:
o Initially, the agent is in the first room or on the square [1,1], and we already know that this
room is safe for the agent, so to represent on the below diagram (a) that room is safe we will
add symbol OK. Symbol A is used to represent agent, symbol B for the breeze, G for Glitter or
gold, V for the visited room, P for pits, W for Wumpus.
o At Room [1,1] agent does not feel any breeze or any Stench which means the adjacent squares
are also OK.

Agent's second Step:

o Now agent needs to move forward, so it will either move to [1, 2], or [2,1]. Let's suppose agent
moves to the room [2, 1], at this room agent perceives some breeze which means Pit is around
this room. The pit can be in [3, 1], or [2,2], so we will add symbol P? to say that, is this Pit
room?

4
 o Now agent will stop and think and will not make any harmful move. The agent will go back to
the [1, 1] room. The room [1,1], and [2,1] are visited by the agent, so we will use symbol V to
represent the visited squares.

Agent's third step:

o At the third step, now agent will move to the room [1,2] which is OK. In the room [1,2] agent
perceives a stench which means there must be a Wumpus nearby. But Wumpus cannot be in
the room [1,1] as by rules of the game, and also not in [2,2] (Agent had not detected any stench
when he was at [2,1]). Therefore agent infers that Wumpus is in the room [1,3], and in current
state, there is no breeze which means in [2,2] there is no Pit and no Wumpus. So it is safe, and
we will mark it OK, and the agent moves further in [2,2].

Agent's fourth step:

o At room [2,2], here no stench and no breezes present so let's suppose agent decides to move to
[2,3]. At room [2,3] agent perceives glitter, so it should grab the gold and climb out of the
cave.

3. LANGUAGE MODELS
Humans use a limited number of conventional signs (smiling, shaking hands) to communicatein much
the same way as other animals. Humans have also developed a complex, structuredsystem of signs
known as language that enables them to communicate most of what they knowabout the world.

One of the actions that is available to an agent is to produce language.

This is called a speechact. "Speech" is used in the same sense as in "free speech," not "talking," so
typing, skywriting,and using sign language all count as speech acts.
English has no neutral word for an agent thatproduces language, either by speaking or writing or
anything else. We will use speaker, hearer, andutterance as generic terms referring to any mode of
communication.

Imagine a group of agents are exploring the wumpus world together. The group gains anadvantage
(collectively and individually) by being able to do the following:
• Inform each other about the part of the world each has explored, so that each agent hasless exploring
to do. This is done by making statements: There's a breeze here in 3 4.
• Query other agents about particular aspects of the world. This is typically done by askingquestions:
Have you smelled the wumpus anywhere?
• Answer questions. This is a kind of informing. Yes, I smelled the wumpus in 2 5.
• Request or command other agents to perform actions: Please help me carry the gold. Itcan be seen
as impolite to make a direct requests, so often an indirect speech act (a requestin the form of a

5
 statement or question) is used instead: / could use some help carrying thisorCould you help me cany
this?
• Promise to do things or offer deals: /'// shoot the wumpus if you let me share the gold.
• Acknowledge requests and offers: OK.
• Share feelings and experiences with each other: You know, old chap, when I get in a spotlike that, I
usually go back to the start and head out in another direction, or Man, thatwumpus sure needs some
deodorant!

Planning problem
The hard part for an agent is to decide when a speech act of some kind is called for,and to decide which
speech act, out of all the possibilities, is the right one. At one level,this is just the familiar planning
problem—an agent has a set of possible actions to choosefrom, and must somehow try to choose
actions that achieve the goal of communicating someinformation to another agent.

Understanding problems
 The problem of understanding speech acts is much like other understanding problems,such as
image understanding or medical diagnosis.
 Part of the speech act understanding problem is specific to language.
 We need to know something about the syntax and semantics of a language to determine why
anotheragent performed a given speech act.
 The understanding problem also includes the more generalproblem of plan recognition.
 Part of the understanding problem can be handled by logical reasoning.
 We will see thatlogical implications are a good way of describing the ways that words and
phrases combine toform larger phrases. Another part of the understanding problem can only be
handled by uncertainreasoning techniques.
 Usually, there will be several states of the world that could all lead to thesame speech act, so
the understander has to decide which one is more probable.

Fundamentals of language
Formal languages—the ones like Lisp and first-order logic that areinvented and rigidly defined
Natural languages—the ones like Chinese, Danish, andEnglish that humans use to talk to one
another.

Formal language
 A formal language is defined as a set of strings, where each string is a sequence of
symbolstaken from a finite set called the terminal symbols.
 most of them are similar in that they are based onthe idea of phrase structure
 Strings are composed of substrings called phrases, whichcome in different categories.
 Categorizing phrases helps us to describe the allowable strings of thelanguage.
 Any of the noun phrases can combine with a verb phrase (or VP) suchas "is dead" to form a
phrase of category sentence
 Grammatical categories are essentially posited as part of ascientific theory of language that
attempts to account for the difference between grammatical andungrammatical categories.

Example:
noun phrase"the wumpus," "the king,"
verb phrase : "is dead"
sentence :"the wumpus is dead"
Grammatical categories :"the wumpus is dead" , "wumpus the dead is"

Categories such as NP, VP, and S are called nonterminal symbols. In the BNF notation,rewrite rules
consist of a single nonterminal symbol on the left-hand side, and a sequence ofterminals or
nonterminals on the right-hand side

The meaning of a rule such as S— NP VP

6
Grammatical formalisms can be classified by their generative capacity: the set oflanguages they can
representfour classes of grammaticalformalisms that differ only in the form of the rewrite rules.

 Recursively enumerable grammars use unrestricted rules: both sides of therewrite rules can
have any number of terminal and nonterminal symbols. Thesegrammars are equivalent to
Turing machines in their expressive power.
 Context-sensitive grammars are restricted only in that the right-hand hand sidemust contain at
least as many symbols as the left-hand side. The name contextsensitivecomes from the fact that
a rule such as
ASB — AXB
says that an S can berewritten as an X in the context of a preceding A and a following B.
 In context-free grammars (or CFGs), the left-hand side consists of a singlenonterminal
symbol. Thus, each rule licenses rewriting the nonterminal as the righthandside in any
context. CFGs are popular for natural language grammars, althoughit is now widely accepted
that at least some natural languages are not context-free.
 Regular grammars are the most restricted class. Every rule has a single nonterminalon the left-
hand side, and a terminal symbol optionally followed by a nonterminalon the right-hand side.
Regular grammars are equivalent in power to finite-statemachines. They are poorly suited for
programming languages because, for example,they cannot represent constructs such as
balanced opening and closing parentheses.

To give you an idea of which languages can be handled by which classes, thelanguagea"b" (a
sequence of n copies of a followed by the same number of b) canbe generated by a context-free
grammar, but not a regular grammar. The languagea"b"c" requires a context-sensitive grammar,
whereas the language a*b* (a sequenceof any number of a followed by any number of b) can be
described by any of the fourclasses. A summary of the four classes follows.

The component steps of communication

A typical communication episode, in which speaker S wants to convey proposition P to hearer H
using words W, is composed of seven processes.
Three take place in the speaker:
Intention: S wants H to believe P (where S typically believes P)
Generation: S chooses the words W (because they express the meaning P)
Synthesis: 5 utters the words W (usually addressing them to H)

Four take place in the hearer:

Perception: H perceives W (ideally W' = W, but misperception is possible)
Analysis: H infers that W has possible meanings P\,..., P,, (words andphrases can have several
meanings)
Disambiguation: H infers that S intended to convey P, (where ideally Pj = P,but misinterpretation is
possible)
Incorporation: H decides to believe P, (or rejects it if it is out of line with
whatH already believes)

7
Intention :
 Speaker decides that there is something that is worth saying tothe hearer.
 This often involves reasoning about the beliefs and goals of the hearer, so that theutterance
will have the desired effect.
 For our example, the speaker has the intention of havingthe hearer know that the wumpus is no
longer alive.
Generation :
 The speaker uses knowledge about language to decide what to say.
 This is harder than the inverse problem of understanding (i.e., analysis and
disambiguation).Generation has not been stressed as much as understanding in AI, mainly
because we humansare anxious to talk to machines, but are not as excited about them talking
back.
 For now, we justassume the hearer is able to choose the words "The wumpus is dead."
Synthesis :
 Most language-based AI systems synthesize typed output on a screen or paper,which is a trivial
task.
 Speech synthesis has been growing in popularity, and some systems arebeginning to sound
human.
 The agent in the below example is synthesizing a string of soundswritten in the phonetic
alphabet "[thaxwahmpahsihzdeyd]."
 The sounds that get synthesized are different fromthe words that the agent generates.
 Also note that the words are run together; this is typical ofquickly spoken speech.
Perception:
When the medium is speech, the perception step is called speech recognition;when it is printing, it is
called optical character recognition. Both have moved from being

esoteric to being commonplace within the last five years. For the example, let us assume that the
hearer perceives the sounds and recovers the spoken words perfectly.

8
Analysis:
Two main parts:
syntactic interpretation (or parsing) and semantic interpretation.

 Semantic interpretation includes both understanding the meanings of words and

incorporating knowledge of the current situation (also called pragmatic interpretation).
 The word parsing is derived from the the Latin phrase pars orationis, or "part of speech," and
refers to the process of assigning a part of speech (noun, verb, and so on) to each word in a
sentence and grouping the words into phrases.
 One way of displaying the result of a syntactic analysis is with a parse tree.
 A parse tree is a tree in which interior nodes represents phrases, links represent applications of
grammar rules, and leaf nodes represent words.
 We define the yield of a node as the list of all the leaves below the node, in left-to-right
order, then we can say that the meaning of a parse tree is that each node with label X asserts
that the yield of the node is a phrase of category X.
 Semantic interpretation is the process of extracting the meaning of an utterance as an
expression in some representation language.
 Two possible semantic interpretations: that the wumpus is not alive, and that it is tired (a
colloquial meaning of dead).
Ambiguity
 Utterances with several possible interpretations are said to be ambiguous.
 lexical ambiguity, where a word has more than one meaning.
 Syntactic ambiguity (also known as structural ambiguity) can occur with or without lexical
ambiguity.
For example, the string "I smelled a wumpus in 2,2" has two parses: one where the propositional
phrase modifies the noun, and one where it modifies the verb.
 semantic ambiguity :The syntactic ambiguity leads to a semantic ambiguity, because one
parse means the wumpus is in 2,2 and the other means that a stench is in 2,2.
 Referential ambiguity occurs because natural languages consist almost entirely of words for
categories, not for individual objects. There is no word for the-apple-I-had-for-lunch-
today,]\\sicategories like apple.
 Pragmatic ambiguity occurs when the speaker and hearer disagree on what the current
situation is. If the speaker says "I'll meet you next Friday" thinking that they're talking about
the 17th, and the hearer thinks that they are talking about the 24th, then there is
miscommunication.
 Local ambiguity :Sometimes a phrase or sentence has local ambiguity, where a substring can
be parsed several ways, but only one of those ways fits into the larger context of the whole
string. For example, in the C programming language, the string *c means "pointer to c" when it
appears in the declaration char *c; but it means "multiply by c" when it appears in the
expression 2 *c.
.
Disambiguation.
 Most speakers are not intentionally ambiguous, but most utterances have several legal
interpretations. Communication works because the hearer does the work of figuring out which
interpretation is the one the speaker probably meant to convey.
 Disambiguation is the first process that depends heavily on uncertain reasoning. Analysis
generates possible interpretations; if more than one interpretation is found, then disambiguation
chooses the one that is best.

1. The world model: the probability that a fact occurs in the world.
2. The mental model: the probability that the speaker forms the intention of communicating this fact
to the hearer, given that it occurs.9 (This combines models of what the speaker believes, what the
speaker believes the hearer believes, and so on.)
3. The language model: the probability that a certain string of words will be chosen, given that the
speaker has the intention of communicating a certain fact.
9
4. The acoustic model: the probability that a particular sequence of sounds will be generated, given
that the speaker has chosen a given string of words

Incorporation.A totally naive agent might believe everything it hears, but a sophisticated agent treats
the words W and the derived interpretation P, as additional pieces of evidence that get considered
along with all other evidence for and against P,.

It only makes sense to use language when there are agents to communicate with who
(a) understand a common language,
(b) have a shared context on which to base the conversation, and
(c) are at least somewhat rational.

TWO MODELS OF COMMUNICATION

Study of communication centers on the way that an agent's beliefs are turned into words and back into
beliefs in another agent's knowledge base (or head). There are two ways of looking at ENCODED
MESSAGE the process.

ENCODED MESSAGE
The encoded message model says that the speaker has
o a definite proposition P in mind, and
o encodes the proposition into the words (or signs) W.
The hearer then tries to
o decode the message W to retrieve the original proposition P (cf. Morse code).

Under this model, the meaning in the speaker's head, the message that gets transmitted, and the
interpretation that the hearer arrives at all ideally carry the same content. When they differ, it is
because of noise in the communication channel or an error in encoding or decoding.

SITUATED LANGUAGE
Limitations of the encoded message model led to the situated language model, which says that the
meaning of a message depends on both the words and the situation in which the words are uttered. In
this model, just as in situation calculus, the encoding and decoding functions take an extra argument
representing the current situation.

This accounts for the fact that the same words can have very different meanings in different situations.

"I am here now" represents one fact when spoken by Peter in Boston on Monday, and quite another
fact when spoken by Stuart in Berkeley on Tuesday.

More subtly, "You must read this book" is a suggestion when written by a critic in the newspaper, and
an assignment when spoken by an instructor to the class. "Diamond" means one thing when the subject
is jewelry, and another when the subject is baseball.

The situated language model points out a possible source of communication failure: if the speaker and
hearer have different ideas of what the current situation is, then the message may not get through as
intended.

For example, suppose agents X, Y, and Z are exploring the wumpus world together. X and Y secretly
meet and agree that when they smell the wumpus they will both shoot it, and if they see the gold, they
will grab it and run, and try to keep Z from sharing it. Now supposeX smells the wumpus, while Y in
the adjacent square smells nothing, but sees the gold.

X yells "Now!," intending it to mean that now is the time to shoot the wumpus, but Y interprets itas
meaning now is the time to grab the gold and run.

10
 4. INFORMATION RETRIEVAL
Information retrieval is the task of finding documents that are relevant to a user’s need for information.
The best-known examples of information retrieval systems are search engines on the World Wide Web. A
Web user can type a query such as [AI book] into a search engine and see a list of relevant pages.

An information retrieval (henceforth IR) system can be characterized by

1. A corpus of documents. Each system must decide what it wants to treat as a document: a paragraph, a
page, or a multipage text.
2. Queries posed in a query language. A query specifies what the user wants to know.
The query language can be just a list of words, such as [AI book]; or it can specify a phrase of words
that must be adjacent, as in [―AI book‖]; it can contain Boolean operators as in [AI AND book]; it can
include non-Boolean operators such as [AI NEAR book] or [AI book site:www.aaai.org].
3. A result set. This is the subset of documents that the IR system judges to be relevant to the query. By
relevant, we mean likely to be of use to the person who posed the query, for the particular information need
expressed in the query.

4. A presentation of the result set. This can be as simple as a ranked list of document titles or as complex as a
rotating color map of the result set projected onto a three- dimensional space, rendered as a two-dimensional
display.
Boolean keyword model
Each word in the document collection is treated as a Boolean feature that is true of a document if the word
occurs in the document and false if it does not. The query language is the language of Boolean expressions
over features. A document is relevant only if the expression evaluates to true.

Advantage
1 . Simple to explain and implement.
Disadvantages
1. The degree of relevance of a document is a single bit, so there is no guidance as to how to order the
relevant documents for presentation.
2. Boolean expressions are unfamiliar to users who are not programmers or logicians.

IR scoring functions
BM25 scoring function

11
A scoring function takes a document and a query and returns a numeric score; the most relevant documents have
the highest scores. In the BM25 function, the score is a linear weighted combination of scores for each of the
words that make up the query.
Three factors affect the weight of a query term:
1. Frequency with which a query term appears in a document (also known as TF for term frequency).
For the query [farming in Kansas], documents that mention ―farming‖ frequently will have higher
scores.
2. Inverse document frequency of the term, or IDF . The word ―in‖ appears in almost every document, so
it has a high document frequency, and thus a low inverse document frequency, and thus it is not as
important to the query as ―farming‖ or ―Kansas.‖
3. Length of the document. A million-word document will probably mention all the query words, but may
not actually be about the query. A short document that mentions all the words is a much better candidate
BM25systems create an index ahead of time that lists, for each vocabulary word, the documents that contain the
word. This is called the hit list for the word. Then when given a query, we intersect the hit lists of the query
words and only score the documents in the intersection.

IR system evaluation
Two measures used in the scoring: recall and precision.
Precision measures the proportion of documents in the result set that are actually relevant.
Recall measures the proportion of all the relevant documents in the collection that are in the result set
A summary of both measures is the F1 score, a single number that is the harmonic mean of precision and
recall, 2P R/(P + R).

IR refinements
There are many possible refinements to the system described here, and indeed Web search engines are
continually updating their algorithms as they discover new approaches and as the Web grows and changes.

pivoted document length normalization scheme; the idea is that the pivot is the document length at which the
old-style normalization is correct; documents shorter than that get a boost and longer ones get a penalty.

Most IR systems do case folding of ―COUCH‖ to ―couch,‖ and some use a stemming algorithm to reduce.
―couches‖ to the stem form ―couch,‖ both in the query and the documents. This typically yields a small increase
in recall but can harm precision

Recognize synonyms, such as ―sofa‖ for ―couch.‖

anytime there are two words that mean the same thing, speakers of the language conspire to evolve the meanings
to remove the confusion.
Related words that are not synonyms also play an important role in ranking—terms like ―leather‖, ―wooden,‖
or ―modern‖ can serve to confirm that the document really is about ―couch.‖

Synonyms and related words can be found in dictionaries or by looking for correlations in documents or in
queries—if we find that many users who ask the query [new sofa] follow it up with the query [new couch], we
can in the future alter [new sofa] to be [new sofa OR new couch]

As a final refinement, IR can be improved by considering metadata—data outside of the text of the document.
Examples include human-supplied keywords and publication data. On the Web, hypertext links between
documents are a crucial source of information

12
The PageRank algorithm
PageRank3 was one of the two original ideas that set Google’s search apart from other Web search engines
when it was introduced in 1997.

PageRank was invented to solve the problem of the tyranny of TF scores: if the query is [IBM], how do we make
sure that IBM’s home page, ibm.com, is the first result, even if another page mentions the term ―IBM‖ more
frequently? The idea is that ibm.com has many in-links (links to the page), so it should be ranked higher: each in-
link is a vote for the quality of the linked-to page.

The HITS algorithm

The Hyperlink-Induced Topic Search algorithm, also known as ―Hubs and Authorities‖ or HITS, is
another influential link-analysis algorithm.

HITS differs from PageRank in several ways.

First, it is a query-dependent measure: it rates pages with respect to a query. That means that it must be
computed anew for each query—a computational burden that most search engines have elected not to
take on.

Given a query, HITS first finds a set of pages that are relevant to the query.
It does that by intersecting hit lists of query words, and then adding pages in the link neighborhood of
these pages—pages that link to or are linked from one of the pages in the original relevant set.

Each page in this set is considered an authority on the query to the degree that other pages in the
relevant set point to it. A page is considered a hub to the degree that it points to other authoritative
pages in the relevant set. Just as with PageRank, we don’t want to merely count the number of links;
we want to give more value to the high-quality hubs and authorities

Question answering
Information retrieval is the task of finding documents that are relevant to a query, where the query may
be a question, or just a topic area or concept. Question answering is a somewhat different task, in
which the query really is a question, and the answer is not a ranked list of documents but rather a short
response—a sentence, or even just a phrase.

Once the n-grams are scored, they are filtered by expected type.
If the original query starts with
―who,‖ then we filter on names of people;
for ―how many‖ we filter on numbers,
13
 for ―when,‖ on a date or time.
There is also a filter that says the answer should not be part of the question;

5. INFORMATION EXTRACTION
Information extraction is the process of acquiring knowledge by skimming a text and looking for
occurrences of a particular class of object and for relationships among objects. A typical task is to extract
instances of addresses from Web pages, with database fields for street, city, state, and zip code; or instances of
storms from weather reports, with fields for temperature, wind speed, and precipitation.

Six different approaches to information extraction

1. Finite-state automata for information extraction

2. Probabilistic models for information extraction
3. Conditional random fields for information extraction
4. Ontology extraction from large corpora
5. Automated template construction
6. Machine reading

Finite-state automata for information extraction

The simplest type of information extraction system is an attribute-based extraction system that assumes that the
entire text refers to a single object and the task is to extract attributes of that object.

The problem of extracting from the text ―IBM ThinkBook 970. Our price: $399.00‖ the set of attributes
{Manufacturer=IBM, Model=ThinkBook970, Price=$399.00}. We can address this problem by defining a
tem- plate (also known as a pattern) for each attribute we would like to extract

Templates are often defined with three parts:

1. A prefix regex,
2. A target regex, and
3. A postfix regex.
For prices, the target regex is as above, the prefix would look for strings such as ―price:‖ and the postfix could
be empty. The idea is that some clues about an attribute come from the attribute value itself and some come
from the surrounding text.

If a regular expression for an attribute matches the text exactly once, then we can pull
out the portion of the text that is the value of the attribute. If there is no match, all we can do is give a default
value or leave the attribute missing; but if there are several matches, we need a process to choose among them.
One strategy is to have several templates for each attribute, ordered by priority.

One step up from attribute-based extraction systems are relational extraction systems, which deal with multiple
objects and the relations among them. Thus, when these systems see the text ―$249.99,‖ they need to
determine not just that it is a price, but also which object has that price. A typical relational-based extraction
system is FASTUS, which handles news stories about corporate mergers and acquisitions

FASTUS consists of five stages:

1. Tokenization: segments the stream of characters into tokens
2. Complex-word handling : collocations such as ―set up‖ and ―joint venture,‖ as well as proper names such
as ―Bridgestone Sports Co.‖
3. Basic-group handling : noun groups and verb groups
4. Complex-phrase handling : combines the basic groups, finite-state and thus can be processed quickly
5. Structure merging : merges structures that were built up in the previous step
Disadvantage:
When information extraction must be attempted from noisy or varied input, simple finite-state approaches fare
poorly. It is too hard to get all the rules and their priorities right;

Probabilistic models for information extraction

14
. The simplest probabilistic model for sequences with hidden state is the hidden Markov model, or HMM.
HMM models a progression through a sequence of hidden states, xt , with an observation et at each step.
The observations are the words of the text, and the hidden states are whether we are in the target, prefix, or post
fix part of the attribute template, or in the background.

HMMs have two big advantages over FSAs for extraction.

 First, HMMs are probabilistic, and thus tolerant to noise. In a regular expression, if a single expected
character is missing, the regex fails to match; with HMMs there is graceful degradation with missing
characters/words, and we get a probability indicating the degree of match, not just a Boolean match/fail.
 Second HMMs can be trained from data; they don’t require laborious engineering of templates, and
thus they can more easily be kept up to date as text changes over time.

HMM templates:
 Consist of one or more target states,
 Prefix states must precede the targets,
 Postfix states most follow the targets, and
 other states must be background.

With sufficient training data, the HMM automatically learns a structure

Once the HMMs have been learned, we can apply them to a text

Conditional random fields for information extraction

 conditional random field, or CRF, which models a conditional probability distribution of a set of target
variables given a set of observed variables
 Like Bayesian networks, CRFs can represent many different structures of dependencies among the
variables. One common structure is the linear-chain conditional random field for repre- senting
Markov dependencies among variables in a temporal sequence.
 Conditional Random Fields (CRFs) are the state of the art approaches taking the sequence
characteristics to do better labeling. However, as the information on a Web page is two-dimensionally
laid out, previous linear-chain CRFs have their limitations for Web information extraction.
 The feature functions are the key components of CRF.
 The general form of a feature function is fi(zn−1, zn, x1:N , n), which looks at a pair of adjacent states
zn−1, zn, the whole input sequence x1:N , and where we are in the sequence. The feature functions
produce a real value
 Features are not limited to binary functions. Any real-valued function is allowed. Designing the features
of an CRF is the most important task. In CRFs for real applications it is not uncommon to have tens of
thousands or more features.
 CRF formalism gives us a great deal of flexibility in defining them. This flexibility can lead to
accuracies that are higher than with less flexible models such as HMMs

Ontology extraction from large corpora

Information extraction is finding a specific set of relations (e.g., speaker, time, location) in a specific text (e.g.,
a talk announcement). A different applica- tion of extraction technology is building a large knowledge base or
ontology of facts from a corpus.

1 . First it is open-ended—we want to acquire facts about all types of domains, not just one specific
domain.
2 . Second, with a large corpus, this task is dominated by precision, not recall—just as with question
answering on the Web.
3. Third, the results can be statistical aggregates gathered from multiple sources, rather than being extracted
from one specific text.

templates:
NP such as NP (, NP )* (,)? ((and | or) NP )? .
NP is a variable standing for a noun phrase

This template matches the texts ―diseases such as rabies affect your dog‖ and ―supports network protocols such

15
as DNS,‖ concluding that rabies is a disease and DNS is a network protocol.

With a large corpus we can afford to be picky; to use only the high-precision templates. We’ll miss many
statements of a subcategory relationship, but most likely we’ll find a paraphrase of the statement somewhere
else in the corpus in a form we can use.

Automated template construction

The subcategory relation is so fundamental that is worthwhile to handcraft a few templates to help identify
instances of it occurring in natural language text.

It is possible to learn templates from a few examples, then use the templates to learn more examples, from
which more templates can be learned, and so on. In one of the first experiments of this kind, Brin (1999) started
with a data set of just five examples:
(―Isaac Asimov‖, ―The Robots of Dawn‖)
(―David Brin‖, ―Startide Rising‖)
(―James Gleick‖, ―Chaos—Making a New Science‖)
(―Charles Dickens‖, ―Great Expectations‖) (―William
Shakespeare‖, ―The Comedy of Errors‖)
Clearly these are examples of the author–title relation, but the learning system had no knowl- edge of authors or
titles.
Each match is defined as a tuple of seven strings,
(Author, Title, Order, Prefix, Middle, Postfix, URL) ,
where Order is true if the author came first and false if the title came first, Middle is the characters between the
author and title, Prefix is the 10 characters before the match, Suffix is the 10 characters after the match, and URL
is the Web address where the match was made.

 Given a set of matches, a simple template-generation scheme can find templates to explain the
matches.
 The biggest weakness in this approach is the sensitivity to noise. If one of the first few templates is
incorrect, errors can propagate quickly.
 One way to limit this problem is to not accept a new example unless it is verified by multiple
templates, and not accept a new template unless it discovers multiple examples that are also found by
other templates.

Machine reading
 Automated template construction is a big step up from handcrafted template construction, but it still
requires a handful of labeled examples of each relation to get started.
 To build a large ontology with many thousands of relations, even that amount of work would be onerous;
 we would like to have an extraction system with no human input of any kind—a system that could read on
its own and build up its own database.
 Such a system would be relation-independent; would work for any relation.
 These systems work on all relations in parallel, because of the I/O demands of large corpora. They
behave less like a traditional information- extraction system that is targeted at a few relations and more
like a human reader who learns from the text itself; because of this the field has been called machine
reading.
 Given a set of labeled examples of this type, T EX T RUNNER trains a linear-chain CRF to extract further
examples from unlabeled text. The features in the CRF include function words like ―to‖ and ―of‖ and
―the,‖ but not nouns and verbs (and not noun phrases or verb phrases). Because T EX T RUNNER is
domain-independent, it cannot rely on predefined lists of nouns and verbs.

16
 6. NATURAL LANGUAGE PROCESSING
1. Natural Language Understanding

 Taking some spoken/typed sentence and working out what it means

2. Natural Language Generation
 Taking some formal representation of what you want to say and working out a way to
express it in a natural (human) language (e.g., English)
Applications of NLP
 Machine Translation
 Database Access
 Information Retrieval
– Selecting from a set of documents the ones that are relevant to a query
 Text Categorization
– Sorting text into fixed topic categories
 Extracting data from text
– Converting unstructured text into structured data
 Spoken language control systems
 Spelling and grammar checkers

Natural language understanding

Raw speech signal
 Speech recognition
Sequence of words spoken
 Syntactic analysis using knowledge of the grammar
Structure of the sentence
 Semantic analysis using info. about meaning of words
Partial representation of meaning of sentence
 Pragmatic analysis using info. about context
Final representation of meaning of sentence

L ANGUAGE M OD ELS
Formal languages, such as the programming languages Java or Python, have precisely defined language models.
A language can be defined as a set of strings; ―print(2 + 2)‖ is a legal program in the language Python,
whereas ―2)+(2 print‖ is not. Since there are an infinite number of legal programs, they cannot be

17
enumerated; instead they are specified by a set of rules called a grammar. Formal languages also have rules that
define the meaning or semantics of a program; for example, the rules say that the ―meaning‖ of ―2 + 2‖ is 4,
and the meaning of ―1/0‖ is that an error is signalled

N-gram character models

Ultimately, a written text is composed of characters—letters, digits, punctuation, and spaces in English
A model of the probability distribution of n-letter sequences is thus called an n-gram model
An n-gram model is defined as a Markov chain of order n – 1
P (ci | c1:i−1 ) = P (ci | ci−2:i−1 ) .

For a trigram character model in a language with 100 characters, P(Ci |Ci−2:i−1 ) has a million entries, and can be
accurately estimated by counting character sequences in a body of text of
10 million characters or more. We call a body of text a corpus (plural corpora), from the
Latin word for body

language identification: given a text, determine what natural language it is written in One approach to language
identification is to first build a trigram character model of each candidate language, P (ci | ci−2:i−1 , ), where the
variable ranges over languages. For each the model is built by counting trigrams in a corpus of that language.
Genre classification
Genre classification means deciding if a text is a news story, a legal document, a scientific article, etc. While
many features help make this classification, counts of punctuation and other character n-gram features go a
long way (Kessler et al.,
1997).
Named-entity recognition
Named-entity recognition is the task of finding names of things in a document and deciding what class they
belong to. For example, in the text ―Mr. Sopersteen was prescribed aciphex,‖ we should recognize that ―Mr.
Sopersteen‖ is the name of a person and ―aciphex‖ is the name of a drug. Character-level models are good for this
task because they can associate the character sequence ―ex ‖ (―ex‖ followed by a space) with a drug name and
―steen ‖ with a person name, and thereby identify words that they have never seen before

Smoothing n-gram models

The major complication of n-gram models is that the training corpus provides only an esti- mate of the true
probability distribution
we will adjust our language model so that sequences that have a count of zero in the training corpus will be
assigned a small nonzero probability (and the other counts will be adjusted downward slightly so that the
probability still sums to 1). The process od adjusting the probability of low-frequency counts is called
smoothing.
simplest type of smoothing was suggested by Pierre-Simon Laplace in the 18th cen- tury: he said that, in the lack of
further information, if a random Boolean variable X has been false in all n observations so far then the estimate
for P (X = true ) should be 1/(n + 2).
A better approach is a backoff model, in which we start by estimating n-gram counts, but for
any particular sequence that has a low (or zero) count, we back off to (n − 1)-grams. Linear
interpolation smoothing is a backoff model that combines trigram, bigram, and unigram models by linear
interpolation.

Model evaluation
With so many possible n-gram models—unigram, bigram, trigram, interpolated smoothing with different values
of λ, etc.—how do we know what model to choose? We can evaluate a model with cross-validation Split the
corpus into a training corpus and a validation corpus. Determine the parameters of the model from the training
data. Then evaluate the model on the validation corpus.

18
The evaluation can be a task-specific metric, such as measuring accuracy on language identification.
Alternatively we can have a task-independent model of language quality: cal- culate the probability assigned to
the validation corpus by the model; the higher the proba- bility the better. This metric is inconvenient because
the probability of a large corpus will be a very small number, and floating-point underflow becomes an issue.
A different way of describing the probability of a sequence is with a measure called perplexity, defined as

Perplexity (c1:N ) = P (c1:N ) N .

Perplexity can be thought of as the reciprocal of probability, normalized by sequence length

N-gram word models

Word n-gram models need to deal with out of vocabulary words With character mod- els, we didn’t have to
worry about someone inventing a new letter of the alphabet.1 But with word models there is always the
chance of a new word that was not seen in the training corpus, so we need to model that explicitly in our
language model.

Text categorization
NLP techniques have proven successful in a related task: sorting text into fixed topic categories.
There are several commercial services that provide access to news wire stories in this manner.
A subscriber can ask for all the news on a particular industry, company, or geographic area,
for example. The providers of these services have traditionally used human experts to assign

the categories. In the last few years, NLP systems have proven to be just as accurate, correctly
categorizing over 90% of the news stories. They are also far faster and more consistent, so there
has been a switch from humans to automated systems.
Text categorization is amenable to NLP techniques where IR is not because the categories
are fixed, and thus the system builders can spend the time tuning their program to the problem.
For example, in a dictionary, the primary definition of the word "crude" is vulgar, but in a large
sample of the Wall Street Journal, "crude" refers to oil 100% of the time.

7. MACHINE TRANSLATION
Machine translation (MT) is automated translation. It is the process by which computer software is
used to translate a text from one natural language (such as English) to another (such as Spanish).

To do translation well, a translator (human or machine) must read the original text, understandthe
situation to which it is referring, and find a corresponding text in the target language that does a
good job of describing the same or a similar situation.

Often this involves a choice. For example, the English word "you" can be translated into French as
either the formal "vous" or the informal "tu." There is just no way that one can refer to the concept of
19
"you" in French without also making a choice of formal or informal. Translators (both machine and
human) sometimes find it difficult to make this choice.

To process any translation, human or automated, the meaning of a text in the original (source)
language must be fully restored in the target language, i.e. the translation. While on the surface this
seems straightforward, it is far more complex.

Translation is not a mere word-for-word substitution. A translator must interpret and analyze all of the
elements in the text and know how each word may influence another. This requires extensive expertise
in grammar, syntax (sentence structure), semantics (meanings), etc., in the source and target languages,
as well as familiarity with each local region.

Human and machine translation each have their share of challenges. For example, no two individual
translators can produce identical translations of the same text in the same language pair, and it may
take several rounds of revisions to meet customer satisfaction. But the greater challenge lies in how
machine translation can produce publishable quality translations.

A representation language that makes all the distinctions necessary for a set of languages is called an
interlingua. Some systems attempt to analyze the source language text all the way into an interlingua knowledge
representation and then generate sentences in the target language from that representation.

This is difficult because it involves three unsolved problems:

1. Creating a complete knowledge representation of everything;
2. Parsing into that representation; and
3. Generating sentences from that representation.

Rule-Based Machine Translation Technology

 Rule-based machine translation relies on countless built-in linguistic rules and millions of bilingual
dictionaries for each language pair.
 The software parses text and creates a transitional representation from which the text in the target
language is generated.
 This process requires extensive lexicons with morphological, syntactic, and semantic information,
and large sets of rules. The software uses these complex rule sets and then transfers the
grammatical structure of the source language into the target language.
 Translations are built on gigantic dictionaries and sophisticated linguistic rules. Users can improve
the out-of-the-box translation quality by adding their terminology into the translation process. They
create user-defined dictionaries which override the system’s default settings.

In most cases, there are two steps:

 An initial investment that significantly increases the quality at a limited cost, and
 An ongoing investment to increase quality incrementally.
 While rule-based MT brings companies to the quality threshold and beyond, the quality
improvement process may be long and expensive.

Statistical Machine Translation Technology

20
Statistical machine translation utilizes statistical translation models whose parameters stem from the
analysis of monolingual and bilingual corpora.

Building statistical translation models is a quick process, but the technology relies heavily on
existing multilingual corpora.

A minimum of 2 million words for a specific domain and even more for general language are required.
Theoretically it is possible to reach the quality threshold but most companies do not have such large
amounts of existing multilingual corpora to build the necessary translation models.

Additionally, statistical machine translation is CPU intensive and requires an extensive hardware
configuration to run translation models for average performance levels.

Rule-Based MT vs. Statistical MT

 Rule-based MT provides good out-of-domain quality and is by nature predictable.
 Dictionary-based customization guarantees improved quality and compliance with corporate
terminology. But translation results may lack the fluency readers expect.
 In terms of investment, the customization cycle needed to reach the quality threshold can be long
and costly. The performance is high even on standard hardware.

Statistical MT provides good quality when large and qualified corpora are available. The translation is
fluent, meaning it reads well and therefore meets user expectations. However, the translation is neither
predictable nor consistent. Training from good corpora is automated and cheaper. But training on
general language corpora, meaning text other than the specified domain, is poor. Furthermore,
statistical MT requires significant hardware to build and manage large translation models.

Transfer model
They keep a database of translation rules (or examples), and whenever the rule (or example) matches, they
translate directly.
Transfer can occur at the lexical, syntactic, or semantic level.
For example, a strictly syntactic rule maps English [Adjective Noun] to French [Noun Adjective]. A mixed
syntactic and lexical rule maps French [S1 ―et puis‖ S2] to English [S1 ―and then‖ S2].

21
All that remains is to learn the phrasal and distortion probabilities
1. Find parallel texts
2. Segment into sentences
3. Align sentences
4. Align phrases
5. Extract distortions
6. Improve estimates with EM
EXAMPLE
TAUM-METEO system, developed by the University of Montreal, which translates weather reports
from English to French. It works because the language used in these government weather reports is
highly stylized and regular.

A representative system is SPANAM (Vasconcellos and Leon, 1985), which can translate a Spanish
passage into English of this quality.

 This is mostly understandable, but not always grammatical and rarely fluent.

22
 Another possibility is to invest the human effort on pre-editing the original document.
 If the original document can be made to conform to a restricted subset of English (or whatever
the original language is), then it can sometimes be translated without the need for post-editing.
 This approach is particularly cost-effective when there is a need to translate one document into
many languages, as is the case for legal documents

The problem is that different languages categorize the world differently. A majority of the situations
that are covered by the English word "open" are also covered by the German word "offen," but the
boundaries of the category differ across languages.
In English, we extend the basic meaning of "open" to cover open markets, open questions, and open
job offerings.
In German, the extensions are different. Job offerings are "freie," not open, but the concepts of loose
ice, private firms, and blank checks all use a form of "offen."

8. DATABASE ACCESS
The first major success for natural language processing (NLP) was in the area of database access.
The disadvantage is that the user never knows which wordings of a query will succeed and which are
outside the system's competence some commercial systems have built up large enough grammars and
lexicons to handle a fairly wide variety of inputs. The main challenge for current systems is to follow
the context of an interaction. The user should be able to ask a series of questions where some of them
implicitly refer to earlier questions or answers:
What countries are north of the equator?
How about south?
Show only the ones outside Australasia.
What is their total area?
Some systems (e.g., TEAM (Grosz et al., 1987)) handle problems like this to a limited degree,

Natural Language Inc. and Symantec are still selling database access tools that use natural language,
but customers are less likely to make their buying decisions based on the strength of the natural
language component than on the graphical user interface or the degree of integration of the database
with spreadsheets and word processing.
Natural language is not always the most natural way to communicate: sometimes it is easier to point
and click with a mouse to express an idea.
The emphasis in practical NLP has now shifted away from database access to the broad
interpretation.

In part, this is a reflection of a change in the computer industry.

In the early 1980s, most online information was stored in databases or spreadsheets. Now the majority
of online information is text: email, news, journal articles, reports, books, encyclopedias. Most
computer users find there is too much information available, and not enough time to sort through it.
Text interpretation programs help to retrieve, categorize, filter, and extract information from text. Text
interpretation systems can be split into three types: information retrieval, text categorization, and data
extraction.

9. SPEECH RECOGNITION
It is the task of mapping from a digitally encoded acoustic signal to a string of words. Speech
understanding is the task of mapping from the acoustic signal all the way to an interpretation of the
meaning of the utterance.
A speech understanding system must answer three questions:
1. What speech sounds did the speaker utter?
2. What words did the speaker intend to express with those speech sounds?
3. What meaning did the speaker intend to express with those words?

23
To answer question 1, we have to first decide what a speech sound is. It turns out that all human
languages use a limited repertoire of about 40 or 50 sounds, called phones. Roughly speaking, a phone
is the sound that corresponds to a single vowel or consonant, but there are some complications:

Combinations of letters such as "th" and "ng" produce single phones, and some letters produce
different phones in different contexts (for example, the "a" in rat and rate. Once we know what the
possible sounds are, we need to characterize them in terms of features that we can pick out of the
acoustic signal, such as the frequency or amplitude of the sound waves.

Question 2 is conceptually much simpler. You can think of it as looking up words in a dictionary that
is arranged by pronunciation.
We get a sequence of three phones, [k], [ce], and ft], and find in the dictionary that this is the
pronunciation for the word "cat."

Two things make this difficult.

The first is the existence of homophones, different words that sound the same, like "two" and "too."1
The second is segmentation, the problem of deciding where one word ends and the next begins.

Question 3 we already know how to answer—use the parsing and analysis algorithms
Some speech understanding systems extract the most likely string of words and pass them directly to
an analyzer. Other systems have a more complex control structure that considers multiple possible
word interpretations so that understanding can be achieved even if some individual words are not
recognized correctly.

Signal processing
Sound is an analog energy source. When a sound wave strikes a microphone, it is converted to an
electrical current, which can be passed to an analog-to-digital converter to yield a stream of bits
representing the sound.

We have two choices in deciding how many bits to keep.

First, the sampling rate is the frequency with which we look at the signal. For speech, a sampling rate
between 8 and 16 KHz (i.e., 8 to 16,000 times per second) is typical. Telephones deliver only
about 3 KHz.
Second, the quantization factor determines the precision to which the energy at each sampling point
is recorded. Speech recognizers typically keep 8 to 12 bits. That means that a low-end system,
sampling at 8 KHz with 8-bit quantization, would require nearly half a megabyte per minute of speech.

The first step in coming up with a better representation for the signal is to group the samples
together into larger blocks called frames. This makes it possible to analyze the whole frame for the
appearance of speech phenomena such as a rise or drop in frequency, or a sudden onset or cessation of
energy. Within each frame, we represent what is happening with a vector of features.
Figure 24.33 shows frames with a vector of three features. Note that the frames overlap;
this prevents us from losing information if an important acoustic event just happens to fall on a
frame boundary.

24
The final step in many speech signal processing systems is vector quantization. If there are n features
in a frame, we can think of this as an n-dimensional space containing many points.
Some information is lost in going from a feature vector to a label that summarizes a whole
neighborhood around the vector, but there are automated methods for choosing an optimal quantization
of the feature vector space so that little or no inaccuracy is introduced (Jelinek, 1990).

First, we end up with a representation of the speech signal that is compact. But more importantly, we
have a representation that is likely to encode features of the signal that will be useful for word
recognition. A given speech sound can be pronounced so many ways: loud or soft, fast or slow, high-
pitched or low, against a background of silence or noise, and by any of millions of different speakers
each with different accents and vocal tracts. Signal processing hopes to capture enough of the
important features so that the commonalities that define the sound can be picked out from this
backdrop of variation.
The dual problem, speaker identification, requires one to focus on the variation instead of the
commonalities in order to decide who is speaking.

Defining the overall speech recognition model

Speech recognition is the diagnostic task of recovering the words that produce a given acoustic signal.
It is a classic example of reasoning with uncertainty. We are uncertain about how well the
microphones (and digitization hardware) have captured the actual sounds, we are uncertain about
which phones would give rise to the signal, and we are uncertain about which words would give rise to
the phones. As is often the case, the diagnostic task can best be approached with a causal model—the
words cause the signal.

We can break this into components with Bayes' rule:

Given a signal, our task is to find the sequence of words that maximizes P(words\signal). Of the three
components on the right-hand side, P(signal) is a normalizing constant that we can ignore. P(words) is
known as the language model. It is what tells us, when we are not sure if we heard

"bad boy" or "pad boy" that the former is more likely. Finally, P(signal\words) is the acoustic model.
It is what tells us that "cat" is very likely to be pronounced [kcet].

The language model: P( words)

25
Assign a probability to each of the (possibly infinite) number of strings. Context-free grammars are no
help for this task, but probabilistic context-free grammars (PCFGs) are promising.
Bigram model : it says that the probability of any given word is determined solely by the previous
word in the string

Trigrammodel that provides values for />(w;|w,-_iw/_2). This is a more powerful language model,
capable of determining that "ate a banana" is more likely than "ate a bandana." The problem is that
there are so many more parameters in trigram models that it is hard to get enough training data to come
up with accurate probability estimates.

The acoustic model: P(signallwords)

The acoustic model is responsible for saying what sounds will be produced when a given string of
words is uttered. We divide the model into two parts. First, we show how each word is described as a
sequence of phones, and then we show how each phone relates to the vector quantization values
extracted from the acoustic signal.
Markov model
Two sources of phonetic
First, different dialects have different pronunciations.
The top of Figure. gives an example of this: for "tomato," you say [tow mey tow] and I say [tow maa
tow]. The alternative pronunciations are specified as a Markov model.

In general, a Markov model is a way of describing a process that goes through a series of states.
The model describes all the possible paths through the state space and assigns a probability to each
one. The probability of transitioning from the current state to another one depends only on the current
state, not on any prior part of the path.

Markov model with seven states (circles), each corresponding to the production of a phone. The
arrows denote allowable transitions between states, and each transition has a probability associated
with it.3 There are only two possible paths through the model, one corresponding to the phone
sequence [t ow m ey t ow] and the other to [t ow m aa tow]. The probability of a path is the product of
the probabilities on the arcs that make up the path. In this case, most of the arc probabilities are 1 and
we have

P([towmeytow] ("tomato") = P([r«w»iaa?ow] ("tomato") = 0.5

The second source of phonetic variation is coarticulation.

26
Similar models would be constructed for every word we want to be able to recognize. Now if the
speech signal were a list of phones, then we would be done with the acoustic model. We could take a
given input signal (e.g., [towmeytow]) and compute P(signal\words) for various word strings (e.g.,
"tomato," "toe may tow," and so on). We could then combine these with P(words) values taken from
the language model to arrive at the words that maximize P(words\signal).

hidden Markov model or HMM

Unfortunately, signal processing does not give us a string of phones. So all we can do so far is
maximize P(words\phones). Figure 24.36 shows how we can compute P(signal\phone) using a model
called a hidden Markov model or HMM. The model is for a particular phone,[m], but all phones will
have models with similar topology.

A hidden Markov model is just like a regular Markov model in that it describes a process that goes
through a sequence of states. The difference is that in a regular Markov model, the output is a sequence
of state names, and because each state has a unique name, the output uniquely determines the path
through the model.
In a hidden Markov model, each state has a probability distribution of possible outputs, and the same
output can appear in more than one state.4
HMMs are called hidden models because the true state of the model is hidden from the observer. In
general, when you see that an HMM outputs some symbol, you can't be sure what state the symbol
came from.

Actually, most phones have a duration of 50-100 milliseconds, or 5-10 frames at 10 msec/frame. So
the [C1,C4,C6J sequence is unusually quick. Suppose we have a more typical speaker who generates
the sequence [C1,C1,C4,C4,C6,C6J while producing the phone. It turns out there are two paths
through the model that generate this sequence. In one of them both C4s come from the Mid state (note
the arcs that loop back), and in the other the second C4 comes from the End state. We calculate the
probability that this sequence came from the [m] model in the same way: take the sum over all possible
paths of the probability of the path times the probability that the path generates the sequence.

P([Cl,C\,C4, C4, C6, C6]jLm]) =

(0.3 x 0.7 x 0.9 x 0.1 x 0.4 x 0.6) x (0.5 x 0.5 x 0.7 x 0.7 x 0.5 x 0.5) +
(0.3 x 0.7 x 0.1 x 0.4 x 0.4 x 0.6) x (0.5 x 0.5 x 0.7 x 0.1 x 0.5 x 0.5)
= 0.0001477

27
Putting the models together
We have described three models.
The language bigram model gives us P(wordi\wordi-\}.
The word pronunciation HMM gives us P(phones\word).
The phone HMM gives us P(signal\phone).
If we want to compute P(words\signal), we will need to combine these models in some way.
Oneapproach is to combine them all into one big HMM. The bigram model can be thought of as an
HMM in which every state corresponds to a word and every word has a transition arc to every other
word. Now replace each word-state with the appropriate word model, yielding a bigger model in which
each state corresponds to a phone. Finally, replace each phone-state with the appropriate phone model,
yielding an even bigger model in which each state corresponds to a distribution of vector quantization
values.

Some speech recognition systems complicate the picture by dealing with coarticulation effects at either
the word/word or phone/phone level. For example, we could use one phone model for [ow] when it
follows a [t] and a different model for [ow] when it follows a [g]. There are many trade-offs to be
made—a more complex model can handle subtle effects, but it will be harder to train. Regardless of
the details, we end up with one big HMM that can be used to computeP(words\signal).

10. ROBOT (HARDWARE –PERCEPTION –PLANNING –MOVING)

Robots are physical agents that perform ROBOT tasks by manipulating the physical world. To do so
they are equipped with effectors such as legs, wheels, joints, and grippers Effectors have a single
purpose: to assert physical forces on the environment.

Robots are also equipped with sensors, which allow them to perceive their environment. Present day
robotics employs a diverse set of sensors, including cameras and lasers to measure the environment,
and gyroscopes and accelerometers to measure the robot’s own motion.
Manipulators, or robot arms are physically anchored to their workplace, for example in a factory
assembly line or on the International Space Station.

Mobile robot
Mobile robots move about their environment using wheels, legs, or similar mechanisms. They have
been put to use delivering food in hospitals, moving containers at loading docks, and similar tasks.
Unmanned ground vehicles, or UGVs, drive autonomously on streets, highways, and off-road. The
planetary rover explored Mars for a period of 3 months in 1997. Other types of mobile robots include
unmanned air vehicles (UAVs), commonly used for surveillance, crop-spraying,

Autonomous underwater vehicles (AUVs) are used in deep sea exploration. Mobile robots deliver
packages in the workplace and vacuum the floors at home.
The third type of robot combines mobility with manipulation, and is often called a mobile
manipulator. Humanoid robots mimic the human torso.

The field of robotics also includes prosthetic devices (artificial limbs, ears, and eyes for humans),
intelligent environments (such as an entire house that is equipped with sensors and effectors), and
multibody systems, wherein robotic action is achieved through swarms of small cooperating robots.
Real robots must cope with environments that are partially observable, stochastic, dynamic, and
continuous.

Robotics brings together many of the concepts, including probabilistic state estimation, perception,
planning, unsupervised learning, and reinforcement learning.

ROBOT HARDWARE

28
Sensors
Sensors are the perceptual interface between robot and environment.
Passive sensors, such as cameras, are true observers of the environment: they capture signals that are
generated by other sources in the environment.
Active sensors, such as sonar, send energy into the environment.

They rely on the fact that this energy is reflected back to the sensor. Active sensors tend to provide
more information than passive sensors, Range finders are sensors that measure the distance to nearby
objects.

Sonar sensors emit directional sound waves, which are reflected by objects, with some of the sound
making it back into the sensor. The time and intensity of the returning signal indicates the distance
to nearby objects.

Sonar is the technology of choice for autonomous underwater vehicles. Stereo vision (see Section
24.4.2) relies on multiple cameras to image the environment from slightly different viewpoints,
analyzing the resulting parallax in these images to compute the range of surrounding objects. For
mobile ground robots, sonar and stereo vision are now rarely used, because they are not reliably
accurate.

Time of flight camera :Most ground robots are now equipped with optical range finders. Just like
sonar sensors, optical range sensors emit active signals (light) and measure the time until a reflection

Scanning lidars tend to provide longer ranges than time of flight cameras, and tend to perform better in
bright daylight.

End of range sensing are tactile sensors such as whiskers, bump panels, and touch-sensitive skin.
These sensors measure range based on physical contact, and can be deployed only for sensing objects
very close to the robot.
location sensors
Most location sensors use range sensing as a primary component to determine location. Outdoors, the
Global Positioning System (GPS) is the most common solution to the localization problem. GPS the
distance to satellites that emit pulsed signals. At present, there are 31 satellites in orbit, transmitting
signals on multiple frequencies. GPS receivers can recover the distance to these satellites by analyzing
phase shifts. By triangulating signals from multiple satellites, GPS receivers can determine their
absolute location on Earth to within a few meters.

Differential GPS involves a second ground receiver with known location, providing millimeter
accuracy under ideal conditions. Unfortunately, GPS does not work indoors or underwater.

Proprioceptive sensors, which inform the robot of its own motion. To measure the exact
configuration of a robotic joint, motors are often equipped with shaft decoders that count the
revolution of motors in small increments. On mobile robots, shaft decoders that report wheel
revolutions can be used for odometry—the measurement of distance traveled.

Inertial sensors, such as gyroscopes, rely on the resistance of mass to the change of velocity. They
can help reduce uncertainty.

Other important aspects of robot state are measured by force sensors and torque sensors. These are
indispensable when robots handle fragile objects or objects whose exact shape and location is
unknown. Imagine a one-ton robotic manipulator screwing in a light bulb. It would be all too easy to
29
apply too much force and break the bulb. Force sensors allow the robot to sense how hard it is gripping
the bulb, and torque sensors allow it to sense how hard it is turning. Good sensors can measure forces
in all three translational and three rotational directions. They do this at a frequency of several hundred
times a second, so that a robot can quickly detect unexpected forces and correct its actions before it
breaks a light bulb.

Effectors
Effectors are the means by which robots move and change the shape of their bodies. To understand the
design of effectors, it will help to talk about motion and shape in the abstract, using the concept of a
degree of freedom (DOF)

 For nonrigid bodies, there are additional degrees of freedom within the robot itself.
 Robot joints also have one, two, or three degrees of freedom each.
 Six degrees of freedom are required to place an object, such as a hand, at a particular point in a
particular orientation.

the car has three effective degrees of freedom but two control DOF lable degrees of freedom.
Differential drive robots possess two independently actuated wheels (or tracks), one on each side, as
on a military tank.

Legs, unlike wheels, can handle rough terrain. However, legs are notoriously slow on flat surfaces, and
they are mechanically difficult to build.

A robot that can remain STABLE upright without moving its legs is called statically stable. A robot is
statically stable if its center of gravity is above the polygon spanned by its legs.

C omplete robot also needs a source of power to drive its effectors. The electric motor is the most
popular mechanism for both manipulator actuation and locomotion, but pneumatic actuation using
compressed gas and hydraulic actuation using pressurized fluids also have their application niches.

11. ROBOTIC PERCEPTION

Perception is the process by which robots map sensor measurements into internal representations of the
environment. Perception is difficult because sensors are noisy, and the environment is partially
observable, unpredictable, and often dynamic.

30
The robot’s own past actions as observed variables in the model. Figure 25.7 shows the notation used
in this chapter: Xt is the state of the environment (including the robot) at time t, Zt is the observation
received at time t, and At is the action taken after the observation is received.

We would like to compute the new belief state, P(Xt+1 | z1:t+1, a1:t), from the current
belief state P(Xt | z1:t, a1:t−1) and the new observation zt+1.
we condition explicitly on the actions as well as the observations, and we deal with continuous rather
than discrete variables.

The probability P(Xt+1 | xt, at) : is the transition model or motion model, and
P(zt+1 | XMOTION MODEL t+1) : is the sensor model.

Localization and mapping

Localization is the problem of finding out where things are—including the robot itself. Knowledge
about where things are is at the core of any successful physical interaction with the environment. For
example, robot manipulators must know the location of objects they seek to manipulate; navigating
robots must know where they are to find their way around.

Monte Carlo localization

Localization using particle filtering is called Monte Carlo localization, or MCL. The MCL algorithm
is an instance of the particle-filtering algorithm. All we need to do is supply the appropriate motion
model and sensor model.

The operation of the algorithm is illustrated in as the robot finds out where it is inside an office
building. In the first image, the particles are uniformly distributed based on the prior, indicating global
uncertainty about the robot’s position. In the second image, the first set of measurements arrives and
the particles form clusters in the areas of high posterior belief. In the third, enough measurements are
available to push all the particles to a single location.

Kalman filter
The Kalman filter is the other major way to localize. A Kalman filter represents the posterior P(Xt |
z1:t, a1:t−1) by a Gaussian. The mean of this Gaussian will be denoted μ t and its covariance Σt. The
main problem with Gaussian beliefs is that they are only closed under linear motion models f and
linear measurement models h. localization algorithms using the Kalman LINEARIZATION filter
linearize the motion and sensor models.

31
The problem of needing to know the identity of landmarks is an instance of the data association
problem

In some situations, no map of the environment is available. Then the robot will have to acquire a map.
This is a bit of a chicken-and-egg problem: the navigating robot will have to determine its location
relative to a map it doesn’t quite know, at the same time building this map while it doesn’t quite know
its actual location. This problem is important for many robot applications, and it has been studied
extensively under the name simultaneous localization and mapping, abbreviated as SLAM.

Monte Carlo localization, a particle filtering algorithm for mobile robot localization.
(a) Initial, global uncertainty.

(b) Approximately bimodal uncertainty after navigating

in the (symmetric) corridor.

(c) Unimodal uncertainty after entering a room and finding

32
it to be distinctive.

Machine learning in robot perception

Machine learning plays an important role in robot perception. One common approach is to map high
dimensional sensor streams into lower-dimensional spaces using unsupervised machine learning
methods

Another machine learning technique enables robots to continuously adapt to broad changes in sensor
measurements. Adaptive perception techniques enable robots to adjust to such changes.

Methods that make robots collect their own training data (with labels!) are called selfsupervised. In
this instance, the robot SELF-SUPERVISED uses machine learning to leverage a short-range sensor
that works well for terrain classification into a sensor that can see much farther.

12. ROBOT : MOVING

All of a robot’s deliberations ultimately come down to deciding how to move effectors. The point-to-
point motion problem is to deliver the robot or its end effector to a designated target location.

We begin by finding a suitable representation in which motion-planning problems can be described

and solved. It turns out that the configuration space—the space of robot states defined by location,
orientation, and joint angles—is a better place to work than the original PATH PLANNING 3D space.
The path planning problem is to find a path from one configuration to another in configuration space.

There are two main approaches: cell decomposition and skeletonization.

Configuration space
This raises the question of how to map between workspace coordinates and configuration space.
Transforming configuration space coordinates into workspace coordinates is simple: it involves a
series of straightforward coordinate transformations. These transformations are linear for prismatic
joints and trigonometric for revolute joints. This chain of coordinate transformation is known as
kinematics.

The inverse problem of calculating the configuration of a robot whose effector location is specified in
workspace coordinates is known as inverse kinematics. The second problem with configuration space
representations arises from the obstacles that may exist in the robot’s workspace

Cell decomposition methods

first CELL approach to path planning uses cell decomposition—that is, it decomposes the free space
into a finite number of contiguous regions, called cells. The path-planning problem then becomes a
discrete graph-search problem, very much like the search problems Such a decomposition has the
advantage that it is extremely simple to implement, but it also suffers from three limitations.

First, it is workable only for low-dimensional configuration spaces,

Second, there is the problem of what to do with cells that are ―mixed‖—that is, neither entirely within
free space nor entirely within occupied space.
Third, any path through a discretized state space will not be smooth. It is generally difficult to
guarantee that a smooth solution exists near the discrete path.

Skeletonization methods
The second major family of path-planning algorithms is based on the idea of skeletonization. These
algorithms reduce the robot’s free space to a one-dimensional representation, for which the planning
33
problem is easier. This lower-dimensional representation is called a skeleton of the configuration
space.

It is easy to show that this can always be achieved by a straight-line motion in configuration space.
it is a Voronoi graph of the free space—the set of all points that are equidistant to two or more
obstacles. To do path planning with a Voronoi graph, the robot first changes its present configuration
to a point on the Voronoi graph. It is easy to show that this can always be achieved by a straight-line
motion in configuration space. Second, the robot follows the Voronoi graph until it reaches the point
nearest to the target configuration. Finally, the robot leaves the Voronoi graph and moves to the target.
Again, this final step involves straight-line motion in configuration space.

An alternative to the Voronoi graphs is the probabilistic roadmap, a skeletonization approach that
offers more possible routes, and thus deals better with wide-open spaces.

13. PLANNING UNCERTAIN MOVEMENTS

None of the robot motion-planning algorithms discussed thus far addresses a key characteristic of
robotics problems: uncertainty. In robotics, uncertainty arises from partial observability of the
environment and from the stochastic (or unmodeled) effects of the robot’s actions.

The field of robotics has adopted a range of techniques for accommodating uncertainty. If the robot
faces uncertainty only in its state transition, but its state is fully observable, the problem is best
modeled as a Markov decision process (MDP). The solution of an MDP is an optimal policy, which
tells the robot what to do in every possible state.

For example, the coastal navigation heuristic requires the robot to stay near known landmarks to
decrease its uncertainty. Another approach applies variants of the probabilistic roadmap planning
method to the belief space representation. Such methods tend to scale better to large discrete POMDPs.

Robust methods
Uncertainty can also be handled using so-called robust control methods rather than probabilistic
methods. A robust method is one that assumes a bounded amount of uncertainty in each aspect of a
problem, but does not assign probabilities to values within the allowed interval. A robust solution is
one that works no matter what actual values occur, provided they are within the assumed interval.

Fine-motion planning involves moving a robot arm in very close proximity to a static environment
object.

A fine-motion plan consists of a series of guarded motions. The termination conditions are contact
with a surface. To model uncertainty in control, we assume that instead of moving in the commanded
direction, the robot’s actual motion lies in the cone Cv about it.

Dynamics and control

The transition model for a dynamic state representation includes the effect of forces on this rate of
change. the dynamic state has higher dimension than the kinematic space, and the curse of
dimensionality would render many motion planning algorithms inapplicable for all but the most simple
robots. For this reason, practical robot system often rely on simpler kinematic path planners.

A common technique to compensate for the limitations of kinematic plans is to use a CONTROLLER
separate mechanism, a controller, for keeping the robot on track.

Potential-field control
34
potential fields as an additional cost function in robot motion planning, but they can also be used for
generating robot motion directly, dispensing with the path planning phase altogether.

Define an attractive force that pulls the robot towards its goal configuration and a repellent potential
field that pushes the robot away from obstacles.
No planning was involved in generating the potential field
The potential field can be calculated efficiently for any given configuration. Moreover, optimizing the
potential amounts to calculating the gradient of the potential for the present robot configuration. These
calculations can be extremely efficient, especially when compared to path-planning algorithms, all of
which are exponential in the dimensionality of the configuration space (the DOFs) in the worst case.

Potential field control is great for local robot motion but sometimes we still need global planning.

Reactive control
control decisions that require some model of the environment for constructing either a reference path
or a potential field.
First, models that are sufficiently accurate are often difficult to obtain, especially in complex or remote
environments, such as the surface of Mars, or for robots that have few sensors. Second, even in cases
where we can devise a model with sufficient accuracy, computational difficulties and localization error
might render these techniques impractical.

A reflex agent architecture using reactive control is more appropriate

It is possible, nonetheless, to specify a controller directly without an explicit environmental model.

Variants of this simple feedback-driven controller have been found to generate remarkably robust
walking patterns, capable of maneuvering the robot over rugged terrain

Reinforcement learning control

One particularly exciting form of control is based on the policy search form of reinforcement
Learning. Policy search needs an accurate model of the domain before it can find a policy.

The controls are the manual controls of of the helicopter: throttle, pitch, elevator, aileron, and rudder.
All that remains is the resulting state—how are we going to define a model that accurately says how
the helicopter responds to each control? The answer is simple: Let an expert human pilot fly the
helicopter, and record the controls that the expert transmits over the radio and the state variables of the
helicopter. About four minutes of human-controlled flight suffices to build a predictive model that is
sufficiently accurate to simulate the vehicle.

35