1

INTRODUCTION

Dave Bowman: Open the pod bay doors, HAL.

HAL: I'm sorry Dave, I'm afraid I can't do that.

Stanley Kubrick and Arthur C. Clarke,

screenplay of 2001: A Space Odyssey

This book is about a new interdisciplinary field variously called computer speech and language processing or human language technology or natural language processing or computational linguistics. The goal of this new field is to get computers to perform useful tasks involving human language, tasks like enabling human-machine communication, improving human-human communication, or simply doing useful processing of text or speech.

One example of a useful such task is a conversational agent. The HAL 9000 computer in Stanley Kubrick's film 2001: A Space Odyssey is one of the most recognizable characters in twentieth-century cinema. HAL is an artificial agent capable of such advanced language-processing behavior as speaking and understanding English, and at a crucial moment in the plot, even reading lips. It is now clear that HAL's creator Arthur C. Clarke was a little optimistic in predicting when an artificial agent such as HAL would be available. But just how far off was he? What would it take to create at least the language-related parts of HAL? We call programs like HAL that converse with humans via natural language conversational agents or dialogue systems. In this text we study the various components that make up modern conversational agents, including language input (automatic speech recognition and natural language understanding) and language output (natural language generation and speech synthesis).

Let's turn to another useful language-related task, that of making available to non-English-speaking readers the vast amount of scientific information on the Web in English. Or translating for English speakers the hundreds of millions of Web pages written in other languages like Chinese. The goal of machine translation is to automatically translate a document from one language to another. Machine translation is far from a solved problem; we will cover the algorithms currently used in the field, as well as important component tasks.

Many other language processing tasks are also related to the Web. Another such task is Web-based question answering. This is a generalization of simple web search, where instead of just typing keywords a user might ask complete questions, ranging from easy to hard, like the following:

• What does “divergent” mean?

• What year was Abraham Lincoln born?

• How many states were in the United States that year?

How much Chinese silk was exported to England by the end of the 18th century?

• What do scientists think about the ethics of human cloning?

Some of these, such as definition questions, or simple factoid questions like dates and locations, can already be answered by search engines. But answering more complicated questions might require extracting information that is embedded in other text on a Web page, or doing inference (drawing conclusions based on known facts), or synthesizing and summarizing information from multiple sources or web pages. In this text we study the various components that make up modern understanding systems of this kind, including information extraction, word sense disambiguation, and so on.

Although the subfields and problems we've described above are all very far from completely solved, these are all very active research areas and many technologies are already available commercially. In the rest of this chapter we briefly summarize the kinds of knowledge that is necessary for these tasks (and others like spell correction, grammar checking, and so on), as well as the mathematical models that will be introduced throughout the book.

1.1 KNOWLEDGE IN SPEECH AND LANGUAGE PROCESSING

What distinguishes language processing applications from other data processing systems is their use of knowledge of language. Consider the Unix wc program, which is used to count the total number of bytes, words, and lines in a text file. When used to count bytes and lines, wc is an ordinary data processing application. However, when it is used to count the words in a file it requires knowledge about what it means to be a word, and thus becomes a language processing system.

Of course, wc is an extremely simple system with an extremely limited and impoverished knowledge of language. Sophisticated conversational agents like HAL, or machine translation systems, or robust question-answering systems, require much broader and deeper knowledge of language. To get a feeling for the scope and kind of required knowledge, consider some of what HAL would need to know to engage in the dialogue that begins this chapter, or for a question answering system to answer one of the questions above.

HAL must be able to recognize words from an audio signal and to generate an audio signal from a sequence of words. These tasks of speech recognition and speech synthesis tasks require knowledge about phonetics and phonology; how words are pronounced in terms of sequences of sounds, and how each of these sounds is realized acoustically.

Note also that unlike Star Trek’s Commander Data, HAL is capable of producing contractions like I’m and can’t. Producing and recognizing these and other variations of individual words (e.g., recognizing that doors is plural) requires knowledge about morphology, the way words break down into component parts that carry meanings like singular versus plural.

Moving beyond individual words, HAL must use structural knowledge to properly string together the words that constitute its response. For example, HAL must know that the following sequence of words will not make sense to Dave, despite the fact that it contains precisely the same set of words as the original.

I'm I do, sorry that afraid Dave I'm can't.

The knowledge needed to order and group words together comes under the heading of syntax.

Now consider a question answering system dealing with the following question:

How much Chinese silk was exported to Western Europe by the end of the 18th century?

In order to answer this question we need to know something about lexical semantics, the meaning of all the words (export, or silk) as well as compositional semantics (what exactly constitutes Western Europe as opposed to Eastern or Southern Europe, what does end mean when combined with the 18th century). We also need to know something about the relationship of the words to the syntactic structure. For example, we need to know that by the end of the 18th century is a temporal end-point, and not a description of the agent, as the by-phrase is in the following sentence:

• How much Chinese silk was exported to Western Europe by southern merchants?

We also need the kind of knowledge that lets HAL determine that Dave's utterance is a request for action, as opposed to a simple statement about the world or a question about the door, as in the following variations of his original statement.

REQUEST: HAL, open the pod bay door.

STATEMENT: HAL, the pod bay door is open.

INFORMATION QUESTION: HAL, is the pod bay door open?

Next, despite its bad behavior, HAL knows enough to be polite to Dave. It could, for example, have simply replied No or No, I won't open the door. Instead, it first embellishes its response with the phrases I'm sorry and I'm afraid, and then only indirectly signals its refusal by saying I can't, rather than the more direct (and truthful) I won't. $ ^{1} $ This knowledge about the kind of actions that speakers intend by their use of sentences is pragmatic or dialogue knowledge.

Another kind of pragmatic or discourse knowledge is required to answer the question

• How many states were in the United States that year?

What year is that year? In order to interpret words like that year a question answering system need to examine the earlier questions that were asked; in this case the previous question talked about the year that Lincoln was born. Thus this task of coreference resolution makes use of knowledge about how words like that or pronouns like it or she refer to previous parts of the discourse.

To summarize, engaging in complex language behavior requires various kinds of knowledge of language:

• Phonetics and Phonology — knowledge about linguistic sounds

Morphology — knowledge of the meaningful components of words

• Syntax — knowledge of the structural relationships between words

• Semantics — knowledge of meaning

• Pragmatics — knowledge of the relationship of meaning to the goals and intentions of the speaker.

• Discourse — knowledge about linguistic units larger than a single utterance

1.2 AMBIGUITY

A perhaps surprising fact about these categories of linguistic knowledge is that most tasks in speech and language processing can be viewed as resolving ambiguity at one of these levels. We say some input is ambiguous if there are multiple alternative linguistic structures that can be built for it. Consider the spoken sentence I made her duck. Here's five different meanings this sentence could have (see if you can think of some more), each of which exemplifies an ambiguity at some level:

(1.1) I cooked waterfowl for her.

(1.2) I cooked waterfowl belonging to her.

(1.3) I created the (plaster?) duck she owns.

(1.4) I caused her to quickly lower her head or body.

(1.5) I waved my magic wand and turned her into undifferentiated waterfowl.

These different meanings are caused by a number of ambiguities. First, the words duck and her are morphologically or syntactically ambiguous in their part-of-speech. Duck can be a verb or a noun, while her can be a dative pronoun or a possessive pronoun. Second, the word make is semantically ambiguous; it can mean create or cook. Finally, the verb make is syntactically ambiguous in a different way. Make can be transitive, that is, taking a single direct object (1.2), or it can be ditransitive, that is, taking two objects (1.5), meaning that the first object (her) got made into the second object (duck). Finally, make can take a direct object and a verb (1.4), meaning that the object (her) got caused to perform the verbal action (duck). Furthermore, in a spoken sentence, there is an even deeper kind of ambiguity; the first word could have been eye or the second word maid.

We will often introduce the models and algorithms we present throughout the book as ways to resolve or disambiguate these ambiguities. For example deciding whether duck is a verb or a noun can be solved by part-of-speech tagging. Deciding whether make means "create" or "cook" can be solved by word sense disambiguation. Resolution of part-of-speech and word sense ambiguities are two important kinds of lexical disambiguation. A wide variety of tasks can be framed as lexical disambiguation problems. For example, a text-to-speech synthesis system reading the word lead needs to decide whether it should be pronounced as in lead pipe or as in lead me on. By contrast, deciding whether her and duck are part of the same entity (as in (1.1) or (1.4)) or are different entity (as in (1.2)) is an example of syntactic disambiguation and can

be addressed by probabilistic parsing. Ambiguities that don't arise in this particular example (like whether a given sentence is a statement or a question) will also be resolved, for example by speech act interpretation.

1.3 MODELS AND ALGORITHMS

One of the key insights of the last 50 years of research in language processing is that the various kinds of knowledge described in the last sections can be captured through the use of a small number of formal models, or theories. Fortunately, these models and theories are all drawn from the standard toolkits of computer science, mathematics, and linguistics and should be generally familiar to those trained in those fields. Among the most important models are state machines, rule systems, logic, probabilistic models, and vector-space models. These models, in turn, lend themselves to a small number of algorithms, among the most important of which are state space search algorithms such as dynamic programming, and machine learning algorithms such as classifiers and EM and other learning algorithms.

In their simplest formulation, state machines are formal models that consist of states, transitions among states, and an input representation. Some of the variations of this basic model that we will consider are deterministic and non-deterministic finite-state automata and finite-state transducers.

Closely related to these models are their declarative counterparts: formal rule systems. Among the more important ones we will consider are regular grammars and regular relations, context-free grammars, feature-augmented grammars, as well as probabilistic variants of them all. State machines and formal rule systems are the main tools used when dealing with knowledge of phonology, morphology, and syntax.

The third model that plays a critical role in capturing knowledge of language is logic. We will discuss first order logic, also known as the predicate calculus, as well as such related formalisms as lambda-calculus, feature-structures, and semantic primitives. These logical representations have traditionally been used for modeling semantics and pragmatics, although more recent work has focused on more robust techniques drawn from non-logical lexical semantics.

Probabilistic models are crucial for capturing every kind of linguistic knowledge. Each of the other models (state machines, formal rule systems, and logic) can be augmented with probabilities. For example the state machine can be augmented with probabilities to become the weighted automaton or Markov model. We will spend a significant amount of time on hidden Markov models or HMMs, which are used everywhere in the field, in part-of-speech tagging, speech recognition, dialogue understanding, text-to-speech, and machine translation. The key advantage of probabilistic models is their ability to solve the many kinds of ambiguity problems that we discussed earlier; almost any speech and language processing problem can be recast as: "given N choices for some ambiguous input, choose the most probable one".

Finally, vector-space models, based on linear algebra, underlie information retrieval and many treatments of word meanings.

Processing language using any of these models typically involves a search through

a space of states representing hypotheses about an input. In speech recognition, we search through a space of phone sequences for the correct word. In parsing, we search through a space of trees for the syntactic parse of an input sentence. In machine translation, we search through a space of translation hypotheses for the correct translation of a sentence into another language. For non-probabilistic tasks, such as state machines, we use well-known graph algorithms such as depth-first search. For probabilistic tasks, we use heuristic variants such as best-first and A* search, and rely on dynamic programming algorithms for computational tractability.

For many language tasks, we rely on machine learning tools like classifiers and sequence models. Classifiers like decision trees, support vector machines, Gaussian Mixture Models and logistic regression are very commonly used. A hidden Markov model is one kind of sequence model; other are Maximum Entropy Markov Models or Conditional Random Fields.

Another tool that is related to machine learning is methodological; the use of distinct training and test sets, statistical techniques like cross-validation, and careful evaluation of our trained systems.

1.4 LANGUAGE, THOUGHT, AND UNDERSTANDING

To many, the ability of computers to process language as skillfully as we humans do will signal the arrival of truly intelligent machines. The basis of this belief is the fact that the effective use of language is intertwined with our general cognitive abilities. Among the first to consider the computational implications of this intimate connection was Alan Turing (1950). In this famous paper, Turing introduced what has come to be known as the Turing Test. Turing began with the thesis that the question of what it would mean for a machine to think was essentially unanswerable due to the inherent imprecision in the terms machine and think. Instead, he suggested an empirical test, a game, in which a computer's use of language would form the basis for determining if it could think. If the machine could win the game it would be judged intelligent.

In Turing's game, there are three participants: two people and a computer. One of the people is a contestant and plays the role of an interrogator. To win, the interrogator must determine which of the other two participants is the machine by asking a series of questions via a teletype. The task of the machine is to fool the interrogator into believing it is a person by responding as a person would to the interrogator's questions. The task of the second human participant is to convince the interrogator that the other participant is the machine, and that they are human.

The following interaction from Turing's paper illustrates the kind of interactions he had in mind. Clearly a convincing simulation of humanity doesn't require being an expert in all domains:

Q: Please write me a sonnet on the topic of the Forth Bridge.

A: Count me out on this one. I never could write poetry.

Q: Add 34957 to 70764.

A: (Pause about 30 seconds and then give answer as) 105621.

Turing predicted that roughly by the end of the twentieth century a machine with 10

gigabytes of memory would have around a 30% chance of fooling a human interrogator after 5 minutes of questioning. Given the fact that you can fool some of the people all the time, it is not clear how rigorous this particular standard is. Nevertheless, the critical issue for Turing was that using language as humans do is sufficient, by itself, as an operational test for intelligence.

Surprisingly, many of the issues raised by Turing became relevant as early as 1966 in the context of the ELIZA program (Weizenbaum, 1966). ELIZA was an early natural language processing system capable of carrying on a limited form of conversation with a user. Consider the following session with a version of ELIZA that imitated the responses of a Rogerian psychotherapist.

User $ _{1} $: You are like my father in some ways.

ELIZA $ _{1} $: WHAT RESEMBLANCE DO YOU SEE

User2: You are not very aggressive but I think you don't want me to notice that.

ELIZA $ _{2} $: WHAT MAKES YOU THINK I AM NOT AGGRESSIVE

User3: You don't argue with me.

ELIZA $ _{3} $: WHY DO YOU THINK I DON'T ARGUE WITH YOU

User4: You are afraid of me.

ELIZA $ _{4} $: DOES IT PLEASE YOU TO BELIEVE I AM AFRAID OF YOU

As we will discuss in more detail in Ch. 2, ELIZA is a remarkably simple program that makes use of pattern-matching to process the input and translate it into suitable outputs. The success of this simple technique in this domain is due to the fact that ELIZA doesn't actually need to know anything to mimic a Rogerian psychotherapist. As Weizenbaum notes, this is one of the few dialogue genres where the listener can act as if they know nothing of the world.

ELIZA's deep relevance to Turing's ideas is that many people who interacted with ELIZA came to believe that it really understood them and their problems. Indeed, Weizenbaum (1976) notes that many of these people continued to believe in ELIZA's abilities even after the program's operation was explained to them. In more recent years, Weizenbaum's informal reports have been repeated in a somewhat more controlled setting. Since 1991, an event known as the Loebner Prize competition has attempted to put various computer programs to the Turing test. Although these contests seem to have little scientific interest, a consistent result over the years has been that even the crudest programs can fool some of the judges some of the time (Shieber, 1994). Not surprisingly, these results have done nothing to quell the ongoing debate over the suitability of the Turing test as a test for intelligence among philosophers and AI researchers (Searle, 1980).

Fortunately, for the purposes of this book, the relevance of these results does not hinge on whether or not computers will ever be intelligent, or understand natural language. Far more important is recent related research in the social sciences that has confirmed another of Turing's predictions from the same paper.

Nevertheless I believe that at the end of the century the use of words and educated opinion will have altered so much that we will be able to speak of machines thinking without expecting to be contradicted.

It is now clear that regardless of what people believe or know about the inner workings of computers, they talk about them and interact with them as social entities. People act

toward computers as if they were people; they are polite to them, treat them as team members, and expect among other things that computers should be able to understand their needs, and be capable of interacting with them naturally. For example, Reeves and Nass (1996) found that when a computer asked a human to evaluate how well the computer had been doing, the human gives more positive responses than when a different computer asks the same questions. People seemed to be afraid of being impolite. In a different experiment, Reeves and Nass found that people also give computers higher performance ratings if the computer has recently said something flattering to the human. Given these predispositions, speech and language-based systems may provide many users with the most natural interface for many applications. This fact has led to a long-term focus in the field on the design of conversational agents, artificial entities that communicate conversationally.

1.5 THE STATE OF THE ART

We can only see a short distance ahead, but we can see plenty there that needs to be done.

Alan Turing.

This is an exciting time for the field of speech and language processing. The startling increase in computing resources available to the average computer user, the rise of the Web as a massive source of information and the increasing availability of wireless mobile access have all placed speech and language processing applications in the technology spotlight. The following are examples of some currently deployed systems that reflect this trend:

• Travelers calling Amtrak, United Airlines and other travel-providers interact with conversational agents that guide them through the process of making reservations and getting arrival and departure information.

• Luxury car makers such as Mercedes-Benz models provide automatic speech recognition and text-to-speech systems that allow drivers to control their environmental, entertainment and navigational systems by voice. A similar spoken dialogue system has been deployed by astronauts on the International Space Station.

• Blinkx, and other video search companies, provide search services for million of hours of video on the Web by using speech recognition technology to capture the words in the sound track.

• Google provides cross-language information retrieval and translation services where a user can supply queries in their native language to search collections in another language. Google translates the query, finds the most relevant pages and then automatically translates them back to the user's native language.

• Large educational publishers such as Pearson, as well as testing services like ETS, use automated systems to analyze thousands of student essays, grading and assessing them in a manner that is indistinguishable from human graders.

• Interactive tutors, based on lifelike animated characters, serve as tutors for children learning to read, and as therapists for people dealing with aphasia and Parkinsons disease. (? , ?)

• Text analysis companies such as Nielsen Buzzmetrics, Umbria, and Collective Intellect, provide marketing intelligence based on automated measurements of user opinions, preferences, attitudes as expressed in weblogs, discussion forums and user groups.

1.6 SOME BRIEF HISTORY

Historically, speech and language processing has been treated very differently in computer science, electrical engineering, linguistics, and psychology/cognitive science. Because of this diversity, speech and language processing encompasses a number of different but overlapping fields in these different departments: computational linguistics in linguistics, natural language processing in computer science, speech recognition in electrical engineering, computational psycholinguistics in psychology. This section summarizes the different historical threads which have given rise to the field of speech and language processing. This section will provide only a sketch; see the individual chapters for more detail on each area and its terminology.

1.6.1 Foundational Insights: 1940s and 1950s

The earliest roots of the field date to the intellectually fertile period just after World War II that gave rise to the computer itself. This period from the 1940s through the end of the 1950s saw intense work on two foundational paradigms: the automaton and probabilistic or information-theoretic models.

The automaton arose in the 1950s out of Turing's (1936) model of algorithmic computation, considered by many to be the foundation of modern computer science. Turing's work led first to the McCulloch-Pitts neuron (McCulloch and Pitts, 1943), a simplified model of the neuron as a kind of computing element that could be described in terms of propositional logic, and then to the work of Kleene (1951) and (1956) on finite automata and regular expressions. Shannon (1948) applied probabilistic models of discrete Markov processes to automata for language. Drawing the idea of a finite-state Markov process from Shannon's work, Chomsky (1956) first considered finite-state machines as a way to characterize a grammar, and defined a finite-state language as a language generated by a finite-state grammar. These early models led to the field of formal language theory, which used algebra and set theory to define formal languages as sequences of symbols. This includes the context-free grammar, first defined by Chomsky (1956) for natural languages but independently discovered by Backus (1959) and Naur et al. (1960) in their descriptions of the ALGOL programming language.

The second foundational insight of this period was the development of probabilistic algorithms for speech and language processing, which dates to Shannon's other contribution: the metaphor of the noisy channel and decoding for the transmission of language through media like communication channels and speech acoustics. Shannon

also borrowed the concept of entropy from thermodynamics as a way of measuring the information capacity of a channel, or the information content of a language, and performed the first measure of the entropy of English using probabilistic techniques.

It was also during this early period that the sound spectrograph was developed (Koenig et al., 1946), and foundational research was done in instrumental phonetics that laid the groundwork for later work in speech recognition. This led to the first machine speech recognizers in the early 1950s. In 1952, researchers at Bell Labs built a statistical system that could recognize any of the 10 digits from a single speaker (Davis et al., 1952). The system had 10 speaker-dependent stored patterns roughly representing the first two vowel formants in the digits. They achieved 97–99% accuracy by choosing the pattern which had the highest relative correlation coefficient with the input.