“I'm sorry Dave, I'm afraid I can't do that”:
Linguistics, Statistics, and Natural Language Processing circa 2001¹

Lillian Lee, Cornell University
In Computer Science: Reflections on the Field, Reflections from the Field (Report of the National Academies' Study on the Fundamentals of Computer Science), pp. 111-118, 2004.

It's the year 2000, but where are the flying cars? I was promised flying cars.
- Avery Brooks, IBM commercial

According to many pop-culture visions of the future, technology will eventually produce the Machine that Can Speak to Us. Examples range from the False Maria in Fritz Lang's 1926 film Metropolis to Knight Rider's KITT (a talking car) to Star Wars' C-3PO (said to have been modeled on the False Maria). And, of course, there is the HAL 9000 computer from 2001: A Space Odyssey; in one of the film's most famous scenes, the astronaut Dave asks HAL to open a pod bay door on the spacecraft, to which HAL responds, “I'm sorry Dave, I'm afraid I can't do that”.

Natural language processing, or NLP, is the field of computer science devoted to creating such machines -- that is, enabling computers to use human languages both as input and as output. The area is quite broad, encompassing problems ranging from simultaneous multi-language translation to advanced search engine development to the design of computer interfaces capable of combining speech, diagrams, and other modalities simultaneously. A natural consequence of this wide range of inquiry is the integration of ideas from computer science with work from many other fields, including linguistics, which provides models of language; psychology, which provides models of cognitive processes; information theory, which provides models of communication; and mathematics and statistics, which provide tools for analyzing and acquiring such models.

The interaction of these ideas together with advances in machine learning (see [other chapter]) has resulted in concerted research activity in statistical natural language processing: making computers language-enabled by having them acquire linguistic information directly from samples of language itself. In this essay, we describe the history of statistical NLP; the twists and turns of the story serve to highlight the sometimes complex interplay between computer science and other fields.

Although currently a major focus of research, the data-driven, computational approach to language processing was for some time held in deep disregard because it directly conflicts with another commonly-held viewpoint: human language is so complex that language samples alone seemingly cannot yield enough information to understand it. Indeed, it is often said that NLP is “AI-complete” (a pun on NP-completeness; see [other chapter]), meaning that the most difficult problems in artificial intelligence manifest themselves in human language phenomena. This belief in language use as the touchstone of intelligent behavior dates back at least to the 1950 proposal of the Turing Test² as a way to gauge whether machine intelligence has been achieved; as Turing wrote, “The question and answer method seems to be suitable for introducing almost any one of the fields of human endeavour that we wish to include”.

The reader might be somewhat surprised to hear that language understanding is so hard. After all, human children get the hang of it in a few years, word processing software now corrects (some of) our grammatical errors, and TV ads show us phones capable of effortless translation. One might therefore be led to believe that HAL is just around the corner.

Such is not the case, however. In order to appreciate this point, we temporarily divert from describing statistical NLP's history -- which touches upon Hamilton versus Madison, the sleeping habits of colorless green ideas, and what happens when one fires a linguist -- to examine a few examples illustrating why understanding human language is such a difficult problem.

Ambiguity and language analysis

At last, a computer that understands you like your mother.
- 1985 McDonnell-Douglas ad

The snippet quoted above indicates the early confidence at least one company had in the feasibility of getting computers to understand human language. But in fact, that very sentence is illustrative of the host of difficulties that arise in trying to analyze human utterances, and so, ironically, it is quite unlikely that the system being promoted would have been up to the task. A moment's reflection reveals that the sentence admits at least three different interpretations:

1. The computer understands you as well as your mother understands you.
2. The computer understands that you like your mother.
3. The computer understands you as well as it understands your mother.

That is, the sentence is ambiguous; and yet we humans seem to instantaneously rule out all the alternatives except the first (and presumably the intended) one. We do so based on a great deal of background knowledge, including understanding what advertisements typically try to convince us of. How are we to get such information into a computer?

A number of other types of ambiguity are also lurking here. For example, consider the speech recognition problem: how can we distinguish between this utterance, when spoken, and “... a computer that understands your lie cured mother”? We also have a word sense ambiguity problem: how do we know that here “mother” means “a female parent”, rather than the Oxford English Dictionary-approved alternative of “a cask or vat used in vinegar-making”? Again, it is our broad knowledge about the world and the context of the remark that allows us humans to make these decisions easily.

Now, one might be tempted to think that all these ambiguities arise because our example sentence is highly unusual (although the ad writers probably did not set out to craft a strange sentence). Or, one might argue that these ambiguities are somehow artificial because the alternative interpretations are so unrealistic that an NLP system could easily filter them out. But ambiguities crop up in many situations. For example, in “Copy the local patient files to disk” (which seems like a perfectly plausible command to issue to a computer), is it the patients or the files that are local?³ Again, we need to know the specifics of the situation in order to decide. And in multilingual settings, extra ambiguities may arise. Here is a sequence of seven Japanese characters:

Since Japanese doesn't have spaces between words, one is faced with the initial task of deciding what the component words are. In particular, this character sequence corresponds to at least two possible word sequences, “president, both, business, general-manager” (= “a president as well as a general manager of business”) and “president, subsidiary-business, Tsutomu (a name), general-manager” (= ?). It requires a fair bit of linguistic information to choose the correct alternative.⁴

To sum up, we see that the NLP task is highly daunting, for to resolve the many ambiguities that arise in trying to analyze even a single sentence requires deep knowledge not just about language but also about the world. And so when HAL says, “I'm afraid I can't do that”, NLP researchers are tempted to respond, “I'm afraid you might be right”.

Firth things first

But before we assume that the only viable approach to NLP is a massive knowledge engineering project, let us go back to the early approaches to the problem. In the 1940s and 1950s, one prominent trend in linguistics was explicitly empirical and in particular distributional, as exemplified by the work of Zellig Harris (who started the first linguistics program in the USA). The idea was that correlations (co-occurrences) found in language data are important sources of information, or, as the influential linguist J. R. Firth declared in 1957, “You shall know a word by the company it keeps”.

Such notions accord quite happily with ideas put forth by Claude Shannon in his landmark 1948 paper establishing the field of information theory; speaking from an engineering perspective, he identified the probability of a message's being chosen from among several alternatives, rather than the message's actual content, as its critical characteristic. Influenced by this work, Warren Weaver in 1949 proposed treating the problem of translating between languages as an application of cryptography (see [other chapter]), with one language viewed as an encrypted form of another. And, Alan Turing's work on cracking German codes during World War II led to the development of the Good-Turing formula, an important tool for computing certain statistical properties of language.

In yet a third area, 1941 saw the statisticians Frederick Mosteller and Frederick Williams address the question of whether it was Alexander Hamilton or James Madison who wrote some of the pseudonymous Federalist Papers. Unlike previous attempts, which were based on historical data and arguments, Mosteller and Williams used the patterns of word occurrences in the texts as evidence. This work led up to the famed Mosteller and Wallace statistical study which many consider to have settled the authorship of the disputed papers.

Thus, we see arising independently from a variety of fields the idea that language can be viewed from a data-driven, empirical perspective -- and a data-driven perspective leads naturally to a computational perspective.

A “C” change

However, data-driven approaches fell out of favor in the late 1950's. One of the commonly cited factors is a 1957 argument by linguist (and student of Harris) Noam Chomsky, who believed that language behavior should be analyzed at a much deeper level than its surface statistics. He claimed,

It is fair to assume that neither sentence (1) [Colorless green ideas sleep furiously] nor (2) [Furiously sleep ideas green colorless] ... has ever occurred .... Hence, in any [computed] statistical model ... these sentences will be ruled out on identical grounds as equally “remote” from English.⁵ Yet (1), though nonsensical, is grammatical, while (2) is not.

That is, we humans know that sentence (1), which at least obeys (some) rules of grammar, is indeed more probable than (2), which is just word salad; but (the claim goes), since both sentences are so rare, they will have identical statistics -- i.e., a frequency of zero -- in any sample of English. Chomsky's criticism is essentially that data-driven approaches will always suffer from a lack of data, and hence are doomed to failure.

This observation turned out to be remarkably prescient: even now, when billions of words of text are available on-line, perfectly reasonable phrases are not present. Thus, the so-called sparse data problem continues to be a serious challenge for statistical NLP even today. And so, the effect of Chomsky's claim, together with some negative results for machine learning and a general lack of computing power at the time, was to cause researchers to turn away from empirical approaches and toward knowledge-based approaches where human experts encoded relevant information in computer-usable form.

This change in perspective led to several new lines of fundamental, interdisciplinary research. For example, Chomsky's work viewing language as a formal, mathematically-describable object has had lasting impact on both linguistics and computer science; indeed, the Chomsky hierarchy, a sequence of increasingly more powerful classes of grammars, is a staple of the undergraduate computer science curriculum. Conversely, the highly influential work of, among others, Kazimierz Adjukiewicz, Joachim Lambek, David K. Lewis, and Richard Montague adopted the lambda calculus, a fundamental concept in the study of programming languages, to model the semantics of natural languages.

The empiricists strike back

By the '80s, the tide had begun to shift once again, in part because of the work done by the speech recognition group at IBM. These researchers, influenced by ideas from information theory, explored the power of probabilistic models of language combined with access to much more sophisticated algorithmic and data resources than had previously been available. In the realm of speech recognition, their ideas form the core of the design of modern systems; and given the recent successes of such software -- large-vocabulary continuous-speech recognition programs are now available on the market -- it behooves us to examine how these systems work.

Given some acoustic signal, which we denote by the variable $a$, we can think of the speech recognition problem as that of transcription: determining what sentence is most likely to have produced $a$. Probabilities arise because of the ever-present problem of ambiguity: as mentioned above, several word sequences, such as “your lie cured mother” versus “you like your mother”, can give rise to similar spoken output. Therefore, modern speech recognition systems incorporate information both about the acoustic signal and the language behind the signal. More specifically, they rephrase the problem as determining which sentence $s$ maximizes the product $P(a\vert s)\times P(s)$. The first term measures how likely the acoustic signal would be if $s$ were actually the sentence being uttered (again, we use probabilities because humans don't pronounce words the same way all the time). The second term measures the probability of the sentence $s$ itself; for example, as Chomsky noted, “colorless green ideas sleep furiously” is intuitively more likely to be uttered than the reversal of the phrase. It is in computing this second term, $P(s)$, where statistical NLP techniques come into play, since accurate estimation of these sentence probabilities requires developing probabilistic models of language. These models are acquired by processing tens of millions of words or more. This is by no means a simple procedure; even linguistically naive models require the use of sophisticated computational and statistical techniques because of the sparse data problem foreseen by Chomsky. But using probabilistic models, large datasets, and powerful learning algorithms (both for $P(s)$ and $P(a\vert s)$) has led to our achieving the milestone of commercial-grade speech recognition products capable of handling continuous speech ranging over a large vocabulary.

But let us return to our story. Buoyed by the successes in speech recognition in the '70s and '80s (substantial performance gains over knowledge-based systems were posted), researchers began applying data-driven approaches to many problems in natural language processing, in a turn-around so extreme that it has been deemed a “revolution”. Indeed, now empirical methods are used at all levels of language analysis. This is not just due to increased resources: a succession of breakthroughs in machine learning algorithms has allowed us to leverage existing resources much more effectively. At the same time, evidence from psychology shows that human learning may be more statistically-based than previously thought; for instance, work by Jenny Saffran, Richard Aslin, and Elissa Newport reveals that 8-month-old infants can learn to divide continuous speech into word segments based simply on the statistics of sounds following one another. Hence, it seems that the “revolution” is here to stay.

Of course, we must not go overboard and mistakenly conclude that the successes of statistical NLP render linguistics irrelevant (rash statements to this effect have been made in the past, e.g., the notorious remark, “Every time I fire a linguist, my performance goes up”). The information and insight that linguists, psychologists, and others have gathered about language is invaluable in creating high-performance broad-domain language understanding systems; for instance, in the speech recognition setting described above, a better understanding of language structure can lead to better language models. Moreover, truly interdisciplinary research has furthered our understanding of the human language faculty. One important example of this is the development of the head-driven phrase structure grammar (HPSG) formalism -- this is a way of analyzing natural language utterances that truly marries deep linguistic information with computer science mechanisms, such as unification and recursive data-types, for representing and propagating this information throughout the utterance's structure. In sum, computational techniques and data-driven methods are now an integral part both of building systems capable of handling language in a domain-independent, flexible, and graceful way, and of improving our understanding of language itself.

Acknowledgments

Thanks to the members of the CSTB Fundamentals of Computer Science study -- and especially Alan Biermann -- for their helpful feedback. Also, thanks to Alex Acero, Takako Aikawa, Mike Bailey, Regina Barzilay, Eric Brill, Chris Brockett, Claire Cardie, Joshua Goodman, Ed Hovy, Rebecca Hwa, John Lafferty, Bob Moore, Greg Morrisett, Fernando Pereira, Hisami Suzuki, and many others for stimulating discussions and very useful comments. Rie Kubota Ando provided the Japanese example. The use of the term “revolution” to describe the re-ascendance of statistical methods comes from Julia Hirschberg's 1998 invited address to the American Association for Artificial Intelligence. I learned of the McDonnell-Douglas ad and some of its analyses from a class run by Stuart Shieber. All errors are mine alone. This paper is based upon work supported in part by the National Science Foundation under ITR/IM grant IIS-0081334 and a Sloan Research Fellowship. Any opinions, findings, and conclusions or recommendations expressed above are those of the authors and do not necessarily reflect the views of the National Science Foundation or the Sloan Foundation.

Bibliography

1

Adjukiewicz, Kazimierz. 1935. Die syntaktische Konnexität. Studia Philosophica, 1:1-27. English translation available in Storrs McCall, editor, Polish Logic 1920-1939, Clarendon Press (1967).

2

Chomsky, Noam.
1957. Syntactic Structures. Number IV in Janua Linguarum. Mouton, The Hague, The Netherlands.

3

Firth, John Rupert. 1957. A synopsis of linguistic theory 1930-1955. In the Philological Society's Studies in Linguistic Analysis. Blackwell, Oxford, pages 1-32. Reprinted in Selected Papers of J. R. Firth, edited by F. Palmer. Longman, 1968.

4

Good, Irving J. 1953. The population frequencies of species and the estimation of population parameters. Biometrika, 40(3,4):237-264.

5

Harris, Zellig. 1951. Methods in Structural Linguistics. University of Chicago Press. Reprinted by Phoenix Books in 1960 under the title Structural Linguistics.

6

Lambek, Joachim. 1958. The mathematics of sentence structure. American Mathematical Monthly, 65:154-169.

7

Lewis, David K. 1970. General semantics. Synthèse, 22:18-67.

8

Montague, Richard. 1974. Formal Philosophy: Selected Papers of Richard Montague. Yale University Press. Edited by Richmond H. Thomason.

9

Mosteller, Frederick and David L. Wallace. 1984. Applied Bayesian and Classical Inference: The Case of the Federalist Papers. Springer-Verlag. First edition published in 1964 under the title Inference and Disputed Authorship: The Federalist.

10

Pollard, Carl and Ivan Sag. 1994. Head-driven phrase structure grammar. Chicago University Press and CSLI Publications.

11

Saffran, Jenny R., Richard N. Aslin, and Elissa L. Newport. 1996. Statistical learning by 8-month-old infants. Science, 274(5294):1926-1928, December.

12

Shannon, Claude E. 1948. A mathematical theory of communication. Bell System Technical Journal, 27:379-423 and 623-656.

13

Turing, Alan M. 1950. Computing machinery and intelligence. Mind, LIX:433-60.

14

Weaver, Warren. 1949. Translation. Memorandum. Reprinted in W.N. Locke and A.D. Booth, eds., Machine Translation of Languages: Fourteen Essays, MIT Press, 1955.

For further reading

Charniak, Eugene. 1993. Statistical Language Learning. MIT Press.

Jurafsky, Daniel and James H. Martin. 2000. Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition. Prentice Hall. Contributing writers: Andrew Kehler, Keith Vander Linden, and Nigel Ward.

Manning, Christopher D. and Hinrich Schütze. 1999. Foundations of Statistical Natural Language Processing. The MIT Press.

About this document ...

“I'm sorry Dave, I'm afraid I can't do that”: Linguistics, Statistics, and Natural Language Processing circa 2001¹

This document was generated by hand-editing the output of LaTeX2HTML translator Version 2002-2-1 (1.70), with $USE_CURLY_QUOTES = 1; $HTML_VERSION = "4.0,unicode";

Copyright © 1993, 1994, 1995, 1996, Nikos Drakos, Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999, Ross Moore, Mathematics Department, Macquarie University, Sydney.

The command line arguments were:
latex2html -t 'I'm sorry Dave, I'm afraid I can't do that: Linguistics, statistics, and natural language processing circa 2001' -split 0 cmplg.tex

The translation was initiated by Lillian Lee on 2007-06-18

Footnotes

... 2001¹

Appears in Computer Science: Reflections on the Field, Reflections from the Field (Report of the National Academies' Study on the Fundamentals of Computer Science), pp. 111-118, 2004.

... Test²

Roughly speaking, a computer will have passed the Turing Test if it can engage in conversations indistinguishable from that of a human's.

... local?³

Or, perhaps, the files themselves are patient? But our knowledge about the world rules this possibility out.

... alternative.⁴

To take an analogous example in English, consider the non-word-delimited sequence of letters “theyouthevent”. This corresponds to the word sequences “the youth event”, “they out he vent”, and “the you the vent”.

... English.⁵

Interestingly, this claim has become so famous as to be self-negating, as simple web searches on “Colorless green ideas sleep furiously” and its reversal will show.