See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/228868826

Assessing linguistic complexity

Article · January 2008

DOI: 10.1075/slcs.94.07juo

CITATIONS READS
103 2,573

1 author:

Patrick Juola
Duquesne University
95 PUBLICATIONS 2,863 CITATIONS

SEE PROFILE

All content following this page was uploaded by Patrick Juola on 26 May 2014.

The user has requested enhancement of the downloaded file.

 Assessing Linguistic Complexity

Patrick Juola

Duquesne University

[email protected]

Abstract

The question of ”linguistic complexity” is interesting and fruitful.

Unfortunately, the intuitive meaning of ”complexity” is not amenable

to formal analysis. This paper discusses some proposed definitions

and shows how complexity can be assessed in various frameworks.

The results show that, as expected, languages are all about equally

“complex,” but further that languages can and do differ reliably in

their morphological and syntactic complexities along an intuitive con-

tinuum.

I focus not only on the mathematical aspects of complexitgy, but

on the psychological ones. Any claim about ”complexity” is inherently

1
 about process, including an implicit description of the underlying cog-

nitive machinery. By comparing different measures, one may better

understand on human language processing and similarly, understand-

ing psycholinguistics may drive better measures.

1 Introduction

Most people with any background in language have at least an informal un-

derstanding of language complexity — a language is “complex” to the extent

that you have to study in order to pass the exam on it, and in particular to the

amount of stuff you simply have to memorize, such as lists of irregular verbs,

case systems and declension patterns, and apparently arbitrary aspects of

words such as gender. But to compare languages objectively requires a more

formal specification of complexity, ideally one suited to a unified numerical

measurement. Many ad-hoc complexity measures have been proposed, of

which (Nichols 1986) is an obvious example; she counts the number of points

in “a typical sentence” that are capable of receiving inflection.

McWhorter’s definition (McWhorter 2001; McWhorter 2005) encompasses

a number of similar ad-hoc measures (e.g., a language is more complex if it

has more marked members in its phonemic inventory, or if it makes more

2
extensive use of inflectional morphology), but he ties this, at least in theory,

to a single numerical measure – the length of the grammar that a language

requires. Despite the obvious practical difficulties (how do you compare two

different inflectional paradigms, or how do you balance simple morphology

with complex phonology), this provides a reasonable formulation for judging

complexity.

However, as will be discussed in the remainder of this chapter, the ques-

tion of “length” itself raises the issue of in what language (in a mathematical

sense, i.e. set of primitive operations) the description should be made. It

is argued here that information theory provides several different approaches

that yield different answers to questions of “linguistic complexity” — and

that analysis of data from natural languages can shed light on the psycho-

logical and cognitive processes appropriate to such description.

2 Information Theory Basics

2.1 Zipf and Shannon

The clearest statement of the motivation of the current work on an information-

theoretic basis can be found in (Zipf 1949), in his argument about applica-

3
tions of words:

Man talks in order to get something. Hence man’s speech may

be likened to a set of tools that are engaged in achieving objectives.

True, we do not yet know that whenever man talks, his speech is

invariably directed to the achievement of objectives. Nevertheless, it

is thus directed sufficiently often to justify our viewing speech as a

likely example of a set of tools, which we shall assume to be the case.

Human speech is traditionally viewed as a succession of words to

which “meanings” (or “usages”) are attached. We have no quarrel

with this traditional view which, in fact, we adopt. Nevertheless in

adopting this view of ‘words with meanings’ we might profitably com-

bine it with our previous view of speech as a set of tools and stated:

words are tools that are used to convey meanings in order to achieve

objectives. . .

Now if we concentrate our attention upon the possible internal

economies of speech, we may hope to catch a glimpse of their inherent

nature. Since it is usually felt that words are “combined with mean-

ings” we may suspect that there is latent in speech both a more and a

less economical way of “combining words with meanings,’ both from

the viewpoint of the speaker and from that of the auditor.

4
 Information theory provides a way, unavailable to Zipf, of resolving the

two viewpoints he distinguishes. The speaker’s economy requires that he

express messages in the most compact form possible, up to a single word

that can be used in any situation for any meaning. The hearer’s economy

requires that the speaker be easily understandable, and thus that the amount

of message reconstruction effort – including the effort of listening to the

statement –be minimized. A certain minimum amount of information must

be conveyed in the speaker’s ‘signal’ so that a listener can distinguish his

messages, but at the same time, too much useless information will clog things

up. One can thus see that both the speaker and hearer have incentives to

make the channel as efficient and easily understood as possible.

This framework can be placed on a firm mathematical footing. (Shannon

1948; Shannon 1951). Shannon analyzed all communications as a series of

messages along a channel between an information source and a listener, and

established mathematical bounds for the maximum amount of information,

measured in bits, that can be carried along such a channel.1 If less than

this information is sent, some messages will not be distinguishable from each

other. If more is sent, the “extra” is wasted resources. This is, for any source,

a measure of the information content of that source and a lower bound on

5
the amount of time/space/bandwidth necessary to send messages from that

source and be understood. And in particular, to achieve this optimum use

requires a Zipf-like framework, where the most frequently sent messages have

the least “information” associated with them.

This can provide a framework for mathematical analysis of the intuitive

notion of language complexity. All language is a communication between

the speaker and hearer; a message is “complex” if it has a large information

content, and a language is “complex” if sending the message in that language

requires much more bandwidth than the information content of the message.

In practical terms, there are a number of problems with the direct ap-

plication of Shannon’s formalism. It requires the researcher to enumerate

beforehand all possible “messages” that might be sent along with their prob-

abilities. But more serious problems lurk in the theoretical underpinnings.

In the simplest formulation, Shannon’s message probabilities are indepen-

dent, meaning that the chance of a message being sent does not change,

depending upon what other messages have been sent. In natural language,

context matters, and understanding a message may depend crucially on the

content of the previous messages. Finally, the assumption that the receiver

may need to be able to recover the message perfectly might be problematic,

6
since the knowledge I wish to transmit may be less than the language struc-

ture demands; if all I wish to say is that one of my siblings has red hair,

the sex of the relevant sibling is irrelevant, as is their comparative age, but

some languages (e.g. Japanese) may force me not only to specify the sex

of my sibling, but whether they are older or younger — or to engage in a

long, roundabout, marked, complex paraphrasing. In short, the “messages”

of Japanese are different than those of English, suggesting another approach

may be more fruitful.

2.2 Kolomogorov and other complexity definitions

Another important measurement of complexity is that of Kolmogorov com-

plexity (Li and Vitányi 1997). Kolmogorov complexity measures the infor-

mativeness of a given string (not, as in Shannon’s formulation, a message

source) as the length of the algorithm required to describe/generate that

string. Under this formulation, a string of a thousand alternating ‘a’s and

‘b’s would be easily (and quickly) described, while a (specific) random collec-

tion of a thousand ‘a’s and ‘b’s would be very difficult to describe. For a large

corpus of messages, this could be used as an operationalization of the average

amount of information contained per message. In particular, notice that for

7
any given set of messages produced by a Shannon source, it is a very efficient

use of the channel to transmit, instead of a stream of individually coded

messages, an algorithmic generator for the specific stream of interest. This

illustrates the close relationship between Shannon entropy and Kolmogorov

complexity: Shannon’s entropy is an upper bound on (and asymptotically

equal to) Kolmogorov complexity. Although the mathematics required to

prove this is non-trivial, the result can be seen intuitively by observing that

a decompression program and a compressed file can be used to (re)generate

the original string.

Unfortunately, Kolmogorov complexity is formally uncomputable, in a

strict technical sense related to the Halting Problem. Despite this technical

limitation, Kolmogorov complexity is of interest as an unattainable ideal. If,

as argued above, Kolmogorov complexity represents the ultimate possible file

compression, a good file compressor can be seen as an attempt to approximate

this kind of complexity within a tractable formal framework. By restricting

the kind of operations permitted, it may be possible to develop a useful

complexity measurement framework.

One example of such a restriction is that of linear complexity.(Massey

1969; Schneier 1996) Linear complexity addresses the issue by assuming that

8
the reconstruction machinery/algorithm is of a specific form, a linear feed-

back shift register (LFSR, see figure 1) composed of an ordered set of (shift)

registers and a (linear) feedback function. The register set acts as a queue,

where the past few elements of the sequence line politely up in order, while

the feedback function generates the next single text element and adds it to

the end of the queue (dropping the element at the head, of course). The

linear complexity of a given sequence is defined as the size of the smallest

LFSR generating a given sequence, and can be efficiently determined by a

simple computer program (Massey 1969). Such systems are widely used as

cryptographic encoding systems and random number generators because of

their tremendous ability to generate long, apparently complex sequences from

very little information.

What are the implications of this form of complexity? In this framework,

the next element of the sequence (be it a word, morpheme, phoneme, etc.)

is completely determined by a deterministic function of the most recent N

elements of the sequence. As will be discussed later, there is no place in this

framework for the notion of context, lexicon, or even long-term memory.

Another commonly-used framework is that of Ziv-Lempel complexity, (Lem-

pel and Ziv 1976; Ziv and Lempel 1977) (the complexity metric that underlies

9
 Feedback function

N-bit shift register

Figure 1: Sample Linear-Feedback Shift Register (LFSR)

10
most of the ZIP family of commercial file compressors) which by contrast in-

volves long-term memory almost exclusively. In this framework, the system

builds a collection (lexicon?) of previously-seen strings and compresses new

strings by describing them in terms of previously seen items. This adaptive

“lexicon” is not confined solely to traditional lexemes, but can also incorpo-

rate short-range collocations such as phrasal verbs or common combinations.

It has been proven that Ziv-Lempel complexity, as instantiated in LZ77 (Ziv

and Lempel 1977) or similar techniques will, given infinite computing power

and memory, compress any communications stream to any desired closeness

to the theoretical maximum density, without prior knowledge of the mes-

sages or probabilities of the information sources. The popularity of the ZIP

program is a testament to the practical effectiveness of this model.

At least in theory, this argument (although not necessarily the psycho-

logical underpinnings) would apply to any other proposed form of file com-

pression.

These techniques provide a method of testing and measuring the amount

of information and the amount of redundancy in any string. In particular, as

will be discussed in the following section, by comparing the analytic results

of a complexity metric on a large set of comparable messages, one can infer

11
a detailed analysis not only of overall complexity, but of its linguistic and

cognitive components.

3 Linguistic Experiments

3.1 Information and language

To perform this analysis, we first must consider the types of information of

interest. Elsewhere (Juola 1997), it has been argued that the information

relevant to a text can be broken down into four major categories :

• the complexity of the idea(s) conveyed

• the complexity of the author’s style

• the complexity mandated by the language in which the author writes

• the shared information omitted between the author and her audience

The third aspect is of course what is traditionally meant by “linguistic

complexity”; by holding other aspects constant and varying the language

in which a particular communication occurs, this can be observed numeri-

cally as an overall language complexity metric. Neglecting for a moment the

question of individual authorial style, we can compare this by measuring the

12
information contained in several expressions of the same ideas. A language

with high “complexity” in McWhorter’s sense will require more measured

information in samples written in that language.

As an example, consider a simple propositional message, such as my

brother has red hair. From a purely semantic analysis of the meanings of

the words, it is apparent that a single person is being spoken of and that

the person is neither myself, nor the listener. It is not necessary to transmit

this explicitly, for example via third person singular verb inflection. The

“information” in this inflection is therefore additional complexity demanded

by the structure of the language in which it is expressed. We can further

distinguish between redundancy and complexity in the nature of this infor-

mation. If English (or any language) were perfectly regular, the nature of the

inflection would be prefectly (and redundantly) predictable from the seman-

tics expressed. In a more realistic setting, the inflection carries additional

information necessary to describe the inflection itself (e.g., the irregularity of

the verb to have) that is irrelevant, unnecessary, and “complex.” In theory,

a perfect compression program would be able to extract all the redundant

information, but would still need to track the complexity.

13
 Other examples of this sort of mandatory complexity would include gen-

der markings, various forms of agreement, different lexical paradigms such as

conjucations or declensions, and different syntactic structures such as word

order, passivization, and so forth. We can continue this analysis by focusing

on the information contained in specific aspects or levels of language. For ex-

ample, a language that is morphologically rich would be expected to contain

much of its mandatory information in the system of inflectional morphology.

By systematically distorting the samples to remove inflectional morphology,

one can determine the importance of this type of complexity in the overall

measure.

3.2 Validating Complexity Measurements

A first requirement for any empirical test of these theories is a suitable

testbed; a collection of the same messages (as far as possible) expressed

in different languages. Fortunately, this material is widely available in trans-

lated literature; the Bible, in particular, is an ideal sample.(Resnik, Olsen,

and Diab 1999) It is relatively large, widely available, for the most part free

of copyright restrictions, and generally well-translated.

In conjunction with the definitions of the previous section, we can then

14
formalize common wisdom (pace McWhorter) regarding linguistic complexity

as follows:

For any complexity metric, the measured complexity of the

Bible should be the same regardless of the language or translation.

A preliminary study (Juola 1998) has provided some data in support

of this hypothesis. Using Bible versions in six languages (Dutch, English,

Finnish, French, Maori, and Russian), it was shown that the variation in

size of the uncompressed text (4242292 bytes, +/- 376471.4, or about 8.86%

variation) was substantially more than the variation in size after compression

via LZ (1300637 bytes, +/- 36068.2, about 2.77%). This strongly suggests

that much of the variance in document size (of the Bible) is from the charac-

ter encoding system, and that the underlying message complexity is (more)

uniform.

We can contrast this with the results found (Juola 2000) by a comparable

study of the linear complexity of the Bible translations. In this study, it was

shown that the linear complexity of the Bible varies directly with the number

of characters in the relevant translation, and that there is no descriptive

advantage to the linear complexity framework. (We are tempted to draw the

15
related conclusion that this similarly illustrates that the underlying process

behind linear complexity does not describe human cognition well.)

Ideally, we would have some way of directly validating compression-based

complexity measurements. One approach to this has been by setting up artifi-

cial environments such as inflectional paradigms and to vary the complexity

of the paradigms. One such experiment (Juola, Bailey, and Pothos 1998)

developed an artificial lexicon and systematically varied the percentage of

words subject to a simple inflectional “rule,” from 0 to 100%. The resulting

word list(s) were subjected to an operation the researchers termed “scram-

pression,” repeated permutation followed by compression using the LZ for-

mulation. The results, replicated here as figure 2, show that, as expected, the

“simplest” system is no inflection at all, that the system where all words were

inflected was more complex (reflecting the additional complexity of the rule

itself), but that the measured complexity varied smoothly and continuously,

and that the most complex system (as predicted by Shannon’s mathematics)

was at intermediate stage where the question of whether a word was subject

to the rule was itself complex.

From a mathematical and methodological perspective, these results in-

dicate that LZ-based complexity measurements are practical to take from

16
 315
estimated C

310

305

300

295

290

285
0 10 20 30 40 50

17
Figure 2: Variations in measured complexity as number of inflected words
varies from 0 (0%) to 55 (100%)
appropriate corpora, and give meaningful measurements. Linear complexity

measurements can also be taken, but do not map to our preliminary intuitions

and cannot be validated against the world. From a psycholinguistic perspec-

tive, this argues strongly that LZ, with its focus on the lexicon and long-term

storage and retrieval of words is a better model of the underlying regular-

ities of language as expressed in corpora than linear complexity is. This is

of course unsurprising; no sane person would believe a theory of language

processing that did not include a lexicon. But this also suggests that careful

choice of a model (and comparison of results) can illustrate other aspects of

the cognition underlying human language, either in ways that it is the same

or different from traditional, information-theory based computer programs.

In particular, there are lots of other non-LZ based compression programs,

many of which apparently work better on language than LZ-compression.

Can we infer properties of the human mind from these programs?

3.3 Further compression analyses : Translation

The previous section discussed three positive factors in linguistic complexity

in conjunction with an apparent negative aspect, “the complexity/information

omitted as shared between the author and her audience.” One observation

18
made by scholars of the translation process is that text in translation tends

to be more explicit than the source material. This is easy to understand

from a theoretical perspective. Information believed to be held in common

between the speaker and hearer need not be expressed. For example, a writer

and a reader who share a common knowledge of the city layout can be less

explicit in giving directions than a writer and her out-of-town visitor.

One of the more obvious example of common knowledge is, of course, the

(shared) knowledge of a language as held between two speakers. Attached

to this formal knowledge, in addition, is a host of other associations, cul-

tural cues, and general knowledge that can be broadly lumped together as

“cultural context.” It is reasonable to assume that two fluent speakers of

the same language/dialect share a tremendous amount of associations and

general knowledge that is not necessarily a formal aspect of the language;

just as a physicist can omit the definition of “momentum” when talking

to another physicist, a native Englishman can omit the observation that

Manchester is north of London when talking to another Englishman, but

not necessarily when speaking to a foreigner. Similarly, these cities’ loca-

tions would probably be known to a British reader but not necessarily to

the readers of the same work in translation. (Baker 1993) suggests that this

19
feature of increased specificity may be a universal aspect of the translation

process and presents a brilliantly clear example, where the (English) sentence

“The example of Truman was always in my mind” is translated, into Arabic,

into a huge paragraph—the English translation of this paragraph containing

four sentences and ninety-one words, explaining Truman’s relationship to

the American Presidency and the end of the Second World War, his political

background, and the metaphorical relevance of this “example.” This sort of

implicit information need not even necessarily be “factual,” but can instead

be habits of thought, expectations, associations, and so forth.

If true, this conjecture and theoretical framework may allow us to measure

the degree of shared knowledge between a reader and writer, even between

the writer and a “general audience.” For the simplest case, that of two

different linguistic communities viewed as separate “general audiences,” the

primary difference is, of course, the social and cultural context of both. This

conjecture has been tested directly (Juola 1997) using the LZ compression

test defined above.

In particular, the Biblical data defined above incorporate the original

language as one of the test set.2 If the translation process adds information

as suggested above, the Hebrew original should be significantly shorter than

20
 Language Size (raw) Size (compressed)
English 859,937 231,585
Dutch 874,075 245,296
Finnish 807,179 243,550
French 824,584 235,067
Maori 878,227 221,101
Russian 690,909 226,453
Mean 822,485 235,775
Deviation 70,317 10,222
HEBREW (506,945) (172,956)

Table 1: Sizes and information content (in bytes) of Pentateuch

any of the translations after compression. As can be seen in table 1, this is

true.

Again, we see that the variation in compressed size is much smaller than

the original, and we see that Hebrew is substantially smaller than either.

Of course, written Hebrew omits vowels, but we can see that this is not the

primary cause of this by observing other texts.

Table 2 shows similar sizes for a variety of translations of Orwell’s 1984,

originally written in English and made available in a variety of languages

by the Copernicus 106 MULTEXT-East Project3 or by the ECI/MCI corpus

from the ACL. Of the nine versions of 1984, the two smallest are the (original)

English. Similarly, the smaller version of Dark Freedom (again distributed

by ECI/MCI) is the Uzbek original.

21
 Work Language Size
1984 ENGLISH (v2) 228,046
1984 ENGLISH 232,513
1984 Slovene 242,088
1984 Croat (v2) 242,946
1984 Estonian 247,355
1984 Czech 274,310
1984 Hungarian 274,730
1984 Romanian 283,061
1984 Bulgarian 369,482

Dark Freedom UZBEK 80,209

Dark Freedom English 90,266

Table 2: Information content of other works

We thus see that the original version of corpora tend to be smaller than

their translations, in keeping with the theoretical predictions of the transla-

tion scholars.

3.4 Complexity via distortion : Preliminary experiments

With this framework in hand, we are able to address more specific questions,

such as the comparative role of syntax and morphology in the complexity of a

particular language. In particular, just as one could measure the complexity

of a morphological rule by systematically varying the degree to which it

applied, so one can distort the morphology of a language sample as a whole

to see the effects of “morphological complexity.”

22
 As before, we start with some definitions. In particular, we treat the

ability to predict parts of a word token as a measure of morphology. For

example, a plural context predicts a plural noun, which in English is typically

marked with the -s suffix. The fact that the suffix -ing often signals a present

participle and therefore something likely to follow the string “I am” is a

morphological regularity. If the noun form is suppletively irregular, then

there is no morphological complexity involved — simply a different lexical

item. A morphologically complex language, under this view, is simply one

where the information conveyed by these processes contributes substantively

to the information conveyed by the entire text: for example, one where the

agent/patient relationships cannot be determined by examination of the word

order but can be determined by inspection of (e.g.) the regularities in the

endings of the individual words.

By systematically distorting these regularities, we can measure the role

of morphological information within a corpus. For example, rearranging

the word order would destroy syntactic relationships (one would no longer

be able to identify agent and patient in English), but not morphological

ones (they would still be identifiable in strongly case-marked languages).

Morphological regularities can be easily hidden by simple type-substitution.

23
 jump walk touch 8634 139 5543
jumped walked touched 15 4597 1641
jumping walking touching 3978 102 6
Figure 3: Example of morphological degradation process

Consider figure 3: by replacing each type in the input corpus with a randomly

chosen symbol (here a decimal number), the simple regularity between the

rows and columns of the left half of the table have been replaced by arbitrary

relationships; where the left half can be easily described by a single row and

column, the entire right half would need to be memorized individually. More

accurately, this process is carried out tokenwise where each token of a given

type is replaced by the same (but randomly chosen) symbol which is unique to

each type. This represents a hypothetical alteration to a language where the

morphological rules have been “degraded” into a situation where all versions

of related words can only be treated as suppletive forms, or alternately where

the morphological regularity has been drastically altered in favor of a very

complex system of lexical choice.

Note, however, that this does not eliminate lexical information. If, in

the original text, “I am” predicted a present participle, the new symbols

that replace “I am” will predict (the set of) symbols which correspond to

and replace present participles. However, because these symbols have been

24
randomly rewritten, there will be no easy and simple properties to determine

which symbols these are. Examining figure 3 again, the third row contains

the present participles. In the left hand side, all entries in this row are

instantly recognizable by their “-ing” morphology; the right hand side has

no similar test. In other words, there is no longer a regular (morphological)

way of predicting anything about the new token. Compressing the resulting

substituted file will show the effects on the “complexity,” i.e. the role of

morphological complexity.

Performing this substitution has shown (Juola 1998) that languages dif-

fer strongly in the way and to the degree that this substitution affects their

(compressed) sizes. The results are attached as table 3. As can be seen,

the resulting r/c ratios sort the languages into the order (of increasing com-

plexity) Maori, English, Dutch, French, Russian, Finnish. It is also evident

that there is significant (phonological) information which is destroyed in the

morphological degradation process, as three of the six samples actually have

their information content reduced.

Validation of these numbers, of course, might be problematic; although

few would argue with any measure of morphological complexity that puts

Russian and Finnish near the top (and English near the bottom), there are

25
 Table 3: Size (in bytes) of various samples
Language Uncompressed Comp.(raw) Comp. (“cooked”) R/C Ratio
Dutch 4,509,963 1,383,938 1,391,046 0.994
English 4,347,401 1,303,032 1,341,049 0.972
Finnish 4,196,930 1,370,821 1,218,222 1.12
French 4,279,259 1,348,129 1,332,518 1.01
Maori 4,607,440 1,240,406 1,385,446 0.895
Russian 3,542,756 1,285,503 1,229,459 1.04

few numbers against which to compare this ranking. Fortunately, three of

the languages in this sample (English, French, and Russian) are also part of

the sample addressed in (Nichols 1992). Nichols finding is that English is

less complex than either Russian or French, which are themselves equivalent.

The ranking here agrees with Nichols’ ranking, placing English below the

other two and French and Russian adjacent. This agreement, together with

the intuitive plausibility, provides a weak validation.

Further examination of the corpora indicates some other interesting find-

ings. These findings, presented as table 4, can be summarized as the ob-

servation that (within the studied samples) the ordering produced by the

complexity-theoretic analysis is identical with the ordering produced by the

number of word types in the samples, and identical-reversed with the ordering

produced by the number of word tokens. (This identity is highly significant;

Spearman’s rank test yields p < 0.0025, while Kendall’s T test yields p <

26
 Table 4: R/C ratios with linguistic form counts
Language R/C Types in sample Tokens in sample
Maori 0.895 19,301 1,009,865
English 0.972 31,244 824,364
Dutch 0.994 42,347 805,102
French 1.01 48,609 758,251
Russian 1.04 76,707 600,068
Finnish 1.12 86,566 577,413

Table 5: Type/Token ratio across cased/uncased languages

Concept/Case English Latin
Nominative (the) night nox
Genitive of (the) night noctis
Dative to (the) night nocti
Accusative into (the) night noctem
Ablative in (the) night nocte
Number of types 5(6) 5
Number of tokens 9(14) 5

0.001.) In other words, languages which are morphologically complex tend

to have a wide variety of distinct linguistic forms, while languages which are

morphologically simple have more word tokens, repeating a smaller number

of types, as in table 5.

This finding is intuitively plausible; the linguistic constructs which in

a language like Russian or Latin are expressed in terms of morphological

variation are expressed in other languages through function words—which

almost by definition are few types but many tokens. However, this finding is

not necessary; it is theoretically possible for a language to exist that makes

27
use of an extraordinarily wide variety of function word types (thus inflating

the number of types) or that inflates the number of tokens (for example

by indicating plurality with repeated tokens). This finding then is further

evidence for the approximate equality of overall linguistic complexity, at least

within this sample.

3.5 Complexity via distortion : Subsequent experiments

A larger-scale version of these results (Juola 2005) confirms these findings.

An analysis of twenty-four Bibles, including nine English translations and six-

teen non-English versions, shows, again, languages are about equally com-

plex, but that they express their complexity differently at different levels.

(Table 6 shows the languages represented in this study.)

Methodologically, this study was slightly different, both the scope of the

levels studied, but also in how distortion was performed. All samples were di-

vided identically into verses (e.g. Genesis 1:2 represented the same “message”

as best the translators could manage). Samples were distorted morphologi-

cally by the deletion of 10% of the letters in each verse at random. Samples

were distorted syntactically by the deletion of 10% of the words (maximal

non-blank sequences), while pragmatic distortion was obtained by deleting

28
 English Versions Non-English Versions
American Standard Version (asv) Bahasa Indonesian (bis)
Authorized (King James) Version (av) Portuguese (brp)
Bible in Basic English (bbe) Haitian Creole (crl)
Darby Translation (dby) French (dby)
Complete Jewish Bible (jps) Finnish (fin)
Revised Standard (rsv) Hungarian (karoli)
Webster’s Revised (rweb) Dutch (lei)
Young’s Literal Translation (ylt) German (lut)
Modern Greek (mgreek)
Modern Greek [unaccented] (mgreeku)
French (neg)
Russian (rst)
German (sch)
German (uelb)
Ukranian (ukraine)
Table 6: Bible samples and languages used

10% of the verses themselves at random. We justify this as an exploration

of “pragmatic” complexity by noting that sentences themselves must be in-

terpreted through context. For example, the use of pronouns hinges on the

availability of antecedents; languages where subjects can be dropped alto-

gether make even heavier use of context. Just as elimination of words can

distort the structure of individual sentences and make them difficult to fol-

low, so eliminating previous sentences or verses, while having no effect on the

syntactic acceptability of the individual sentences, will make the pragmatics

of the discourse difficult to follow.

Although all three procedures will provably delete the same number of

29
expected characters from the source files, the effects of compression on these

files was expected to differ as argued above. Compression was performed

4
using the UNIX gzip utility.

Table 7 shows the compressed and uncompressed sizes (in bytes) of the

various Biblical translations; Table 8 presents the (compressed) sizes both

of the original and of the various distorted versions. Table 9 normalizes all

results by presenting them as multiples of the original size (e.g. the morpho-

logically distorted version of ’asv’ is 1.17 times its original size when both

are compressed). This table is also sorted in order of decreasing value, and

English versions are labeled in ALL CAPS for ease of identification. Exam-

ining these tables together supports the hypotheses presented earlier. First,

all languages appear to be more uniform in size in their compressed than

in their uncompressed sizes. Second, languages appear to differ in a reliable

way in the degree to which their complexity is expressed in morphology or

syntax; languages with high measured morphological complexity (as mea-

sured via morphological distortion) have low measured syntactic complexity

(as measured via syntactic distortion). A new finding of this experiment is

that all languages appear to be about equal in their pragmatic complexity;

the variation in column 3 of table 9 is almost nonexistent. This suggests that

30
 Version Uncompressed Compressed
asv 4280478 1269845
av 4277786 1272510
bbe 4309197 1234147
bis 4588199 1346047
brp 3963452 1273549
crl 4351280 1312252
dby 4227529 1265405
drb 4336255 1337768
fin 4169287 1370042
jps 4309185 1288416
karoli 3957833 1404639
lei 4134479 1356996
lsg 4252648 1347637
lut 4180138 1341866
mgreek 4348255 1468263
mgreeku 4321200 1341237
neg 4188814 1323919
rst 3506401 1269684
rsv 4061749 1253476
rweb 4247431 1262061
sch 4317881 1417428
uelb 4407756 1383337
ukraine 3564937 1315103
ylt 4265621 1265242

Table 7: Uncompressed and compressed sizes (bytes) of various Biblical

translations

31
conversation-level constructions are handled similarly in all languages stud-

ied, perhaps reflecting underlying similarities in overall cognitive processing.

One minor point needs to be addressed. It is intuitively plausible that

deleting words or phrases from a corpus should result in the overall lowering

of information. It is less plausible that deleting letters should result in the

overall increase in information — but that is a clear result of table 9. The ex-

planation is simple, but unhelpful. Essentially, by deleting letters randomly,

the computer is creating a number of variant spelling of the same word, or

alternatively a number of freely varying synonyms for each lexical item. Un-

surprisingly, the compression program needs to determine (and store) which

of the variant spellings is used in each instance. Strongly inflecting languages

such as Finnish and Hungarian, ironically, suffer the least penalty because

they have more word forms, and therefore fewer new variants are introduced

in each one. Thus, the results show both that Finnish and Hungarian have

high morphological complexity but low syntactic complexity, a fact that is

both intuitivel and supported by prior experiments.

It should finally be noted that in this sample, there is a single example of a

creole language present. Sample ’crl’ is Haitian Creole. Under McWhorter’s

hypothesis, this should be the smallest in compressed size. It is, in fact, about

32
Ver.. Normal Morph. Prag. Syn.
asv 1269845.00 1487311.21 1162625.76 1235685.14
av 1272510.00 1483580.35 1164269.82 1236878.18
bbe 1234147.00 1461058.00 1130314.66 1207880.69
bis 1346047.00 1601727.60 1229751.14 1282227.44
brp 1273549.00 1452517.41 1164547.00 1226345.93
crl 1312252.00 1518413.44 1200910.23 1285958.47
dby 1265405.00 1474721.85 1158317.63 1229267.18
drb 1337768.00 1557772.69 1224481.67 1295321.34
fin 1370042.00 1548445.81 1251576.37 1294915.38
jps 1288416.00 1504061.39 1179304.44 1253465.97
karoli 1404639.00 1575233.97 1283223.98 1330816.55
lei 1356996.00 1522644.85 1239991.16 1297103.54
lsg 1347637.00 1560146.19 1233249.90 1300432.48
lut 1341866.00 1529869.88 1226982.62 1285579.01
mgreek 1468263.00 1701221.45 1343161.70 1408603.16
mgreeku 1341237.00 1550600.30 1226936.07 1292009.44
neg 1323919.00 1533468.52 1211325.57 1277609.72
rst 1269684.00 1410919.44 1161101.89 1204901.46
rsv 1253476.00 1445795.94 1147694.76 1215960.88
rweb 1262061.00 1471718.69 1154952.77 1227624.14
sch 1417428.00 1604828.85 1295267.52 1353695.77
uelb 1383337.00 1598355.09 1265317.06 1329331.17
ukraine 1315103.00 1451550.59 1201829.66 1247197.05
ylt 1265242.00 1482038.11 1158189.17 1230766.62

Table 8: Biblical sizes (bytes) after various transformations

33
Ver.. Morph. Prag. Syn.
ukraine 1.10375 0.913867 0.948365
rst 1.11124 0.914481 0.948977
karoli 1.12145 0.913561 0.947444
lei 1.12207 0.913777 0.955864
fin 1.13022 0.913531 0.945165
sch 1.13221 0.913815 0.955037
lut 1.14011 0.914385 0.958053
brp 1.14053 0.914411 0.962936
RSV 1.15343 0.91561 0.970071
uelb 1.15543 0.914685 0.96096
mgreeku 1.1561 0.914779 0.963297
crl 1.15711 0.915152 0.979963
lsg 1.15769 0.91512 0.964972
neg 1.15828 0.914954 0.965021
mgreek 1.15866 0.914796 0.959367
drb 1.16446 0.915317 0.968271
DBY 1.16541 0.915373 0.971442
AV 1.16587 0.91494 0.971999
RWEB 1.16612 0.915132 0.972714
JPS 1.16737 0.915313 0.972874
ASV 1.17125 0.915565 0.973099
YLT 1.17135 0.915389 0.972752
BBE 1.18386 0.915867 0.978717
bis 1.18995 0.913602 0.952587

Table 9: Normalized sizes (bytes) after various transformations

34
average. Readers who wish to regard this as evidence against McWhorter’s

hypothesis are sharply reminded of the dangers of drawing inferences from a

single data point. As discussed in the next section, a much larger experiment

on this scale should be run.

4 Discussion

The results presented above can be viewed as an extended proof of concept

for the idea that “language complexity” is a meaningful concept subject to

empirical testing. In particular, the terminology and concepts of informa-

tion theory can provide a structure for developing such studies. Taking a

functionalist view of language — language serves to pass a message between

the speaker and hearer — one can compare the mathematical properties of

several different expressions of the same message.

In particular, by focusing on different translations of “the same” text,

we can establish whether there are underlying differences in these properties

that characterize the language in which the texts are written. By systemat-

ically distorting a single base text, we can establish whether this distortion

introduces systematic changes in the properties.

The key questions are thus: what are the (mathematical) properties that

35
can be measured, and how do they relate to the (linguistic) properties of

general interest? In particular, how do our intuitive understandings of “com-

plex” relate both to language and to mathematics?

The variety of definitions of complexity, both linguistic and mathematical,

suggest that this relationship is not well-understood within either community.

One possible approach to resolving this is to examine more closely the notion

of process and the relationship of the processes captured in both psycholin-

guistics and the mathematical formalisms. In particular, most complexity

formalisms start from a set of mathematical or algorithmic “primitives,” and

describe the complexity of an object in terms of the number of primitives re-

quired for it to function. For example, the fundamental primitives underlying

Kolmogorov complexity are exactly that underly Turing machines — reading

and writing unanalyzed symbols in an unbounded memory space. The fun-

damental primitives of LZ complexity include storage and retrieval from a

learned “lexicon” of expressions. The “complexity” is defined as the number

of primitives in a description of the process, ignoring completely the time

element (how many primitives need actually be performed). By contrast,

linear complexity assumes a very strong bound on the amount of memory

available, and measures the complexity by the size of this very bound. Since

36
human memory is known to have a large capacity, it makes intuitive sense

that linear complexity should not well measure aspects of human cognition –

while the plausibility of the lexicon strengthens the plausibility of LZ-based

complexity measurements..

The framework developed in this paper thus provides researchers with a

new way of investigating theories of the psycholinguistic processing of lan-

guage. Specifically, by identifying the theoretical “primitives” implicit in the

psychological theory, and developing a text compression method based upon

those primitives, the results of compressing should be at least as plausible and

at least as successful as the results obtained by other compression methods

such as LZ-compression.

Similarly, a good theory of language processing should be able to capture

the underlying regularities of human language in a sufficiently practical way

that will be useful for the quite practical problem of text compression. With

the ever increasing amount of text available (Nerbonne 2004), the ability to

store ideas in a way that matches the efficiency of the human mind is almost

certainly a desirable goal.

Finally, examination of algorithmic representation in terms of linguistic

primitives may provide medical and psychological advantages as well. If our

37
theory represents human performance under ideal conditions, studying how

performance changes when theory is tweaked slightly may illustrate or ex-

plain aspects of human performance under less than ideal conditions, includ-

ing both situational and cognitive degradation. This may improve science’s

ability both to diagnose and to treat linguistic/cognitive disorders.

5 Future Work and Conclusions

Briefly to recap the findings of the previous sections: the question of “lin-

guistic complexity” can be formalized objectively and with relative precision

using information theory. Depending upon the exact definition of “complex-

ity” used, different measurements will be obtained. In particular, by using

multiple translations of the same basic text (for example, the Bible or the

novel 1984), one can see whether “all languages are about equally complex”

as received wisdom requires, whether a particular language or language group

is systematically less complex as McWhorter suggests, and whether there any

interesting variation or pattern in complexity.

Performing this experiment shows that languages do appear to be about

equally complex when compared under the Ziv-Lempel definition of complex-

ity. Less psycholinguistically plausible definitions, such as linear complexity,

38
show correspondingly less plausible results. In particular, the variation in

text lengths of uncompressed versions of the Bible was significantly greater

than the variations in compressed lengths.

Further studies showed that varying particular aspects of language com-

plexity — the regularity or irregularity of morphology — could similarly

be detected by compression-based information theory measurements. Well-

known aspects of language-in-translation, specifically the observation that

translated texts need to be more explicit so that the reader can understand

them against a different cultural background, can be observed by noting that

translated texts are more complex, in this framework, than their originals.

Finally, by systematically distorting particular types of expression, one can

measure the role that type of expression plays in the overall complexity of

language. Using this framework, studies have shown that languages differ

in their morphological and/or syntactic complexity, and in particular, that

languages with high morphological complexity have low syntactic complex-

ity and vice versa. This is compatible with the tool-and-effort framework of

Zipf and of Shannon; a language capable of sending information via morphol-

ogy need not encode the same information syntactically A further finding —

that languages do not appear to differ in their pragmatic complexity — may

39
reflect the underlying cognitive universals of the human mind/brain system.

These findings only scratch the surface of what could be explored un-

der this framework. In particular, McWhorter’s theory that creoles are less

complex than older languages could be tested by collecting a corpus of texts

written in a variety of languages including both creoles and non-creoles. More

carefully, one would want a text sample, written in one language and inde-

pendently translated into a second (creole) and third (non-creole) language.

Assuming that the same concepts are expressed in each translation, any sig-

nificant size differences could only be attributed to the structural properties

of the languges. If the creole translations were systematically smaller than

the non-creole ones, this would be strong evidence supporting McWhorter’s

theoretical observations.

Another area of immediate research could be in the study of other for-

malizations of complexity and their usefulness and plausibility as linguistic

measurements. The three concepts presented above are only a few of the

dozens of definitions proposed and used by information theorists as com-

pression technology. A systematic study of how different measurements vary

would help researchers begin to understand which methods give intuitively

plausible results and which don’t.

40
 Beyond this, the primary difference between the intuitively plausible re-

sults of the Ziv-Lempel metric and the implausible ones of linear complexity

have been argued to be a result of the difference in handling of long-term

memory and the lexicon. Can we find other methods that illustrate or demon-

strate other important aspects of cognition?

This is the ultimate promise of complexity studies. A recent problem in

linguistics has been dealing with the results of corpus-oriented studies; the

problem of inferring linguistic processes from simply collections of products.

The process-oriented analysis implicit in information theory may provide a

useful tool to help in this inference and bridge the gap between the paper

and the mind.

Notes
1
More formally, Shannon demonstrated that, given a source S, capable of sending any

of N messages i = 1, . . . , N , the average length of a message (in bits), is given by

N
X
H(S) = pi · l i (1)
i=1

where pi is the probability that message i is sent, and li is the length of message i.

The optimum lengths are when li is equal to the negative logarithm of pi . Thus,

41
 N
X
H(S) = −pi log2 (pi ) (2)
i=1

2
The Bible as a whole is actually somewhat problematic in this regard as there is no

single original language. For this reason, we momentarily restrict our attention to the first

five books, which are all originally written in Hebrew.

3
This project is described at http://nl.ijs.si/ME/Corpus/mte-D21M/mte-D21M.html

4
Another analysis, using a slightly different compression scheme (bzip2, using the

Burrows-Wheeler transform), obtained similar results but will not be discussed further.

References

Baker, M. (1993). Corpus linguistics and translation studies : Implications

and applications. In M. Baker, G. Francis, and E. Tognini-Bonnelli

(Eds.), Text and Technology : In Honour of John Sinclair, pp. 233–

250. Amsterdam: John Benjamins.

Juola, P. (1997). A numerical analysis of cultural context in translation. In

Proceedings of the Second European Conference on Cognitive Science,

Manchester, UK, pp. 207–210.

Juola, P. (1998). Measuring linguistic complexity : The morphological tier.

Journal of Quantitative Linguistics 5 (3), 206–13.

42
Juola, P. (2000, August). A linear model of complexity (and its

flaws). In Fourth International Conference on Quantitative Linguistics

(QUALICO-2000), Prague, Czech Republic.

Juola, P. (2005, September 24–26). Compression-based analysis of lan-

guage complexity. In Presented at Approaches to Complexity in Lan-

guage, Helsinki, Finland.

Juola, P., T. M. Bailey, and E. M. Pothos (1998, August). Theory-neutral

system regularity measurements. In Proceedings of the Twentieth An-

nual Conference of the Cognitive Science Society (CogSci-98), Madison,

WI.

Lempel, A. and J. Ziv (1976, January). On the complexity of finite se-

quences. IEEE Transactions on Information Theory IT-22 (1), 75–81.

Li, M. and P. Vitányi (1997). An Introduction to Kolmogorov Complexity

and Its Applications (2nd ed.). Graduate Texts in Computer Science.

New York: Springer.

Massey, J. L. (1969, January). Shift-register synthesis and BCH decoding.

IEEE Transactions in Information Theory IT-15 (1), 122–7.

McWhorter, J. H. (2001). The world’s simplest grammars are creole gram-

43
 mars. Linguistic Typology 6, 125–166.

McWhorter, J. H. (2005). Defining Creole. Oxford: Oxford University

Press.

Nerbonne, J. (2004, June). The data deluge. In Proc. 2004 Joint In-

ternational Conference of the Association for Literary and Linguis-

tic Computing and the Association for Computers and the Humanities

(ALLC/ACH 2004), Göteborg, Sweden. To appear in Literary and Lin-

guistic Computing.

Nichols, J. (1986). Head-marking and dependent-marking grammar. Lan-

guage 62, 56–117.

Nichols, J. (1992). Linguistic Diversity in Space and Time. Chicago, IL:

University of Chicago Press.

Resnik, P., M. B. Olsen, and M. Diab (1999). The Bible as a parallel

corpus: Annotating the ‘Book of 2000 tongues’. Computers and the

Humanities 33 ((1-2)), 129–153.

Schneier, B. (1996). Applied Cryptography, Second Edition : Protocols,

Algorithms and Source Code in C. New York: John Wiley and Sons,

Inc.

44
Shannon, C. E. (1948). A mathematical theory of communication. Bell
View publication stats

System Technical Journal 27 (4), 379–423.

Shannon, C. E. (1951). Prediction and entropy of printed English. Bell

System Technical Journal 30 (1), 50–64.

Zipf, G. K. (1949). Human Behavior and the Principle of Least Effort. New

York: Hafner Publishing Company. Reprinted 1965.

Ziv, J. and A. Lempel (1977, May). A universal algorithm for sequen-

tial data compression. IEEE Transactions on Information Theory IT-

23 (3), 373–343.

45