The “handedness” of language: Directional symmetry breaking of sign usage in words
Language, which allows complex ideas to be communicated through symbolic sequences, is a characteristic feature of our species and manifested in a multitude of forms. Using large written corpora for many different languages and scripts, we show that the occurrence probability distributions of signs at the left and right ends of words have a distinct heterogeneous nature. Characterizing this asymmetry using quantitative inequality measures, viz. information entropy and the Gini index, we show that the beginning of a word is less restrictive in sign usage than the end. This property is not simply attributable to the use of common affixes as it is seen even when only word roots are considered. We use the existence of this asymmetry to infer the direction of writing in undeciphered inscriptions that agrees with the archaeological evidence. Unlike traditional investigations of phonotactic constraints which focus on language-specific patterns, our study reveals a property valid across languages and writing systems. As both language and writing are unique aspects of our species, this universal signature may reflect an innate feature of the human cognitive phenomenon.
Language - and by extension, writing - distinguishes humans from all other species . The ability to communicate complex information across both space and time have enabled society and civilization to emerge . The recent availability of publicly accessible “big data”, such as the large digitized corpus on the Google Books website, has revolutionized the quantitative analysis of socio-cultural phenomena and led to new empirical discoveries [3, 4, 5]. Language in its written form is represented as symbolic sequences that convey information. Statistical analysis of such sequences have led to the identification of several quantitative properties that hold across many human languages. For example, one of the best known empirical regularities associated with language is the scaling behavior - referred to as Zipf’s law - that quantifies how some words occur far more frequently than others . Several possible theoretical explanations of the phenomenon have been proposed [7, 8]. Words are themselves composed of signs corresponding to letters, syllabograms or logograms depending on the writing system. It has long been known that the different signs, e.g., letters, also occur with characteristic frequencies - a fact that has been used by cryptographers over the ages to break simple substitution ciphers. This was illustrated dramatically in fiction by Poe (The Gold-Bug) and Conan Doyle (The Adventure of the Dancing Men). For English, the phrase “ETAOIN SHRDLU” has often been used as a mnemonic for recalling the approximate order of the most commonly occurring letters in typical texts. However, a cursory glance through an English dictionary (or encyclopedia) to ascertain, for each letter of the alphabet, the number of pages that are required to list all the words (or entries) that begin with that letter, will alert one to a strong deviation from what is naively expected from the frequency distribution of letters. For instance, one of the letters having the largest number of entries in a dictionary is ‘c’ which does not even appear among the most frequently used letters in English as per the phrase above. This apparent anomaly arises from the fact that the letter ‘c’ has a much higher probability of occurring (relative to other letters) at the beginning of an English word - possibly a result of the specific orthography of English, where it can appear as the initial letter of the words china, can, cent, etc., in all of which it is pronounced differently - but does not occur so frequently at other positions. While it is rarer to come across situations where words are arranged according to their last character, it is possible to ask whether the frequency distribution of the letters that occur at the end of a word will similarly show a distinct character. For example, had we arranged the entries of an English dictionary in the order of the last letter of each word, we would see that this would result in the number of entries corresponding to each letter showing a far more unequal distribution than seen in a conventional dictionary. In other words, the last position in an English word is usually occupied by one of very few letters, suggesting that the final letter is much more tightly constrained than the initial letter. We show that this is not just true for English but holds in at least 23 other languages, including those which use writing systems not based on letters (alphabets and abjads), but instead on signs representing syllables or logograms. We have also applied this property of asymmetry in sign usage patterns between the beginning and end of a sequence to an undeciphered corpus of inscriptions, showing that we can determine the direction in which the sequences were written (left to right or right to left).
Fig 1 (a-c) shows the occurrence probability distribution of the 26 letters of the Latin alphabet at the initial and final positions of English words reconstructed from a large database, and compares it with their probabilities of occurring anywhere in a text. We note immediately that letters may differ greatly in terms of their occurrence probability depending on the position - but most importantly, the distribution for the right terminal character (the last letter) in an English word appears to be much more heterogeneous than the one corresponding to the left terminal character (the first letter). In other words, the choice of letters that can occur as the final character of a word is more restrictive, i.e., the occurrence probabilities are more unequal, with very few accounting for the right terminal position for a major fraction of the words, compared to their position-independent probabilities. In contrast, the probability distribution of letters that occur as the initial character is more egalitarian (in comparison to the final one), implying a somewhat higher degree of freedom of choice at the left terminal position. To ensure that this left-right asymmetry in sign usage distributions - suggesting a “handedness” of words in terms of the letter frequency distributions at their terminal positions - is not an artifact of the corpus one is using, we performed the same analysis with Google Books Ngram data, focusing on words that occur with a frequency of more than in the corpus digitized by Google (see Supporting Information for details). As seen from Fig S1, the qualitative features are similar to that observed in Fig 1, indicating the robustness of the observed left-right asymmetry of sign occurrence probability patterns in words.
In this paper we will argue that this directional asymmetry is not just a feature of a particular language but appears to be universal, holding across different languages and writing systems. Regardless of whether the signs we are considering represent letters (for alphabetic scripts like English), syllabograms (for syllabic scripts such as Japanese Kana) or logograms (for logographic scripts like Chinese or logo-syllabic ones like Sumerian cuneiform), the distribution of the signs that begin a word shows relatively less heterogeneity than that for the ones that occur at its end. We have used measures of inequality (viz. Gini index and information entropy) to quantitatively assess the degree of asymmetry in the sign occurrence distributions for different linguistic corpora. The difference in the two distributions also indicate the differential information contents of the initial and final characters - and links our result to the statistical and information-theoretic analysis of language . This approach was pioneered by Shannon who used the concept of predictability, i.e., the constraints imposed on a letter by those that have preceded it, to estimate the bounds for the entropy (the amount of information per letter) and redundancy in English [10, 11]. Considering the consequences of the most prominent structural patterns of texts - viz., the clustering of letters into words - Schürmann and Grassberger subsequently showed that the the average entropies of letters located inside a word are much smaller than that of the letters at the beginning [12, 13]. However, this is true even if one reverses the word - so that terminal letters of words (whether initial or final) have less predictability than those in other positions. Here we ask the relatively simpler question of whether the statistical properties of the left and right terminal characters are different and find a surprising non-trivial asymmetry in the heterogeneity of the respective distributions. Analysis of correlation between sign occurrences in written texts have traditionally focused on the phonotactic constraints of specific languages, e.g., determining the consonants or consonant clusters that are allowed to occur before and after a vowel in any syllable of a given language. While there is considerable variation between different languages as regards the possible arrangements in which consonants and vowels can be combined to make meaningful words, here we show the existence of general patterns that hold across many different language families.
In order to quantify the heterogeneity in sign usage distribution at the beginning and at the end of a word, we have used the Gini index or coefficient . It measures dispersion in the distribution of a quantity and is widely used in the socio-economic literature to quantify the degree of inequality, e.g., in the distribution of income of individuals or households . The value of the Gini index (see Eq. 1 in Materials and Methods) expresses the nature of the empirical distribution relative to a uniform distribution, with if all values of the variable have the same probability of occurrence (“perfect equality”) while corresponds to the extreme situation with the variable always taking up a single value (corresponding to a delta function probability distribution). Thus, if the probability of occurrence of any sign (e.g., the letters ‘A-Z’ in the case of the English alphabet) at the beginning (or end) of a word is about the same, the corresponding Gini index will be close to zero. Otherwise, it is a finite number () whose exact value depends on the extent of inequality in usage of the different signs. Measures related to the Gini index have previously had limited use in the context of linguistic sequences, e.g., to select attributes for decision tree induction in classification for data mining .
Using the Gini index on the distributions of letters (1-gram) that occur at the left and right terminal positions of words in English, we can quantitatively express the visible difference between patterns of unequal occurrence of signs at the two ends seen in Figs 1 (b-c). Fig 1 (d) shows the Lorenz curve - a graphical representation of the inequality of a distribution - for the occurrence probability of the different 1-grams anywhere in a sequence as well as the two terminal positions. For a set of symbols (, , …) that are indexed according to their probability of occurrence in non-decreasing order (), the curve is obtained by joining using linear segments the points (), , where is the cumulative proportion of the population of symbols and is the cumulative proportion of occurrence probabilities. The diagonal line represents the case of complete equality, so that higher inequality is manifested as greater deviation between the empirical curve - showing the cumulative probability of occurrence of signs arranged in a non-decreasing order of occurrence frequency - and the diagonal. The diagram clearly shows that for different signs the probabilities of occurring at the right terminal position, i.e., the end of a word, is more unequal than their occurrence probability in the beginning (i.e., left terminal position of a word), or indeed, anywhere in a sequence. This quantitatively establishes that there is relatively more variation in the letters at the start of a word - and conversely less so when ending it. It indicates an inherent left-right asymmetry in the sign usage distribution of words in English that is related to the different degrees of freedom associated with choosing letters that begin and end a word. This asymmetry is more pronounced for letters or 1-grams, as measured by the normalized difference of Gini indices (defined as where and are the Gini indices for the signs occurring in the left and right terminal positions, respectively) for the two terminal positions, (with bootstrap confidence intervals ; for details see Methods), than for 2-grams (, bootstrap confidence intervals ) and becomes even less noticeable for higher-order -grams. We have therefore focused our analysis on using 1-grams for the subsequent results reported here.
To ensure that the left-right asymmetry does not arise simply as a result of the use of common prefixes (such as de- or un-) or suffixes (such as -ed or -ly) in English, we have also analyzed the Ogden list of Basic English words comprising two or more characters. This is a set of commonly used English root words obtained after removing all affixes or bound morphemes . For this data-set, we obtain (with bootstrap confidence intervals ) clearly indicating that the relative lack of variation in the right terminal position of a word in English is not an artifact resulting from, say, a large number of words ending with a limited set of suffixes.
While this asymmetry in the usage distribution for signs that begin and end words written in English is certainly striking, it would be even more significant if the phenomenon turned out to be valid for linguistic sequences in general. We have, therefore, carried out a systematic investigation of the inequality in sign usage distributions at the terminal positions of sequences that are chosen from languages spanning a broad array of language families. Fig 2 shows the Lorenz curves corresponding to these different corpora, each indicating the differences in the occurrence probabilities of signs at different positions for that language. The writing systems considered are also quite diverse, ranging from alphabetic to logographic, whose corresponding signaries (i.e., the set of distinct characters used for writing in that system) can vary in size from about two dozen to several thousands. Fig 3 shows the results obtained for the different written corpora we have analyzed, where the degree of asymmetry in sign occurrence at the left and right ends of a sequence is measured by the normalized difference between the respective Gini indices. The most important feature of our results is the clear distinction that can be made between languages that are conventionally written left to right, such as English, and those which are written right to left, such as Arabic, according to the sign of obtained for the corresponding corpus. A negative value of implies that the signs occurring in the left terminal position have a relatively more equitable distribution while the sign usage distribution at the right terminal position is more unequal, and conversely for positive comparatively few signs occur with high frequency at the left end of a sequence than the right end. Thus, our result implies that all languages and writing systems considered here exhibit an asymmetry between the beginning and end of a word in terms of the degree of inequality manifested in their respective sign occurrence probability distributions, the probability in choosing different signs being significantly more heterogeneous at the end than in the beginning.
To ensure that the observed distinction between the sign usage patterns at the two terminal positions of a sequence are significant, we compared our results with those obtained from corpora of randomized sequences, which by design have the same distribution of sign occurrences at all positions. For a rigorous comparison, we have used surrogate datasets that have the same frequency distribution of different signs as the original corpus (see Methods for details) so that any distinction between them arises only from differences in the nature of the distributions of sign occurrence at the terminal positions. As randomized sequences are expected not to have any left-right asymmetry in sign usage patterns, the mean value of for the surrogate data is expected to be zero. However, statistical fluctuations will result in the random corpora belonging to the ensemble having small non-zero values of distributed about and the standard deviation of the distribution (indicated by error bars in Fig 3) indicates whether a observed difference in Gini indices can arise by chance even when there is no asymmetry. As seen in Fig 3 almost all the corpora analyzed by us exhibit asymmetries that are clearly distinct from what might be expected if they were just the result of noise.
As the asymmetry observed in the linguistic sequences should not depend on the particular corpus from which they are chosen, we obtained confidence intervals for the empirical values by bootstrap resampling of the data (see Methods for details). Fig 3 shows that in almost all cases this interval does not have any overlap with the interval obtained for randomized sequences - indicating that our results are robust with respect to variations in the corpus. The accuracy of the estimate, which is inversely related to the length of the confidence interval, appears to become higher as the database size, i.e., the total number of sequences being considered, is increased. Indeed, Fig 4 shows that the database needs to be larger than a minimal size ( words for English) in order for the significance of the observed asymmetry to be established. Using increasingly larger databases, the difference between the empirical and randomized corpora become more pronounced.
Apart from the Gini index, the inequality in sign occurrence distribution at the terminal positions of a sequence can also be measured by other means, e.g., using information entropy (also referred to as Shannon entropy), a key concept in information theory. It measures the amount of uncertainty in a process by quantifying the non-uniformity in the probability distribution of different events. Thus, if all signs have equal probability of occurrence at a particular position in any sequence, this would correspond to the highest value of entropy (while the Gini index would have been the lowest in this case). On the other hand, if only a single sign occurs at this position in all sequences (i.e., the case of extreme inequality for which Gini index is highest), then the entropy is zero. Fig S2 shows that using the normalized difference of entropy , estimated from the sign occurrence distributions at the left and right ends of the different linguistic corpora, yields qualitatively similar results to those obtained by using the Gini index. The sign of in almost all cases is seen to be consistent with the direction of writing, with left-to-right written languages having positive values of (with the lone exception of Egyptian Hieroglyphics, for which we have used a database in which all sequences have been oriented so as to read from left to right for all our analysis) while those written right-to-left have . This is in broad agreement with our earlier conclusion that there is relatively more equality in the probability of occurrence of different signs at the beginning of a word than at its end - reflected in the higher non-uniformity for sign usage distribution for the latter. Thus, it suggests that the asymmetry we observe in linguistic sequences may be robust with respect to the specific measure of inequality being used.
Intriguingly, we find that the asymmetry can also appear in a corpus of inscriptions that are so far undeciphered and whose relation to language is therefore not yet established. As an illustration, we have analyzed sign sequences appearing in the archaeological artifacts (e.g., seals, sealings, pottery, copper tablets, etc.) obtained from excavations carried out at sites of the Indus Valley Civilization (IVC) that existed during 2600-1900 BCE in present day Pakistan and northwestern India [18, 19, 20]. While there is some debate as to whether these inscriptions constitute “writing” in the sense of encoding spoken language [21, 22], there is near unanimity among scholars that these were mostly written from right to left as inferred from the archaeological evidence (e.g., signs get more crowded at the left end of some inscriptions or spill out of an otherwise linear arrangement) [23, 24, 25, 26]. We have used a database where the relatively few sequences which are believed to have been written from left to right have been reversed so as to be oriented in the same direction as the majority, following standard procedure used for constructing concordances for Indus Valley Civilization inscriptions. We observe from Fig 3 that the for sign usage distribution is positive, indicating that the choice of signs is less restricted in the right terminal position than the left. This would suggest, based on the connection previously seen between the sign of and the direction of writing, that the IVC inscriptions are written from right-to-left, which corroborates the consensus view as mentioned above.
We have reported evidence here for a novel universal feature in the empirical statistics of linguistic sequences. Unlike the more well-known Zipf’s law and Heap’s law , which relate to the frequency of word usage, we focus on a more elementary level, viz., that of the signs - corresponding to letters, syllabograms or logograms, depending on the writing system - which constitute individual words. The distribution of occurrence for the different signs at the left and right terminal positions in a word are shown to have a distinct heterogeneous character that are characterized by measures of inequality such as the Gini index or information entropy. We observe that, in general, the information content at the beginning of a sequence tends to be higher than at the end, which is reflected in the significant asymmetry in terms of the restriction of sign usage at these the two positions. This is a pattern that is valid across different languages and scripts, possibly revealing a feature inherent in the information processing and communicating capabilities of the human cognitive apparatus.
The reason for the appearance of the directional asymmetry in sign usage distributions for linguistic sequences is yet to be definitively identified. However, it is not unreasonable to expect that this is related to the phonotactic constraints inherent in different languages. The initial sound of a word can be chosen with greater freedom from the set of all available speech sounds (phonemes) of the language, compared to all subsequent sounds that may depend - to a greater or lesser extent - on the sound(s) preceding them. For example, very few of the three-consonant clusters that can in principle occur in English are actually allowed . Thus, one would expect a higher degree of variability in the initial sound compared to the one at the end of a word. As writing reflects the patterns of spoken language, to greater or lesser extent depending on the system, one would expect this difference between the beginning and end to be manifested in it. An indirect indication that phonotactic considerations may be at least partially responsible for the asymmetry is provided by the degree of the difference between the inequalities of sign usage at the two ends of a word in different writing systems - especially when normalized difference is used as a measure. We observe that, broadly speaking, the magnitude of (as well as, ) is larger for scripts that have a higher proportion of phonetic representation [29, 30]. Thus, alphabetic and syllabic systems which have a much greater phonetic character than logographic or logo-syllabic systems tend to typically show a more pronounced asymmetry (Fig S2). As even an apparently logographic system such as Chinese has some degree of phoneticism , it is not surprising that systems having a high degree of logography also show a difference in the sign usage distribution between the beginning and end of words, although this effect is much less marked than in other (more phonetic) scripts. There are exceptions to this general trend - for example, Hebrew, which is an alphabetic script, shows a low degree of asymmetry that is difficult to distinguish from effects due to stochastic fluctuations arising from sampling effects in a finite corpus. Hawaiian, which also appears to have a very low , however, shows a significant asymmetry when information entropy is used to measure sign usage inequality in place of the Gini index (see Supporting Information). Other indications that a simple phonotactic explanation for the observed asymmetry may not be adequate is shown by the fact that the relative position of some languages in terms of (or ) do not necessarily conform to common perceptions about the degree of phonetic representation in the corresponding scripts used for writing them . For example, the Korean han’gŭl script is considered to have a higher proportion of phoneticism than French ; however, the latter exhibits higher asymmetry in terms of both the measures of inequality used here (see Fig 3 and Supporting Information).
The asymmetry reported here may be used to infer the direction of writing, which is one of the basic pre-requisites for interpreting any linguistic sequence. A variety of possible directions have been seen in different writing systems, both historical and present . The most common, left to right in horizontal lines, is the direction in which all scripts descending from the Greek and Brāhmī systems are written, including English, French, German, Hindi and Tamil. Scripts that are written in the other direction, i.e., right to left in horizontal lines, are also common and are used in ancient and modern Semitic scripts including Arabic and Hebrew. Another common orientation is from top to bottom in vertical columns, which is the direction in which Chinese and scripts influenced by it (such as Japanese) were traditionally written. Other, less common, directions of writing are also known, including bottom to top (the Celtic Ogham script) and boustrophedon, where the direction reverses in successive lines (as in archaic Greek and Luwian hieroglyph inscriptions). In cases where the inscriptions are undeciphered, such as those of IVC, the direction usually has to be inferred by indirect means. The asymmetry in sign usage patterns reported here - which shows that the beginning of sequences can be distinguished from the end by the nature of heterogeneity in the distributions of sign occurrence at these positions - can provide a valuable tool for ascertaining the direction of writing in such cases. Availability of a sufficiently large corpus would, however, be necessary for a reliable determination of the direction of writing in these inscriptions.
Data description. We have analyzed data from written corpora of twenty four languages (twenty two belonging to nine linguistic families, as well as, two language isolates), along with a corpus of undeciphered inscriptions from the Indus Valley Civilization (ca. 2600-1900 BCE). The writing systems considered range from alphabetic (that use only a few dozens of distinct letters) and syllabic to logo-syllabic and logographic (involving thousands of signs). The average corpus size is about ten thousand unique words collected from a variety of sources. Each word considered for our analysis consisted of multiple graphemes, corresponding to letters, logograms, hieroglyph signs or syllables depending on the writing system used. Detailed description of each corpus is provided in the Supporting Information.
Estimation of occurrence probability distribution. Probability distribution of sign occurrences in a corpus of inscriptions are estimated from frequency counts of the distinct signs appearing in the sequences belonging to the database, i.e.,
For establishing the directional asymmetry of sign usage, we focus specifically on the sign occurrence distributions at the left and right terminal positions of a sequence. The inequality of sign usage at these positions, which is reflected in the non-uniform nature of the corresponding distributions, is quantified by measuring the Gini coefficient or the information entropy.
Measuring Gini coefficient. The Gini coefficient or index is a measure of how unequal are the probabilities of all the different events that are possible. A value of zero for the coefficient corresponds to situations where all events are equally probable. Conversely, when only one event out of all possible ones is observed in every instance, the Gini coefficient attains its maximum value of 1. For a discrete probability distribution , where the possible values of the discrete variable are indexed according to their probability of occurrence in non-decreasing order (, ), the Gini coefficient  is measured as
where is the cumulative probability of with and . For a given set of inscriptions, we estimate the cumulative probability distributions and for the signs occurring in the left and right terminal positions, respectively, and use Eq. 1 to compute the corresponding estimated Gini indices and . The normalized difference between these two values,
provides a measure for the asymmetry in the distribution of sign frequencies at the left and right terminal positions of the sequences.
Estimating information entropy. Apart from Gini index, we have used a measure based on information or Shannon entropy for quantifying the nature of the inequality of sign usage distributions at left and right terminal positions in a sequence. As entropy measures the unpredictability of information generated from a source, it can be used to characterize the underlying distribution of any process that produces a discrete sequence of symbols (chosen from a set of possible ones) and is defined as
where is the probability of occurrence of the -th symbol and the use of base 2 logarithm implies that the entropy can be expressed in units of bits . In particular, given any database of inscriptions, we obtain estimates for the probabilities and for a particular sign from the corresponding signary to occur in the left terminal and right terminal positions. After estimating these probabilities for all signs that occur in the corpus of inscriptions, the left and right terminal entropies ( and , respectively) are calculated by using Eq. 2. The normalized difference between the two entropy values, , provides a measure of the degree of asymmetry in sign frequency at the two ends of a sequence.
Bootstrap confidence intervals. To quantify the degree of robustness in the estimates obtained using the empirical databases, we have used a bootstrap method to obtain confidence intervals for the measured values. For each corpus we have created resampled datasets (i.e., bootstrap samples) by random sampling with replacement, containing the same number of sequences as the original dataset. The probability distributions for sign usage at the terminal positions are then calculated for every bootstrap sample. Finally, a confidence interval for the normalized difference between the left and right terminal Gini indices, (or, of information entropy, ) for the corpus is computed using the Bias-Corrected and accelerated (BCa) method which adjusts for bias in the bootstrap sample distributions relative to the actual sampling distribution . The BCa confidence interval adjusts the percentiles of the bootstrap distribution of the parameter according to the calculation of a bias correction coefficient and an acceleration coefficient. The former adjusts for any skewness present in the bootstrap sampling distribution (it is zero if the distribution is symmetric). The latter coefficient adjusts for nonconstant variances (if any) in the resampled data.
Sequence randomization. The statistical significance of the measured asymmetry in sign usage is measured by comparing the results obtained from the empirical database with an ensemble of randomized surrogate sequence corpus. Each ensemble is generated by taking each sequence in turn that belongs to a database and doing a random permutation of the signs. This reordering ensures that the frequency distribution of the signs in each sequence (and thus, also the corpus) is unchanged in the randomized set, although all correlations (that contribute to 2- and higher order -gram distributions) are disrupted. We then perform the same calculations as for the original empirical data for measuring the degree of asymmetric sign usage in left and right terminal positions of these randomized sequences - which, by design, are expected not to have any asymmetry. Significant difference in the results of the two datasets ensures that the measured asymmetry is not arising from stochastic fluctuations.
We thank P. P. Divakaran, Deepak Dhar, Iravatham Mahadevan, Shakti N.
Menon, Adwait Mevada, Ganesh Ramachandran and Chandrasekhar Subramanian for helpful discussions and suggestions. We also thank Bryan K. Wells for allowing the use of the IVC sequence database compiled by him. This work was partially supported by the IMSc PRISM project funded by the Department of Atomic Energy, Government of India.
- 1. Deacon, T. W. The symbolic species: The co-evolution of language and the brain (W. W. Norton, New York, N.Y., 1997)
- 2. Dunbar, R. Grooming, gossip, and the evolution of language (Faber, London, 1996).
- 3. Michel, J.-B. et al. Quantitative analysis of culture using millions of digitized books. Science 331, 176-182 (2011).
- 4. Petersen, A. M., Tenenbaum, J. N., Havlin, S., Stanley, H. E. & Perc, M. Languages cool as they expand: Allometric scaling and the decreasing need for new words. Scientific Reports 2, 943 (2012).
- 5. Dodds, P. S. et al. Human language reveals a universal positivity bias. Proc. Natl. Acad. Sci. USA 112, 2389-2394 (2015).
- 6. Zipf, G. Selected studies of the principle of relative frequency in language (Harvard University Press, Cambridge, Mass., 1932).
- 7. Mitzenmacher, M. A brief history of generative models for power law and lognormal distributions. Internet Mathematics 1, 226-251 (2003).
- 8. Ferrer i Cancho, R. & Solé, R. V. Least effort and the origins of scaling in human language. Proc. Natl. Acad. Sci. USA 100, 788-791 (2003).
- 9. Mumford, D. & Desolneux, A. Pattern Theory: The stochastic analysis of real-world signals (A. K. Peters, Natick, Mass., 2010).
- 10. Shannon, C. E. A mathematical theory of communication. Bell System Tech. J. 27, 379-423 (1948).
- 11. Shannon, C. E. Prediction and entropy of printed English. Bell System Technical Journal 30, 50-64 (1951).
- 12. Schürmann, T. & Grassberger, P. The predictability of letters in written English. Fractals 4, 1-5 (1996).
- 13. Schürmann, T. & Grassberger, P. Entropy estimation of symbol sequences. Chaos 6, 414-427 (1996).
- 14. Gini C. Variabilità e Mutuabilità: Contributo allo studio delle distribuzioni e delle relazioni statistiche (C. Cuppini, Bologna, 1912) [Eng. trans. of extracts in Ceriani, L. & Verme P. The origins of the Gini index. J. Econ. Inequal. 10, 421-433 (2012)].
- 15. Sinha, S., Chatterjee, A., Chakraborti, A. & Chakrabarti, B. K. Econophysics: An Introduction (Wiley-VCH, Weinheim, 2011).
- 16. Breiman, L., Friedman, J. H., Olshen, R. A., & Stone, C. J. Classification and Regression Trees (Wadsworth, Belmont, CA, 1984).
- 17. Plag, I. Word-Formation in English (Cambridge University Press, Cambridge, 2003).
- 18. Possehl, G. The Indus Civilization: A Contemporary Perspective (AltaMira Press, Lanham, MD, 2002).
- 19. Sinha, S., Ashraf, M. I., Pan, R. K. & Wells, B. K. Network analysis of a corpus of undeciphered Indus civilization inscriptions indicates syntactic organization. Computer Speech and Language 25, 639-654 (2011).
- 20. Wells, B. K. The Archaeology and Epigraphy of Indus Writing (Archaeopress, Oxford, 2015).
- 21. Lawler, A. The Indus Script - Write or Wrong ? Science 306, 2026-2029 (2004).
- 22. Rao, R., Yadav, N., Vahia, M., Joglekar, H., Adhikari, R. & Mahadevan, I. Entropic evidence for linguistic structure in the Indus script. Science 324, 1165 (2009).
- 23. Hunter, G. The Script of Harappa and Mohenjodaro and its connection with other scripts (Kegan Paul, London, 1934).
- 24. Mahadevan, I. The Indus Script: Texts, Concordance and Tables (Archaeological Survey of India, Calcutta, 1977).
- 25. Parpola, A. Deciphering the Indus Script (Cambridge University Press, New York, 1994).
- 26. Daniels, P. T. & Bright W., Eds., The World’s Writing Systems (Oxford University Press, New York, 1996).
- 27. Heaps, H. S. Information Retrieval: Computational and Theoretical Aspects (Academic Press, New York, 1978).
- 28. McMahon, A. An Introduction to English Phonology (Edinburgh University Press, Edinburgh, 2002).
- 29. DeFrancis, J. & Marshall Unger, J. Rejoinder to Geoffrey Sampson. “Chinese script and the diversity of writing systems”. Linguistics 32, 549-554 (1994).
- 30. Robinson, A. Writing and Script (Oxford University Press, Oxford, 2009).
- 31. Marshall Unger, J. Ideogram: Chinese Characters and the Myth of Disembodied Meaning (University of Hawai‘i Press, Honolulu, 2004).
- 32. Coulmas, F. The Blackwell encyclopedia of writing systems (Blackwell, Malden, MA, 1996).
- 33. Brown, M. C. Using Gini-style indices to evaluate the spatial patterns of health practitioners: Theoretical considerations and an application based on Alberta data. Social Science & Medicine 38, 1243-1256 (1994).
- 34. Efron, B. (1987) Better bootstrap confidence intervals. J. Amer. Stat. Asso. 82, 171-185 (1987).
Description of the corpora
Arabic: We have used a database of 14867 unique words (that are represented using two or more characters) of Classical (or Quranic) Arabic, a Semitic language written using a consonantal alphabet or ‘abjad’ . The words are obtained from Tanzil, an international project started in 2007 to produce a standard Unicode text for the Qur’an (http://tanzil.net/download/, accessed: 25th March 2015). The signary comprises 36 signs, viz., 28 consonantal signs and 8 consonants with diacritical marks indicating vowels. The words range in length from 2 to 11 characters, the average length being 5.39.
Chinese: We have used a database of 13104 unique words (cì) of the Chinese language (belonging to the Sino-Tibetan language family) that are written using two or more signs (zì)  which have been obtained from a public-domain online Chinese-English dictionary CC-CEDICT (http://www.mdbg .net/chindict/chindict.php?pahe=cc-cedict, accessed: 8th January 2011). The database contains annotations identifying idiomatic expressions and loan words, and indicates proper nouns by capitalization of the corresponding English translation. From the total set of 115430 words and phrases available in the dictionary we have removed all single character words, idiomatic phrases, variants of the same word, hyphenated compound words, proper nouns and loan words. The Chinese writing system has been variously described as either logographic , or, an imperfect phonographic system with additional logographic attributes [37, 38, 39]. The signary for our data comprises 3691 distinct graphemes corresponding to logograms (hanzi). The sign sequences range in length from 2 to 9 signs, the average length being 2.46. Traditionally, Chinese is written top to bottom in vertical columns shifting from right to left; however, in modern times it is more frequently being written left to right in horizontal lines, and this is the convention used in our database. The words in our database are written using traditional Chinese characters.
Dutch: We have used a list of the 10000 most commonly used words in Dutch, a member of the Germanic branch of the Indo-European language family, from which we have chosen the 9146 unique non-hyphenated words comprising two or more characters. The data has been collected from the Wortschatz website maintained by the University of Leipzig (http://wortschatz.uni-leipzig.de/Papers/top 10000nl.txt, accessed: 22nd May 2015). The signary used has 31 distinct alphabetic characters comprising 21 consonants, 5 vowels, 3 vowels with diacritical marks (acute accents or diaeresis), the digraph ‘ij’ that is considered as a letter in the Dutch language and an extra letter from the German alphabet (the Eszett). The words range in length from 2 to 25 letters, the average length being 7.6.
Egyptian (Hieroglyphs): Ancient Egyptian, a member of the Afro-Asiatic (Hamito-Semitic) language family, is written using a mixed system (also referred to as a logoconsonantal system ) with several hundreds of hieroglyph signs that can represent logograms, phonograms and/or determinatives . We have used as data 39933 unique sequences comprising two or more hieroglyph signs of the Middle Egyptian Dictionary compiled by Mark Vygus (updated April 2015, http://www.pyramidtext sonline.com/ MarkVygusDictionary.pdf, accessed: 22nd May, 2015). The hieroglyphic signs are represented using the Gardiner sign list numbering system , the signary for the database used by us comprising 1859 distinct signs. The sign sequences range in length from 2 to 17 hieroglyphs, the average length being 5.12. The conventional direction of reading hieroglyphic sequences is “toward the face of human or animal pictograms, i.e., the signs are turned towards the beginning of the inscription” . In the database used by us all sequences have been oriented so as to read from left to right.
English: We have used the Mieliestronk list of 58109 distinct words (comprising two or more letters) of the English language - belonging to the Germanic branch of the Indo-European language family - that has been compiled by merging several different word-lists ( http://www.mieliestronk.com/wordlist.html, accessed: 4th December 2011). The words vary in length from 2 to 22 letters, the average being 8.34. The signary consists of the 26 lower case letters of the English alphabet. The list excludes spellings that are considered to be non-British. If a word is hyphenated, it is listed in unhyphenated form by removing the punctuation mark. The list contains some multiword phrases that are in common usage, rendered as a single word. Several words are included in both their singular and plural forms. Note that this is the word list for English that is included in the Supplementary Data Sets.
For corroboration of our results obtained by analyzing the above dataset, we have also used statistics of sequence position-specific distributions of letter usage frequencies in a list of 97565 distinct words of the English language compiled from Google books Ngram data (English Version 20120701, in http://storage.googleapis.com/ books/ngrams/books/datasetsv2.html) by Peter Norvig and made freely available for public access (http://norvig .com/tsv/ngrams-all.tsv.zip, accessed: 22nd June 2015) . Only those words are used which occur with a frequency of more than in the corpus of books scanned by Google. Details of the procedure used for compiling the N-gram frequency statistics are given in http://norvig.com /mayzner.html (accessed: 22nd June 2015). Note that, there are some differences in the relative frequency of occurrence of the different letters in the two databases. This is because, unlike in the preceding case where the occurrence probabilities are computed from a database comprising unique words, the Google 1-gram distributions are computed from corpora of books containing multiple occurrences of the same word. The frequencies of each letter is thus weighted by the occurrence frequencies in the corpus of different words (whose distribution follow Zipf’s law) containing that letter. However, despite these differences in details, the inequality of sign usage at the right terminal positions is visibly higher than that in the left terminal positions (Fig S1).
To ensure that our results are not an artifact of commonly used affixes (e.g., prefixes like de- or un- and suffixes like -ed or -ly) in English, we have also used the list of 850 words proposed by C. K. Ogden  as the core words of Basic English, a simplified subset of regular English. This word list is widely used as the beginner’s vocabulary for the teaching of English as a second language. It comprises only word roots, which for regular use are extended with different affixes. We have used the 848 words comprising two or more letters (i.e., omitting the single character words a and i from the Ogden list), which range in length between 2 and 14, the average length being 5.2.
In addition to these three data sets, we have tested several others of varying sizes including a lemmatised list of frequently used words from the British National Corpus , lemma being the canonical form of a word chosen to represent all the forms having the same meaning. All of these data sets exhibit similar left-right asymmetry as the above.
Finnish: We have used a list of the 10000 most commonly used words (all of which use two or more letters) in the Finnish language, belonging to the Finnic branch of the Uralic language family. The data, obtained from the Wikiverb website, has been collected from newsgroup discussions, press and modern literature (http://wiki.verbix.com/Documents/WordfrequencyFi, accessed: 24th June 2015). The signary used has 25 distinct signs - i.e., all vowels and consonants of the modern Latin alphabet excepting “q”,“x” and “w”, along with two additional vowels “ä” and “ö”. The words vary in length from 2 to 25 letters, the average being 7.81.
French: We have used a list of the 10000 most commonly used words in French, a Romance language belonging to the Indo-European family, from which we have chosen the 9189 unique words that have two or more characters. The data has been collected from the Wortschatz website maintained by the University of Leipzig (http://wortschatz.uni-leipzig.de/Papers/top10000fr.txt, accessed: May 22nd 2015). The signary used has 30 distinct alphabetic characters comprising 26 letters of the Latin alphabet along with 3 vowels with diacritical marks (acute accents or diaeresis) and an apostrophe sign. The words range in length from 2 to 19 letters, the average length being 7.7.
German: We have used a list of the 9172 most commonly used words in German, a member of the Germanic branch of the Indo-European language family, from which we have chosen the 9152 distinct words that have two or more characters. The data has been collected from the Wortschatz website maintained by the University of Leipzig (http://wortschatz.uni-leipzig.de/ Papers/top10000de.txt, accessed: May 22nd 2015). The signary used has 32 distinct alphabetic characters comprising the 26 letters of the Latin alphabet along with 4 vowels having diacritical marks (umlauts or acute accents), a ligature (the Eszett or scharfes S) and an apostrophe sign. The words vary in length between 2 to 27 letters, the average being 8.1.
Greek: We have used a list of the 10000 most frequently occurring words - grouped by lemma - in classical Greek literature written in ancient Greek which belongs to the Indo-European language family, compiled by Kyle Johnson from the Thesaurus Linguae Graecae corpus (maintained by University of California, Irvine) using the Classical Language Toolkit (http://cltk.org) and made freely available to the public (http://kyle-p-johnson.com/assets/most-common-greek-words.txt, accessed: 24th June 2015). From this dataset we have used the 9868 distinct words that have two or more characters. The signary has 124 distinct characters as the words are represented in the traditional polytonic orthography used for ancient Greek, involving 24 basic letters used in conjunction with several varieties of diacritical marks (e.g., accents, breathing marks, iota subscript and diaeresis). The words range in length from 2 to 18 characters, the average being 6.9.
Hausa (Boko): Hausa, a Chadic language belonging to the Afro-Asiatic family, is written using Boko, a Latin-based alphabet, which was devised in the 19th century and became the official system in the early part of the 20th century (in earlier periods, it was written in Ajami, an Arabic alphabet). We have used a list of 7062 unique words that have two or more characters obtained from a Hausa online dictionary maintained by the University of Vienna (http://www.univie.ac.at/ Hausa/KamusTDC/CD-ROMHausa/KamusTDC/ARBEIT2.txt, accessed: 19th May, 2015). The signary used has 30 distinct alphabetic characters comprising 23 letters from the Latin alphabet, four additional signs representing glottalized consonants, two digraphs (‘sh’ and ‘ts’) and an apostrophe sign. The words range in length from 2 to 22 characters, the average being 6.0.
Hawaiian: Hawaiian is Polynesian language belonging to the Austronesian family that had no written form until the 19th century when foreign missionaries devised an alphabetic system for recording it based on the Latin script. The data used for our analysis has been collected from the entries of A dictionary of the Hawaiian language (1922) compiled by Lorrin Andrews and revised by Henry H Parker (Board of Commissioners of Public Archives of the Territory of Hawaii, Honolulu) and freely available online (http://ulukau.org/elib/cgi-bin/library?c=parker&l=en, accessed on 28th May 2015). After removing all non-native words that contain characters that do not belong to the Hawaiian alphabet, we have compiled a data-base of 14009 unique words containing two or more characters. The signary comprises 28 distinct characters, with 12 basic letters - representing 5 vowels and 7 consonants - of the Hawaiian alphabet along with vowels used in conjunction with diacritical marks (breve and macron) indicating short or long pronunciation, and a sign to indicate glottal stop (the ‘okina). The words range in length from 2 to 26 characters, the average being 9.0.
Hebrew: We have used a list of the 10000 most commonly used words (compiled from online written texts) in modern Hebrew, a Semitic language written using a consonantal alphabet or ‘abjad’, from which we have chosen the 9993 distinct words that are represented using two or more characters. The data has been collected from a list maintained by Teach Me Hebrew, an online Hebrew language learning site (http://www.teachmehebrew.com/hebrew-frequency-list.html, accessed: 26th December 2013). The signary comprises 31 signs, viz., 27 consonantal signs (comprising 22 letters of which five use different forms - called sofit - when used at the end of a word) and 4 signs used in conjunction with niqqud diacritical marks. The words range in length from 2 to 13 characters, the average length being 5.1.
Hindi: Hindi is an Indo-Aryan language, a branch of the Indo-European family, which is written in the Devanagari script that is sometimes classified as an alphasyllabary  or ‘abugida’ . Like the other writing systems that are descended from the Brāhmī script of ancient India, Devanagari uses as its main functional unit the aksara, which may consist of only a vowel but more frequently represents a syllable consisting of a consonant and an inherent vowel along with diacritical marks that may indicate use of other vowels . We have used a database of 6441 distinct words written using two or more characters, collected from an online dictionary (Shabdanjali) of Hindi developed by the Language Technology Research Center at Indian Institute of Information Technology, Hyderabad (http://ltrc.iiit.ac.in/showfile.php?filename=downloads/ shabdanjali-stardict/index.html, accessed: 20th May 2015). The signary comprises 571 distinct signs, comprising 11 vowels, 33 consonants, their conjunctions with each other and with consonantal sound modifiers (the anusvara, chandrabindu, visarga and halant). The words range in length from 2 to 10 signs, the average length being 3.5.
Japanese (Kana): Japanese, which belongs to the Japonic language family, is written using a combination of the logographic Kanji system (adopted from Chinese characters) and the syllabic kana system. The latter, in turn, consists of a pair of distinct syllabaries: hiragana, used for writing native Japanese words and katakana, which is used for foreign words. For our study, we have focused only on the syllabic writing system for Japanese. We have used a list of 1162 distinct words written using two or more signs from the kana syllabary, which is obtained from a list of common Japanese words collected from textbooks used by foreign learners of the language and maintained by Japanese Words, an online site for learning the Japanese language (http://www.japanesewords.net/36/over-1000-japanese-words-list/, accessed: 29th May 2015). The signary has 103 distinct characters comprising 46 basic signs of Hiragana and 21 basic signs of Katakana, 22 Hiragana and 9 Katakana signs used in conjunction with diacritical marks (the dakuten and handakuten), smaller forms of 4 hiragana characters (viz., of ya, yu and yu which indicate the yōon feature, and the sokuon used to mark a geminate consonant) and a special symbol (chōonpu, the long vowel mark). The words range in length from 2 to 13 characters, the average being 3.8.
Korean: Korean is a language isolate with no established connection to any of the major language families of the world and is written using Han’gŭl, a purely phonetic script, although in earlier times a system based on Chinese characters (Hanja) was used. Each character corresponds to a syllable, the syllabic block being composed of two to six letters (including at least one consonant and one vowel) from the basic alphabet comprising 10 vowels, 14 consonants and 27 digraphs. The number of possible distinct syllabic blocks or characters exceeds 11,000 although a far smaller number is in actual use . We have used a list of 5888 commonly used words compiled by the National Institute of Korean Language in 2004 (publicly accessible from https://en.wiktionary.org/wiki/Wiktionary:Frequency_lists/ Korean_5800, accessed: 24th June 2015) from which we have chosen 5141 distinct words that are represented using multiple syllabic blocks. The signary comprises 966 distinct characters (each corresponding to a syllabic block). The words range in length from 2 to 6 characters, the average length being 2.7.
Linear B: Linear B is a syllabic script, with most of its signs representing consonant-vowel combinations, that was used for writing archaic Greek between 1500 and 1200 BCE. We have used as data 1924 distinct sequences comprising two or more characters from the Linear B Lexicon compiled by Chris Tselentis (https://www.scribd.com/doc/56265843/ Linear-B-Lexicon, accessed: 15th May 2015). The signary comprises 87 distinct signs representing syllables. The words range in length from 2 to 8 signs, the average being 3.8.
Malay (Rumi): We have used a list of the 10000 most commonly used words in Malay, a member of the Austronesian language family, from which we have chosen the 9970 unique words that have two or more characters. All words are written in Rumi or Latin script, which is the most commonly used form for writing Malay at present, although a modified Arabic script (Jawi) also exists. The data has been collected from the list of high frequency words that are publicly available at Invoke IT Blog (https://invokeit.wordpress.com/frequency-word-lists/, accessed: 4th January, 2014). The signary comprises the 26 letters of the Latin alphabet. The words range in length from 2 to 17 letters, the average being 6.8.
Persian: We have used a list of 10000 most commonly used words (each represented using two or more characters) in Persian, a member of the Indo-Iranian branch of the Indo-European language family, which is written using a modified form of the consonantal Arabic alphabet or ‘abjad’. The words are obtained from a list of high-frequency words compiled using the Tehran University for Persian Language corpus and available at Invoke IT Blog (https://invokeit .wordpress.com/frequency-word-lists/, accessed: 4th January 2014). The signary comprises 40 signs, viz., 32 consonantal signs, a long vowel indicator (‘alef madde’), a ligature (‘lām alef’), a diacritic (‘hamze’), 3 consonants with the ‘hamze’ diacritical mark and different forms for the consonants ‘kâf’ and ‘ye’ when they occur in final position. The words range in length from 2 to 13 letters, the average being 5.2.
Russian: We have used a list of 9011 distinct words that use two or more characters in Russian, a member of the Slavic branch of the Indo-European language family and which is written using a Cyrillic alphabet. The data has been collected from Russian Learners’ Dictionary: 10,000 words in frequency order compiled by Nicholas J Brown (Routledge, London, 1996), after removing all words that use characters not in the standard Russian alphabet. The signary comprises the 33 letters of the modern Russian alphabet. The words range in length from 2 to 21 letters, the average being 8.0.
Spanish: We have used a list of 4902 distinct high-frequency words (that use two or more characters) in Spanish, a Romance language belonging to the Indo-European family. The data has been collected from A Frequency Dictionary of Spanish compiled by Mark Davies (Routledge, London, 2006). The signary used has 35 distinct alphabetic characters comprising 26 letters of the basic Latin alphabet along with an additional character ñ and two digraphs (‘ch’ and ‘ll’), as well as, vowels with diacritical marks (acute accents or diaeresis). The words range in length from 2 to 19 letters, the average being 7.4.
Tamil: Tamil is a Dravidian language is written in a script (sometimes classified as an ‘abugida’) derived from Brāhmī script and thus shares a common origin with the Devanagari script used for writing Hindi (see above) although it differs significantly both in appearance and structure . It has 31 basic signs consisting of 12 vowels, 18 consonants and a special character, with combinations of the different vowels and consonants yielding a possible 216 compound letters. Additional characters from the Grantha script and diacritical marks are sometimes used to represent sounds not native to Tamil, e.g., in words borrowed from other languages. As with other Indian scripts, Tamil uses the aksara as its basic unit - however, unlike then it has eliminated most conjuncts, consonant clusters being placed in a linear string. The data used for our analysis has been collected from texts (e.g., Paripaadal, Thiruppavai, Kamba Ramayanam, Sundara Kandam, Akananooru songs, etc.) available in Chennai Library, an online repository of Tamil literature (http://www.chennailibrary.com, accessed: 27th December 2012). From this a data-base of 1991 unique words containing two or more characters was compiled. The signary comprises 187 distinct signs corresponding to different basic and compound letters and the special character. The words range in length from 2 to 9 letters, the average being 3.8.
Turkish: We have used a list of 9909 distinct high-frequency words (that use two or more characters) in Turkish, a member of the Turkic language family. The data has been collected from a Wiktionary word frequency list (https://en.wiktionary.org/wiki/ Wiktionary:Frequency_lists/ Turkish_WordList_10K, accessed: 14th July 2015). The signary used has 32 letters, comprising 29 letters of the Turkish alphabet and 3 vowels used in conjunction with circumflex accents. The words range in length from 2 to 17 letters, the average being 6.9.
Sumerian (Cuneiform): Sumerian, a language isolate spoken in ancient Mesopotamia during the 3rd millennium BCE, was written using a logosyllabic system with several hundred signs of a cuneiform script representing logograms, phonograms and/or determinatives. We have used as data 19221 unique words (comprising two or more cuneiform signs) collected from texts available in the Electronic Text Corpus of Sumerian Literature (ETCSL, http://etcsl.orinst.ox.ac.uk, accessed: 20th October 2010). The signary comprises 1364 distinct transliteration values of Cuneiform signs that represent syllables. The sequences range in length from 2 to 11 signs, with the average being 3.6.
Urdu: We have used a database of 4998 unique words that are represented using two or more characters in Urdu, an Indo-Aryan language belonging to the Indo-European family, that is written using an extended Persian alphabet. The words are obtained from a list of frequently used words maintained by the Center for Language Engineering at Lahore (http://www.cle.org.pk /software/ling_resources/UrduHighFreqWords.htm, accessed: 1st January 2014). The signary comprises 46 signs, viz., 38 consonantal signs, 3 long vowels (‘alef madde’, ‘lām alef madde’ and ‘ya’), 2 semi-consonants (‘hamzah’ used in conjunction with ‘wao’ or ‘ya’), a nasalized consonant (‘noon ghunna’), a ligature (‘lām alef’) and an additional sign (‘ta’ marbuta’) used for writing certain loan-words. The words range in length from 2 to 11 letters, the average being 4.6.
Undeciphered (Indus): As an example of an undeciphered corpus on which to apply our analysis, we have used the set of inscriptions obtained from archaeological excavations at various sites of the Indus Valley civilization (ca. 2600-1900 BCE). The data used for our analysis is collected from the ICIT Database of Indus Writing compiled by Bryan K. Wells [46, 47] and maintained by Andreas Fuls (http://caddy.igg.tu-berlin.de/indus/welcome.htm, accessed: 20th October 2017) from which we have removed all incomplete and multiple-line inscriptions thereby obtaining 1837 unique sequences that contain two or more signs. The Indus signs are represented using the Wells sign list numbering system [46, 47], the signary for the database used by us comprising 568 distinct signs. The sequences range in length from 2 to 13 signs, the average being 4.6. The direction of the sign sequences vary, the majority being written right to left, although examples of left to right also exist, as inferred from external evidence such as signs becoming relatively cramped towards the end of a sequence inscribed on an archaeological artifact (e.g., seals, tablets or potsherds) [48, 51]. In the database used by us , the relatively few sequences which are believed to have been written from left to right have been reversed so as to be oriented in the same direction as the majority (i.e., which are believed to be written from right to left). This follows the standard procedure used also for constructing earlier concordances for Indus Valley Civilization inscriptions [48, 51]. Note that if only the 1779 sequences which are believed to be written from right to left are considered for our analysis, we obtain and which are identical to the values reported in the main text for the database in which all 1837 sequences are oriented right-to-left.
Robustness of results
An important consideration when quantifying the inequality of sign
usage is the size of the signary, i.e., the number
of visually distinct signs (sometimes referred to as
‘graphs’ ) that can be identified in each corpus.
In several scripts, complex characters that are recognizably
the compound of two or more basic characters are quite commonly used -
as in the system of conjunct consonant signs or ligatures in the
Devanagari script used for writing Hindi and other South Asian
languages - and in principle, one could either consider these as
separate signs or decompose them into the constituent signs, which
result in very different signary sizes. Also, Semitic scripts such as
Arabic and Hebrew that are essentially consonantal alphabets often use
diacritical marks for indicating vowel usage. In such cases, the
same consonant is used in conjunction with different diacritics when
the vowel following it is different. The signary size would depend
whether these are considered to be distinct signs or not. In several
other scripts, special marks can be used together with the vowels
and consonants, e.g., the use of apostrophe to indicate the
omission of one or more sounds in European languages such as French or
German, and the use of a glottal stop marker (’okina) in
Hawaiian. Once again, whether these signs are treated as distinct
elements or part of the associated letter will
affect the signary size. We observe that although the numerical values
of the Gini indices (and information entropy) can be affected by
changes in the signary
size, the asymmetry in terminal
sign usage reported here is robust with respect to these choices about
conventions for identifying distinct signs constituting the signary
for a given corpus.
We have also examined our results for their dependence on corpus size, i.e., the number of words included in each database. As shown in Fig 4 in the main text (for English) and Fig S3 in Supplementary Information (for Persian), the scores for each language indicative of their asymmetry converge to a value which is relatively independent of the corpus size as the number of words included in the corresponding database is increased. We have explicitly checked that the scores for the different languages attain asymptotic values (for large enough database sizes) which are distinct from each other. Thus the variation of the scores for the different languages cannot be attributed to corpus size alone.
Possible non-phonotactic mechanisms for emergence of
It is possible that the asymmetry we have reported here could arise in
situations for reasons other than phonotactic constraints. For
instance, in the undeciphered IVC inscriptions, the most frequently
occurring sign - viz., the U-shaped “jar”
symbol  - that appears very frequently at the end of
dominates the probability distribution of signs that
can occur at the left terminal position (accounting for about a third
of all the distinct sequences comprising the corpus).
By contrast, the most frequently occurring sign at the right terminal
position is seen to begin only about of the sequences.
These distinctive sign usage patterns at the two terminal
positions of IVC sequences gives rise to the heterogeneity in the
corresponding occurrence probability distributions. In the absence of
a decipherment, it is purely speculative whether the inequality arises
for phonotactic reasons (as in the linguistic sequences considered
here) or a fortuitous stylistic convention.
Appearance of sign usage asymmetry in writing systems that
are not alphabetic or syllabic
One of the writing systems included in our study, that of Chinese language, is traditionally
considered to be logographic . It is thus worth considering
whether a non-phonotactic mechanism could be responsible for the appearance
of sign usage asymmetry in such a system. It has been reported that, while
single signs or characters (zì) can indeed represent words, the
majority of Chinese words (cì)
are written using multiple signs, some studies suggesting as much as 70%
of commonly used words involve two or more signs .
While some of these words are polysyllabic, several others are compound words
created by affixation, i.e., combining a word with a prefix, infix or suffix.
As usage of suffix in Chinese is much more common than other types of affix
this may possibly be a mechanism by which sign usage asymmetry can arise
(e.g., the use of a limited set of suffixes at the end of a multi-character word
can reduce the diversity of sign usage at that position). However, we have
carried out analysis of the database of Chinese words after eliminating
words with affixes and observed that the sign usage asymmetry still persists.
A plausible hypothesis one may consider is that the remaining words in the database
possess some degree of phonetic character. Thus, some variant of the phonotactic
argument for the observed asymmetry suggested in the main text may arise even
for this system. We note that several authors have indeed argued that, instead of
treating it as a purely logographic system , Chinese writing is better thought
of as an “imperfect phonographic system with additional logographic attributes” . Indeed the importance of phonetic information in the development of the Chinese writing system is considered to be unequivocal .
We also note that other writing systems that use logograms, such as the Egyptian hieroglyphic system, are nevertheless phonological in character to a considerable degree. Indeed, the recognition that Egyptian hieroglyphic writing system is significantly phonetic was crucial to its decipherment by Jean-Francois Champollion . Egyptian represents an apparently unique polyconsonantal system , with extensive use of determinatives that appear at the end of words (following the phonetic signs) supplying specified meanings. It is intriguing to consider whether a variant of the phonological argument that has been suggested to underlie the sign usage asymmetry in alphabetic and syllabic systems could also apply to these apparently logographic and logoconsonantal systems.
Entropy difference in Egyptian hieroglyphs
As can be seen from Fig S2, for almost all sequence databases that are read from left to right (or rendered in that format) we see a positive value of except for Egyptian hieroglyphs. Instead of having a higher entropy for the left terminal sign compared to right terminal sign (as should have been the case given that all sequences have been oriented to be read from left to right in the database), it exhibits a marginally higher entropy for right terminal sign in comparison to left terminal sign. Note that the numerical value of for this data-set lies almost within the error bars of the corresponding randomized data-set, so that the direction cannot be conclusively determined given the data. The discrepancy may also be connected to a distinctive feature of Egyptian hieroglyphic inscriptions, viz., there are many more rare or low-frequency signs that can appear in the right terminal position but never in the left terminal position. Thus, there are 1502 distinct signs that appear in the right terminal position of a sequence while only 798 distinct signs appear in the left terminal position. If we compute the entropy of the characters occurring in the left and right terminal positions by confining our attention to the most frequently occurring signs for either position (so that the number of terms used for computing the two entropy values are comparable) we note that the left terminal sign entropy is consistently higher than the right terminal sign entropy for all (see Fig S4). This is consistent with our hypothesis according to which the left terminal position should have exhibited higher entropy.
Inequality of sign usage distributions at different
positions in a sequence
In the main text we have only considered the difference in the occurrence probability of signs at the two extreme positions of a sequence, viz., at its beginning and at its end. However, one can also consider the sign usage distribution at other positions in a sequence, and in particular, ask whether there is a systematic variation in the inequality measures of sign usage frequencies with the relative position of their occurrence in a sequence. To study this, instead of considering the entire data-set of sequences in a given language (e.g., English words) together, we split it into sub-sets, each having sequences of the same length. This is necessary as otherwise, it is difficult to compare the properties of (say) the 4th character in a sequence when the sequence itself is a 4-letter word (in which case, the character is also the right terminal symbol) and the 4th character in a 7-letter world (in which case it is in the middle of the sequence).
We show in Figs S5 and S6 the variation of the Gini index and the entropy (respectively) as a function of sign position in words of a given length , where varies from 2 to 20. Note that the number of words of different lengths vary considerably in the data-set as shown in the side-panel in the left (showing the frequency of occurrence of words of a given length in the Mieliestronk English corpus data-base described above). Thus the variation in (or ) with position in the much longer words are based on statistics computed over extremely few exemplars. However, broadly we can see that the Gini index for the left-most sign is lower than that for the right-most sign in English words supporting our hypothesis. Note that, it appears that the second position from left appears to have a much higher value than either its left or right neighbour. However, this is unique to the English database and such a feature cannot be seen in other language data-bases that we have analyzed.
When we look at the variation of with sign position for words of a given length, we find that the entropy of the left-most sign is higher than the right-most sign in English words, which is again consistent with our hypothesis. The other positions do not exhibit any consistent variation except for the second sign starting from the left which (consistent with what was seen for ) has an entropy value that is lower than its left or right neighbour. Again, this seems to be an unique feature of the English database as we do not see it occur in the data-bases for other languages.
For example, in Figs S7 and S8 we show and , respectively, as a function of sign position in words of a given length for the Persian language database. As Persian is read from right to left we have numbered the rightmost sign as 1, the one left to it 2 and so on. Consistent with our hypothesis we find that for the left-most sign is higher than that for the right-most sign and the for the left-most sign is lower than that for the right-most sign. However, no other sign appears to exhibit any characteristic trend in the variation of either or . One may therefore conclude that a predictable difference in (and ) for the signs occurring in the left-most and right-most positions of sequences in a given language is possibly the only consistent feature seen across the languages and writing systems considered by us.
- 35. Coulmas, F. Writing Systems: An Introduction to their Linguistic Analysis (Cambridge University Press, Cambridge, 2003).
- 36. Chao, Y. R. Language and Symbolic Systems (Cambridge University Press, Cambridge, 1968).
- 37. DeFrancis, J. The Chinese Language (University of Hawai‘i Press, Honolulu, 1984).
- 38. DeFrancis, J. Visible Speech: The Diverse Oneness of Writing Systems (University of Hawai‘i Press, Honolulu, Honolulu, 1989).
- 39. Sproat, R. A Computational Theory of Writing Systems (Cambridge University Press, Cambridge, 2000).
- 40. Daniels, P. T. and Bright, W., Eds. The World’s Writing Systems (Oxford University Press, New York, 1996).
- 41. Gardiner, A. Egyptian Grammar: Being an Introduction to the Study of Hieroglyphs (3rd edn., Griffith Institute, Oxford, 1957).
- 42. Norvig, P., English letter frequency counts: Mayzner Revisited or ETAOIN SRHLDCU. (2013) Available at: http://norvig.com/mayzner.html. (Accessed: 22nd June 2015)
- 43. Ogden, C. K. Basic English: A General Introduction with Rules and Grammar (Paul Treber, London, 1930). Word list available at: http://ogden.basic-english.org/ (Accessed: 31st May 2016).
- 44. Kilgarriff, A. Putting frequencies in the dictionary. International Journal of Lexicography 10, 135-155 (1997).
- 45. Daniels, P. T. Fundamentals of grammatology. J. Amer. Oriental Soc. 100, 727-731 (1990).
- 46. Sinha, S., Ashraf, M. I., Pan, R. K. & Wells, B. K. Network analysis of a corpus of undeciphered Indus civilization inscriptions indicates syntactic organization. Computer Speech and Language 25, 639-654 (2011).
- 47. Wells, B. K. The Archaeology and Epigraphy of Indus Writing (Archaeopress, Oxford, 2015).
- 48. Mahadevan, I. The Indus Script: Texts, Concordance and Tables (Archaeological Survey of India, Calcutta, 1977).
- 49. Coulmas, F. The Blackwell Encyclopedia of Writing Systems (Blackwell, Malden, MA, 1996).
- 50. Possehl, G. The Indus Civilization: A Contemporary Perspective (AltaMira Press, Lanham, MD, 2002).
- 51. Parpola, A. Deciphering the Indus Script (Cambridge University Press, New York, 1994)
- 52. Yip, P. C. The Chinese Lexicon: A Comprehensive Survey (Routledge, New York, 2000).
- 53. Li, C. N. & Thompson, S. A. Mandarin Chinese: A Functional Reference Grammar (University of California Press, Berkeley, 1981).
- 54. Bard, K. A. An Introduction to the Archaeology of Ancient Egypt (John Wiley, Chichester, 2015).