Language Universals: Their Discovery and Verbalization by Machine

UNIVAUTO was applied to the 30-language sample from the classical paper on the order of meaningful elements by Greenberg (Greenberg 1966). The system generated an article, which was published in a peer-reviewed journal with no post-editing, and without disclosing the “machine” origin of the text (Pericliev 1999a).

By way of illustration, below I give some excerpts from Pericliev (1999a) in italics. Comments may follow in angular brackets. The query to the system was to find non-statistical (or exceptionless) universals which are supported by at least four positive examples. The generated text criticizes some of the findings of Greenberg and lists 59 new universals, which he failed to discover in the database.

Further implicational universals in Greenberg’s data

The goal of this article is to study the implicational universals in the 30 languages sample of Greenberg 1966 and compare the results of the two studies.

Further implicational universals in Greenberg’s data

The goal of this article is to study the implicational universals in the 30 languages sample of Greenberg 1966 and compare the results of the two studies.

<The generated text continues with description of what an implicational universal is, a table of Greenberg’s 30 languages sample, accompanied by the abbreviations used, and a listing of the universals he found. His universals, verbalized by our program, are listed with their numeration in the original publication. An alpha-numeric numeration means that an originally complex universal has been split into elementary ones of the form “If A then B”. Cf. text below>

Below we restate, in our wording, the implicational universals of the type studied here which Greenberg 1966 found in these data, preserving for ease of reference their original numeration. If the original formulation expresses a complex implicational statement, it is split into elementary propositions of the type “If P then Q”:

We confirmed the validity of universals [12, 13, 15-a, 15-b, 21-a, 22-a, 27-a].

Universal [27-b] is also true, however it violates our restriction pertaining to the occurrence of at least 4 positive examples in the dataset. [27-b] is supported in 1 language (Thai).

Universals [16-a, 16-b, 16-c] are uncertain, rather than indisputably valid in the database investigated, since they assume properties in languages, which are actually marked in the database as “unknown or inapplicable” (notated with “*” in Table 4.1). Universal [16-a] would hold only if the feature AuxV/VAux is applicable for Berber, Hebrew, and Maori and in these languages the inflected auxiliary precedes the verb. Universal [16-b] would hold only if the feature AuxV/VAux is applicable for Burmese and Japanese and in these languages the verb precedes the inflected auxiliary. Universal [16-c] would hold only if the feature AuxV/VAux is applicable for Loritja and in this language the verb precedes the inflected auxiliary.

Universal [23-a] is false. It is falsified in Basque, Burmese, Burushaski, Finnish, Japanese, Norwegian, Nubian, and Turkish, in which the proper noun precedes the common noun but in which the noun does not precede the genitive.

We found the following previously undiscovered universals in the data.

Universal 1. If in a language the adjective precedes the adverb then the main verb precedes the subordinate verb.

Examples of this universal are 8 languages: Fulani, Guarani, Hebrew, Malay, Swahili, Thai, Yoruba, and Zapotec.

<There follows a listing of new universals found by the system.>

Universal 58. If a language has an initial yes-no question particle then this language has the question word or phrase placed first in an interrogative word question.

<There follows a conclusion which is a summary of the results.>

The universals presented in the generated text are of historical interest only. To-date there are much larger word order databases (e.g., Hawkins 1983, Dryer 1992) that would provide the empirical material for drawing more adequate conclusions about ordering in the world languages. The system’s discoveries nevertheless are instructive in showing the deficiency of human, manual search for universals even in a rather small sample, comprising only 30 languages, described in terms of 15 features. Thus, we see that the solution reached by Greenberg is neither complete (i.e., giving all universals) nor sound (i.e., giving only the correct universals).

The UCLA Phonological Segment Inventory Database (UPSID) is compiled by Maddieson and colleagues at UCLA and is one of the most detailed collections of segment inventories of the world languages. Originally, the database consisted of 371 languages (Maddieson 1984.). A later corrected and expanded version (Maddieson & Precoda 1991, Maddieson 1991) contains already a 451 language sample, and I use this later version, usually referred to as UPSID-451, for the computations.

UPSID-451 contains phonologically contrastive segments of the world languages. For each language in the database, a segment inventory is included containing segments that have lexically contrastive function.

UPSID-451 is compiled on a genetic principle, classifying the languages into 18 major genetic groupings (= language families). These language families fall into 4 large geographical areas: Eurasia, Americas, Africa, Australia.

UPSID-451 can be assessed as a very representative sample. The primary goal of the database is to provide a sample from which statistically valid statements concerning frequency and co-occurrence can be drawn. The database is thus a reliable empirical source for a computational investigation.

In our quest for plausible universals, I ran UNIVAUTO on the UPSID-451 data, requiring that the system discover implications of the logical form “If Segment A, then Segment B” such that these implications are:

Below are listed the first five universals found and verbalized by UNIVAUTO:

The system found and verbalized 146 highly significant universals of type “If Segment A, then Segment B” that are valid in at least 90% of the languages in UPSID-451 and are supported by languages from at least two different language families from at least two different geographical areas. Of these 146 universals, 50 or 1/3 are exceptionless.

The vast majority of listed patterns do not appear in Maddieson 1984 (or any other source known to me). These patterns seem to have remained unnoticed by previous researchers, showing, again, the computational complexity of the task of discovering universals. Thus, to find all valid patterns (relative to a database) of the form “If Segment A, then Segment B” implies, first, finding all logically possible combinations from all segments used in the database, and secondly, testing all such hypotheses against the database. More technically, this reduces to finding all ordered pairs (= implications) of segments from the segments used and testing them. The number of ordered pairs of segments from a set of N segments is computed by the formula: N² – N. Thus, given that the number of segments used in UPSID-451 is N = 919 the number of potential hypotheses to construct and subsequently test equals 919² – 919 = 843,642. Additionally, the found putative universals must be statistically significant to minimize mere chance as the reason for their occurrence, a task also difficult to achieve manually. Most previous research on phonological universals in contrast seems to have been appeased with those universals that just happen to be noticed by individual linguists, instead of trying an exhaustive (machine) search of the databases studied, and not infrequently posited universals without explicit concern for their statistical significance.

UNIVAUTO has so far produced around 60 pages of text, covering about 250 new universals from the fields of word order and phonology. It has found (cf. Pericliev 1999a, 2000) that two of the proposed word order universals in the classical article by Greenberg (1966) are actually false and that seven others are exceptionless relative to the database investigated rather than statistical, as claimed by Greenberg.

Three other of Greenberg’s ordering universals were shown to tacitly assume feature values for some languages which are actually unknown to the database.

All these circumstances have remained unnoticed by previous human researchers, and ironically, some of the problematic universals are widely disseminated in the linguistic community (cf. e.g., the complete enumeration of Greenberg’s (1966) ordering universals in The Linguistics Encyclopedia, London and N.Y., 1991).

Inspecting two further word order databases from Greenberg (1966) and Hawkins (1983), which are really small 24x4 tables, the system also managed to find patterns that have escaped these authors, considered to be the authorities in the field.

Similarly, many novel phonological universals were found in the UPSID database in comparison with Maddieson’s (1984) findings, as well as some problems in these and other related proposals in the literature (lack of statistical significance and/or low level of validity and/or insufficiently diverse language support). (Cf. Pericliev 2008)

Three design properties of the system enhance the chances of finding novel knowledge. The first is the system’s ability to explicitly check its own discoveries against those of a human agent exploring the same data. More generally, this strategy is not impractical in a linguistic discovery system on universals in view of the availability of universals archives, such as the Konstanz Archive. The second is the exhaustive search of a combinatorial space that the system performs. Such comprehensive searches of combinatorial spaces, that are furthermore dense with solutions, are known to be very difficult, if not completely beyond the reach of a human investigator, a trite circumstance in computer science (but, unfortunately, not so in many domain sciences as linguistics). As a corollary of the exhaustive search, the system can make meta-scientific claims to the effect that “These are all universals of the studied type (relative to the database)”. The third design property is the ability of the system to handle diverse queries (esp. those concerning different logical types of universals), some of which may not have been seriously posed or pursued before.

The interestingness of UNIVAUTO’s findings is partly derived from the interestingness of the task it automates. Indeed, linguistics has always considered the discovery or falsification of a universal an achievement.

From a purely design perspective, the system attempts to enhance the discovery of interesting universals by outputting only the stronger claims and discarding the weaker ones (cf. Pericliev 2012). Thus, if Universal 1 logically implies Universal 2, the first is retained and the second is ignored. E.g., “All languages have stops” implies “If a language has a fricative, it also has a stop” and the second claim must therefore be dismissed as a pseudo-universal. (Ironically, this claim has been actually made more than 60 years ago in a celebrated book by Jakobson (1941), another linguistic luminary, and has never been refuted.)

The plausibility of posited universals has been a major concern for UNIVAUTO. Universals are inductive generalizations from an observed sample to all human languages and as such they need substantial corroboration. The system disposes with two principled mechanisms to this end. The first is the mechanism ensuring statistical plausibility, allowing the user to specify a significance threshold for the system’s inferences. It is embodied in two diverse methods, the chi-square test and the permutation test, which can alternatively be used. The second plausibility mechanism pertains to the need for qualitatively different languages to provide support for a hypothetical universal for it to be outputted by the program. The specific measure of “typological diversity” of the supporting languages is chosen by the user of the system, by selecting the minimum number of language families and geographical areas to which the supporting languages must belong.

The plausibility of (eventual) criticisms of a human agent’s discoveries is even less problematic. Indeed, one can definitely (and not only plausibly) say when a proposition is false relative to a known database, and that is exactly what the system does.

With some discovery systems, the user/designer may encounter difficulties in interpreting the program’s findings. With other systems, typically those that model previously defined domain-specific problems, and hence systems searching conventional problem spaces, the findings would as a rule be more intelligible. However, intelligibility is a matter of degree and UNIVAUTO seems unique in producing an understandable English text to describe its discoveries.

UNIVAUTO thus both states in English its discoveries (new universals + problems) and the supporting evidence that makes these discoveries plausible/valid. Additionally, it provides a general context into which it places these discoveries (in the introductory parts of the generated text), as well as a summary of the findings (in the conclusion part of the generated text). The readability and self-contained nature of the texts the system normally produces must not be overstated. Some users may prefer to use the output as a “skeleton article” to be subsequently enlarged and edited to fit further stylistic and linguistic needs.

Some discovery systems model general scientific tasks (for induction, classification, explanation, etc.) and would therefore be readily portable to diverse problems in diverse scientific domains. UNIVAUTO is such a system. It mimics the general task of discovery of (logic) patterns from data, and hence would be applicable not only to language universals discovery, where the objects described in the data are languages, but to any database describing any type of objects, be they linguistic or not. This however applies primarily to its discovery module. The text generation module, as it stands, is less flexible and most probably unportable to a domain outside of universals.

The degree of formalization discovery programs require may result in our deeper understanding of the tasks modeled, esp. if the sciences from which the task is originally taken are not sufficiently formalized. Another source of insightfulness may be the outcomes of discovery programs, in the case when they make conspicuous some overlooked aspects of the results. Both the implementation and use of UNIVAUTO have triggered a number of linguistically important insights.

Some of these are worth mentioning here.

First, in the application of the system to linguistic typologies it was consistently found that a set of non-statistical (in contrast to statistical) universals exists that describes all and only the actually attested types, whereas previous influential authors (Hawkins 1983), although strong proponents of exceptionless universals, have claimed them insufficient to do the job. This consistency of the system’s results could not be chance of course, so that it was only a short step finding the explanation. Indeed, a linguistic typology is equivalent to a propositional function, and therefore, as known from propositional logic, for any propositional function there exists a propositional expression that generates it. As a corollary, for any linguistic typology there exists a set of non-statistical universals, describing all and only its attested types (Pericliev 2002).

And, secondly, exploring the 451 language database UPSID with UNIVAUTO has led to the formulation of a phonological principle to the effect that if Phoneme 1 implies Phoneme 2, then both phonemes share at least one feature and, besides, Phoneme 2 never has more features than Phoneme 1. This formulation was made possible only after the system’s discovery of all universals of this type valid in the database. The subsequent (machine-aided) representation of the phonemes in terms of their feature structure highlighted this statistically significant pattern, holding in 94.5 per cent of the cases (Pericliev 2008).