CLICS²: An improved database of cross-linguistic colexifications assembling lexical data with the help of cross-linguistic data formats

2.1 Harvesting cross-linguistic data with Concepticon

In order to make semantic association patterns comparable across the world’s languages, it is clear that we must base our analysis on a rigorous collection of comparative concepts. The Concepticon reference catalog project (List, Cysouw & Forkel, 2016; List, Cysouw, Greenhill & Forkel, 2018) is an attempt to provide consistent links across the multitude of lexical questionnaires that linguists have used to elicit words. Concepticon works by defining specific concept sets based on published datasets (whether these are from field work, or from historical or typological studies), and then linking the labels used by researchers to these defined conceptsets. For example, the “dull, blunt” vs. “dull, stupid” error is solved by linking the “dull” label in the list of Chén (1996) to the concept set stupid 1518 while linking that of Swadesh (1952) to the concept set blunt³⁷⁹.^[4] If possible, these links are further checked against the original data, that is, the words in the target languages that were elicited in the end, in order to make sure that what is glossed as blunt is indeed reflecting the comparative concept stupid.

The immediate advantage of linking data to the Concepticon is that it enables us to merge data from different sources quickly and safely. We now know which word lists contain lexemes for stupid1518 and which contain lexemes for blunt³⁷⁹. And we can now directly ask which languages contain lexemes that colexify stupid¹⁵¹⁸ and blunt³⁷⁹, and colexify these lexemes with any of the other 3144 concept sets defined in Concepticon. In order to avoid errors, we have striven for rigor and strictness when linking concepts to Concepticon. Concepticon does not tolerate “fuzzy” matchings and deliberately avoids linking one elicitation gloss in a single dataset to more than one concept set in the Concepticon resource. If no ideal concept set could be found to link a given elicitation gloss in a questionnaire, it was left unlinked rather than linking to a semantically “close” concept.

Concepticon concept sets can further be linked among each other with help of a simplifying ontology that identifies concept sets which are broader or narrower with respect to their denotation range. The concept set arm or hand²¹²¹, for example, is useful for languages such as Russian or Irish, where the words рука rukà and làmh respectively refer not only to the part of the arm which other languages denote as hand, but also to the entire upper limb. The concept set arm or hand²¹²¹ is thus considered broader than either hand¹²⁷⁷ or arm¹⁶³⁷. While scholars might object to this procedure, preferring to represent a comparative concept reflecting the semantics of Russian рука rukà or Irish làmh by linking them to both hand and arm, it is important to emphasize that this practice, which may seem counterintuitive from the perspective of a given language, is critical if one wishes to guarantee a rigorous mapping of word elicitation glosses in questionnaires to lexical comparative concepts. If a given questionnaire contains the gloss arm/hand (as we can find across many questionnaires which have been used to assemble a large number of data points) and we linked it to both arm¹²⁷⁷ and hand¹⁶³⁷, we would lose the essential information that the original questionnaire was asking for the word expressing the concept that covers both concept sets in a single term. Since the ontology allows us to derive the information that arm or hand is semantically broader than arm and hand, we can choose, over the course of our analysis, to link the elicitation gloss arm/hand to both narrower concept sets.

Each Concepticon gloss is also linked to additional metadata e.g. a semantic field, ontological categories (reflecting the more language-specific notion of part of speech), as well as additional metadata derived from norm datasets in psycholinguistics and natural language processing, including age-of-acquisition information for individual languages (Kuperman, Stadthagen-Gonzalez & Brysbaert, 2012), ontologies like WordNet (Princeton University, 2010), or word frequency counts, again for individual languages (Brysbaert & New, 2009).

To illustrate how lexical comparative concepts are organized in the Concepticon, Figure 2 provides a small excerpt of the data which is linked to the concept set fat (organic substance)³²³. As we can see from the figure, this concept set itself is narrower than the concept set organic fat or oil²⁵⁵¹, which shows that many languages do not explicitly distinguish oil from fat. On the other hand, the concept set is broader than fat (for nourishment)²⁰⁹⁵. The definition at the top of the figure indicates that the comparative concept should only be linked to those elicitation glosses which target the organic substance as opposed to potential non-organic variants. The five exemplary elicitation glosses in the table at the bottom of the figure are only a small excerpt of what can be found in the whole data linked by the Concepticon. According to the current version (Concepticon-1.1.0, List, Cysouw, Greenhill & Forkel 2018), the concept set fat (organic substance) recurs in 107 different questionnaires and surfaces in the form of 34 distinct elicitation glosses. As we can see from the five examples in the figure, elicitation glosses can vary drastically, not only because they may be given in different languages, but also because authors do not always pay much attention to consistency. Thus, Swadesh used two different glosses, fat (organic substance) in his list of 1952, and fat (grease) in his later list from 1955, but when inspecting other articles written by Swadesh, we can see that these were absolutely not the only two variants he used, and we find fat in Swadesh (1950) and grease in Swadesh (1971).

Figure 2:

Example for the representation of data in the Concepticon project. Note that the three different separators in the elicitation glosses for fat/grease are given as such in the data, thus reflecting the high degree of inconsistency we find in linguistic practice.

2.2 Using, expanding, and improving Concepticon

We have created a simple automated mapping algorithm to quickly link new wordlists into Concepticon. After applying this algorithm, all links are manually checked to avoid embarrassing errors. In order to further facilitate the task of concept mapping, we wrote a small web-based standalone application which serves as a straightforward lookup tool, including a fuzzy search, across all 3042 concept sets which are currently defined and which can currently be used in seven languages (English, German, Chinese, French, Spanish, Russian, and Portuguese). This web-application, which can be used offline from common web browsers, is provided along with the supplementary material (SI:A) accompanying this paper (see below for further information on the supplementary material). It can also be accessed at http://calc.digling.org/concepticon, where the most recent version is listed. Figure 3 gives a brief example illustrating how it can be used.

Figure 3:

Example for the web-based lookup tool for Concepticon mapping.

Given the complexity of the lexical semantics of natural languages, it is obvious that resources like Concepticon or datasets that link to it will contain uncertainties, less-than-perfect links, and even straightforward errors.^[5] Concepticon is designed to be easily correctable and welcomes contributions and additions submitted in form of GitHub issues^[6] or email inquiries.^[7] Reference catalogs such as Concepticon are community efforts that can only be enhanced if the research community actively takes part in improving them and it is advantageous for our field if scholars help correct existing problems.

3.1 General statistics

Our new database automatically derives colexifications for 1220 language varieties (of which 1029 are distinguished by Glottolog) and a total of 2487 distinct Concepticon concept sets. Based on a strict threshold that only accepts a given colexification if it occurs in at least three different language families (as defined by Glottolog),^[8] this results in a total of 2638 different colexification patterns, corresponding to 66140 individual instances of colexifications in the languages in our sample. Partitioning the colexification network with the help of the Infomap algorithm (Rosvall & Bergstrom, 2008) resulted in 248 different communities consisting of more than one concept.^[9] Given that some of the concepts in our data are never colexified (probably due to the sparseness of the data) our colexification network consists of 1534 different concepts.

Figure 4:

The current coverage of our colexification database in terms of language varieties. Colors indicate the major genetic subgroupings of the languages, following the Glotto- log classification.

In Table 2, we list the most frequently recurring colexifications in our database, sorted by the attestation per language family along with detailed counts on distinct languages and distinct words that attest the colexification. As can be seen from the table, the results are not particularly surprising. It is well-known that many languages use the same word for moon and month or for wood and tree. That kinship terms are particularly heavily colexified clearly results from underspecification and does not reflect any potential instances of semantic shift. After manually inspecting the data to look for potential errors, we are quite confident that the unbalanced distribution of our data did not lead to any major errors, providing the same look and feel as the original CLICS database, while offering a drastically increased quantity of data.

The increase in data is also reflected when one inspects the geographic distribution of languages covered in our colexification database. As can be seen from the map in Figure 4, our colexification data has drastically reduced the number of empty areas, which were so characteristic of the original CLICS database. However, this does not mean that the data could not be further improved. We can still find many areas on the map, especially in Africa and North America, but also in the Pacific and South Asia, in which coverage is poor to non-existent. We hope to improve the geographical coverage in further releases.

Figure 5:

Bird’s eye view of our new colexification data in CLICS². The graph shows all connected components (113) with some of the communities highlighted.

3.2 Exploring colexifications through web-based applications

Our application (hosted at http://clics.clld.org) allows users to explore the data from various additional perspectives, including geographic maps, the inspection of the data of individual languages, or the distribution of concepts for which we find translations in our data. In addition, each data point can be traced back to its original source, allowing the users to rigorously check whether the automatic findings we present can be confirmed through qualitative research. In order to illustrate the new application, a number of examples follow.

3.2.1 Bird’s eye view of the colexification data

Before going into details, a bird’s eye view of the data is given in Figure 5. While we can see that most of the nodes in our graph form a large connected component, we can also see that the community detection algorithm singles out communities, adding structure which would otherwise be difficult to spot.

Figure 6:

The largest community in our sample, with the concept say¹⁴⁵⁸ showing most connec- tions to the other concepts.

3.2.3 Colexifications of wheel and foot

As a further example, let us consider a case of regional colexification that was already mentioned by Mayer, List, Terhalle & Urban (2014) and can also be found in our new colexication database: the colexification of foot¹³⁰¹ and wheel⁷¹⁰ in some South-American languages. In contrast to the example of say, word, and language in the preceding section, this colexification does not reflect a global pattern which could be identified when looking into the partitions based on the Infomap community detection analysis, which places wheel and foot into distinct communities. An additional view of the colexification data, introduced by Mayer, List, Terhalle & Urban (2014), however, allows one to find areal patterns, provided they are frequent enough and recur across different language families. This view (called subgraph by Mayer, List, Terhalle & Urban 2014) presents the subgraph derived from the closest neighbors of a given query concept. Neighbors of the starting concept are identified by setting a frequency threshold. In consecutive steps, more nodes (the neighbors of the neighbors) can be added to the subgraph, depending on the size of the network, which should not exceed a certain number of nodes to allow for convenient inspection.

Thus, while the colexification between wheel and foot does not show up in our community analysis, we find it in the subgraph view, as shown in Figure 7. As we can see from the different concrete word forms reflecting the colexification, we are not dealing with a direct borrowing that spread among the languages. Instead, the colexification either reflects an instance of loan transfer (in the terminology of Weinreich 1974) or an indirect metaphorical extension. What may substantiate the latter hypothesis is the fact that the wheel-foot colexification is not restricted to Southern America, but seems to be also reflected in some African languages located on the Eastern coast of Africa (Gilman, 1986; Heine & Fehn, 2017), but our current version of CLICS² does not contain data on these particular languages. The explanation for this particular colexification can thus be sought in historical factors, as a metaphorical extension linked to the introduction of the wheel as a widespread technology in a colonial context.

Figure 7:

Detecting regional colexification patterns with help of the subgraph explorer. The col- ors for the languages in the top-left table indicate their genetic relationship. As can be seen, all languages belong to different language families (with three of them being isolates), although they are geographically close.

This again has immediate implications for ongoing debates on linguistic paleography. First, the wheel-foot metaphor shows that concrete historical events may be reflected in languages. Second, however, it also shows that we need to be very careful when evaluating this evidence. As we can see from the subgraph in Figure 7, there are plenty of colexifications for circle¹⁴⁶⁷ and wheel in our data as well (our data counts 26 concrete colexifications across 11 different language families). Assuming that societies usually have a way to express the concept ‘circle’, while ‘wheel’ may be missing, our data suggests that the most straightforward strategy to express a new concept ‘wheel’ starts from the word for ‘circle’. Since this can easily happen independently, as we can again see from our data, these findings might be of importance for on-going debates on the origin of terms for ‘wheel’, especially in Indo-European (Hock, 2017; Anthony & Ringe, 2015). Further studies on lexical typology, including studies on independently recurring patterns of semantic shift as well as the frequency of loan transfer, are required before this linguistic data can be reliably used to reconstruct ancestral cultures. Our extended colexification data may serve as a starting point for these investigations.

4.1 Shortcomings of the previous version of CLICS

The original CLICS database by List, Mayer, Terhalle & Urban (2014) was compiled in a mostly automatic manner. The data were assembled from four different sources (Key & Comrie, 2007; Group, 2008; Haspelmath & Tadmor, 2009; Borin, Comrie & Saxena, 2013) and were mostly already linked to the same set of comparative concepts, originally based on Buck (1949). The lexical entries were automatically cleaned, using regular expressions and similar standard techniques for text manipulation, and then compared for colexifications. The resulting network was analyzed with help of the Infomap algorithm for community detection (Rosvall & Bergstrom, 2008) in order to decrease the complexity, single out colexifications that point to instances of homophony, and split the semantic network into a meaningful set of subgraphs, representing areas of high semantic affinity, close to the notion of semantic fields (Anttila, 1990).

Since then CLICS has enjoyed considerable popularity among scholars working in the field of lexical typology. On the one hand, this is reflected in studies that mention the database in a favorable manner (Östling, 2016; Georgakopoulos & Polis, 2018; Šipka, 2015), as an inspiration source for similar or enhanced analyses (Söderqvist, 2017; Brochhagen, 2015; Dellert, 2016; Pericliev, 2015; Gast & Koptjevskaja-Tamm, 2018), or as a potential dataset for additional studies (Youn, Sutton, Smith, Moore, Wilkins, Maddieson, Croft & Bhattacharya, 2016). On the other hand, it is reflected in a couple of studies that make direct use of the data provided in CLICS (Schapper, Roque & Hendery, 2016; Koptjevskaja-Tamm & Liljegren, 2017; Staffanson, 2017; Regier, Carstensen & Kemp, 2016). It seems that the general strategy of using an interactive interface that allows scholars to explore the actual data, offering both a bird’s eye view on colexification patterns while keeping in touch directly with the original data fulfils a certain need for studies on lexical typology, serving as an example for computer-assisted as opposed to purely computer-based frameworks (List, 2016).

As mentioned, the earlier CLICS database suffered from a number of shortcomings. Not only was the coverage in terms of languages rather small, with a sample of only 220 languages heavily biased towards Southern America and North Eurasia (Östling, 2016), but the data was also not easy to expand, as new word lists would have demanded a considerable amount of overlap with the 1280 comparative concepts in CLICS. A factor further complicating the expansion of the existing CLICS database was the difficulty in its curation. Given that the data came from independent sources, and that the small number of developers did not have the linguistic expertise to check all wordlists systematically, it was impossible to correct errors in the data itself. Although such curation would have also gone against the original policy of the database, insofar as it was originally built on the idea of providing a different view of already curated datasets, it constituted a serious problem for further development. An additional problem was the transparency of the algorithms underlying CLICS: while the source code was online and freely available, it was difficult to use in its previous state, as it was largely undocumented and provided an unfortunate mix of code written for data deployment and code written for data analysis.

4.2 Software implementation

While the original CLICS framework was deployed via PHP, we have integrated the original code for data visualization (Mayer, List, Terhalle & Urban, 2014) into a Python-based CLLD application (Forkel & Bank, 2018). CLLD offers not only more granular access to the data, but also provides the look-and-feel of well-known typological databases like the Atlas of Pidgin and Creole Language Structures (Michaelis, Maurer, Haspelmath & Huber, 2013) or the World Atlas of Language Structures (Dryer & Haspelmath, 2013). The new framework allows users to inspect the inferred colexifications online via the CLLD framework, and also allows them to curate their own colexification datasets, to analyze, inspect, and even deploy them, thanks to a standalone application which they can use to convert their own colexification data to an interactive visualization very similar to the look and feel of the original CLICS database.

The new database of cross-linguistic colexifications comes along with a simple interface to compute the colexifications and network statistics. To identify and plot colexifications, we follow the strategy employed by Mayer, List, Terhalle & Urban (2014), but have significantly refactored the code, leading to a drastic increase of speed when searching for colexifications. In contrast to the regular expressions by which the data was automatically cleaned in the original CLICS framework, we have decided to use an even more rigorous approach by stripping off all metalinguistic information that can often be found in linguistic datasets (morpheme boundary markers, brackets for scholars’ comments or to indicate pronunciation alternatives) and representing all lexical entries internally with help of ASCII letters. While this carries the danger of introducing errors, our tests indicate that most of these problems can be singled out by only considering colexifications with a frequency above a certain average. Furthermore, since we show the original values as they appear in the data to the users, scholars wishing to work with the data in concrete form can easily check with the original entry or even go back to the original sources of each colexification that we identify with our automated procedure.

We provide platform-independent versions of the Python code which can be used on a Mac, Windows, or Linux computers running either Python 2 or 3. The package provides useful command line tools which we describe in detail at the projects GitHub page at https://github.com/clics/clics2. The software package providing the colexification API is further hosted with Zenodo (see https://doi.org/10.5281/zenodo.1299093 for the most recent version), as well as the 15 datasets (see https://zenodo.org/communities/clics/).

In addition to the CLLD application that provides the data online, we also provide code that exports colexification data to a standalone application, purely based on JavaScript, which can be used locally (and offline) in a web browser, or shared online using a static web server.^[11] We assume that this service may turn out to be useful especially for users who want to run their own datasets through our framework, but don’t have the technical means or expertise to set up a complex CLLD application. The new CLICS² software package also offers information how the standalone application can be computed.

4.3 Cross-linguistic data formats

In order to increase the comparability of cross-linguistic data and to ease the curation and reuse of existing datasets, we follow the standards and recommendations of the Cross Linguistic Data Framework (CLDF). CLDF is a standard to capture different data types often encountered in cross-linguistic research, such as wordlists, typological features, parallel texts, and dictionaries (Forkel, Greenhill & List, 2017). The major features of the CLDF specification are: (A) a simple text format for data-storage, based on CSV (comma-separated values), extended by recent recommendations by the World Wide Web Consortium, allowing metadata to be incorporated and data to be linked across multiple CSV files (Pollock, Tennison, Kellogg & Herman, 2015; Tennison, Kellogg & Herman, 2015), (B) a flexible software API which allows validation of whether a given dataset conforms to the specifications, (C) an ontology which allows frequently recurring objects and properties in comparative linguistics to be recognized, and (D) the rigorous integration of reference catalogs in order to increase the comparability of data across datasets.

4.4 Cross-linguistic reference catalogs

CLDF strongly encourages the usage of reference catalogs, such as Glottolog or Concepticon, when preparing linguistic data. The advantage of organizing language varieties not only by their common name, but also by adding the identifiers offered by Glottolog are obvious. Since Glottolog harvests various types of information regarding language varieties all over the world, ranging from geographical coordinates via references in the literature up to genealogical classifications, scholars linking the languages in their data to Glottolog identifiers can automatically dispose of this information when carrying out additional studies. Disadvantages may result from incorrect links to Glottolog or from problems that experts may encounter when checking the information provided by Glottolog. If, for example, scholars do not agree with the genealogical classification provided by Glottolog, they may prefer to add their own classification to their dataset. However, even if specific information turns out to be erroneous or not satisfying enough for scholars to help in their application, it is still useful to try to provide a link to Glottolog, as it will make it much easier for other scholars to find their data. Furthermore, Glottolog is curated in public and changes can be proposed and made in a transparent manner in the form of GitHub issues^[12] or by contacting the editors via email.

As outlined in Section 2, these advantages also hold for the usage of Concepticon as a reference catalog for comparative concepts. While scholars may still use and embrace their individual questionnaires with their preferred elicitation glosses and concept definitions, linking them to Concepticon guarantees that their data is cross-linguistically comparable and easily accessible to other researchers as well.

4.5 Choosing a representative sample using Average Mutual Coverage

A crucial issue when assembling different datasets in the way this is done in our updated version of the CLICS database is whether the overlap in terms of comparative concepts across datasets is sufficient and representative enough. While the data aggregated from the 15 different datasets should be interesting enough for manual inspection and analysis, some analysis strategies (for example those based on hypothesis testing, as mentioned by Roberts 2018) will require balanced and representative samples.

A seemingly straightforward way to identify a representative sample would be to determine the subset that provides the optimal overlap in terms of languages and concepts. In order to get a better idea of how skewed the data in our updated version of CLICS actually is, we carried out a detailed investigation of the different subsets of the data, employing the concept of average mutual coverage (AMC) in multilingual wordlists, as provided in the most recent version of the LingPy software package (List, Greenhill & Forkel, 2017, Version 2.6).

Here, the AMC of a given wordlist is defined as the average of the number of concepts shared between all pairs of languages in a given wordlist divided by the number of concepts in total. Assuming we have a concept list of 100 concepts and three different languages A, B, and C, which have translations for 90, 70, and 60 concepts each, we can determine the average mutual coverage by first checking the individual overlap among the languages (which is not necessarily equal to the number of the concepts translated in the “smaller” of two varieties), and then divide these numbers by the total amount of concepts. If we assume a mutual coverage of 65 between A and B, of 55 between A and C, and 45 between B and C, we can sum up and average the mutual coverage between all pairs. In this example, this would result in an AMC score of 0.55 (0.65+0.55+0.453).^[13]

Therefore, if users want a representative sample, they can make use of our AMC statistics, which are provided along with the source code to compute them in the supplementary material, to extract their sample of choice of the datasets we provide in full (SI:B).

We can use this metric to evaluate how well-balanced a given selection of languages and concepts is. This can be done by dividing the data into different subsets in which concepts and languages are consecutively deleted from the data, using their average size (number of concepts, or number of languages which provide a translation for a concept) as a criterion. We can then compute the AMC for each of these subsets and plot the data in order to see how the AMC scores change when reducing the number of languages and concepts in the data.

We carried out this analysis both for our new collection of datasets in CLDF format and the data underlying the original CLICS database. The results can be seen in Figure 8. As can be seen from these plots, our new data collection is heavily unbalanced, with extremely low mutual coverage scores for samples of a large number of concepts and languages. Comparing our data with the AMC statistics for different subsets of the original CLICS database, however, we can also see that our data supersedes the original CLICS coverage for a subset of about 1200 concepts and about 300 languages.

Figure 8:

Average mutual coverage in A CLICS, and B our new colexification database. The legend shows the number of concepts for which the AMC has been calculated.