Illustrating linguistic data reuse: a modest database for semantic distance

Besides new algorithms and tools that facilitate established workflows, one change prompted by computer-assisted approaches to language comparison is a distinct relationship between scientists and their data. A critical part of our work, and perhaps the one with the most lasting impact, is to promote an approach in which the data life-cycle is not constrained within the limits of planning and publishing a study. Data are organized and planned for reuse in investigations perhaps not even considered during collection, with the output of one project becoming the input of another.

An ongoing project at our department, the Cross-Linguistic Data Formats (CLDF) initiative (Forkel et al. 2018), embodies this life cycle of data in line with the FAIR principles of data management (Wilkinson et al., 2016). Part of the ambitious umbrella project Glottobank, a global research consortium to document and understand linguistic diversity and evolution,  datasets collected in CLDF formats can be quickly aggregated from different sources to form new datasets. An example of how CLDF stands at the heart of a data life-cycle is our CLICS project (Rzymski, Tresoldi et al., 2019), in which we carry out a comprehensive analysis of colexifications (i.e., words used to express two different concepts, either due to polysemy or to homonymy), automatically inferred from lexical sources that we converted to CLDF format and aggregated. CLICS is a derivative database, and the success and interest among the public, from psycholinguists to conlangers, also owes to the coding framework that backs it. Using the Python APIs that we have developed, anybody can recompute the network of colexifications and its clusters, experiment with other designs and algorithms, or apply the published network to alternative objectives.

We can illustrate this progression with the case of a less ambitious personal project in which, for two distinct goals, I lacked some measure of semantic distance (however approximate) between pairs of concepts from our Concepticon project (List et al. 2019). I needed distances both for predicting the prevalence of accidental resemblances (Tresoldi, 2019), and for simulating linguistic change to test novel phylogenetic techniques. There are several interpretations of “semantic distance,” as in hierarchic or otherwise structural ontologies (for example, the amount of nodes that needs to be traversed in WordNet, cf. Princeton University, 2010) or in complementary occurrences in corpora (as in vogue with word vectors, in the often repeated “king – man + woman = queen” equation, but see Nissim 2019). Neither method was adequate because, more as with typologies presented in semantics and semiology (cf. Traugott & Dasher 2001, Bloomfield 1933, Blank 1999), I needed an approximate probability for a word changing its meaning following a given “semantic path” (as in “awesome”, formerly “inspiring awe” and now “exciting”, or “villain”, formerly “countryman” and now “vile”) as well as the chance for apparently unrelated lexemes having a common, albeit remote, source (such as “mouse” and “muscle”). A network linking all concepts, but where the path between WOOD and TREE was shorter than the one between CUSHION and COOKING.

Before running community detection algorithms such as “infomap” (Rosvall et al. 2009), CLICS first normalizes lexeme transcriptions and then imposes thresholds of occurrence. These steps and limits are helpful in a study for establishing global and areal patterns, but results in a graph with singletons (words not linked to any other) and independent subgraphs. There is no guarantee that it will link all concept pairs, no matter how long the path between them. Likewise, edge weights are computed for the same goal of community identification rather than for approximate semantic closeness.

As a tool developed with principles of FAIR data in mind, the CLICS library allows to deal with the first issue by allowing to collect any and every occurrence of colexification. The table below presents a sample of the output, with each pair of concepts followed by its counts, ranging from the most common, MOON/MONTH, to spurious associations (like LOOK/BURNING, but notice how, among the minimum counts, we find semantically similar concepts, such as FLOWER/SPROUT (VERB)):

Concept A Concept B Family count Language count Word count
MOON MONTH 57 320 328
WOOD TREE 57 298 405
CLAW FINGERNAIL 55 217 225
LOOK BURNING 1 1 1
STEPDAUGHTER CHILD-IN-LAW 1 1 1
FLOWER SPROUT (VERB) 1 1 1

This new output, in a textual tabular format, can be combined with the communities published with CLICS to get the desired distance matrix. With the script available on GitHub, we can define different correction parameters for the counts of each pair (favoring families over languages, and languages ​​over words), as well as reducing the distance between concepts of the same cluster. The result is what we call for: a graph of which all Concepticon concepts are a part (i.e., a partially connected network), with edge weights based on colexification counts and corrected according to cluster membership. The hypothesis is that meanings change according to semantic distance, with changes within the same cluster meaning more likely. Our script generates a file in standard GML format, available here which can be explored with ordinary graph tools.

Despite not all concept pairs being linked by means of a single edge, it is now possible to find a path between every pair, with the distance between the concepts, taken as a surrogate for the semantic distance, corresponding to the sum of the edge weights involved. We can obtain the shortest and best path, or even sub-optimal ones, with pathfinding algorithms such as the ones of Dijkstra (1959) and Yen (1971). Computing the shortest path for the over 3 million pairs in this experiment is not particularly useful, as many of the links would be spurious and an actual account would need to look for the mean distance of non-overlapping shortest path (as performed in network analysis). However, it is still fun to play and look for the distance between a handful of random pairs (here the defined as the mean distance of the three shortest paths), along with the best overall solution for each pair:

Concept A Concept B Distance Shortest Path Steps in shortest path
BALL HALF 36.70 BALL/INSIDE/CENTER OR MIDDLE/HALF 2
FODDER MAPLE TREE 188.73 FODDER/HIDE/FEAR (TO BE AFRAID)/MAPLE TREE 2
COCONUT FINE (PENALTY) 154.25 COCONUT/WET/ANSWER/FINE (PENALTY) 2
BOLT (MOVE IN HASTE) CABBAGE 326.74 BOLD (MOVE IN HASTE)/FALL/FRUIT/VEGETABLE/CABBAGE 3
BRAID (VERB) SLEEP 155.17 BRAID (VERB)/SWALLOW/SLEEP 1

While these semantic distances are tentative and far from ideal, as a quick exploration can confirm, it is worth looking at how FAIR data made it feasible to promptly and easily obtain a database to test both the prototypes mentioned in the beginning. As part of a scientific cycle, it not only suggested that it was worth to continue investing both ideas, but it further allowed to generate data that other scientists can use just as promptly and quickly, spending even less time in data collection and processing (thus also saving computing power). We could use these data, for example, to bootstrap projects for developing  similar but better data, perhaps combining the different approaches to “semantic distance“ mentioned above, and which likewise would not need to be restricted within the boundaries of a single publication.

References

Blank, Andreas. 1999. “Why do new meanings occur? A cognitive typology of the motivations for lexical semantic change” in Blank, Andreas; Koch, Peter (eds.), Historical Semantics and Cognition. Berlin/New York: Mouton de Gruyter, pp. 61-90.

Bloomfield, Leonard. 1933. Language. New York: Allen & Unwin.

Dijkstra, Edsger W. 1959. “A note on two problems in connexion with graphs.” Numerische mathematik 1, no. 1: 269-271.

Forkel, Robert; List, Johann-Mattis; Greenhill, Simon J.; Rzymski, Christoph; Bank, Sebastian; Cysouw, Michael; ammarström, Harald; Haspelmath, Martin; Kaiping, Gereon A.; and Gray, Russell D. 2018. “Cross-Linguistic Data Formats, advancing data sharing and re-use in comparative linguistics.” Scientific data 5.

List, Johann Mattis; Greenhill, Simon J.; Rzymski, Christoph; Schweikhard, Nathanael; Forkel, Robert (eds.). 2019. Concepticon 2.1.0. Jena: Max Planck Institute for the Science of Human History. (Available online at http://concepticon.clld.org, Accessed on 2019-10-28.)

Nissim, Malvina; van Noord, Rik; van der Goot, Rob. 2019. Fair is Better than Sensational:Man is to Doctor as Woman is to Doctor. arXiv:1905.09866

Princeton University “About WordNet.” WordNet. Princeton University. 2010.

Rosvall, Martin; Axelsson, Daniel; and Bergstrom, Carl T. 2009. “The map equation.” in The European Physical Journal Special Topics 178, no. 1: 13-23.

Rzymski, C., T. Tresoldi, S. Greenhill, M. Wu, N. Schweikhard, M. Koptjevskaja-Tamm, V. Gast, T. Bodt, A. Hantgan, G. Kaiping, S. Chang, Y. Lai, N. Morozova, H. Arjava, N. Hübler, E. Koile, S. Pepper, M. Proos, B. Epps, I. Blanco, C. Hundt, S. Monakhov, K. Pianykh, S. Ramesh, R. Gray, R. Forkel, and J. List. 2019. The Database of Cross-Linguistic Colexifications, reproducible analysis of cross-linguistic polysemies. Manuscript. 1-24. Preprint, currently under review.

Traugott, Elizabeth C.; Dasher, Richard B. 2001. Regularity in Semantic Change. Cambridge: Cambridge University Press.

Tresoldi, T. 2019. “A cross-linguistic computational approach on chance resemblances”. Presented at the workshop Computer-assisted approaches in historical and typological language comparison, organized as part of the Annual Meeting of the Societas Linguistica Europea (2019/08/21-24, Leipzig).

Wilkinson, Mark D., Dumontier, Michel; Aalbersberg, IJsbrand J.; Appleton, Gabrielle; Axton, Myles; Baak, Arie; Blomberg Niklas et al. 2016. “The FAIR Guiding Principles for scientific data management and stewardship.” in Scientific data, 3.

Yen, Jin Y. 1971. “Finding the k shortest loopless paths in a network.” Management Science 17, no. 11: 712-716.


Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.