In lexicography, there are two commonly used types of semantic data categorization: semantic domains and semantic labels. The difference between the two approaches is simple: semantic domains presume that each lexical meaning belongs to one and only one group (or sub-group of a larger group). Semantic labels do not have such limitations.
Most recently published cross-linguistic wordlist data are often organized into semantic domains, cf. the CLLD-friendly lists, including WOLD (Haspelmath and Tadmor 2009), IDS (Key and Comrie 2005), LexCauc (Belyaev and Forker, unpublished) and others. The same type of data organization has also been used in earlier works, such as Kibrik and Kodzasov (1988, 1990) thesauri of East-Caucasian verbs and nouns or SIL Comparative African Wordlist by Snider and Roberts (2006). The WOLD semantic categorization has also been adopted in the Concepticon project (List et al. 2020) that features maps different wordlists using a set of clearly defined meta-concepts to which the entries in the lists are linked.
Below I provide two examples of such notation from Haspelmath and Tadmor (2009) and from Kibrik and Kodzasov (1990):
(1) WOLD example:
Meaning 5.7: the potato
Semantic field: Food and Drink
(2) Kibrik and Kodzasov (1990) Example:
1. Тело (человека, животного) [Body (human, animal)]
I.1. Голова и шея [Head and neck]
1. Голова [Head]: <words for ‘head’ in different languages>
This type of categorization, however, may lead to peculiar results. For example, in the WOLD database the words for potato, olive and pepper belong to ‘food and drink’, while pumpkin, mushroom and banana are ‘agriculture and vegetation’. The structure of the database does not allow to get all the six words with a single query using semantic annotation: one may only get all the words belonging to these categories, including e.g. farmer and field, which obviously do not belong to ‘food and drink’. Mathematically speaking, this structure only allows for a set union operation, but does not allow for a set intersection (simply because sets do not intersect).
This problem is relevant for most of the wordlists simply because objects and actions usually belong to multiple domains, just like potato is both ‘food’ and ‘plant’, a farmer belongs to ‘agriculture and vegetation’ (a farmer cultivates plants) and simultaneously is a type of ‘social relation’ (a farmer has a certain position in the social hierarchy), while field belongs both to ‘agriculture and vegetation’ (because it is where the plants are being cultivated) and ‘spatial relations’ (because it is a place). In the WOLD type of classification, however, one is always forced to decide, which single semantic domain each word belongs to, even though more than one may be relevant. Thus, if one wants to make a claim about borrowability or other properties of particular semantic domains, domain-based classification may be inapplicable, not to say useless, and the data may require a lot of additional markup and re-annotation.
Traditional dictionaries, on the other hand, rarely use this type of categorization and usually rely on labels. Each word in these dictionaries can be annotated with multiple categorical labels based on its semantics and other properties. For example, in Oxford Learner’s Dictionary the word chap is simultaneously marked as British English, informal and old-fashioned, something completely impossible in the domain-based approach:
(3) Oxford Learner’s Dictionary Example:
(British English, informal, old-fashioned)
Before computational approaches have been widely introduced to linguistics, semantic domains had a major advantage: searching for data and analyzing it was much simpler when working with lexical data organized into semantic domains. If one wanted to make a statement about lexical processes (e.g. borrowing) and somehow involve semantics, domain-based structure was much more convenient because a linguist would be able to use a pre-existing classification rather than manually look through the whole dictionary looking for each word with a specific label.
With today’s technology, however, it has become very easy to search and subset data by labels. Since most of the data are stored in digital table formats (e.g. csv, tsv, csvw, xls, etc.) and not on paper, label-based annotation no longer causes problems for searching: one may simply filter the data set using an “IF x IN y” statement.
Thanks to this major simplification, instead of assigning one category per lexeme, one may list the categories it belongs to in a corresponding cell of the column where semantic attribution is stored, similarly to how hashtags are used in Twitter or Instagram. The WOLD categorization could be improved as follows:
|Lexical Meaning||Semantic Tags||Lexical Meaning||Semantic Tags|
|to eat||food and drink||pumpkin||agriculture and vegetation; food and drink|
|food||food and drink||mushroom||agriculture and vegetation; food and drink|
|potato||agriculture and vegetation; food and drink||banana||agriculture and vegetation; food and drink|
|olive||agriculture and vegetation; food and drink||farmer||agriculture and vegetation; social and political relations; basic actions and technology|
|pepper||agriculture and vegetation; food and drink||field||agriculture and vegetation; the physical world; spatial relations|
We thus allow semantic domains to intersect and assign the words for potato, olive, pepper, pumpkin, mushroom and banana to all categories that seem to be relevant. A query for ‘food and drink’ will result in everything that is related to eating and drinking and the query for ‘agriculture and vegetation’ will result in all plants and everything else related to agriculture. To compare, currently used domain-based structure will yield some of the things that can be eaten and some of the things that can be planted, leading to obvious difficulties in the analysis.
A similar approach has already been implemented by several scholars in their annotation of wordlist data. One such example is a meta-wordlist by Starostin (2000), supposedly used by Sergei A. Starostin to determine semantic matches between words for cognate detection. Each gloss in the list was assigned a set of semantic tags. To my knowledge, Starostin’s idea was to compare two glosses by their “senses”, i.e. by the lists of tags assigned to each concept. The advantages of this approach are obvious: using Starostin’s annotation one may automatically track cognates by partial semantic matching, which significantly speeds up cognate detection.
This system worked somehow differently from the one proposed here, mainly because of the different purposes. Starostin did not use semantic categorization but assigned what appears to be primitive meanings to the entries, e.g. ‘food’ was tagged with “to eat; grain; fat; belly; fruit”. Searching for similar words in Starostin’s system thus implied simply searching for words with similar sets of semantic tags and deciding how many and which tags two entries have to share in order to be ‘similar’. The STARLING similarity judging system has been revived in the pysen Python library recently published by List (for details see previous post in this blog, see List’s previous post in this blog on Concept similarity in STARLING).
The system that I propose in this paper is different from Starostin’s in that an ‘X’ would have to be a type of ‘Y’ to be assigned this tag, so since ‘food’ is not a type of ‘grain’, it could not be tagged with this category. The ‘grain’ tag could be used for meanings ‘wheat’, ‘oats’, ‘millet’ and the like. Importantly, the two approaches do not exclude each other and can be successfully work together. Since Starostin’s tagging system is already there and does not require any reworking, it can be treated as a parallel layer of annotation.
The approach I propose here can be implemented in a relatively easy way: the annotation system can be constructed by aggregating the existing categorical annotations from different lexical databases. For example, many lists linked to Concepticon provide a semantic categorization, which can easily be transformed into tag annotation by assigning all the possible categories from the donor lists to the Concepticon entries. Additionally, by linking a system with tags, such as the SENSE database of the STARLING software package to Concepticon, one could use these annotations as metadata for Concepticon concept sets. The suggested implementation is imperfect in many respects, but it could still work as a temporary solution for improving the semantic searchability of Concepticon. A perfect solution would, of course, be to annotate the database manually, but given the size of the database, it would require a significant amount of work just to develop the tagging system, let alone the annotation itself.
To sum up, when using the label-based approach to semantic classification, linguists are no longer forced to lose information in their annotation. By (re-)introducing label-based categorization to cross-linguistic lexicography and combining it with modern data formats, we are able to take the best from both approaches and create a markup that is both flexible and meaningful as well as easily searchable.
Haspelmath, Martin & Tadmor, Uri (eds.) 2009. World Loanword Database. Leipzig: Max Planck Institute for Evolutionary Anthropology. URL: https://wold.clld.org.
Key, Mary Ritchie and Bernard Comrie (eds.) 2015. The Intercontinental Dictionary Series. Leipzig: Max Planck Institute for Evolutionary Anthropology. URL: http://ids.clld.org.
List, Johann Mattis &Rzymski, Christoph & Greenhill, Simon & Schweikhard, Nathanael & Pianykh, Kristina & Tjuka, Annika & Wu, Mei-Shin & Forkel, Robert (eds.) 2020. Concepticon 2.3.0. Jena: Max Planck Institute for the Science of Human History. URL: http://concepticon.clld.org.
Roberts, J., & Snider, K. (2006). SIL comparative African wordlist (SILCAWL). Dallas: SIL International.
Starostin, Sergei A. (2000): The STARLING database program. Moscow: RGGU. URL: https://starling.rinet.ru/program.php?lan=en.
Кibrik, А. Je., & Kodzasov, S. V. (1988). Sopostavitel’noe izučenie dagestanskix jazykov. Glagol [A comparative study of Daghestanian languages. The verb]. Moscow: Lomonosov University Publishing.
Kibrik, K., A. Je. Kibrik and S. V. Kodzasov (1990): Sopostavitel’noe izučenie dagestanskix jazykov. Imja, fonetika [A comparative study of Daghestanian languages: Noun. Phonetics]. Moscow: Lomonosov University Publishing.