Studying language contact within a computer-assisted framework
Talk held at the 64th Annual Conference of the International Linguistic Association (2019-05-30/2019-06-01, Universidad Nacional de San Martín, Buenos Aires).
Group “Computer-Assisted Language Comparison” Department of Linguistic and Cultural Evolution Max-Planck Institute for the Science of Human History Jena, Germany 2019-06-01 very long title P(A|B)=P(B|A)... 1 / 32
(1821-1868) “These assumptions, which follow logically from the results of our re- search, can be best illustrated by the image of a branching tree.” (Schle- icher 1853: 787) 3 / 32
(1843-1901) “I want to replace [the tree] by the im- age of a wave that spreads out from the center in concentric circles be- coming weaker and weaker the far- ther they get away from the center.” (Schmidt 1872: 27, my translation) 5 / 32
Contact Hugo Schuchardt (1842-1927) “We connect the branches and twigs of the tree with countless horizon- tal lines and it ceases to be a tree.” (Schuchardt 1870 [1900]: 11) 7 / 32
Cantonese [tai³³-iœŋ²¹] (Mandarin tàiyáng) phylogeny-related English mountain vs. French montagne, Spanish montaña trait-related German Damm vs. English dam distribution-based German Job, Joker, Junkie, Journal . List (forthcoming) 10 / 32
lookalikes between Chinese and Hmong-Mien as borrowings from Chinese, for historical reasons (Ratliff 2010) assume all vocabulary from a specific semantic field to be borrowed (e.g., religion, seafaring, etc.) 11 / 32
early 21st century with phylogenetic approaches (Gray and Atkinson 2003, Ringe et al. 2002) accompanied by pioneering work on sequence comparison (Kondrak 2000) later followed by more and more approaches on different topics (phylogenetic networks, Nakhleh et al. 2005, automatic cognate detection, Hauer and Kondrak 2011), now a fully established sub-field of historical linguistics 12 / 32
solutions: conflicts in the phylogeny, explain them by invoking borrowings (MLN approach, Nelson-Sathi et al. 2011, List et al. 2014) similar words among unrelated languages (Mennecier et al. 2016) tree reconciliation methods (Willems et al. 2016) borrowability statistics (Sergey Yakhontov, as reported by Starostin 1990, Chén 1996, McMahon et al. 2005) 13 / 32
of proposed solutions: conflicts in the phylogeny tend to overestimate the amount of borrowing, since there are multiple reasons for conflicts in phylogenies, not only borrowing (Morrison 2011) sequence comparison on unrelated languages seem solid, but one needs to be careful with chance resemblances based on onomatopoetic words etc. (mama, papa, etc., Jakobson 1960, Blasi et al. 2016) tree reconciliation methods are unrealistic if word trees are derived from simple edit distances sublist-approaches may be useful, but they require large accounts on known borrowings, which we usually lack 13 / 32
data in linguistics are steadily increasing our qualitative methods reach their practical limits we need to take computational methods into account but computational methods are not very accurate and may yield wrong results 15 / 32
Starting Grant (2017-2022) Host: MPI-SHH (Jena) Current team: 2 post-docs, 2 docs, and myself Objectives go beyond historical linguistics and Sino-Tibetan (but they are our starting point) http://calc.digling.org 17 / 32
Solving 1 identify the core class of your problem (modeling, inference, analysis) 2 formalize the problem in a way that allows one to test it (specify data and techniques for evaluation) 3 do not hesitate to define sub-problems, given that qualitative solutions are often holistic 4 look at existing qualitative solutions 5 search for inspiration in neighboring disciplines (graph theory, computer science, evolutionary biology) by looking for similar processes that could be addressed in an analogous or similar way 6 accept a qualitative or semi-automatic solution for inference processes, but make sure that the results are annotated in a machine-readable way 7 insist on transparent output (no black boxes) to allow for an immediate review of results by experts 19 / 32
Problems (1-3) borrowing is a process that happened during different stages in time, reflected in form of borrowing or contact layers identification of source and target of borrowings is almost impossible without knowing the history of a given area distinguishing borrowing from inheritance, chance, and typological patterns of denotation is also difficult for classical linguistics contact areas may overlap 21 / 32
Existing Solutions (4-6) recent borrowings can be detected with automatic sequence comparison approaches (Mennecier et al. 2016) searching for borrowings in unrelated languages spoken in similar regions can control for inheritance (Mennecier et al. 2016) highly advanced techniques for cognate detection are available by now (List 2014, List et al. 2017) methods for clustering and partitioning are well advanced, but need to be applied in a correct fashion 22 / 32
Transparent Output (7) lift the data to a high standard of phonetic transcriptions use interactive applications to transparently share the findings rigorous testing and training on datasets from different languages of the world prefer direct output (concrete items identifying a contact area) for initial studies 23 / 32
languages from three different families (Sino-Tibetan, Hmong-Mien, Tai-Kadai), aggregated from four different sources (Beijing University 1964, Sun et al. 1991, Chen 2012, Castro 2015) unified phonetic transcriptions following the Cross-Linguistic Transcription System framework (Anderson et al. forthcoming, https://clts.clld.org) unification of elicitation glosses with help of Concepticon (List et al. 2016, https://concepticon.clld.org) data curation following the principles of the Cross-Linguistic Data Formats initiative (Forkel et al. 2018, https://cldf.clld.org) first inspection of data with help of EDICTOR (List 2017, http://edictor.digling.org) 25 / 32
cognate detection using LexStat as implemented in LingPy (List 2014, List et al. 2017) B cross-family borrowing detection using a new feature-based prosody-aware approach for pronunciation distance calculation and flat clustering approach C interactive analysis of inferences D partition cognate sets into groups indicative of a contact zone 26 / 32
Calculation pronunciation distance depends on prosody (with weak and strong positions in each word, see List 2014) feature systems for huge numbers of sounds were lacking so far, but are available now with CLTS (Anderson et al. forthcoming) alignment methods are well-developed and can be used to compare words beforehand (List 2014) The approach is work in progress, contact me for more information. 27 / 32
two major contact areas (Hmong-Mien and Sui, Sinitic/Bai and Hmong-Mien) not all languages under similar influence inspection shows that most borrowings can be confirmed 28 / 32
partitioning cognate sets and their associated meanings based on their distribution across languages yields about 6 groups in which five and more concepts are consistently shared the groups show different distributions and offer additional insights into the distribution of shared lexical traits as some problems are not yet handled (missing data, specific coding errors), a manual analysis should ideally start from here 28 / 32
by checking the purity of cognates sets with respect to the families across which they occur, we can derive rankings of concepts, according to their relative borrowability in our dataset borrowability often thought of as a stable characteristics of concepts, also due to Swadesh’s doctrine of basic vocabulary, but it is clear that concepts evolve with culture, and terms for technical innovations may therefore be highly borrowable, as long as they are new, but they would later not be borrowed again therefore, all statistics on borrowability have to be taken with care, as they also reflect the history of a given region and not necessarily general patterns of language change 29 / 32
a weak (Spearman rank: -0.18, p<0.005) negative correlation between the purity of concepts with respect to potential borrowings and the borrowing statistics of the World Loanword Project (WOLD, Haspelmath and Tadmor 2008) a weak (Spearman rank: 0.19, p<0.005) positive correlation with the WOLD project’s age score the new ranks based on concept purity could be used to expand the limited scope of the WOLD project systematically 29 / 32
apply to more language families (esp. South-American languages) work on inference of more ancient borrowings work on inference of borrowing directions enhance the interactive output 31 / 32
lingulist
shpigford
bluesmoon
speakerdeck
davidbonilla
jponch
samlambert
mclloyd
addyosmani
zakiyama
phodgson
marcelosomers
gr2m
