Slide 1
Slide 1 text
Beyond Cognacy
Challenges of Representing and Analyzing Etymological Data
of South-East Asian Languages
Nathan W. Hill and Johann-Mattis List
Slide 2
Slide 2 text
Background
Slide 3
Slide 3 text
Traditional Approaches to Etymological Data
● traditional etymology looks back on great success story
● huge dictionaries have been published (Pokorny 1959,
Kluge 1883, Mayerhofer 1986–2001, Meyer-Luebke 1911)
● thousands of word histories have been reconstructed
“Chaque mot a son histoire!”
(attr. to Jules Gillieron, 1854-1926)
Slide 4
Slide 4 text
Drawbacks of Traditional Approaches
Etymological dictionaries are:
1. extremely time consuming to produce and to use
2. insufficiently formalized, untransparent, and idiosyncratic
3. difficult if not unrealistic to produce for understudied
languages
Slide 5
Slide 5 text
Quantitative Approaches to Etymological Data
Current approaches to quantitative historical linguistics:
● rely on wordlists of basic vocabulary (Atkinson & Gray 2006)
● show an increase in breadth (more languages)
● show a decrease in depth (fewer words per language)
● usually ignore morphology (important in traditional approaches)
● show an untransparent motivation for cognate judgements
● usually never reach the etymological dictionary level of annotation
Slide 6
Slide 6 text
Challenges for Computers
“the belly” in Chinese and Tibetan:
● Old Chinese *puk
● Written Tibetan (1) grod-pa
● Written Tibetan (2) gsus-pa
● Lhasa Tibetan [tʂʰo¹³-ko ²]
Slide 7
Slide 7 text
Challenges for Computers
“the belly” in Chinese and Tibetan:
● Old Chinese *puk
● Written Tibetan (1) grod-pa
● Written Tibetan (2) gsus-pa
● Lhasa Tibetan [tʂʰo¹³-ko ²]
Slide 8
Slide 8 text
Challenges for Computers
“the belly” in Chinese and Tibetan:
● Old Chinese *puk
● Written Tibetan (1) grod-pa
● Written Tibetan (2) gsus-pa
● Lhasa Tibetan [tʂʰo¹³-ko ²]
Slide 9
Slide 9 text
Challenges for Humans
“There is a severe imbalance of
being data-rich and theory-poor.”
(William S.-Y. Wang, 1996)
● many datasets on South-East Asian languages have been published
(Sidwell 2015, Wang 2004, Huang 1992, etc.)
● large digitized collections have been made available via the STEDT
project (Matisoff 2011)
● but the majority of these data is unprocessed (not further checked
by linguists), lacking etymologies, cognate judgments, phonetic
transcriptions, or concept annotations
Slide 10
Slide 10 text
The Best of Two Worlds?
Can we combine the advantages
of traditional and quantitative
approaches to profit from
computational efficiency and
human insight?
Which challenges do we face when
pursuing integrated frameworks in
South-East Asian languages?
Slide 11
Slide 11 text
Etymological Database of Burmish Languages
Background:
● part of ERC synergy grant 'Beyond Boundaries' (SOAS,
British Museum, British Museum)
Goal:
● creating a classical etymological dictionary, taking full
advantage of computational approaches with an openly
published database online
Slide 12
Slide 12 text
The Burmish Family
classification following Hill 2013
geographic distribution
(Hammaström et al. 2015)
Slide 13
Slide 13 text
Previous Research
Burlig (1967): pioneering and rigorous, but data too sparse
Bradley (1979): inexplicit, not Burmish (Loloish)
Mann (1998): no use of Old Burmese, no relative
chronology of changes, morphemes as cognate sets
Nishi (1999): no reconstruction, very clean organization into
cognate sets, larger dataset than predecessors
Slide 14
Slide 14 text
Challenges
Slide 15
Slide 15 text
Challenges: Suprasegmental Correspondences
Not all sound correspondences occur between sounds which
are in the same prosodic position of a word. Notably
processes like tonogenesis and various patterns of aspiration
and voicing often co-occur with other sound changes. In
these cases, a simple alignment of the words under
consideration is usually not enough, but an analysis of the
patterns of sound correspondences needs to be carried out.
Slide 16
Slide 16 text
Challenges: Suprasegmental Correspondences (1)
Burmish tonal split in checked syllables:
Aspirate initials corresponding to Lashi tone [55]
Wbur. khwak 'bowl', Lashi khuʔ
OBur. khlup 'sew', Lashi khju:p
WBur. khrok 'six', Lashi khjuk
Slide 17
Slide 17 text
Challenges: Suprasegmental Correspondences (1)
Burmish tonal split in checked syllables:
Non-aspirated initials corresponding to Lashi [31]
WBur. kok 'paddy rice', Lashi kuk³¹
WBur. klyap 'ten cents; kyat', Lashi kjɔ³¹
WBur. krok 'be afraid', Lashi kju:k³¹
Slide 18
Slide 18 text
Challenges: Suprasegmental Correspondences (2)
Burling's (1967) law (loss of preglottalization consonants):
Examples to be reconstructed *C-:
OBur. -khuiwḥ < *kuiwḥ 'smoke', Lashi -khouH
OBur. khlyap < *klap 'flat object', Lashi khjapH
Slide 19
Slide 19 text
Challenges: Suprasegmental Correspondences (2)
Burling's (1967) law (loss of preglottalization consonants):
Examples to be reconstructed *ˀC-:
OBur. khat < *ˀkat 'put in (to); pack', Lashi kạ:tH
OBur. khraṅ < *ˀkraŋ 'mosquito', Lashi kjạŋ
Slide 20
Slide 20 text
Challenges: Partial Cognates
‘Cognacy is not a binary relation which is either present or not. Instead, we
can distinguish different subtypes of cognacy, just as biologists can identify
specific types of homology between genes.’ (List 2016: 133)
partial cognacy in Chinese dialects (List et al. 2016)
Slide 21
Slide 21 text
Challenges: Partial Cognates
Binarisation of Partial Cognate Relations:
● strict (only fully identical words are considered cognate)
● loose (words sharing a cognate morpheme are cognate)
Problems of Binarisation:
● not realistic with respect to lexical change
● over- or underestimates the amount of shared cognates
Slide 22
Slide 22 text
Challenges: Partial Cognates
Under-Estimation in Strict Encoding: “yesterday”
● Bola [a³¹ŋji³ nɛʔ³¹]
● Lashi [a³¹ŋjei nap³¹]
● Rangoon [mɑ ³ne ³kɑ ³]
● Xiandao [n³¹m
̥ an³ ]
Slide 23
Slide 23 text
Challenges: Partial Cognates
Under-Estimation in Strict Encoding: “yesterday”
● Bola [a³¹ŋji³ nɛʔ³¹]
● Lashi [a³¹ŋjei nap³¹]
● Rangoon [mɑ ³ne ³kɑ ³]
● Xiandao [n³¹m
̥ an³ ]
Slide 24
Slide 24 text
Challenges: Partial Cognates
Over-Estimation in Loose Encoding: “the leaf”
● Bola [sak faʔ ]
● Lashi [a³¹fu
̱ ʔ ]
● Rangoon [ɑ ³jwɛʔ ]
● Xiandao [a³¹xʐoʔ ]
Slide 25
Slide 25 text
Challenges: Partial Cognates
Over-Estimation in Loose Encoding: “the leaf”
● Bola [sak faʔ ]
● Lashi [a³¹fu
̱ ʔ ]
● Rangoon [ɑ ³jwɛʔ ]
● Xiandao [a³¹xʐoʔ ]
Slide 26
Slide 26 text
Challenges: Language-Internal Cognates
When using alignments to derive statistics from sound
correspondences, dependencies inside a language need to
be taken into account to avoid an overscoring of regularities.
Language-internal cognates are invaluable evidence in
classical cognate judgments and reconstruction. Current
computational approaches ignore them completely.
Slide 27
Slide 27 text
Challenges: Language-Internal Cognates
prefix a- in Old Burmese
● “the branch” a khak
● “the mother” a miy
● “the flower” a po₁ṅʔ
● “the feather” a muyḥ
● “the father” a phiy
● “the leaf” a ro₁k
“the dog” in Atsi
● “the wolf” [vam ¹kʰui²¹mo ]
● “the dog” [kʰui²¹]
● “the fox” [tan kʰui²¹]
Slide 28
Slide 28 text
Preliminary Solutions
Slide 29
Slide 29 text
Algorithms
LingPy (Python library for historical linguistics)
● automatic cognate detection (2014)
● automatic detection of partial cognates (List et al. 2016)
● multiple phonetic alignment (2014)
Slide 30
Slide 30 text
Tools
EDICTOR (Etymological Dictionary Editor)
● data editing (X-Sampa input, automatic consistency tests)
● data annotation (cognates, partial cognates, alignments,
morpheme tagging)
● data inspection (frequency analysis, structural analysis)
Slide 31
Slide 31 text
Tools
Slide 32
Slide 32 text
Materials
● data taken from Huang (1992)
● currently 8 Burmish varieties
● 248 concepts selected (basic vocabulary, and
etymologically important words)
● partial cognates were automatically inferred and then
manually corrected
● alignments were automatically computed (will be
manually corrected)
Slide 33
Slide 33 text
Partial Cognates: Annotation with the EDICTOR
Slide 34
Slide 34 text
Partial Cognates: Annotation with the EDICTOR
Slide 35
Slide 35 text
Partial Cognates: Annotation with the EDICTOR
Slide 36
Slide 36 text
Partial Cognates: Annotation with the EDICTOR
Slide 37
Slide 37 text
Partial Cognates: Annotation with the EDICTOR
Slide 38
Slide 38 text
Partial Cognates: Annotation with the EDICTOR
Slide 39
Slide 39 text
Partial Cognates and Strict vs. Loose Coding
Slide 40
Slide 40 text
Partial Cognates: Strict versus Loose Coding
Slide 41
Slide 41 text
Partial Cognates: Strict versus Loose Coding
Slide 42
Slide 42 text
Sound Correspondences: Searching for Patterns
patterns of initial sounds in 337 aligned partial cognate sets:
Slide 43
Slide 43 text
Sound Correspondences: Searching for Patterns
patterns of initial sounds in aligned partial cognate sets:
Slide 44
Slide 44 text
Sound Correspondences: Searching for Patterns
patterns of initial sounds in aligned partial cognate sets:
Slide 45
Slide 45 text
Sound Correspondences: Searching for Patterns
patterns of initial sounds in aligned partial cognate sets:
Slide 46
Slide 46 text
Sound Correspondences: Searching for Patterns
patterns of initial sounds in aligned partial cognate sets:
Slide 47
Slide 47 text
Sound Correspondences: Searching for Patterns
What shall we do with morpheme alignments?
● if the Neogrammarians are right,
○ a given proto-form in a given context should always
yield the same reflex in a given descendant language
● this means,
○ compatible patterns in aligned cognate sets will hint to
specific proto-sounds or proto-sounds in specific
contexts
Slide 48
Slide 48 text
Sound Correspondences: Searching for Patterns
What is compatibility?
● take all alignments for a given dataset, and select one
common sound position (e.g., initial of each morpheme)
● when plotting for each language in our sample, which
sound occurs in a given cognate set in the position, we
can make a first step to compare these patterns
Slide 49
Slide 49 text
Sound Correspondences: Searching for Patterns
What is compatibility?
compatible
Cognate set L1 L2 L3 L4 L5 L6 L7 L8
morpheus-1 p p p Ø f f Ø p
morpheus-2 p Ø p p Ø f p p
Slide 50
Slide 50 text
Sound Correspondences: Searching for Patterns
What is compatibility?
compatible
Cognate set L1 L2 L3 L4 L5 L6 L7 L8
morpheus-1 p p p Ø f f Ø p
morpheus-3 Ø p p p f f p p
Slide 51
Slide 51 text
Sound Correspondences: Searching for Patterns
What is compatibility?
NOT COMPATIBLE
Cognate set L1 L2 L3 L4 L5 L6 L7 L8
morpheus-1 p p p Ø f f Ø p
morpheus-4 Ø p f p f p p p
Slide 52
Slide 52 text
Sound Correspondences: Searching for Patterns
What is compatibility?
● compatibility of two identical positions in different
alignments is a necessary requirement to assume that the
two alignments represent a common proto-sound in a
common proto-context
● it is not sufficient, as we have to deal with missing data,
which may sufficiently blur the picture
Slide 53
Slide 53 text
Sound Correspondences: Searching for Patterns
Building a compatibility network of aligned cognate sets:
● take the same position (e.g., initial consonant) in all
alignments (called a “site” of the alignment)
● make a network in which the alignment sites are nodes
● edges in the network are drawn between two nodes if
these are compatible with each other
● weights between the edges are determined by counting
the positions without a gap in both alignment sites
Slide 54
Slide 54 text
Sound Correspondences: Searching for Patterns
Slide 55
Slide 55 text
Sound Correspondences: Searching for Patterns
Slide 56
Slide 56 text
Sound Correspondences: Searching for Patterns
Search for maximal cliques to increase cluster compatibility:
● A clique in a network is a group of nodes which are all connected with
each other.
● Cliques of compatible alignment sites represent the strongest
evidence for a coherent group of regular correspondences pointing to
the same proto-sound in a given context.
● We use a simple method to search for non-overlapping cliques by
maximizing their size, which is done in an iterative manner.
Slide 57
Slide 57 text
Sound Correspondences: Searching for Patterns
Slide 58
Slide 58 text
Sound Correspondences: Searching for Patterns
Search for maximal cliques to increase cluster compatibility:
● What looks chaotic is less scary, if we look at the patterns!
● Of 638 cognate sets:
○ 337 occur in at least 3 taxa
○ 317 start with an initial consonant
○ 234 could be assigned to 35 transitive groups of minimal size 2
○ 74% (234 / 337) of the cognate sets can be seen as “regular”. The
remaining cognate sets will be checked and either corrected or
their incompatibility will be explained.
Slide 59
Slide 59 text
Sound Correspondences: Searching for Patterns
Slide 60
Slide 60 text
Sound Correspondences: Searching for Patterns
Slide 61
Slide 61 text
Sound Correspondences: Searching for Patterns
Slide 62
Slide 62 text
Sound Correspondences: Searching for Patterns
Slide 63
Slide 63 text
Sound Correspondences: Searching for Patterns
Slide 64
Slide 64 text
Sound Correspondences: Searching for Patterns
Slide 65
Slide 65 text
Sound Correspondences: Searching for Patterns
Proto-Burmish *k-
Slide 66
Slide 66 text
Sound Correspondences: Searching for Patterns
Proto-Burmish *ˀk-
Proto-Burmish *kh-
Slide 67
Slide 67 text
Sound Correspondences: Searching for Patterns
Be careful with the interpretation of compatibility networks:
● remember, one clique in our network does not necessarily correspond
to one proto-sound in the proto-language (maybe, our alignments are
wrong, our cognates are wrong, the words are borrowed…)
● but: if a proto-sound in a certain number of identical contexts has
derived regularly from proto-language to descendant languages, it
should form a clique in our data!
● compatibility networks of prosodically similar alignment sites are just a
first step towards computer-assisted language reconstruction
Slide 68
Slide 68 text
Language-Internal Cognates: Improving Annotation
Atsi
● “the moon” [lo
̱ mo ]
● “the tiger” [lo²¹mo ]
● “the wolf” [vam ¹kʰui²¹mo ]
● “the dog” [kʰui²¹]
● “the fox” [tan kʰui²¹]
Bola
● “the wolf” [mjaŋ kʰui³ ]
● “the dog” [kʰui³ ]
● “the thunder” [mau³¹mjaŋ ]
● “the sky” [mau³¹kʰauŋ ]
Slide 69
Slide 69 text
Language-Internal Cognates: Improving Annotation
Language Concept Word Motivation
Atsi the moon lo
̱ mo moon m-suffix
Atsi the tiger lo²¹mo tiger m-suffix
Atsi the wolf vam ¹kʰui²¹mo bear dog m-suffix
Atsi the dog kʰui²¹ dog
Atsi the fox tan kʰui²¹ fox(?) dog
Slide 70
Slide 70 text
Language-Internal Cognates: Improving Annotation
Lexeme motivation annotation in the EDICTOR:
● simplified schema, all glosses are allowed, as long
they do not contain white-space
● question marks can be used to express doubt
● identically annotated morphemes can be inspected
(they indicated language-internal cognates)
● partial and full colexifications can be used to
assist the morphological analysis
● data can be visualized with help of partial
colexification graphs
Slide 71
Slide 71 text
Language-Internal Cognates: Improving Annotation
“black” , “dark”, and “the mouse or rat” in Bola
Slide 72
Slide 72 text
Language-Internal Cognates: Improving Annotation
“black” , “dark”, “early”, and “the mouse or rat” in Atsi
Slide 73
Slide 73 text
Language-Internal Cognates: Improving Annotation
“bad”, “good”, and “the man” in Bola
Slide 74
Slide 74 text
Language-Internal Cognates: Improving Annotation
“good”, “the man”, and “the son” in Atsi
Slide 75
Slide 75 text
Language-Internal Cognates: Improving Annotation
automatically computed
partial colexification graph
for “noʔ²¹” in Atsi
Slide 76
Slide 76 text
Language-Internal Cognates: Improving Annotation
partial colexification graph
for manually annotated
cluster of “tree” in Bola
Slide 77
Slide 77 text
Availability
● Our GitHub (listing issues, plans, and software)
○ https://github.com/digling/burmish
● Our database in the EDICTOR (current state,
constantly evolving):
○ https://dighl.github.io/burmish/
● Results of our Alignment Site Analysis:
○ https://dighl.github.io/burmish/plot-corrs.html
● For more questions, email us:
○ [email protected]
Slide 78
Slide 78 text
Outlook
Slide 79
Slide 79 text
Outlook
Chances are there!
● working with partial cognates increases the realism of our analyses
● working with alignments opens new horizons for computer-assisted
consistency analysis
● annotation of language-internal cognacy opens exciting research
avenues for the investigation of semantic change, lexical typology,
and language relationship
Slide 80
Slide 80 text
Outlook
Challenges remain!
● How can we get from compatibility clusters in our alignments to first
reconstructions?
● Can we use compatibility to test the consistency of reconstruction
systems?
● How can we formalize the assignment of cross-semantic cognates in
our wordlists?
Slide 81
Slide 81 text
Thanks for Your Attention!
Thanks to:
● the Équipe AIRE (UPMC, Paris) for inspiration and help with network analyses
● Guillaume Jacques and Laurent Sagart for helpful feedback on EDICTOR tool
and active help in creating the first concept list
● Doug Cooper for helping out with parts of the data that we used
Slide 82
Slide 82 text
References
Bradley, David (1979). Proto-Loloish. London: Curzon Press.
Burling, Robbins (1967). Proto-Lolo-Burmese. Bloomington: Indiana University.
Hill, Nathan W. (2013) 'The merger of Proto-Burmish *ts and *č in Burmese.'
SOAS Working Papers in Linguistics 16: 334-345.
Kluge, Friedrich (1883). Etymologische Wörterbuch der deutschen Sprache.
Strassburg: K. J. Trübner.
List, Johann-Mattis (2014): Sequence comparison in historical linguistics.
Düsseldorf: Düsseldorf University Press.
Slide 83
Slide 83 text
References
List, Johann-Mattis, Philippe Lopez, and Eric Bapteste (2016): Using sequence
similarity networks to identify partial cognates in multilingual wordlists.
Proceedings of the Annual Meeting of the ACL. Berlin: Association of
Computational Linguistics. 599–605.
Mann, Noel Walter (1998). A phonological reconstruction of Proto Northern
Burmic. Unpublished thesis. Arlington: The University of Texas.
Mayerhofer, Manfred (1986-2001). Etymologisches Wörterbuch des
Altindoarischen. Heidelberg: Carl Winter.
Slide 84
Slide 84 text
References
Meyer-Luebke, Wilhelm (1911). Romanisches etymologisches Wörterbuch.
Heidelberg: Winter.
Nishi, Yoshio (1999). Four Papers on Burmese: Toward the history of Burmese
(the Myanmar language). Tokyo: Institute for the study of languages and cultures
of Asia and Africa, Tokyo University of Foreign Studies.
Pokorny, Julius (1959). Indogermanisches etymologisches Wörterbuch. Bern and
Münich: Francke.