Modelling language evolution

1
Modelling language evolution

• Tandy Warnow
• The University of Texas at Austin

2
Species phylogeny
From the Tree of the Life Website,University of
Arizona
Orangutan
Human
Gorilla
Chimpanzee
3
Possible Indo-European tree(Ringe, Warnow and
Taylor 2000)
4
Controversies for Indo-European history

• Subgrouping Other than the 10 major subgroups,
  what is likely to be true? In particular, what
  about
• Italo-Celtic,
• Greco-Armenian,
• Anatolian Tocharian,
• Satem Core?

5
This talk

• Empirical evidence of how estimated phylogenies
  depend upon both the data and the method - and
  can be wrong
• Models of language evolution (from the earliest
  ones to more recent ones), why we need them, and
  what we still need to do. Note simulations and
  estimation methods both depend upon model
  assumptions!
• Results of simulation studies based upon some new
  models
• Comments

6
Nakhleh et al., Transactions of the Philological
Society 2005

• Methods studied UPGMA (lexico-statistics),
  Neighbor joining, maximum parsimony, maximum
  compatibility, weighted MP, weighted MC, and
  GrayAtkinson.
• Datasets Four versions of the RingeTaylor IE
  data (lexical, morphological, and phonological
  characters) lexical only vs. all, screened vs.
  unscreened
• Observations
• UPGMA (lexico-statistics) does the worst - it
  splits known subgroups.
• Other than UPGMA, all methods reconstruct the ten
  major subgroups, Anatolian Tocharian, and
  Greco-Armenian. Nothing else is consistently
  reconstructed.
• When using lexical data only, all methods group
  Italic, Celtic, and Germanic together.
• Some methods (not all) will reconstruct different
  trees on different datasets. Screening datasets
  to remove obvious homoplasy can result in better
  (?) trees.

7
Question how to determine which phylogenies are
reliable?

• Data need high quality data!
• Phylogenetic reconstruction methods need to be
  tested before being trusted! Examples of
  possible tests
• Benchmark real datasets (need good benchmarks!
  Are there any?)
• Simulated datasets (need good models!)

8
Simulation study (cartoon)
9
Simulation study (cartoon)
FN
FN false negative (missing edge) FP false
positive (incorrect edge) 50 error rate
FP
10
Modelling language evolution

• Models of evolution allow reconstruction methods
  to be evaluated in simulation. This allows us to
  understand the conditions under which each method
  will perform well.
• Models of evolution (for simulation purposes)
  need to reflect good scholarship, and should be
  able to reproduce the properties of real data.
• Models of evolution are also present in
  estimation methods, whether explicitly (as in ML
  or Bayesian) or implicitly.

11
Issues in modelling language evolution

• Character evolution model.
• Variation between characters.
• Cladogenesis model tree vs. network vs. dialect
  continuum?

12
Modelling the evolution of single linguistic
characters

• Types of linguistic characters
• Phonological (sound changes)
• Lexical (meanings based on a wordlist)
• Morphological (especially inflectional)
• Modelling issues state space, lexical clock,
  homoplasy, and polymorphism
• Easy lexical clock not believed, and most
  linguistic characters have infinite number of
  possible states.
• More interesting homoplasy, polymorphism, and
  variation between characters.

13
Homoplasy-free evolution

• When a character changes state, it changes to a
  new state not in the tree
• In other words, there is no homoplasy (character
  reversal or parallel evolution)
• First inferred for weird innovations in
  phonological characters and morphological
  characters in the 19th century.

0
0
1
0
0
0
0
1
1
14
Lexical characters can also evolve without
homoplasy

• For every cognate class, the nodes of the tree in
  that class should form a connected subset - as
  long as there is no undetected borrowing nor
  parallel semantic shift.
• However, our research suggests that 15 of
  lexical characters evolve homoplastically.

1
1
1
0
0
0
1
1
2
15
Polymorphism

• Polymorphism means two or more states exhibited
  by the same language for a character.
• Most common examples are lexical two or more
  words for the same basic meaning. Examples
  big/large, little/small, rock/stone.
• Lexical polymorphism results primarily from
  semantic shift, but polymorphism due to borrowing
  also occurs.
• Incidence lexical polymorphism is very common
  but transient (almost all polymorphisms lost
  within a millenium). Less frequent for other
  types of characters.

16
Modelling variation between characters
Rates-across-sites

• If a site (i.e., character) is twice as fast as
  another on one edge, it is twice as fast
  everywhere.

B
D
A
C
B
D
A
C
17
Modelling variation between characters The
no-common-mechanism model

• In this model, there is a separate random
  variable for every combination of site and edge -
  the underlying tree is fixed, but otherwise there
  are no constraints on variation between sites.

C
A
D
B
B
D
A
C
18
Homoplasy-free models without polymorphism

• The earliest models were all tree models,
  homoplasy-free and obeyed the lexical clock.
• Ringe-Warnow PP (perfect phylogeny - i.e.,
  homoplasy-free, no common mechanism,
  non-parametric tree model)

19
Cladogenesis

• The speciation model ranges from trees all the
  way to dialect continuums. Intermediate models
  include horizontal transfer (borrowing) and
  hybridization (creoles).

20
Modelling borrowing Networks and Trees within
Networks

21
Perfect Phylogenetic Network model

• Nakhleh et al. Perfect Phylogenetic Network (PPN)
  model all characters evolve without homoplasy
  down a tree contained within the network.
  Published in Language, 2005.
• Warnow-Evans-Ringe-Nakhleh (2004) extends PPN
  model to allow for limited and identifiable
  homoplasy.

22
Perfect Phylogenetic Network for IENakhleh et
al., Language 2005
23
What about polymorphism?

• Our first model of polymorphism (Bonet et al.,
  1996) was a non-parametric model for
  homoplasy-free characters, no-common-mechanism
  model, with polymorphism due to semantic shift.
• Three problems (1) because it is non-parametric,
  it cannot be used for simulation (2) homoplasy is
  fairly frequent for lexical characters (15 of
  characters) (3) what about polymorphism due to
  borrowing?

24
Nichols and Gray model for polymorphism

• Geoff Nichols and Russel Gray (2006)
  Homoplasy-free, rates-across-sites, parametric
  model in which the character adds and loses
  states under a stochastic process. The number of
  states in a lineage can go up and down (including
  down to 0 and then back up).
• Problems (1) homoplasy is frequent in lexical
  characters (2) what is the linguistic process?

25
What needs to be done in modelling

• We need parametric models of character evolution
  that include reasonable levels of homoplasy, in
  which polymorphism arises due to semantic shift
  (conflation of two characters), by borrowing, or
  due to other linguistic processes.
• We also need cladogenesis models that incorporate
  population-level processes, and can represent
  dialect continuums.

26
Simulation study (Barbancon et al.)

• Simulated evolution down networks with 30 leaves,
  three contact edges, and with moderate levels of
  homoplasy and borrowing for 300 lexical
  characters and 60 morphological characters.
• Compared trees constructed by various methods to
  the genetic tree contained in the network, for
  topological accuracy.
• Methods compared NJ, UPGMA, weighted and
  unweighted MP and MC.

27
Standard Model Conditions

• Screened dataset
• Lexical characters 4 homoplastic, 10 evolve
  with borrowing
• Morphological characters no homoplasy nor
  borrowing
• Unscreened dataset
• Lexical characters 20 homoplastic, 20 borrowed
• Morphological characters 5 homoplastic, no
  borrowing
• Molecular clock for the cladogenesis model
• No-common-mechanism model with moderate variation
  between characters
• Lexical weight1, morphological weight50

28
Simulation study (cartoon)
FN
FN false negative (missing edge) FP false
positive (incorrect edge) 50 error rate
FP
29
Clocklike data
30
Clocklike data
31
Points

• Screening the data helps to improve the
  phylogenetic accuracy of most methods.
• When data are generated under network models,
  methods which reconstruct trees do not perform
  well.
• Modelling helps us predict the conditions under
  which different methods will perform well, or
  poorly. The more accurate the models, the more
  relevant the predictions. We need better models!

32
Future research

• Testing other methods in simulation (including
  some network construction methods)
• Formulating improved (more realistic) models of
  language evolution
• Implementing simulation tools under these
  improved models
• Developing estimation methods under these
  improved models
• Reanalyzing IE, and looking at some new families
  (or subfamilies)

33
Acknowledgements

• Funding NSF, the David and Lucile Packard
  Foundation, the Radcliffe Institute for Advanced
  Studies, The Program for Evolutionary Dynamics at
  Harvard, and the Institute for Cellular and
  Molecular Biology at UT-Austin.
• Collaborators Don Ringe, Steve Evans, Luay
  Nakhleh, and Francois Barbancon.

34
For more information

• Please see the Computational Phylogenetics for
  Historical Linguistics web site for papers, data,
  and additional material http//www.cs.rice.edu/na
  khleh/CPHL

35
Differences between characters

• Lexical most easily borrowed (most borrowings
  detectable), and homoplasy relatively frequent
  (we estimate about 25-30 overall for our
  wordlist, but a much smaller percentage for
  basic vocabulary). Also, lexical characters have
  a high incidence (80) of transient polymorphism.
• Phonological can still be borrowed but much less
  likely than lexical. Complex phonological
  characters are infrequently (if ever)
  homoplastic, although simple phonological
  characters very often homoplastic.
• Morphological least easily borrowed, least
  likely to be homoplastic. Rarely polymorphic.