Post on 20-Aug-2020
transcript
Semantic Maps: An Attempt at Catalogization and Comparison
Natalia Levshina F.R.S. – FNRS Université catholique de Louvain
Outline 1. Semantic maps: a typology 2. Semantic maps with R: a case study of European analytic
causatives
Semantic maps • a popular tool in constructional (lexical) and semantic
typology
Semantic maps • a popular tool in constructional (lexical) and semantic
typology • represent polyfunctionality of linguistic expressions
Semantic maps • a popular tool in constructional (lexical) and semantic
typology • represent polyfunctionality of linguistic expressions • a convenient tertium comparationis of form-function
mappings across languages
Haspelmath 2003: Datives
Haspelmath 2003: Datives
Main types of semantic maps
Connectivity (links between objects)
Proximity (distances between
objects)
Type-based
(pre-determined functions, frequencies do NOT count)
Manual/Automatic
Automatic
Token-based
(corpus exemplars, frequencies COUNT)
?
Automatic
Types of semantic maps
Connectivity (links between objects)
Proximity (distances between
objects)
Type-based
(pre-determined functions, frequencies do NOT count)
Manual/Automatic
Automatic
Token-based
(corpus exemplars, frequencies COUNT)
?
Automatic
Traditional maps • Non-hierarchical approaches - non-directional - directional • Hierarchical approaches
Traditional maps • Non-hierarchical approaches - non-directional - directional • Hierarchical approaches
Haspelmath 2003: Datives
Main principles • Nodes: A function is put on a map when there’s at least one
pair of languages which differ wrt. this function (Haspelmath 2003)
Main principles • Nodes: A function is put on a map when there’s at least one
pair of languages which differ wrt. this function (Haspelmath 2003)
• Links: the principle of connectivity (adjacency/contiguity): if a construction has more than one function, they should be connected (see van der Auwera 2013)
Example 1
Function A
Function B
Function C
Construction X
Example 2
Function A
Function B
Function C
Construction Y
Example 3
Function A
Function B
Function C
Construction Z
Example 4
Function A
Function B
Function C
Wrong: the connectivity principle is not observed!
Construction Q
A fix
Function A
Function B
Function C
Construction Q
Traditional maps • Non-hierarchical approaches - non-directional - directional • Hierarchical approaches
Narrog 2010: Goal-Recipient Domain
Traditional maps • Non-hierarchical approaches - non-directional - directional • Hierarchical approaches
Rakhilina et al. (In press) EMPTINESS
Serbian
Types of semantic maps
Connectivity (links between objects)
Proximity (distances between
objects)
Type-based
(pre-determined functions, frequencies do NOT count)
Manual/Automatic
Automatic
Token-based
(corpus exemplars, frequencies COUNT)
?
Automatic
CLICS (List et al. 2014): HAND
http://clics.lingpy.org/
Types of semantic maps
Connectivity (links between objects)
Proximity (distances between
objects)
Type-based
(pre-determined functions, frequencies do NOT count)
Manual/Automatic
Automatic
Token-based
(corpus exemplars, frequencies COUNT)
?
Automatic
Majid et al. 2008: cutting and breaking
Types of semantic maps
Connectivity (links between objects)
Proximity (distances between
objects)
Type-based
(pre-determined functions, frequencies do NOT count)
Manual/Automatic
Automatic
Token-based
(corpus exemplars, frequencies COUNT)
?
Automatic
Wälchli & Cysouw (2012): motion verbs
SEMANTIC MAPS WITH R
Analytic Causatives: Examples • English: make + Vinf, let + Vinf, have + Vinf, cause + to-Vinf • German: lassen + Vinf • Dutch: laten + Vinf, doen + Vinf • Russian: zastavljat’ “force” + Vinf, davat’ “give” + Vinf • French: faire + Vinf, laisser + Vinf • Romanian: face + să + Vsubj, lasă + să + Vsubj
DISTANCE- & TOKEN-BASED MAP
Multilingual corpus of film subtitles
An example of .srt format … 646 00:51:27,880 --> 00:51:32,920 <i>For always evil will look to find a foothold in this world.</i> 647 00:51:39,440 --> 00:51:42,603 Not good. Not good at all. 648 00:51:50,040 --> 00:51:51,326 Eww.
649 00:52:06,760 --> 00:52:09,081 Oh, no. Sebastian. 650 00:52:12,800 --> 00:52:13,847 Good gracious. 651 00:52:34,720 --> 00:52:35,767 Come on. …
Data set • All instances of ACs in 12 languages • 302 multilingual contexts, in which at least one doculect
contains an AC • Alignment: Jörg Tiedemann’s software subalign
Constructional types • ACs with various auxiliaries (e.g. lassen + V, fazer + V) • Transitives and ditransitives • Causative verb + Clause (e.g. dejar + que) • Causal prepositions (because of) • Causal and resultative subordinate clauses • Resultative cxs, e.g. make + Adj • Modals • Particles (niech in Polish) • Reflexives • Insubordination (que + Subj) • …
Matrix
Gower Distances cluster::daisy Situation (row) A EN: And we make them do it… ...or we kill them. make_V IT: E glielo facciamo fare ... o lo uccidiamo. fare_V CZ: Donutíme je to udělat, nebo je zabijeme. donutit_V Situation (row) B EN: Pick up someone my height and build and make them believe it is me. make_V IT: Individua una della mia corporatura e fa credere loro che sia io. fare_V CZ: Vyber někoho, kdo je mi podobný a přesvědč je, že jsem to já. Trans
Distance (A, B) = 1 – 2/3 ≈ 0.33
Multidimensional Scaling • Multidimensional Scaling of the distance matrix (isoMDS()
in MASS package) • 5 dimensions (stress < 0.15) • Visualization with bubble charts (googleVis) to interpret
the semantic functions • Other techniques to interpret and compare form-meaning
mapping…
Intepretation • The main principle: the closer two points on the map, the
more overlapping constructions they share across the languages. From the isomorphism principle it follows that the corresponding situations are more semantically similar (on average), since more authors of the doculects chose identical constructions to represent these causative situations.
faire + V: points
Kriging in fields::Krig()
lassen + V: points
Kriging
TYPE- & LINK-BASED MAPS
Matrix
Selecting Links • the total number of links between all 6 nodes is 15 (these
functions co-occur in at least one construction in the data set)
Selecting Links • the total number of links between all 6 nodes is 15 (these
functions co-occur in at least one construction in the data set) • However, not all of them are necessary because traditional
semantic maps allow for indirectly connected functions
Selecting Links • the total number of links between all 6 nodes is 15 (these
functions co-occur in at least one construction in the data set) • However, not all of them are necessary because traditional
semantic maps allow for indirectly connected functions • A solution: semmap() function in Rling (available upon
request) to induce a semantic map which observes the principle of connectivity
Selecting Links • the total number of links between all 6 nodes is 15 (these
functions co-occur in at least one construction in the data set) • However, not all of them are necessary because traditional
semantic maps allow for indirectly connected functions • A solution: semmap() function in Rling (available upon
request) to induce a semantic map which observes the principle of connectivity
• igraph package for graph building and visualization
Selecting Links • the total number of links between all 6 nodes is 15 (these
functions co-occur in at least one construction in the data set) • However, not all of them are necessary because traditional
semantic maps allow for indirectly connected functions • A solution: semmap() function in Rling (available upon
request) to induce a semantic map which observes the principle of connectivity
• igraph package for graph building and visualization • several solutions may be possible, even with the minimum
number of links (7) , use random sampling to check
A possible semantic map
Frequency of ‘co-constructification’
faire + V
lassen + V
HIERARCHICAL TREES
Tree
make + V and let + V
DIACHRONIC LINKS
Links (Narrog 2010) Edge sequence:
[1] Goal -> Beneficiary
[2] Goal -> Recipient
[3] Goal -> Purpose
[4] Recipient -> Patient
[5] Beneficiary -> Purpose
[6] Recipient -> Experiencer
[7] Recipient -> Possessor
[8] Purpose -> Cause
[9] Goal -> Patient
[10] Location -> Recipient
[11] Beneficiary -> Recipient
[12] Goal -> Location
Diachronic links
Narrog 2010: Goal-Recipient Domain
Narrog 2010: Goal-Recipient Domain
Recipient –Purpose link is not mentioned anywhere in the
data
Narrog 2010: Goal-Recipient Domain
Two missing links from Recipient to Experiencer and
Possessor (listed as uncontroversial links)
Let machines do the job!
Related tools • Componential Analysis: emic grids (cf. Evans 2010) • Cognitive Semantics: radial networks of polysemy • Distributional semantics: Semantic vector space visualizations
(Heylen, Wielfaert) • Variationist MDS or MCA exemplar-based maps (Levshina
2011; Levshina et al. 2013)