Simulating the Shift Towards Semantic Gender in Dutch A Multi-Agent Language Game Approach Promotor: Prof. Dr. Katrien Beuls
Roxana Rădulescu
Advisor: Dr. Remi van Trijp
Introduction – Gender in Dutch
Methods – Language Game Description
• Nominal
Anaphoric Reference Language Game
gender: common (‘de’), neuter (‘het’) • Pronominal gender: masculine (‘hij’), feminine (‘zij’), neuter (‘het’)
• Population
turnover: turnover fraction, transmission rate • World: Dutch nouns distributed over the semantic space: abstract – concrete, countable – uncountable, animate – inanimate, ontological category (none – animal – collective) • Construction based vocabulary: V = {Ck |1 ≤ k ≤ n} • Gender mapping: random, best representative, history
(1) De hond achtervolgde de kat, maar The-SG.C dog-SG.C chased the-SG.C cat-SG.C, but hij kon haar niet vangen. he-SG.M could her-SG.F not catch. ‘The dog chased the cat, but he could not catch her.’
• Construction similarity, with X and Y binary feature vectors of Cx, Cy ∈ V :
X ·Y = sim(Cx, Cy ) = ||X|| · ||Y ||
• Syntactic
strategy – lexical gender of the noun (old opaque strategy) • Semantic strategy – map gender onto the semantic properties of the noun (new transparent strategy)
• Competitor
n P X i × Yi i=1 v v u u u n u n u P u P 2 u u (X ) × (Yi)2 t t i i=1 i=1
selection: random, highest score, probability (τ ) P robability(Ck ) =
score(Ck ) exp( τ ) score(Ci) n exp( τ ) i=1 P
• Scoring
Results
Conclusion masculine
0.9 0.7 0.6 0.5 0.4
0
500
1000
1500
2000
2500
3000
3500
4000
Games/agent Highest score Probability (0.1) Probability (0.3)
masculine
neuter
agents with the necessary cognitive abilities for dealing with gender knowledge loss and mapping the system to the semantic space
80
• Formed
60
clusters lack in quality and mapping does not correspond to real world observed results
40
• Future
feminine
Best representative
'milk'
A-C-I 'day'
History 50
neuter
A-U-I
'help'
History 100
Random
References
Figure 3: Gender distribution for mapping strategies
C-C-A-A 'dog'
Semantic space division by gender - probability (0.1)
Figure 2: Mapping gender onto the semantic space
Audring, J. (2006). Pronominal gender in spoken Dutch. Journal of Germanic Linguistics, 18(02):85–116. De Vos, L. (2013). On variation in gender agreement: The neutralization of pronominal gender in Dutch. Synchrony and Diachrony: A dynamic interface, 133:237.
brief 100 Usage percentage
'letter'
C-U-I
History 5
Lexicon division by gender
80 60 40
Contact Information
20 0
C-C-I
work: extend semantic space, introduce salient semantic dimensions, individuation hierarchy, impose constraints to improve local clustering for each agent
20 0
Probability (0.5) Probability (1.0) Random
Tangibility (C/A) - Countability (C/U) - Animacy (A/I) - Ontological category (-/A/C) Composition percentage
language game to simulate the shift from a syntactic gender agreement system to a semantic one
• Equip
Figure 1: Solving competition
18 16 14 12 10 8 6 4 2 0
feminine
• Stochastic
100 Composition percentage
Alignment success
1 0.8
strategies: lateral inhibition, success rate, frequency
0
500
1000
1500
2500
3000
3500
Games/agent
A-C-I-C 'army'
2000
masculine
feminine
neuter
Figure 4: Gender competition at word level
unknown
4000
• Web:
http://ai.vub.ac.be/members/roxana-radulescu • Email:
[email protected]