| dc.description.abstract | In this research we wanted to investigate whether the effects observed during simultaneous
processing of semantic and syntactic lexical ambiguity could be interpreted within the framework
of error-driven learning. To investigate the interaction of the two types of ambiguity we focused
on polysemy in a highly inflected Serbian noun system and attempted to simulate their processing
effects using naive discriminative learning (Baayen et al., 2011).
Polysemy is the type of lexical ambiguity where one word can have multiple related
senses. For example, isolated word paper could refer to the writing paper, i.e., paper as the
material, but also to scientific paper, and even a daily paper. Polysemous words are a complex
phenomenon, whose processing is affected by number of senses, probability distribution of those
senses, and degree of relatedness among the senses (Filipović Đurđević & Kostić, 2009; 2017,
under revision; Klepousniotou, 2002; Rodd et al., 2002). Additionally, in Serbian, words can take
up to seven inflected forms. Syntactic ambiguity of an isolated inflected form is reflected in the
multitude of syntactic roles the given inflected form can take in the sentence. For example,
inflected masculine noun konja (horse) can indicate the subject in the sentence (Dva konja su
trčala / Two horses were running), but also the object (Jahao sam konja / I rode the horse). It has
been demonstrated that different aspects of syntactic ambiguity affect lexical processing as well:
information load based on relative frequency of the inflected form within its inflectional class and
the number of syntactic functions and meanings (Kostić, 1991), inflectional entropy (Baayen et al,
2006), relative entropy (Milin et al., 2009), etc.
By definition, polysemous words are equally ambiguous in all of the inflected forms
(Gortan-Premk, 2004). Hence, the inflected form does not serve as the cue for their true meaning.
However, some research suggested that meaning can shape the way a noun is used in a
sentence (Kostić et al., 2003).
We presented 35 polysemous nouns of masculine gender in a visual lexical decision task
to 74 participants (data collection still ongoing). Each noun was presented in one of its seven
forms in a latin-square design. Entropy of the sense frequency distribution was estimated in a
norming study (Filipović Đurđević & Kostić, 2017). Relative frequencies of inflected forms and the
number of syntactic functions and meanings were taken from Kostić (1965). Discrimination
learning based predictors were derived from cue-outcome weights matrix calculated by
equilibrium equations (Danks, 2003), implemented in the ndl package (Arppe, et al. 2015). Cues
were bigrams of the polysemous words’ inflected forms presented in the experiment. Outcomes
were their lemmata and 1000 co-occurring context words.
We modelled processing times by applying GAMMs (Wood, 2006) and compared two
models. Information-theoretic model revealed the expected facilitatory effect of entropy of word
senses, but only in the nominative form. Discriminative predictors (Milin et al, 2017) affected
processing in some word forms, again revealing the interaction. AIC comparison showed that the
discrimination-based model was superior to the information-theoretic one. Correlations between
the two sets of measures revealed some interesting relations. Activation was more sensitive to
semantic ambiguity, whereas diversity captured both semantic and syntactic ambiguity. This
suggests a more complex interplay of semantics and morpho-syntax than previously thought and
the possibility of capturing such an interaction with discrimination-based diversity measures. | sr |
