This work introduces SYMPAThy, a data representation model in which the combinatorial properties of a lexical item are described by merging surface and deeper linguistic information. The proposed approach is then evaluated by comparing, for a sample list of verbal idioms, a set of SYMPAThy-based fixedness indexes against the relevant speaker-elicited indexes available in the descriptive norms collected by Tabossi et al. (2011).
By “Word Combinations” (WoCs) we broadly
refer to the range of constructions typically
associated with a lexical item. In Construction
Grammar, constructions (Cxn) are
conventionalized form-meaning pairings that can vary in
both complexity and schematicity
Each lexeme can thus be described as having a
combinatory potential to be defined and observed
at a more constrained, surface POS-pattern level
(P-level) and at the more abstract level of
syntactic structure (S-level). These two levels are often
kept separate, not only theoretically, but also
computationally, as their performance varies according
to the different types of combinations that we want
to track
We advocate a unified and integrated view of a lexeme’s combinatory potential, in order to capture both fixed combinations (MWEs of various types) and more productive aspects of the lexeme’s distributional behaviour. The theoretical premises lie in the constructionist view of the mental lexicon outlined above, whereas a proposal for a computational implementation is illustrated here. Specifically, we i) present SYMPAThy, a model of data representation that takes into account both surface and deeper linguistic information; ii) develop and test an index of productivity for Italian WoCs based on SYMPAThy. 2
We argue that to obtain a comprehensive picture of the combinatory potential of a word and enhance extracting efficacy for WoCs, the P-based approach (which exploits sequences of POS-patterns and association measures) and the S-based approach (which exploits syntactic dependencies and association measures) should be combined. We illustrate this point with an example based on the Target Lexeme (TL) gettare ‘throw’ (V).1
We want to use S-based methods to capture the fact that V occurs typically within some syntactic Frames and not others, that for each Frame we have typical Fillers (lexical items) instantiating Frame slots, and that each slot is associated with certain semantic (ontological) classes:2
1All data is from a version of the “la Repubblica” corpus
2Data extracted by LexIt
So, fine-grained differences do not emerge with the S-method, while the P-based method fails to capture the higher-level generalizations we get with the S-method. In order to get the best of both worlds, we extracted corpus data into SYMPAThy (SYntactically Marked PATterns), a database where information on both levels is stored and accessible jointly: • syntactic frames with argument slots and fillers; • linear order of all elements for each TL; • POS tag for each element (simple preposition vs. preposition with article, definite vs. indefinite article, modal vs. full verb, etc.); flexive form gettarsi ’throw oneself’ and objectless forms are excluded. • morphosyntactic features: finiteness, tense, etc. gender, number, 3
Since constructions span along a continuum between fixedness and productivity, there have been various attempts at measuring how fixed a given WoC is, mostly based on surface features. Nissim and Zaninello (2011) assess the fixedness of a subset of complex nominals by comparing inflected and lemmatized forms, and taking into account the proportion of elements that undergo variation in a given MWE. Inflection is also used by Squillante (2014) on noun-adjective expressions, and is combined with two other measures, interruptibility and substitutability. Zeldes (2013) extends Baayen’s morphological productivity approach to argument structure and estimates the productivity of a syntactic slot from the number of its hapax noun fillers. Wulff (2009) uses a set of morphosyntactic indexes of variations and a collocation-based index of compositionality as variables in a regression study to determine fixedness.
We extend the state of the art of the quantitative approach to construction fixedness by exploiting the potentialities of SYMPAThy to develop a series of corpus-based indexes able to describe the fixedness of some idiomatic expressions. Our approach is then evaluated by comparing, for a sample list of expressions, a composition of our indexes against the behavioral judgments of syntactic flexibility collected by Tabossi et al. (2011). 3.1
In the SYMPAThy model, the combinatory space of a Target Lexeme is assumed to be formed by a network of Cxns, varying for their degree of fixedness/productivity. For any given TL such a representation is built by means of the following fourstep procedure: 1. its SYMPAThy patterns are extracted from a reference corpus; 2. the set of single and multiple slot Cxns that TL combines with are semi-automatically identified. An example for the verb gettare is reported and explained in Appendix 1; 3. each construction is associated with a variational profile formed by a number of statistics extracted from the SYMPAThy pattern to estimate: i) the variability of the fillers that instantiate the syntactic slots of constructions; ii) the morphological variability of the constructions’ components; iii) the variability with respect to determiners; iv) the variability with respect to adjectival and adverbial modifications; v) the variability in the linear order. 4. variational profiles are then used to measure the lexical, morphological and syntactic degrees of freedom of Cxns, providing a multidimensional quantitative characterization of their level of fixedness. 3.2
In what follows, we devise a way to encode the
variation possibilities shown by Cxns, as well as
a meaningful way to combine them. Specifically,
we distinguish a series of dimensions of variation
and propose to exploit Entropy
Entropy is a measure of randomness, calculated as the average uncertainty of a single variable: H(X) = −
X p(x) log2(p(x)) x∈X (1) This measure of randomness can be adapted to our needs by taking the variable X as being a Cxn of interest, and the states of the system x as its values on one dimension of variation. Lower entropy values are to be understood as evidence of fixedness, while higher values suggest a more variable distribution of the states of a given variable, i.e. the target construction tends to be freer.
Observed entropy values, however, can span from 0 to the logarithm of the number of values that X can assume. As a consequence, entropy values related to different dimensions of variation are not comparable, and cannot be combined into a single fixedness index. We overcome this limitation by following Wulff (2008) and describing the randomness of each variability dimension in terms of relative entropy, computed as the ratio between the observed entropy from eq.1 and the maximum entropy Hmax for the variable X:
Hrel(X) =
H(X) Hmax(X) =
H(X) log2(|X|) This measure, that ranges from 0 to 1, has been employed as a flexibility measure to describe the flexibility of a given set of target Cxns along the following dimensions of variation: LEXICAL VARIABILITY. The entropy of the lexical instantiation of the slot positions of a Frame is calculated by assuming that the states x of the random variable X are all the possible fillers that can instantiate a given slot in Cxn (e.g. in subj#gettare#obj:luce#su X, X can be filled by vicenda ‘matter’, mistero ‘mystery’, etc.).
culated as the entropy of the morphological features manifested by the fillers of a Cxn (e.g., gettare#ombra-fs ‘cast shadow-singular’; gettare#ombra-fp ‘cast shadow-plural’).
how variable is the presence or absence of articles determining the available slots in a Cxn, and, if appropriate, their type (DEFinite vs. INDefinite): for instance, gettare#∅+acqua#su DEF+fuoco.
codes how variable is the presence or absence of adjectives, adverbs or prepositional phrases modifying the available slots. In this way, it is possible to account for patterns like:gettare#molta+acqua#su ∅+fuoco.
information on linear order available in SYMPAThy to estimate how variable is the distance in tokens between a TL and the other constituents of a given lexically specified Cxn.
In the experiment reported in the next section, we have combined the single variability measures Hrel(X) into an overall flexibility index F (X) corresponding to four possible combinations: • SUM: F (X) is obtained by summing over all the single Hrel(X) values; • AVERAGE: F (X) is the mean of the single
Hrel(X) values; • AVERAGEP OS : F (X) is the mean of the positive Hrel(X) values; • MAX: F (X) is the highest Hrel(X) value.
We leave to future research the investigation of further ways to combine the variability indexes. (2) 4
In order to evaluate our approach, we set out to test if our indexes can mimic the intuitive judgments of native speakers about the fixedness of fully lexically specified constructions. To do so, we selected a subset of the idioms in the norms collected by Tabossi et al. (2011), and tested to what degree the speaker-elicited flexibility judgments available in this repository can be modeled by a composition of our variability indexes.
Tabossi et al. (2011) collected several normative measures for 245 Italian verbal idiomatic expressions. Using a group of 740 Italian speakers, they collected a minimum of 40 elicited judgments for each idiom on several psycholinguically relevant variables.
Among the different kinds of ratings, those concerning syntactic flexibility have been collected by inserting each idiomatic expression in a sentence in which one of the following vfie syntactic modifications occurred: adverb insertion, adjective insertion, left dislocation, passive and movement. Participants were asked to evaluate, on a 7-point scale, how much the meaning of the idiomatic expression in the syntactically modified sentence was similar to its unmarked meaning as expressed in a paraphrase prepared by the authors. 4.2
Out the 245 expressions in Tabossi et al., we selected the 23 target idioms reported in Appendix 2. Each such idiom can be represented, in our approach, as a fully lexically specified transitive Cxn headed by a given verbal TL, for which the subject slot is underspecified (e.g. gettare#obj:maschera). We built the variational profiles of our target idioms by adopting an adapted version of the procedure described in Section 3: 1. for each TL, we extracted the SYMPAThy patterns from the “la Repubblica” corpus; 2. the patterns involving one of our target idioms were identified and selected; 3. for each idiom, the variability indexes described in Section 3.2 were calculated. Note that, given the nature of our experimental stimuli, the lexical variability index is not relevant; 4. we built a fixedness index for each idiom, according to the four composition methods in the previous section. 4.3
In order to test the cognitive plausibility of the fixedness indexes extracted from SYMPAThy, we calculated the Pearson’s Product-Moment Correlation strength between them and the syntactic
SUM AVERAGE AVERAGE P OS MAX r flexibility ratings in Tabossi et al. (2011). Correlation values are reported in Table 1. In all cases, there is a significant ( p < .05) positive correlation, ranging between .44 and .47, thus supporting the psycholinguistic plausibility of our corpus-based variability indexes.
These results, albeit preliminary, look promising especially given the different nature of the behavioral and corpus-based indexes. On the one hand, the speakers’ ratings are semantically driven, since they are thought to model how much the figurative meaning of a given idiom is sensitive to its syntactic form. On the other hand, the automatically corpus-derived information exploited by our indexes does not take meaning into account. SUch indexes describe a lexically specified Cxn that can in principle have an idiomatic as well as a compositional, literal meaning (even if, presumably, the latter case is rare in the corpus). 5
In this study we presented a procedure for characterizing the combinatorial potential of a lexical item and the degree of fixedness of the Cxns it occurs in. Such a procedure has been preliminary tested on a small sample of idiomatic expressions and the resulting representation has been evaluated against the subject-elicited judgments collected by Tabossi et al. (2011). In the future, we are planning to extend the inventory of variability dimensions (addressing also the question of the semantic compositionality of Cxns), to study their relative weight and their interactions, and to develop more sophisticated ways to combine them.
This research was carried out within the CombiNet
project
ilnks ) ofrm emanig #acqu #sul #Fuoc [[SUBJ]NP gettare (ADV) (ADJ) acqua sul fuoco]
Person, Event,...
SUBJ: OBJ: SU: COMP: fuoco acqua sul ‘defuse, minimize a situation’ raget ‘add fuel to the fire’
Person, Event,...
benzina [[SUBJ]NP gettare (ADV) (ADJ) benzina sul fuoco] lt = GETTARE ‘tHroW’ ofrm emanig ofrm emanig ofrm emanig ofrm [CAUSE (OBJ, [GO (AWAY)])]
Frame ... ... ... ... ... ...
... ... ...
... ... ... [[SUBJ]NP gettare (ADV) (ADJ) fango su [COMP]NP] SUBJ: OBJ:
Person, Event,...
fango (⇒ SG; bare | partitive) COMP: Person, Institution, ... ‘defame, discredit, blacken the name of’ ofrm ofrm emanig ... Questo getta una pesantissima ombra sulla legittimità ... ... rischia di gettare ulteriore fango sul calcio ... ‘This casts a serious shadow on the legitimacy...’ ... Il rivale getta ombra sulla salute del leader ... ‘His opponent casts a shadow on the leader’s health’ ‘(it) may sully football even more’ ... Hanno sempre gettato fango su di noi ...
‘They have always sullied us’ ... Gli amici hanno gettato sulla bara garofani rossi ... ‘Friends threw red carnations on his coffin’ ... getta un sasso sull’autostrada ... ‘(s/he) throws a stone in the highway’ The verb gettare ‘to throw’ combines with the highly schematic subj#obj#comp-su Cxn, whose slots can freely vary with respect to linear order, presence of determiners, modifiers, etc. A semi-productive instance of this construction is the subj#obj:ombra#comp-su Cxn, with a fixed object slot and a partially variable oblique slot, which can appear with a semantically limited range of arguments. A fully lexically specified instance of the same construction is instead the subj#obj: acqua#comp-su:sul-fuoco Cxn, which has both slots instantiated and limited degree of variability.
Appendix 2: List of idioms used as experimental stimuli Gettare la maschera (‘to reveal oneself ’)
Gettare acqua sul fuoco (‘to defuse a situation’) Gettare olio sul fuoco (‘to inflame a situation ’) Mettere la mano sul fuoco (‘to stake one’s life on
Prendere una cotta (‘to get a crush on somebody’) Mettersi il cuore in pace (‘to resign oneself to sth’) Mettere nero su bianco (‘to put sth down in black Mettere il dito sulla piaga (‘to hit someone where
Mettere i puntini sulle i (‘to be nitpicking’)
Perdere il treno (‘to miss an opportunity’) Perdere la bussola (‘to lose one’s bearings’)