Argument structure constructions in a Natural Language Processing
environment
Alba Luzondo-Oyón
(Corresponding author)
Dpto. de Filologías Extranjeras y sus Lingüísticas
Facultad de Filología
Universidad Nacional de Educación a Distancia
c/ Senda del Rey 7
28040, Madrid, Spain
aluzondo@flog.uned.es
Landline: +34 913988699
Francisco Ruiz de Mendoza-Ibáñez
University of La Rioja
Philology Building
c/San José de Calasanz 33
26004, Logroño, La Rioja, Spain
francisco.ruizdemendoza@unirioja.es
Abstract: This paper contributes to the field of computational Construction Grammar
(cf. Steels, 2012; van Trijp, 2011) by providing a linguistically-oriented formalized
treatment of argument structure constructions within the architecture of a
multipurpose lexico-conceptual knowledge base for Natural Language Processing
systems known as FunGramKB (Periñán, 2013). More concretely, we analyze three
members of the family of the English resultative, namely, the resultative (e.g. He
hammered the metal flat/into a flat sheet), the caused-motion construction (e.g. He
shoved the canoe into the water) and the way construction (e.g. He fought his way
free/to freedom), paying special attention to lexical-constructional integration
variability. Thus, the present article offers a description of the format of
constructional schemata in FunGramKB as machine-tractable representations of
constructions and proposes a ‘splitting-like’ solution (à la Boas, 2003) to handle the
mismatches resulting from lexical-constructional fusion. We finally argue that, in the
case of FunGramKB, a feasible computational implementation must be based on
constructional subtypes rather than on broad-scale constructions of the Goldbergian
kind.
Keywords: Construction Grammar; argument-structure constructions; Natural
Language Processing; FunGramKB.
Highlights:




We show how constructions can be handled in a computational environment.
We analyze three members of the family of the English resultative.
We offer a description of the format of constructional schemata.
Lexical-constructional fusion mismatches require a ‘splitting-like’ solution.
1
1. Introduction
Work on Construction Grammar within Cognitive Linguistics is almost three decades
old now. It has given rise to a number of approaches, which are largely compatible
with one another, and whose differences are more often than not a matter of goals and
areas of special focus (for an overview of the different perspectives see Dirven and
Ruiz de Mendoza, 2010; Hoffmann and Trousdale, 2013; and Butler and Gonzálvez,
2014). In the context of this productive research program, the literature has examined
a large number of linguistic constructions (or entrenched form-meaning/function
pairings) of varied nature and complexity. Among them we find the following: ‘Let
alone’ (e.g. Who would like this crap, let alone buy it?; Fillmore, Kay and O’Connor,
1988); ‘What’s X Doing Y?’ (e.g. What’s a physicist doing studying marine life?;
Kay and Fillmore, 1999); ‘contrastive reduplication’ (e.g. My car isn’t MINE-mine,
it’s my parents; Ghomeshi et al., 2004); ‘X is so N(P)’ (e.g. That’s so Monica;
Gonzálvez, 2014). Constructions like these arise from attaching special meaning
implications, of a personal and interpersonal kind, to their basic predication. Thus,
What’s a physicist doing studying marine life? is not an information question, despite
its Wh-interrogative form, but an expression of complaint about a given state of
affairs that the speaker finds to be wrong or odd. A different class of constructional
configurations, known as argument-structure constructions (cf. Goldberg, 1995), is
based on the linguistic expression of predicate-argument relationships. Some of the
members of this latter class are the ditransitive construction (e.g. John gave Mary a
book; cf. Barðdal et al., 2011), the resultative construction (e.g. The boy ate himself
sick; cf. Boas, 2003), the subjective-transitive construction (e.g. I find her so sweet;
cf. Gonzálvez, 2009), and the caused-motion construction (e.g. He pushed the canoe
into the water; cf. Peña, 2009). Argument-structure constructions are the concern of
the present paper.
Clearly, argument-structure constructions only represent a small portion of the
full inventory of form-meaning pairings of a given language or its constructicon (cf.
Boas, 2014, p. 94). However, the amount of work devoted to their analysis —in which
complementary aspects of their complex nature have been emphasized— is
impressive. Many issues have been covered in the literature, such as the universal or
language-specific nature of constructions, how to determine analytical categories and
their grounding in experience, or the challenges that constructions pose for specific
frameworks like Role and Reference Grammar (RRG; cf. Nolan and Diedrichsen,
2
2013; see also Nolan, 2014). A particular area of controversy is the understanding of
the principles that regulate the integration of lexical structure into constructions. In
this respect, three broad analytical perspectives can be distinguished. One, which is
typical of Goldberg’s (1995) work, focuses its attention on capturing broad-scale
generalizations on lexical-constructional behavior. A second one, which arises from
work on the resultative by Boas (2003, 2008, 2011), tips the scale in favor of
including more detailed information in lexical description in order to compensate for
the lack of constraining power of Goldbergian generalizations. A third, more recent
approach, called the Lexical Constructional Model (LCM; cf. Ruiz de Mendoza,
2013; Ruiz de Mendoza and Galera, 2014), has shed light on the constraining role of
some cognitive processes, including metaphor and metonymy, for the integration of
lexical structure into constructional characterizations.
Beyond purely theoretical pursuits, some Construction Grammar accounts
increasingly show greater interest in the formalization and computational
implementation of many of the notions put forward in Construction Grammar and
Cognitive Linguistics. Fluid Construction Grammar (Steels, 2012), Embodied
Construction Grammar (Bergen and Chang, 2005), and Sign-Based Construction
Grammar (Boas and Sag, 2012) share this goal (see also Chang, De Beule and Micelli
2012, and van Trijp, 2013 for comparative descriptions). This paper adds to this
research trend by presenting readers with a linguistically-oriented computational
treatment of argument-structure constructions. More specifically, we study several
members of the resultative family of constructions (cf. Goldberg and Jackendoff,
2004; Luzondo, 2014; Ruiz de Mendoza and Luzondo, 2014).1 This family codes
changes of state or location resulting from verbal action (e.g. He kicked Bob black
and blue, He kicked the ball into the net, etc.). Van Trijp (2011), for example, has
illustrated how argument-structure characterizations can be handled in the context of
Fluid Construction Grammar. In his proposal, van Trijp offers a “design pattern” (see
Steels, 2012, p. 16) that is capable of dealing with many of the complexities of
argument structure and he operationalizes such a pattern within Fluid Construction
Grammar. The design pattern is aimed at capturing the multilayered and indirect
nature of form-meaning mappings (van Trijp, 2011, pp. 2-3). To illustrate, note that
the same syntactic role may appear in a different surface form depending on the
context. Thus, in He saw him crossing the street, the masculine pronouns that
function as Subject of the main clause and as Subject of the subclause are respectively
3
expressed as he and him. To account for this and other cases of variable behavior, van
Trijp proposes a two-stepped solution. In the first step, lexical and phrasal
constructions specify their semantic and combinatorial potential. In the second,
argument-structure constructions select an actual value from this potential and
implement the linking between semantics and syntax, thus yielding specific
realizations (van Trijp, 2011, pp. 8-9). For example, for the verb send, the transitive
construction requires the selection of the following values: (i) at the semantic pole,
‘sender-Agent’ and ‘sent-Patient’ (while the semantic roles ‘Recipient/Goal’ are
discarded); (ii) at the syntactic pole, Subject and Direct Object (DO) (the Indirect
Object (IO) and the Oblique are also left out). This yields utterances like Jill sent a
letter. The ditransitive construction, on the other hand, selects Agent, Patient and
Recipient, which are then mapped onto Subject, IO and DO in John sent Mary a
letter. What this approach does not seem to cover, however, is the fact that when a
given verb fuses with a particular argument-structure construction, the syntactic
pattern may impose different values for each realization. A case in point is that of
sweep and the transitive-resultative construction. In He swept the floor clean, the
construction selects the semantic-syntactic pairs “Agent-Subject” and “Patient-DO”,
which arise from lexical projection (cf. He swept the floor), and adds the result
ingredient (i.e. clean). In He swept the broom to pieces, by contrast, both “the broom”
and “to pieces” are arguments supplied by the construction, as evidenced by the
ungrammaticality of *He swept the broom. Here, the Patient role, which is mapped
onto the object of the causal action, is the instrument employed to perform the goaloriented activity of sweeping. This phenomenon cannot be treated computationally by
merely specifying combinatorial potentials in the verbal construction, since there may
not always be a straightforward match between lexical and constructional properties,
even in cases in which the same verb and the same construction interact with each
other (cf. She ate the cob clean2, My youngest brother ate himself sick with the
grapes3). From a linguistic perspective, the mismatch requires a solution based on the
specification of the factors that license the incorporation of “the broom” as an object
into the resultative construction. A plausible linguistic solution is offered by the
LCM, which postulates the generic-level metaphor A GOAL-ORIENTED
ACTIVITY IS AN EFFECTUAL ACTION, whereby the activity of sweeping is seen
as if it were an action with an impact on the object: the instrument employed to
perform the activity (the broom) is conceptualized as the object of a causal action (see
4
Luzondo, 2014 for more details). However, to our knowledge, the computational
tractability of the metaphor-based solution has not been investigated yet, probably due
to its highly generic format. An alternative approach, which is compliant with
computational requirements, requires a finer-grained constructional specification
dealing with all realization possibilities.
One knowledge engineering project that is sufficiently equipped to account for
constructional realization variability of the kind mentioned above is FunGramKB,
which has been described by its proponents as a “user-friendly online environment for
the semiautomatic construction of a multipurpose lexico-conceptual knowledge-base
for NLP systems” (Periñán and Arcas, 2010a, p. 2667; see also Periñán and Arcas,
2005, 2007ab, 2010b). FunGramKB, is grounded in two largely compatible linguistic
theories (see Nolan and Periñán, 2014): (i) RRG (Van Valin and LaPolla, 1997; Van
Valin, 2005), which is a functional model of language; and (ii) the LCM, a
comprehensive model of meaning construction through language in context which
reconciles functionally-oriented projectionist approaches and cognitively-oriented
constructionist accounts (Ruiz de Mendoza, 2013, p. 232; cf. Ruiz de Mendoza and
Mairal, 2008, 2011; Mairal and Ruiz de Mendoza, 2009). Although, as observed in
Periñán (2013, p. 206), “NLP applications which […] work with no foundation in any
linguistic theory” are easier to build, a symbolic approach (cf. Mairal, 2012) like the
one followed by FunGramKB avoids the problem of being “deceptively intelligent,”
thus ensuring natural language understanding.4 It is for this reason that the LCM and
RRG lie at the heart of the linguistic level of FunGramKB (see Mairal, fc.; Periñán
and Arcas, 2014; Van Valin and Mairal, 2014), as will be shown in the following
sections.
In this context, the present article provides a description of the format of
constructional schemata as machine-tractable representations of constructions and
proposes a ‘splitting-like’ solution to account for the intricacies of lexicalconstructional fusion (cf. sweep above).5 In FunGramKB, constructional schemas,
which are depicted through Attribute-Value Matrices (AVMs), are formalized by
means of descriptors and constraints. In this respect the kind of analysis presented
here is closer to the approach adopted by computational Construction Grammars than
to Goldberg’s (1995, 2006) Cognitive Construction Grammar, in which there is a
clear dearth of formalization.
5
With this in mind, the structure of the rest of this paper is as follows. In
section 2 an outline of the architecture of FunGramKB is presented. Section 2.1.
introduces the Grammaticon, which contains constructicons of a number of
languages. This is the FunGramKB component on which we focus our attention.
Likewise, this section briefly discusses the computational suitability of syntactic
alternations (Levin, 1993) in light of the design of the grammatical level. In Section
2.1.1 we offer a necessarily succinct critical revision of the notion of construction in
FunGramKB. Section 3 deals with the computational treatment of three members of
the resultative family, namely, the resultative, the caused-motion, and the way
constructions. This section argues for the need to posit separate, albeit related, subtypes of constructional schemas of the same general construction; this is an issue that
has not been addressed in the FunGramKB literature as yet. In Section 4, we discuss
which of the three constructionist accounts considered in this paper, i.e. the
approaches by Goldberg (1995), Boas (2003), and Ruiz de Mendoza (2013), is more
readily tractable computationally. Section 5 summarizes the main points made in the
paper.
2. FunGramKB: A general introduction
As shown in Figure 1, the architecture of this knowledge base comprises three major
knowledge levels: the lexical and grammatical levels, which are language-dependent,
and the conceptual level, which is language-independent and therefore shared by all
the languages currently supported in the knowledge base (i.e. Spanish, English,
French, and Italian). In turn, each of these knowledge levels consists of several selfstanding, although interconnected, modules.
6
Figure 1: The architecture of FunGramKB (source: www.fungramkb.com).
What follows is a brief description of each of the components specified in Figure 1:
1.
The lexical level is made up of a Morphicon and a Lexicon (see Mairal and
Periñán, 2009; Luzondo and Jiménez, 2014, p. 210-212 for more details). While the
former handles cases of inflectional morphology, the latter, which preserves some of
the basic assumptions of RRG, stores information regarding morphosyntactic and
collocational information about lexical units, each of which is linked to a concept in
the Ontology. More concretely, lexical entries in FunGramKB are provided with the
following information: (i) basic (i.e. headword, index and language); (ii)
morphosyntactic features (i.e. graphical variant, abbreviation, phrase constituents,
category, number, gender, countability, degree, adjectival position, verb paradigm
with its constraints, and pronominalization); (iii) miscellaneous (i.e. dialect, style and
example); and (iv) core grammar (i.e. essentially syntactic projection based on the
attribution of roles to argument-structure elements). In the case of verbal predicates,
the most important component is the core grammar, which contains a list of
7
“attributes whose values allow the system to build the basic logical structure of verbs
automatically” (Periñán, 2013, p. 212). The following is an outline of the attributes in
the core grammar:
(a) Aktionsart: the most representative RRG verb class in which a given verb
may occur (e.g. activity, accomplishment, causative accomplishment, etc.).
(b) The lexical template of the verb includes: (i) the number of variables that a
verbal predicate may take (i.e. x, y, z); (ii) the so-called thematic frame
mapping, through which the variables of the predicate are mapped onto one and
only one participant in the thematic frame of the concept to which a given sense
of a lexical unit is connected; (iii) the idiosyncratic features or macroroles
assigned to the verb, namely, actor and undergoer.
(c) Moreover, the core grammar displays a repertoire of constructions in which
a given verb may be embedded. That is, lexical entries contain pointers to the
grammatical constructions in the constructicon, through which the projection
from syntax to semantics takes place (Periñán, 2013, p. 217). Thus, although as
we will see in Section 2.1, knowledge on constructions is not stored in the
Lexicon but in the Grammaticon, knowledge engineers working with a verbal
predicate will specify those syntactic patterns with which the verb may fuse. For
example, if kick is being described in the Lexicon, the following constructions
will be selected on the basis of corpus evidence: the resultative (e.g. Pat kicked
Bob black and blue), the caused-motion construction (e.g. Pat kicked the
football into the stadium), the way construction (e.g. Pat kicked his way out of
the operating room), etc. (cf. Bencini and Goldberg, 2000, p. 641). Similarly,
float in Figure 2 below participates in the intransitive motion construction
(IMOT; e.g. The bottle floated into the cave), the caused-motion construction
(CMOT; e.g. Currents floated the slick onto 162 miles of beaches6), the way
construction (WAY; e.g. The ship floated its way under the bridge), etc.
Furthermore, besides the constructions stored in the Grammaticon, verbs in the
Lexicon are provided with one Kernel construction, which is built on the basis
of the features described in (a) and (b) above. Thus, depending on the variables
in the lexical template, a given verb will typically occur in a Kernel-1
Construction (intransitive, e.g. Pat coughed), a Kernel-2 Construction
(monotransitive, e.g. He boiled the water) or a Kernel-3 Construction
(ditransitive, e.g. Anna sent Peter a letter) (cf. Periñán, 2013, p. 213). These are
8
the only kinds of constructions that are not formalized at the grammatical level
of FunGramKB, but are modeled within the lexical entries of verbs (Periñán and
Arcas, 2014, p. 177). As one of such, float in Figure 2 is automatically provided
with a Kernel-1 construction (i.e. ‘something floats’). Note that in FunGramKB
an input text like The bottle floated into the lock would therefore consist of two
argument-structure constructions: “the bottle floated” (Kernel-1) and “into the
lock” (intransitive motion construction) (cf. section 2.1.1.).
At this point, it is important to emphasize that lexical entries and
constructional schemas make use of the same formal representation system, i.e. both
of them are represented through AVMs which eventually merge via unification
processes following the paradigm of constraint-based grammars (Mairal, fc.; Periñán
and Arcas, 2014, p. 179-181). Consider, by way of example, the AVM of the lexical
unit float, which captures part of the information detailed in (a)-(c) above.
Figure 2: Core grammar of float (English Lexicon).
Although inspired in a different theoretical model, i.e. RRG, it follows that the
Lexicon of FunGramKB is consistent with the proposals made by Boas (2008, 2013),
who rightly claims that it is necessary to include more detailed semantic information
in the lexical entries of verbs. This contrasts with Goldberg’s (1995) account, in
9
which verbs are merely reduced to a set of participant roles (e.g. float <floating
entity>), which fuse with the argument roles of the construction on the basis of a set
of constraints (see Section 4).
2.
The grammatical level, or Grammaticon, which is our focus of attention in this
paper, stores constructional schemata helping RRG to build the syntax-to-semantics
linking algorithm. The details of this level are discussed later in section 2.1.
3.
At the conceptual level the same formal language (i.e. COREL, see Periñán
and Mairal, 2010) is employed so that information sharing can take place effectively
among its three conceptual sub-modules: (i) the Cognicon, which stores procedural
knowledge by means of scripts; (ii) the Onomasticon, which keeps information about
instances of entities and events; and (iii) the Ontology or the hierarchical catalogue of
concepts that a person has in mind. It should be noted that the Ontology is the pivot
around which the whole architecture of the knowledge base revolves, which is why
FunGramKB proponents claim that their approach is conceptual. In the Ontology,
concepts are provided with semantic properties in the form of ‘thematic frames’ and
‘meaning postulates’. Both are taken to be language-independent semantic knowledge
representations. On the one hand, a thematic frame is defined as “a conceptual
construct stating the number and type of participants involved in the prototypical
cognitive situation codified by the event” (Periñán and Arcas, 2010b, p. 30). For
example, the thematic frame of the concept +FLOAT_00, to which the lexical units
float, flotar (Spanish), flotteur (French), galleggiare (Italian), etc. are connected,
describes a prototypical scenario in which an entity or “(x1)Theme” lies in the water
(i.e. (x2: +WATER_00)Location). On the other hand, meaning postulates are sets of
one or more logically connected predications (e1, e2, e3, etc.), that is, conceptual
constructs carrying the generic features of concepts (Mairal and Periñán, 2009, p.
224). Continuing with our example, the meaning postulate of the concept
+FLOAT_00 is provided in (1) together with a translation:
(1)
+(e1: +LIE_00 (x1)Theme (x2: +WATER_00)Location (f1: (e2: n
+SINK_00
(x3)Agent
(x1)Theme
(x2)Location
(x4)Origin
(x5)Goal))Scene) = “Floating means that something (x1) lies in the water
(x2) without sinking in it (f1)”.
10
The COREL metalanguage is also employed to code the semantics of
argument-structure constructions at the grammatical level, to which we turn our
attention now.
2.1. The Grammaticon
The Grammaticon in FunGramKB is the repository of constructional schemata. It
comprises several constructicons (i.e. L1-Constructicon, L2-Constructicon, L3Constructicon, and L4-Constructicon), which are inspired in the four constructional
layers of the LCM. For the sake of clarity, we shall briefly discuss the four levels of
description postulated by the LCM. The first is structured around low and high-level
(i.e. abstract) non-situational cognitive models.7 At this level, lexical structure, which
is based on low-level non-situational cognitive models, is integrated into argumentstructure constructions like the resultative, caused-motion, etc. Argument-structure
constructions, which have been created by abstracting away properties that are shared
by lexical predicates, are made up of high-level non-situational cognitive models.
Low-level situational cognitive models are part of the implicational layer, or level 2,
where we find constructions such as ‘What’s X doing Y?’ (Kay and Fillmore, 1999),
which conveys the conventionalized meaning implication that there is a situation that
bothers the speaker (Ruiz de Mendoza, 2001). The X and Y variables receive partial
structure from level 1. In turn, level 3 is grounded in high-level situational models
such as ‘begging’, ‘requesting’, ‘offering’, which abstract common conceptual
material away from low-level situations. Such models underlie illocutionary
constructions, which, like their level-2 counterparts, consist of fixed and variable
elements, as in ‘Can You X?’, ‘You Shall Have X’, ‘Let’s X’. Finally, relations
among non-situational high-level cognitive models are treated at the fourth level. This
layer deals with discourse structure in which level-1 configurations are coupled
through relational patterns like ‘cause-effect’, ‘condition-consequence’, and ‘actionresult’, among others (see Ruiz de Mendoza and Galera, 2014).8
The L1-Constructicon of FunGramKB, which is the object of the present
article, is represented in Figure 3 via the intransitive motion construction:9
11
Figure 3: The interface of the L1-constructicon: the case of the intransitive motion
construction.
Before proceeding with an explanation of the L1-Construction, let us first
discuss the computational suitability of syntactic alternations within FunGramKB.
The reader may have noticed the pair of examples located at the top of Figure 3 (cf.
The bottle floated > The bottle floated into the cave). Since, to this day, there has been
no attempt to produce an exhaustive list of argument-structure constructions, the
range of configurations stored in the L1-Constructicon of FunGramKB was first
grounded in Levin’s (1993) seminal compilation of syntactic alternations.
Nevertheless, it should be immediately pointed out that alternations are not useful
computationally, at least in light of the architecture of the L1-Constructicon. There
are two reasons for this. The first relates to the fact that what the machine needs to
recognize and process is the input text but not the capacity of a given verb to
participate in two alternating constructions. In other words, as previously pointed out,
12
constructional information in FunGramKB crucially depends on descriptors and
constraints on the basis of which each constructional schema is defined on its own,
thus allowing the machine to identify an input text as an example of a particular
construction. For that matter, whether the verb of the input text can express its
arguments in different syntactic positions is irrelevant to the machine. Second,
Levin’s alternating pairs are to be viewed as two closely related constructions, which
may not even share the same number of variables and/or whose meaning and
Aktionsart may in fact be different. These are precisely some of the features through
which constructional schemas are defined in FunGramKB. These features allow us to
distinguish one construction from another (see explanation in (a)-(b) below). To
illustrate this point, consider the ‘induced action’ alternation discussed in Levin
(1993, p. 31):
(2)
a. The horse jumped over the fence.
b. Sylvia jumped the horse over the fence.
Whereas in (2a) jump is embedded in an intransitive motion configuration in which
only one entity is in motion (i.e. ‘X MOVES Z’, following Goldberg’s (1995)
notation), in (2b) the activity of jumping is seen as part of a causal event, in which the
horse is both the object of the subject’s causal action (cf. somebody causes the horse
to jump) and the actor of the activity (cf. the horse jumps). Thus, (2b) is but a case of
the caused-motion construction (‘X CAUSES Y TO MOVE Z’). From this it follows
that the number of constructional arguments in (2a) and (2b) is clearly not the same,
since the former consists of two, while the latter comprises three. Besides the fact that
each linguistic expression realizes a distinct, but related, construction, the Aktionsart
is different in each example. In (2a) we have an active accomplishment. Conversely,
the Aktionsart of the structure realized by Sylvia jumped the horse over the fence is a
causative active accomplishment. Suffice it to say that the same rationale applies to
the examples in Figure 3. Thus, the need to maintain an individualized analysis of
grammatical configurations leads us to discard alternations from the FunGramKB
Editor in Figure 3. In this way, the machine will be able to match a given input text
(i.e. the linguistic realization) with one construction or the other.10
13
Let us now go into the characterization of constructional schemas, as illustrated
in Figure 3. We mentioned in the introduction that FunGramKB constructional
schemas are defined by means of two elements: (i) descriptors (i.e. Aktionsart,
number and type of variables, the thematic role of the variable(s) contributed by the
grammatical construction, macroroles, and the COREL schema); and (ii) constraints,
which focus on phrase realizations and selection preferences (i.e. ontological concepts
like +WATER_00 in (1) above). We shall illustrate the ensuing description of these
elements through the intransitive motion construction (e.g. The bottle floated into the
cave), represented in Figure 3. This construction is a member of the resultative family
in English:
(a) Following RRG’s classification, constructional schemas, as much as verbs in the
Lexicon, are provided with an Aktionsart type, which, in the case of the intransitive
motion construction, is an active accomplishment (ACA, in Figure 3). Recall that, by
contrast, from the point of view of its Aktionsart characterization, the caused-motion
construction (cf. Sylvia jumped the horse over the fence) is a causative active
accomplishment.
(b) At the right-hand side of the Aktionsart box, we state the number and type of
constructional variables involved (i.e. x = ‘”the bottle” and z = “into the cave”, in the
grammatical pattern at hand). Conversely, caused-motion examples display three
constructional arguments. Thus, an approach based on syntactic alternations of the
type The bottle floated into the cave > The water floated the bottle into the cave
precludes an analysis of argument-structure characterizations as discrete entities with
different formal and semantic properties.
(c) For each new variable contributed by the construction (z, in the case of the
intransitive motion), the following information should be specified:
(c.i) Its thematic role: Agent, Theme, Referent, Result, etc. For example, the
thematic role of z is Goal or the destination of motion (cf. “into the cave”).
(c.ii) The macrorole (if any) assigned to the new variable: either actor or
undergoer. This semantic notion of macrorole, which is taken from RRG (Van
Valin, 2005, p. 60), largely corresponds to what other theories label as ‘logical
subject’ (actor) and ‘logical object’ (undergoer), terms not employed in RRG
since they typically express syntactic instead of semantic relations. In fact, they
are called macroroles because they subsume a number of specific thematic
relations. For example, the actor is the most agent-like argument, whereas the
14
undergoer is the most patient-like argument. In FunGramKB these two
macroroles play a crucial role in the linkage between the lexico-grammatical
modules and the conceptual module. The z variable of the intransitive motion
construction, however, is not assigned any macrorole.
(c.iii) The categorial information of the variable. Here, z functions as a
Prepositional Phrase (PP).
(c.iv) The box labeled ‘syntax’ in Figure 3 accounts for the status of the
variables of the construction as argument or nucleus. Again, these are two key
concepts in the RRG notion of clause structure or the layered structure of the
clause (Van Valin, 2005, pp. 4-5). For RRG all languages semantically
distinguish between predicates, their arguments, and non-argument elements.
Syntactically speaking, the nucleus contains the predicate, and the core includes
the nucleus and its arguments. With respect to the construction variables, they
generally function as new arguments of the main nucleus, as in He laughed Paul
out of the room, which is an example of the caused-motion construction.
However, in some other cases, such as the resultative, the intransitive motion or
the way construction, the variable incorporated by the construction functions as
another nucleus, sharing an argument of the main nucleus and forming with it a
single, complex nucleus or nuclear juncture in the terminology of RRG. For
example, in Tobias licked the bowl clean (Google Books Corpus, 2014), the
nucleus is made up of two nuclei: ‘lick’ and ‘clean’, which, together with the
two arguments ‘Tobias’ and ‘the bowl’, form the core of the clause. “Into the
cave” in the intransitive motion also functions as nucleus.
(c.v) Prepositions: this box records those idiosyncratic prepositions that
introduce a particular variable, e.g. ‘as’ in the “as construction” (e.g. The
president appointed Smith as press secretary), or ‘into’, ‘to’, ‘over’, etc. in the
intransitive motion construction.
(c.vi) As for the constraints, these take the form of what FunGramKB calls
‘selectional preferences’. These are understood as conceptual constraints
prototypically related to cognitive situations (cf. Jiménez and Pérez, 2011).
Therefore, basic concepts from the Ontology are employed to code them. For
instance, +LOCATION_00 appears as the selectional preference of the z
variable in the intransitive motion configuration.
15
(d) Finally, the COREL schema of the construction captures the conceptual or
cognitive content of the configuration through the COREL metalanguage introduced
in Section 2. Such a representation employs ontological concepts together with their
meaning postulates, which contain the thematic frame. The difference from the
meaning postulate in (1) above is that in the COREL schema the variables of the
construction are associated with one participant role. As shown below, the COREL
schema of the intransitive motion construction is repeated in (3a) for the reader’s
convenience. A translation is supplied in (3b):
(3)
a. +(e1: <EVENT> (f1: (e2: +MOVE_00 (x1)Agent (x2: x)Theme
(x3)Location (x4)Origin (x5: z)Goal))Result)
b. ‘There is an event and, as a result, x (which is mapped onto the
Theme or entity in motion) moves to another location (z, which is
mapped onto the Goal or destination)’.
The COREL metalanguage in (3a) proves to be fine-grained enough to capture
the semantics of the construction, which authors like Goldberg and Jackendoff (2004,
p. 540) capture through the following skeletal structure, “X1=NP GO Path2=PP.
MEANS: [VERBAL SUBVENT]”. The difference, however, is that an event of motion
(cf. +MOVE_00) in a knowledge base cannot be understood cognitively unless five
participants are borne in mind, i.e. Agent, Theme, Location, Origin and Goal, despite
the fact that only two of them (i.e. “(x2: x)Theme” and “(x5: z)Goal”) are realized
linguistically in the intransitive motion construction.
The AVM for Figure 3 is supplied in Figure 4:
16
Figure 4: AVM of the intransitive motion construction.
To conclude, take again the example The bottle floated into the cave. The
AVM of the predicate float in Figure 2 merges onto the constructional AVM in Figure
4, which inherits monotonically the AVM of the x variable (cf. “the bottle”) from the
Lexicon, while adding the z constructional argument (cf. “into the cave”).
So far we have addressed the structure of the L1-Constructicon and described
the way in which one specific constructional schema is represented in FunGramKB.
However, nothing has been said about what a construction is in FunGramKB. The
following section is devoted to a critical assessment of such a notion.
2.1.1. What is a construction in FunGramKB?
The network of “loosely connected” models known as the family of Construction
Grammars (cf. Östman and Fried, 2004, p. 1) has given rise to concrete views of what
constitutes a construction, their differences being mostly a mere matter of varying
goals and emphasis.11
For the LCM, for example, a construction is:
17
A form-meaning (or function) paring where form affords access to meaning and meaning is
realized by form to the extent that such processes have become entrenched, through sufficient
use, in the speaker’s mind and are generally recognized by competent speakers of the
language in question to be stably associated or at least potentially replicable by other
competent speakers of the same language with immaterial variation in its form and meaning.
(Ruiz de Mendoza, 2013, p. 238)
Authors like Boas (2008, p. 21) contend that in the case of the resultative,
constructions should be better conceived of as ‘mini-constructions’, i.e. form-meaning
pairings representing an individual sense of a verb. For instance, Boas (2011, p. 1288)
claims that the “make-safe” collocation (e.g. He made his pistol safe) is a miniconstruction, that is, a paring of a particular semantics/pragmatics with a very specific
syntactic frame [NP V NP safe]. In turn, Goldberg’s (1995) delineation of
construction was modified and expanded in Goldberg (2006). The original
formulation revolved around the criteria of idiosyncrasy and non-compositionality as
essential requirements for a construction to be recognized as such. This is clearly the
case of What’s X doing Y? (Kay and Fillmore, 1999), which conveys the idea that the
state of affairs referred to by the predication bothers the speaker. This meaning
ingredient is not directly derivable from the components of the construction.
However, owing to the rise of usage-based linguistics (Langacker, 1987; Bybee, 2006,
2013), Goldberg’s 2006 definition no longer considers compositionality to be the sole
requirement for constructional status, and transparent configurations (e.g. Hello) are
listed as constructions as long as they occur with sufficient frequency:
Any linguistic pattern is recognized as a construction as long as some aspect of its form or
function is not strictly predictable from its components parts or form other constructions
recognized to exist. In addition, patterns are stored as constructions even if they are fully
predictable as long as they occur with sufficient frequency.
(Goldberg, 2006, p. 5)
In this statement little is said about what actually counts as “sufficient frequency”, an
issue which the LCM has sorted out by arguing in favor of the replicability criterion
(Ruiz de Mendoza, 2013; Ruiz de Mendoza and Galera, 2014, p. 36-37). Following
upon this logic, an original form-meaning pairing can be regarded as a construction
even if such a pattern is only uttered once by a competent native speaker with the
18
proviso that an equally competent native hearer understands it and perceives it as a
highly natural output. Likewise, some criticisms have been voiced against the lack of
felicity of Goldberg’s (2006) approach on the grounds that virtually almost anything
can be a construction (see Östman and Fried, 2004, p. 1; Bod, 2009, p. 130; Mairal
and Gonzálvez García, 2010, p. 125). FunGramKB joins in with this position and
contends that Goldberg’s (2006, p. 5) view that “all levels of grammatical analysis
involve constructions” (e.g. from morphemes like –ing, going through words,
complex words, to abstract structures such as the passive, etc.) cannot be applied
computationally.12 Consequently, in the computational approach to constructional
meaning pursued in this knowledge-base, a clear-cut distinction is made between
“constructs” and “constructions”. According to Periñán (2013, p. 215; see also
Periñán and Arcas, 2014, p. 172), the former alludes to “any form-meaning pairing
which serves as a building block in the compositionality of sentential semantics”. 13 In
turn, the latter is taken to be any linguistic construct where the meaning of the whole
exceeds the sum of the parts.14 It is in this sense that not all levels of grammatical
analysis can be said to include constructions in FunGramKB, since constructional
meaning must be located in one of the constructicons (the four levels of the LCM),
“where constructions are not found below the argumental layer” (Periñán, 2013, p.
217, our emphasis).
Given this claim, the following theoretical issues need to be addressed. First, if
constructions in FunGramKB are to be understood from a holistic perspective à la
Goldberg (1995) and constructions can only exist in the Grammaticon, it seems
somewhat incongruent to provide “Kernel Constructions”, which reside in the
Lexicon (cf. Section 2), with the status of “constructions”. A simple transitive action
like Bob kicked the ball is fully compositional, and yet, it is given the status of a
construction (a Kernel-2 construction, to be more precise). Or consider the following
example taken from Periñán and Arcas (2014, p. 172): John pounded the nail flat into
the wall. According to these authors, such an input text combines three argumentstructure constructions: a Kernel-2 construction (“John pounded the nail”), the
transitive-resultative (“flat”) and the caused-motion configuration (“into the wall”).
Surprisingly, the Kernel-2 is claimed to be an argument-structure construction,
although not only does it reside in the Lexicon, but also its meaning as a whole is
clearly not greater than the sum of the parts involved in the utterance. Second,
ditransitive structures or Kernel-3, in the terminology of FunGramKB, are confined to
19
the Lexicon and, obviously, no ditransitive construction exists in the Grammaticon. If
this is so, how could the machine process an input text like Pat kicked Bob the
football? Evidently, such a non-compositional structure would need to be accounted
for at the argumental layer (the L1-Constructicon), since the AVM of kick consists of
a Kernel-2 and therefore, only selects two participants (i.e. ‘somebody kicks an
object’). Moreover, kick is a semelfactive verb derived from an activity, but, when the
verb is incorporated into the ditransitive, its Aktionsart is converted through coercion
into a causative accomplishment. These problems call for a revision of the status of
the notion of Kernel constructions in FunGramKB (which should probably be labeled
“Kernel-constructs” instead), while the inclusion of the ditransitive in the L1Constructicon is imperative.
A final query concerns which of the three linguistic approaches to the notion
of construction, to wit, the LCM’s, Boas’s or Goldberg’s is closer to the
computational account of argument-structure constructions provided in FunGramKB.
Before dealing with this issue, the next section is devoted to a computational
treatment of the resultative, caused-motion and way constructions; these three
patterns, which are related through family resemblance (in the sense of Wittgenstein
1955), express the result of the action denoted by the verb.
3. A formalized analysis of three argument structure constructions in
FunGramKB
The resultative construction has been the object of a considerable number of studies
from different perspectives (e.g. Hoekstra, 1988; Levin, 1993, 2006; Rappaport and
Levin, 2001, inter alios). More recently, this pattern has been investigated in a more
productive way in constructionist approaches such as the ones by Boas (2003, 2008,
2011), Broccias (2003), Goldberg and Jackendoff (2004), Iwata (2006), and Hampe
(2010). Broadly speaking, the resultative phrase designates the endpoint of a change
of state undergone by the patient, as exemplified by He broke the box open (Google
Books Corpus, 2010) and The wind broke the windows to pieces (Google Books
Corpus, 2009). In the former instance, the resultative phrase is realized by an
adjectival phrase (AP). In the latter, it takes the form of a Prepositional Phrase (PP),
whereby the change of state is metaphorical perceived as a change of location on the
grounds of the metaphor A CHANGE OF STATE IS A CHANGE OF LOCATIONS
(cf. Lakoff, 1987, 1993; Lakoff and Johnson, 1999).
20
The formalized representation of this construction in FunGramKB is given in
Figure 5:
Figure 5: AVM of the transitive resultative construction.
As shown in Mairal (2012), the properties of this constructional AVM specify the
Aktionsart characterization of the construction as a causative accomplishment. The
construction comprises three participants, i.e. x, y and z, the latter being the variable
contributed by the configuration, whose thematic role is ‘result’ and whose selectional
preference is the concept +STATE_00. As previously noted, this argument is formally
realized as an AP (ADJP) or a PP, syntactically functioning as a nucleus in the
layered structure of the clause. Finally, the semantics of the construction in the
COREL schema reads as follows: ‘there is an event in which something (x1: x) causes
the affected object (x3: y) to result in (x4: z)’. With an input text like the oft-quoted
example He hammered the metal flat, the constructional AVM in Figure 5 would
straightforwardly merge with the AVM of the predicate ‘hammer’, to which z is
added. That is, hammer has two arguments, one acting as the agent (x) and the other
21
as the object (y), but the result phrase comes from the higher-level structure (i.e. the
constructional AVM) with which this verb is compatible. The same rationale applies
to the example Kelly painted the walls red. Here, Agent-Subject and Patient-DO
(following van Trijp’s (2011) notation) are the values selected or inherited from the
AVM of paint in the Lexicon and, thus, they need not be further specified in the
constructional AVM of the transitive-resultative, which only requires the overt
expression of a result component. However, Mairal’s (2012) operationalization of the
transitive resultative does not to do justice to the actual complexity of lexicalconstructional integration. This is evidenced by the existence of realizations like He
painted the brush to pieces (Broccias, 2003, p. 134), in which once again, it is the
instrument employed to perform the action that is construed as the patient undergoing
a change of state. In this connection, Boas (2003) has shown that the resultative is a
highly irregular pattern, which goes far beyond the rather simplistic “hammering”
example. To begin with, verbal predicates tend to exhibit idiosyncratic behavior when
incorporated into the resultative syntactic frame, and resist subtle lexical or syntactic
changes, as shown in (4):
(4)
a. He hammered the metal flat.
b. He hammered the metal into a flat sheet (Culicover, 2009, p. 302).
c. ?He hammered the metal to flatness.
d. *He hammered the metal safe.15
e. *He hammered the metal tubular.
Likewise, more often than not, when the verb collocates with direct objects such as
those in (5a)-(5d), a figurative interpretation of the whole utterance is rendered:
(5)
a. The joggers ran the pavement thin (Levin and Rappaport, 1995, p.
75).
b. She cried herself sick (Google Books Corpus, 2010).
c. They drank the pub dry (Broccias, 2003, p. 198).
d. He swept the broom to pieces (Boas, 2003, p. 7).
Clearly, one does not “run the pavement” or “cry oneself” but rather, quite
exaggeratedly, one runs to a point in which the pavement is made thinner, cries so
22
much that ends up sick, etc. In all these realizations, whether the verb is intransitive
(e.g. cry (x)) or transitive (e.g. sweep (x, y)), neither the direct objects in (5), nor the
result ingredient arise from the argument-structure characterization of these verbal
predicates, and must therefore be seen as part of the construction (cf. *She cried
herself, *He ran the pavement, etc.). By contrast in He hammered the metal flat/into a
flat sheet, the construction only supplies the result element (i.e. ‘z = become flat’).
Moreover, note that, as already mentioned, the same construction can equally impose
a result component alone (cf. Tom swept the floor clean, Tom painted the walls red),
or an object plus the outcome event (cf. Tom painted the brush to pieces) upon the
same verbal predicate.
Let us now see how this variable behavior can be handled in FunGramKB. At
this point of our discussion the following query is in order: how many variables
should be left underspecified in the constructional AVM (e.g. x, y in Figure 5) for the
machine to understand that the information contained in such variables (x, y) is to be
retrieved from the lexical entry of the verb? Or, in more simple words, how can the
machine recognize when the argument-structure of a given verb will be augmented
through the addition of one or two constructional arguments? In FunGramKB, the
system understands that a verb can participate in a construction like the transitiveresultative by means of a default procedure called ‘pattern-matching’, which is based
on the immediate recognition of a constructional pattern when the number and type of
lexical and constructional arguments coincide. For example, in the AVM of Figure 5,
z is the sole argument provided by the construction. However, run, cry, sweep and
drink in the input texts in (5) require that two constructional arguments be specified in
the constructional AVM to account for the fact that neither their objects nor their
results are part of the argument structure of these predicates. Otherwise, the matching
between the construction and the verbs in (5) will not take place. However, if we
further detail the y constructional variable in the AVM of Figure 5, that is, if both y
and z are taken to be the arguments added by the construction, pattern-matching will
again be blocked for cases in which only the result component is contributed, e.g. He
hammered the metal flat, They painted the walls red, She broke the vase to pieces, etc.
Here, “the metal”, “the walls” or “the vase” are retrieved from the Lexicon so that
there is no need to describe a new variable y in the construction.
This situation brings about the creation of two distinct, yet related, subconstructions in the Grammaticon of FunGramKB. This is an issue that has thus far
23
gone unnoticed in the FunGramKB literature (cf. Mairal, 2012, fc.; Periñán, 2013;
Mairal and Periñán, 2014; Periñán and Arcas, 2014; Van Valin and Mairal, 2014).
That is, since pattern-matching must be satisfied, the resulting picture is that separate
constructional schemas need to be listed in order for the machine to be able to process
different realizations of the same general construction. On the one hand, Figure 5,
which we shall rename “transitive-resultative construction (type 1)”, will be able to
account for instances such as I froze the cake solid (Google Books, 2011), She fried
the meat black (Google Books, 2008), He smashed the cup to bits16 (Google Books,
2013), etc. In these cases, only the z constructional argument is detailed in the AVM.
On the other, the AVM in Figure 6 (i.e. the transitive-resultative (type 2)) will be
activated when the input text is a realization in which both y and z are arguments of
the construction, as in (5) above.17
Figure 6: AVM of the transitive resultative (type 2).
24
Another case is point is that of the caused-motion construction, a
configuration whose basic semantic layout expresses a caused change of location
along a path designated by the directional phrase (cf. Goldberg, 1995, p. 152; see also
Torre, 2012; Ruiz de Mendoza and Luzondo, 2014). As with the resultative, the
caused-motion in (6a) may contribute the directional phrase (“into the net”), or, in the
examples in (6bc), it can supply both the theme argument (i.e. “the rats/him”) and the
destination of motion (i.e. “into the maze/room”), which are not part of the
intransitive verb run or of the two-place predicate show.
(6)
a. He kicked the ball into the net.
b. The scientist ran the rats into the maze (Levin, 1993, p. 31).
c. He showed him into the room (Goldberg, 1995, p. 162).
Once again, if the maximum number of added arguments (i.e. y and z) is detailed
within a single construction schema, the construction will not be triggered for an input
text such as (6a) or for others like He shoved the boat into the water (Google Books
Corpus, 2000), He poured the batter into the pan (Google Books Corpus, 2011), He
drove the car into the parking lot (Google Books Corpus, 2006), etc. Conversely, the
description of the z argument alone is insufficient for The scientist ran the rats into
the maze, He walked the horse to the stables (Google Books Corpus, 2013), among
others. Thus, two caused-motion types must be incorporated into the L1-Construction.
These are respectively reproduced in Figures 7 and 8. In both AVMs, the Aktionsart
specification is a causative active accomplishment, z maintains the same descriptors
and constraints, while the COREL scheme is also shared (i.e. ‘there is an event and as
a result the agent causes the theme to move towards a destination’). The difference
between these AVMs lies in the specification of one constructional argument (Figure
7) or two (Figure 8):
25
Figure 7: AVM of the caused-motion construction (type 1).
26
Figure 8: AVM of the caused-motion construction (type 2).
Now consider the way construction (cf. Kuno and Takami, 2004; Christy,
2011; Luzondo, 2013): He elbowed his way into the lottery room; He gambled his
way to fortune; He sighed his way to sleep again. Goldberg (1995, p. 199) represents
the way construction as [SUBJi [V [POSSi way] OBL]], where V is a non-stative verb,
OBL codes a directional and the noun way is a fixed or non-parameterizable element,
in the terminology of the LCM. In contrast to the caused-motion construction, one
essential characteristic of the way construction, according to Goldberg (1995), relates
to the fact that motion must take place along an either literal or metaphorical selfcreated, non pre-established path in which some kind of external obstacle is present.
Moreover, while the resultative and caused-motion constructions can either be literal
(e.g. He painted the wall red, He kicked the ball into the net) or figurative (e.g. She
drank herself silly/into oblivion), the combination of verb and the fixed-element
[POSSi way] always renders a figurative interpretation (e.g. one cannot “elbow a
27
way”, but rather create a way by pushing people aside with one’s elbows). A final
difference among these structures is the following. Recall that in the intransitive
motion and the two types of caused-motion constructions described herein the result
is that of a literal change of location. In the resultative, types 1 and 2, the outcome is a
change of state, whether expressed through and adjective (e.g. He painted the wall
red), or through a PP used metaphorically (e.g. He broke the vase to pieces), as we
noted above. The way construction, however, behaves differently, since it may realize
a result of the action denoted by the verb either in the form of an AP and/or a PP as a
change of state (e.g. (7a)-(7b)), or through a PP expressing a literal change of
location, as in (7c):
(7)
a. Jeffrey and Lemondrop chewed their way loose (Google Books
Corpus, 2013).
b. They snored the way into oblivion (Google Books Corpus, 2010).
c. Gilbert Rasher pushed his way into the room (Google Books Corpus,
2004).
This is a problem for the current design of the Grammaticon. If we go back to the
interface of the L1-Constructicon in Figure 3, we notice that each new variable
contributed by the construction must select one and only one descriptor and one and
only one constraint. For instance, in (7a)-(7c), the construction collaborates with the
verb by adding both the fixed element (i.e. “X’s way”) and the result component (z =
“loose”; “into oblivion”; “into the room”). Thus, y and z need to be spelled out in the
interface. In (7ab), however, the thematic role of z would be that of “result” and
+STATE_00 would work as its selectional preference. By contrast, z in (7c) requires
that the descriptor labeled thematic role be “goal” while its selectional preference
must express a change of location (i.e. +LOCATION_00). Since, currently, the
FunGramKB
editor
does
not
allow
the
specification
of
different
descriptors/constraints for the same constructional variable (i.e. z), we are once again
faced with a situation in which two types of way construction must be posited. Both
sub-AVMS share the following elements: (i) the Aktionsart specification (i.e.
causative active accomplishment); (ii) the y variable; (iii) the COREL schema (see
(10) below). Their differences are shown in (8) and (9):
28
(8) Kowalski fought his way free/to freedom (Google Books Corpus,
2007/2006).
- Constructional variable = z
- Role = Result
- Phrase = ADJP | PP
- Syntax = Nucleus
- Preference = +STATE_00
(9) The crusaders fought their way into the palace (Google Books Corpus,
2000).
- Constructional variable = z
- Role = Goal
- Phrase = PP
- Syntax = Nucleus
- Preference = +LOCATION_00
Even if two constructional schemas are provided, the editor lacks a box in which the
fixed element can be recorded. For this reason, it is still a pending task to determine
how the machine will be able to recognize the existence of this component, which
cannot be explicitly stated anywhere in the L1-Constructicon.
A final issue concerns the meaning of the construction, which is captured in
the COREL schema in (10). Here, we have tried to account for the fact the
construction combines creation expressions (e.g. He made a path) with the
intransitive motion construction (cf. Goldberg, 1995, p. 207):
(10)
a. *(e1: +MOVE_00 (x1: x)Agent (x2: x)Theme (x3)Location
(x4)Origin (x5: z)Goal (f1: +DIFFICULT_00)Manner (f2:
<EVENT>)Means (f3: (e2: +CREATE_00 (x1)Theme (x6:
y)Referent))Result)
b. An entity moves towards a destination with difficulty by means of
EVENT (e.g. fighting) and as a result the entity in motion creates y.
Goldberg contends that the way construction “is used to convey that the
subject moves despite some external difficulty, or in some indirect way: the path is not
29
already established, but must be (…) created by the mover” (Goldberg, 1995, p. 204,
our emphasis). On the basis of corpus data, Luzondo (2013) showed that this
statement is true for prototypical examples of the way construction (e.g. He elbowed
her way into the room), but sometimes there is no such impediment. By way of
illustration, consider the following attested realizations:
(11)
a. Santas jingled and ho-ho-hoed their way into the home bearing gifts
(Corpus of Contemporary American English, 2008).
b. They giggled their way out of the restaurant sometime around 11
PM (Google Books Corpus, 2008).
c. (…) The four of them ate in a cheap dinner where Annete was so put
off by the food that she left her spaghetti untouched and smoked her
way through the meal (Google Books Corpus, 2005).
d. Her hand went up to my shoulder and her fingers ran their way
around my neck (Google Books Corpus, 1990).
As the examples in (11) demonstrate, the way construction can be employed to
construe activities such as ‘giggling’, ‘smoking’, etc. in terms of unimpeded motion
forward along a path. Thus, in these cases, the function of the fixed element is to
highlight the progressive, continuous aspect of the motion event. In the COREL
representation in (10), the external “difficulty” element described in Goldberg (1995)
has been coded via a Manner satellite (i.e. “(f1: +DIFFICULT_00)Manner”).
Nonetheless, it is necessary to allow for the fact that not all realizations of the way
construction involve an obstacle. In FunGramKB, a predication (e1) can follow two
kinds of reasoning operations, to wit, strict (+) or defeasible (*). According to Periñán
and Arcas (2010a) the former allow no exceptions (e.g. islands are surrounded by
water, a cat is a mammal, etc.), while the latter may be overridden in the light of
contradictory information, as is the case of the impediment element of the way
construction. We thus propose to mark the predication in (10) with an asterisk instead
of a plus symbol (see, by contrast, the COREL scheme in Figure 8).
Finally note that although the level of formalization in computational
linguistics is high, FunGramKB does not suffer from one of the drawbacks pointed
out by Goldberg (2006). For this author, unification-based approaches “are not
sufficiently amenable to capturing detailed lexical semantic properties” (Goldberg,
30
2006, p. 216). In this view, the use of feature-based systems for representing
semantics cannot account for frame-based or encyclopedic knowledge. FunGramKB
does employ unification operations to merge lexical and constructional AVMs, but
since an ontological approach is adopted, both the Grammaticon and the Lexicon
retrieve information from the Ontology. This ensures that encyclopedic knowledge is
in fact captured. For example, through the two reasoning mechanisms employed in
FunGramKB, i.e. inheritance and inference (cf. Periñán and Mairal, 2010), the
concept +MOVE_00 in the COREL scheme of the caused-motion construction is
connected to the meaning postulates (i.e. the semantics) of concepts such as
+PATH_00, +SPEED_00, +DIRECTION_00, +FOLLOW, +RUN_00, +STOP_00,
etc., thus leading to the creation of rich frames of knowledge such as those one can
find in FrameNet (see Luzondo and Jiménez, 2014 for a discussion).
4. Linking linguistic theory with FunGramKB
Having explained how constructions are handled in FunGramKB, we would now like
to close this article by establishing a connection between the way FunGramKB
approaches constructional schemas and the distinct, yet complementary, linguistic
treatment of argument-structure constructions proposed by Goldberg and by Boas, or
in a meaning-construction model like the LCM (cf. Section 2.1.1).
In light of the constraints imposed by the machine, which result in the creation
of constructional subtypes (e.g. way construction type 1, way construction type 2,
etc.), it is interesting to observe that such a procedure brings FunGramKB somewhat
closer to a mini-constructionist approach of the kind propounded by Boas (2003), than
to the Goldbergian (Goldberg 1995, 2006) account. To illustrate, take Goldberg’s
(1995, p. 190) representation of the anatomy of the resultative construction (Figure 9):
31
Figure 9 shows that the construction is associated with the semantics ‘X
CAUSES Y TO BECOME Z’, here represented as CAUSE-BECOME <agt pat
result-goal>, where the elements between brackets are the argument roles of the
construction. PRED, in turn, is the variable that will be filled by a given verb together
with its participant roles (e.g. RUN <runner>). These participants fuse with the
argument roles of the construction in a principled manner.18 Solid lines between the
argument roles and the participant roles of the verb indicate that such roles must be
obligatorily fused, while dashed lines are employed to specify roles that can be
contributed by the construction, in this case, “patient” and “result-goal”. At the
explanatory level, Goldberg (1995, p. 190) contends that the construction can add the
result-goal argument when the construction incorporates a verb like wipe, for
example, He wiped the table clean; alternatively, both the patient and result-goal roles
may be supplied, as in He talked himself blue in the face. Nevertheless, the dashed
lines in Figure 9 show that, at the descriptive level, Goldberg’s schematic
representation accounts for the possibility of contributing both the y and z arguments
(instead of only the latter) under the same general pattern. This skeletal structure
corresponds to the AVM of the resultative-type 2, leaving the resultative-type 1 out of
the picture and, with it, the possibility of matching an input text that only contributes
the result component. Since both general principles (i.e. the ‘Semantic Coherence
Principle’ and the ‘Correspondence Principle’) and construction-specific constraints
regulate fusion mechanisms, Goldberg’s (1995) approach does not need to postulate
sub-types of constructions, thus maintaining a level of generalization that cannot be
32
translated into the Grammaticon of FunGramKB for the reasons specified thus far.
Therefore, in line with scholars like Boas (2003, 2013), Iwata (2008), Nemoto (1998),
inter alios, Goldberg’s abstract meaningful constructions turn out to be too broadranging and, in consequence, for our purposes here, they are not always suitable for
an NLP environment which requires different constructional schemas to process input
texts that linguistically realize the same construction.
A similar situation holds for the LCM. In this model, subsumption, which is a
cognitive process whereby lower-level structure is built into higher-level structure, is
regulated by a number of both ‘internal’ and/or ‘external’ constraints. The former
“work on the basis of the compatibility between the conceptual characterizations of
lexical predicates and argument-structure constructions” (Ruiz de Mendoza, 2013, p.
256). Likewise, cognitive phenomena such as high-level metaphor and metonymy are
treated as external constraining factors whose recurrent presence in many of the
argument-structure constructions discussed in the literature also affect subsumption
processes by either permitting or disallowing them (see Ruiz de Mendoza and Pérez,
2001; Peña, 2009 for examples and details).19 In other words, external constraints are
based on how lexical structure can be re-construed in order to make it fit into a given
construction. For example, the verb stare can be used with the caused-motion
construction in She stared me out of the room if the target of staring is re-construed as
the object of an effectual action (cf. Ruiz de Mendoza and Mairal, 2008, p. 378 for a
discussion of related examples). In this sense, the LCM takes sides with Goldberg’s
proposal in that they both aim to provide an economic solution to the task of
delimiting lexical-constructional integration through the postulation of broad
constraints. To better understand this position, take two more instances of the causedmotion construction: They scared the horses out of the horse lot (Google Books
Corpus, 2012) and He laughed himself into the house (Google Books Corpus, 2011).
In the former example, which would correspond to the AVM in Figure 7, the
construction contributes the goal argument alone, while in the latter, both “himself”
and “into the house” are supplied by the grammatical pattern, thus coinciding with the
AVM in Figure 8. For the LCM, by contrast, there is no need to consider the existence
of separate constructions, but instead each realization is licensed via external
constraints. More concretely, They scared the horses out of the horse lot activates the
high-level metaphor AN EXPERIENTIAL ACTION IS AN EFFECTUAL ACTION,
whereby the horses, which are the targets of non-physical impact, are figuratively
33
understood as if they were the objects of actual physical impact. The grammatical
metaphor AN ACTIVITY IS AN EFFECTUAL ACTION is at work in He laughed
himself into the house. Here, the fake-reflexive is not treated as an affected object but
as the object of change (Ruiz de Mendoza and Mairal, 2007, p. 44). In other words,
whereas in The audience laughed the actor off the stage/out of the room, the actor is
both the target of the action and the object of a change, in He laughed himself into the
house, “himself” is not the one being laughed at, but simply the one laughing at
something while moving into the house. As things stand now, even though the
Grammaticon is grounded in the four constructional levels of the LCM, it seems
largely impossible to endow the machine with the apparatus of highly abstract
external constraints put forward by this linguistic model.
In conclusion, the constructionist account that better fits the necessities of the
machine, or the one that is better aligned with the computational analysis carried out
herein, is the one provided by Boas (2003, 2007, 2008, 2010, 2011), who proposes
mini-constructions, i.e. parings of a particular semantics/pragmatics with a very
specific syntactic frame. In FunGramKB, mini-constructions translate as follows:
each constructional sub-schema is conceived of as sharing certain properties (e.g. its
Aktionsart adscription, COREL schema, etc.) but also as differing in the number and
nature of the variables contributed by the construction, resulting in bundles of specific
constraints and specific descriptors, through which form and function are codified.
This means that, as is done in Boas (2011), our focus must be on a more concrete
level of analysis (i.e. constructional subtypes) in order for machine-based natural
language processing to be feasible. Note that in an NLP environment, however, form
and function are divided into several components, the form end of the construction
being captured by the boxes labeled “phrase realization” and “syntax”, while the
meaning end is accounted for by the rest of the descriptors and constraints.
5. Conclusions
This article has presented readers with the essentials of the relatively recent NLP
project called FunGramKB and, more concretely, with the architecture of the L1Constructicon, the component where constructional schemas are represented by
means of AVMs consisting of a set of highly specific descriptors and constraints. Our
aim was to show how argument-structure constructions are handled in a
computational environment. In order to do so, we have examined in some detail three
34
well-known grammatical patterns, namely, the resultative, the caused-motion and the
way construction. Upon analyzing the complex behavior of such configurations in
light of the restrictions imposed by the machine, we have concluded that it is
necessary to descend to a more concrete level of analysis in which general
constructions are split into more specific sub-types of constructional schemas, much
in line with Boas’s approach to constructional behavior. The study of additional
argument-structure patterns will confirm or discard this proposal.
Finally, although one may argue that the account presented here may lead to
the over-generation of constructional schemas, the ultimate computational goal is not
to find a highly economic solution to lexical-constructional merging, à la Goldberg,
but rather to ensure that natural language understanding is in fact accomplished.
Acknowledgement
The research on which this paper is based has received financial support from the
Spanish Ministry of Economy and Competitiveness, grants no. FFI2011-29798-C0201 and FFI2013-43593-P.
References
Barðdal, J., Kristoffersen, K., Sveen, A., 2011. West Scandinavian ditransitives as a
family of constructions: With a special attention to the Norwegian V-REFL-NP
construction. Linguistics 49(1), 53–104.
Basili, R., Pazienza, M. T., Velardi, P., 1996. An empirical symbolic approach to
natural language processing. Artificial Intelligence 85(1996), 59–99.
Bencini, G. M. L., Goldberg, A., 2000. The contribution of argument structure
constructions to sentence meaning. Journal of Memory and Language 43, 640–
651.
Bergen, B. K., Chang, N., 2005. Embodied Construction Grammar in simulationbased language understanding. In: Östman, J. O., Fried, M. (Eds.), Construction
Grammar(s): Cognitive Grounding and Theoretical Extensions. John
Benjamins, Amsterdam/Philadelphia, pp. 147–190.
Boas, H. C., 2003. A Constructional Approach to Resultatives. CSLI Publications,
Stanford.
Boas, H. C., 2007. Construction Grammar in the twenty-first century. English
Language and Linguistics 11(3), 569–585.
Boas, H. C., 2008. Determining the structure of lexical entries and grammatical
constructions in Construction Grammar. Annual Review of Cognitive
Linguistics 6, 113–144.
35
Boas, H. C., 2010. Linguistically relevant meaning elements of English
communication verbs. Belgian Journal of Linguistics 24, 54–82.
Boas, H. C., 2011. Coercion and leaking argument structure in Construction
Grammar. Linguistics 49(6), 1271–1303.
Boas, H. C., 2013. Cognitive Construction Grammar. In: Hoffmann, T., Trousdale,
G., (Eds.), The Oxford Handbook of Construction Grammar. Oxford University
Press, Oxford, pp. 233–254.
Boas, H. C., 2014. Lexical and phrasal approaches to argument structure: Two sides
of the same coin. Theoretical Linguistics 40(1-2), 89-112.
Boas, H. C., Sag, I., (Eds.). 2012. Sign-Based Construction Grammar. CSLI
Publications, Stanford.
Bod, R., 2009. Constructions at work or at rest? Cognitive Linguistics 20(1), 129-134.
Broccias, C., 2003. The English Change Network: Forcing Changes into Schemas.
Mouton de Gruyter, Berlin/New York.
Butler, C. S., Gonzálvez-García, F., 2014. Exploring Cognitive-Functional Space.
John Benjamins, Amsterdam/Philadelphia.
Bybee, J., 2006. From usage to grammar: The mind’s response to repetition.
Language 82(4), 711-733.
Bybee, J., 2013. Usage-based theory and exemplar representations of constructions.
In: Hoffmann, T., Trousdale, G. (Eds.), The Oxford Handbook of Construction
Grammar. Oxford University Press, Oxford, pp. 49–69
Chang, N., De Beule, J., Micelli, V., 2012. Computational Construction Grammar:
Comparing ECG and FCG. In Steels, L. (Ed.), Computational Issues in Fluid
Construction Grammar. Springer, Berlin, pp. 259–288.
Christy, E., 2011. Investigating the differences between the English Way-construction
and the fake reflexive resultative construction. In: Amstrong, L. (Ed.),
Proceedings of the 2011 annual conference of the Canadian Linguistic
Association, pp. 1-14.
Cimiano, P., Unger, C., McCrae, J., 2014. Ontology-Based Interpretation of Natural
Language. Morgan & Claypool Publishers, San Rafael.
Culicover, P., 2009. Natural Language Syntax. Oxford University Press, Oxford.
Diedrichsen, E., 2010. Towards a reconsideration of constructional schemas in RRG:
Are all constructions driven by “constructions”? In: Nakamura, W. (Ed.),
Proceedings of the 10th International Conference on Role and Reference
Grammar
(RRG
2009),
pp.
52–79.
<http://
linguistics.buffalo.edu/people/faculty/vanvalin/rrg.html>
Diedrichsen, E., 2011. The theoretical importance of constructional schemas in RRG.
In: Nakamura, W. (Ed.), New Perspectives in Role and Reference Grammar.
Cambridge Scholars, Newcastle upon Tyne, pp. 168–197.
Dirven, R., Ruiz de Mendoza, F., 2010. Looking back at 30 years of Cognitive
Linguistics. In: Tabakowska, E., Choiński, M., Wiraszka, Ł. (Eds.), Cognitive
Linguistics in Action: From Theory to Application and Back. Mouton de
Gruyter, Berlin/New York, pp. 13-70.
Fillmore, C. J., Kay, P., O’Connor, M. C., 1988. Regularity and idiomaticity in
grammatical constructions: The case of let alone. Language 64, 501–538.
Goldberg, A., 1995. Constructions. A Construction Grammar Approach to Argument
Structure. University of Chicago Press, Chicago/London.
Goldberg, A., 2006. Constructions at Work: The Nature of Generalization in
Language. Oxford University Press, Oxford.
36
Goldberg, A., Jackendoff, R., 2004. The English resultative as a family of
constructions. Language 80(3), 532–568.
Ghomeshi, J., Jackendoff, R., Rosen, N., Russell, K., 2004. Contrastive Focus
Reduplication in English (the salad-salad paper). Natural Language and
Linguistic Theory 22, 307-357.
Gonzálvez-García, F., 2008. Construction Grammar works: An interview with Adele
E. Goldberg. Annual Review of Cognitive Linguistics 6, 345-360.
Gonzálvez-García, F., 2009. The family of object-related depictives in English and
Spanish: Towards a constructionist, usage-based analysis. Language Sciences
31(5), 663–723.
Gonzálvez-García, F., 2014. “That’s so a construction!”: Some reflections on
innovative uses of “so” in Present-day English. In: Gómez González, M. A.,
Ruiz de Mendoza Ibáñez, F., Gonzálvez-García, F. (Eds.), Theory and Practice
in Functional-Cognitive Space. John Benjamins, Amsterdam/Philadelphia, pp.
271–294.
Hampe, B., 2010. Metaphor, constructional ambiguity and the causative resultatives.
In: Handl, S., Schmid, H.-J. (Eds.), Windows to the Mind: Metaphor,
Metonymy and Conceptual Blending. Mouton de Gruyter, Berlin/New York,
pp. 185–215.
Heyvaert, L., 2003. A Cognitive-Functional Approach to Nominalization in English.
Mouton de Gruyter, Berlin/New York.
Hoekstra, T., 1988. Small clause results. Lingua 74, 101–139.
Hoffmann, T., Trousdale, G. (Eds.). 2013. The Oxford Handbook of Construction
Grammar. Oxford University Press, Oxford.
Iwata, S., 2006. Argument resultatives and adjunct resultatives in a lexical
constructional account: The case of resultatives with adjectival result phrases.
Language Sciences 28(5), 449–496.
Iwata, S., 2008. Locative Alternation. A Lexical-Constructional Approach. John
Benjamins, Amsterdam/Philadelphia.
Jiménez, R. Luzondo, A., 2013. Constructions in Role and Reference Grammar: The
case of the English resultative. In: Nolan, B., Diedrichsen, E. (Eds.); Linking
Constructions into Functional Linguistics. The Role of Constructions in
Grammar. John Benjamins, Amsterdam/Philadelphia, pp. 179-204.
Jiménez, R., Pérez Cabello de Alba, M. B., 2011. An account of selection restrictions
in Role and Reference Grammar. Revista Canaria de Estudios Ingleses 62, 99122.
Kay, P., 2013. The limits of Construction Grammar. In: Hoffmann, T., Trousdale, G.
(Eds.), The Oxford Handbook of Construction Grammar. Oxford University
Press, Oxford, pp. 32–48.
Kay, P., Fillmore, C. J., 1999. Grammatical constructions and linguistic
generalizations: The What’s X doing Y? construction. Language 75(1): 1-33.
Kuno, S., Takami, K.-i., 2004. Functional Constraints in Grammar: On the
Unergative-Unaccusative
Distinction.
John
Benjamins,
Amsterdam/Philadelphia.
Lakoff, G., 1987. Women, Fire, and Dangerous Things: What Categories Reveal
about the Mind. The University of Chicago Press, Chicago/London.
Lakoff, G., 1993. The contemporary theory of metaphor. In: Ortony, A. (Ed.),
Metaphor and Thought (2nd ed.). Cambridge University Press, Cambridge, pp.
202–251.
37
Lakoff, G., Johnson, M., 1999. Philosophy in the Flesh: The Embodied Mind and its
Challenge to Western Thought. Basic Books, New York.
Langacker, R. W., 1987. Foundations of Cognitive Grammar, vol. 1: Theoretical
Prerequisites. Stanford University Press, Stanford.
Levin, B., 1993. English Verb Classes and Alternations. A Preliminary Investigation.
University of Chicago Press, Chicago/London.
Levin, B., 2006. English object alternations: A unified account. Unpublished ms..
Stanford.
Levin, B., Rappaport Hovav, M., 1995. Unaccusativity: At the Syntax-Lexical
Semantics Interface. MIT Press, Cambridge, MA.
Luzondo, A., 2013. Revisiting Goldberg’s semantic constraints on the way
construction. Revista Española de Lingüística Aplicada 26(2013), 349-364.
Luzondo, A., 2014. Constraining factors on the family of resultative constructions.
Review of Cognitive Linguistics 12(1), 30-63.
Luzondo, A., Jiménez, R., 2014. FrameNet and FunGramKB: A comparison of two
computational resources for semantic knowledge representation. In: Nolan, B.,
Periñán, C. (Eds.), Language Processing and Grammars. The Role of
Functionally
Oriented
Computational
Models.
John
Benjamins,
Amsterdam/Philadelphia, pp. 197-231.
Mairal, R., 2012. La arquitectura de una base de conocimiento léxico conceptual:
Implicaciones lingüísticas. In: Giammatteo, M., Ferrari, L., Albano, H.
(Eds.), Léxico y Sintaxis. FFyL, UNCuyo, pp. 183-210.
Mairal, R., fc. Constructional meaning representation within a knowledge
engineering framework. Review of Cognitive Linguistics.
Mairal, R., Gonzálvez-García, F., 2010. Verbos y construcciones en el espacio
cognitivo-funcional del siglo XXI. In: Val Álvaro, J. F., Horno Chéliz, M. C.
(Eds.), La Gramática del Sentido: Léxico y Sintaxis en la Encrucijada. Prensas
Universitarias de Zaragoza, Zaragoza, pp. 123-152.
Mairal, R., Periñán, P., 2009. The anatomy of the lexicon component within the
framework of a conceptual knowledge base. Revista Española de Lingüística
Aplicada 22, 217–244.
Mairal, R., Periñán, C., 2014. Representing constructional schemata in FunGramKB
Grammaticon. In: Fleischhauer, J., Latrouite, A., Osswald, R. (Eds.), Exploring
the Syntax-Semantics Interface. Düsseldorf University Press.
Mairal, R., Ruiz de Mendoza, F., 2009. Levels of description and explanation in
meaning construction. In Butler, C. S., Martín Arista, J. (Eds.), Deconstructing
Constructions. John Benjamins, Amsterdam/Philadelphia, pp. 153–198.
Nemoto, N., 1998. On the polysemy of ditransitive save: The role of frame semantics
in Construction Grammar. English Lingusitics 15, 219–242.
Nolan, B., 2011a. Meaning construction and grammatical inflection in the layered
structure of the Irish word: An RRG account of morphological constructions. In:
Nakamura, W. (Ed.), New Perspectives in Role and Reference Grammar,
Cambridge Scholars, Newcastle upon Tyne, pp. 64–101.
Nolan, B., 2011b. Constructions as grammatical objects: A new perspective on
constructions in RRG. Paper read at the 11th International Conference on Role
and Reference Grammar, Pontificia Universidad Católica de Chile, Santiago de
Chile, August 11–13.
Nolan, B., 2014. Theoretical and computational considerations of linking
constructions in Role and Reference Grammar. Review of Cognitive Linguistics
12(2), 410-442.
38
Nolan, B., Diedrichsen, E. (Eds.), 2013. Linking Constructions into Functional
Linguistics. The Role of Constructions in Grammar. John Benjamins,
Amsterdam/Philadelphia.
Nolan, B., Periñán, C. (Eds.), 2014. Language Processing and Grammars. The Role of
Functionally
Oriented
Computational
Models.
John
Benjamins,
Amsterdam/Philadelphia.
Östman, J.-O., Fried, M., 2004. Historical and intellectual background of
Construction grammar. In: Fried, M., Östman, J.-O. (Eds.), Construction
Grammar
in
a
Cross-Language
Perspective.
John
Benjamins,
Amsterdam/Philadelphia, pp. 1-10.
Peña, S., 2009. Constraints on subsumption in the caused-motion construction.
Language Sciences 31(6), 740–765.
Periñán, C., 2013. Towards a model of constructional meaning for natural language
understanding. In: Nolan, B., Diedrichsen, E. (Eds.), Linking Constructions into
Functional Linguistics: The Role of Constructions in Grammar. John
Benjamins, Amsterdam/Philadelphia, pp. 205–230
Periñán, C., Arcas, F., 2005. Microconceptual-Knowledge Spreading in FunGramKB.
Proceedings of the 9th IASTED International Conference on Artificial
Intelligence and Soft Computing. ACTA Press, pp. 239–244.
Periñán, C., Arcas, F., 2007a. Cognitive modules of an NLP knowledge base for
language understanding. Procesamiento del Lenguaje Natural 39, 197–204.
Periñán, C., Arcas, F., 2007b. Deep semantics in an NLP knowledge base.
Proceedings of the 12th Conference of the Spanish Association for Artificial
Intelligence. Universidad de Salamanca, pp. 279–288.
Periñán, C., Arcas, F.,
2010a. Ontological commitments in FunGramKB.
Procesamiento del Lenguaje Natural 44, 27–34.
Periñán, C., Arcas, F., 2010b. The architecture of FunGramKB. Proceedings of the
7th International Conference on Language Resources and Evaluation. European
Language Resources Association, pp. 2667–2674.
Periñán, C., Arcas, F., 2014. The implementation of the CLS constructor in
ARTEMIS. In: Nolan, B., Periñán, C. (Eds.), Language Processing and
Grammars. The Role of Functionally Oriented Computational Models. John
Benjamins, Amsterdam/Philadelphia, pp. 165-196.
Periñán, C., Mairal, R., 2010. La gramática de COREL: Un lenguaje de
representación conceptual. Onomázein 21, 11–45.
Rappaport Hovav, M., Levin, B.. 2001. An event structure account of English
resultatives. Language 77, 766-797.
Ruiz de Mendoza, F. J., 2001. Lingüística cognitiva: Semántica, pragmática y
construcciones.
Clac
8
[online].
Available
at:
http://www.ucm.es/info/circulo/no8/ruiz.htm.
Ruiz de Mendoza, F. J., 2007. High-level cognitive models: In search of a unified
framework for inferential and grammatical behavior. In: Kosecki, K. (Ed.),
Perspectives on Metonymy. Peter Lang, Frankfurt/Main, pp. 11–30.
Ruiz de Mendoza, F. J., 2013. Meaning construction, meaning interpretation and
formal expression in the Lexical Constructional Model. In: Nolan, B.,
Diedrichsen, E. (Eds.), Linking Constructions into Functional Linguistics: The
Role of Constructions in Grammar. John Benjamins, Amsterdam/Philadelphia,
pp. 231–270.
Ruiz de Mendoza, F. J., 2014. Low-level situational cognitive models within the
Lexical Constructional Model and their computational implementation in
39
FunGramKB. In: Nolan, B., Periñán, C. (Eds.), Language Processing and
Grammars: The Role of Functionally Oriented Computational Models. John
Benjamins, Amsterdam/Philadelphia, pp. 367–390.
Ruiz de Mendoza, F. J., Galera, A., 2014. Cognitive modeling. A Linguistic
Perspective. John Benjamins, Amsterdam/Philadelphia.
Ruiz de Mendoza, F. J., Luzondo, A., 2014. Figurative and non-figurative motion in
the expression of result in English. Language and Cognition, in press.
Ruiz de Mendoza, F. J., Mairal, R., 2007. High-level metaphor and metonymy in
meaning construction. In: Radden, G., Köpcke, K. M., Berg, Th., Siemund, P.
(Eds.), Aspects of Meaning Construction in Lexicon and Grammar. John
Benjamins, Amsterdam/Philadelphia, pp. 33–49.
Ruiz de Mendoza, F. J., Mairal, R., 2008. Levels of description and constraining
factors in meaning construction: An introduction to the Lexical Constructional
Model. Folia Linguistica 42(2), 355–400.
Ruiz de Mendoza, F. J., Mairal, R., 2011. Constraints on syntactic alternation:
lexical-constructional subsumption in the Lexical Constructional Model. In:
Guerrero, P. (Ed.), Morphosyntactic Alternations in English: Functional and
Cognitive Perspectives. Equinox, pp. 62–82.
Ruiz de Mendoza, F. J., Pérez, L., 2001. Metonymy and the grammar: Motivation,
constraints, and interaction. Language and Communication 21, 321–357.
Steels, L., 2012. Design methods for Fluid Construction Grammar. In: Steels, L.
(Ed.), Computational Issues in Fluid Construction Grammar. Springer, Berlin,
pp. 3–36.
Torre, E., 2012. Symmetry and asymmetry in Italian caused-motion constructions. An
Embodied Construction Grammar approach. Constructions 1, 1–38.
Van Trijp, R., 2011. A design pattern for argument structure constructions. In: Steels,
L. (Ed.), Design Patterns in Fluid Construction Grammar. John Benjamins,
Amsterdam/Philadelphia, pp. 115–145.
Van Trijp, R., 2013. A Comparison between Fluid Construction Grammar and SignBased Construction Grammar. Constructions and Frames 5(1), 88-116.
Van Valin, R., 2005. The Syntax-Semantics-Pragmatics Interface: An Introduction to
Role and Reference Grammar. Cambridge University Press, Cambridge.
Van Valin, R., LaPolla, R., 1997. Syntax: Structure, Meaning and Function.
Cambridge University Press, Cambridge.
Van Valin, R., Mairal, R., 2014. Interfacing the lexicon and an ontology in a linking
system. In: Gómez González, M. A., Ruiz de Mendoza Ibáñez, F. J.,
Gonzálvez-García, F. (Eds.), Theory and Practice in Functional-Cognitive
Space. John Benjamins, Amsterdam/Philadelphia, pp. 205-228.
Wechsler, S., 2005. Resultatives under the 'event-argument homomorphism' model of
telicity. In: Erteschik-Shir, N., Rapoport, T. (Eds.), The Syntax of Aspect.
Deriving Thematic and Aspectual Interpretation. Oxford University Press,
Oxford, pp. 255-273.
Wittgenstein, L. 1955. Philosophical Investigations. Basil Blackwell, Oxfrod.
40
1
The interested reader is further referred to Jiménez and Luzondo (2013) for an RRG account of the
English resultative. Departing from Van Valin’s (2005, p. 239) description of the resultative pattern,
the article offers a finer-grained analysis of such a configuration by drawing on the work on
constructional schemas carried out in Diedrichsen (2010, 2011), and Nolan (2011ab).
2
Google Books. Grow Fruits & Vegetables the Way They Used to Taste. John Festus Adams. 1988.
Accessed on Oct. 16, 2014.
3
Google Books. White Turtle. Merlinda Bobis. 1999. Accessed on Oct. 18, 2014.
4
In the field of NLP, probabilistic techniques have been successfully applied to a variety of natural
language processing problems (e.g. automatic translation, semantic disambiguation, etc.) and, as far as
coverage is concerned, they clearly outperform rationalist or symbolic methods. However, as noted in
Basili et al. (1996, p. 59), probabilistic models are “intrinsically unable to provide insights into the
mechanisms of human communication” and “eventually a human analyst is to make sense of the data”
(see also Cimiano et al., 2014, p. 6).
5
In FunGramKB, by a splitting approach (see Gonzálvez, 2008, p. 35; Boas, 2010) we specifically
refer to the need to generate finer-grained constructional schemas in order to adequately capture the
various ways in which verbs and constructions interact to yield specific input texts.
6
http://www.vivaveracruz.com/blog/?m=201010. Accessed on Oct. 18, 2014.
7
While low-level cognitive models consist in non-generic semantic structures that result from the
principled connection of elements from our encyclopedic knowledge, high-level cognitive models arise
by abstracting away conceptual material shared by low-level cognitive models. For example, such
disparate notions as ‘kill’, ‘hit’, and ‘cut’ are envisaged as ‘actions’ performed with the help of
instruments and having visible effects (all these are high-level constructs). Non-situational or
propositional cognitive models are those that designate entities and their relations. In turn, situational
cognitive models are conventional, coherently connected series of events such as ‘cooking a meal’,
‘going to the dentist’ or ‘taking a bath’ (cf. Ruiz de Mendoza, 2007, 2014).
8
There are three broad relation types at this level: logical (e.g. cause-effect, evidence-conclusion),
temporal (e.g. precedence, simultaneity), and conceptual (similarity, contrast, conditioning,
concession).
9
For a discussion of constructions belonging to levels 2, 3 and 4, we refer the reader to Mairal and
Periñán (2014) and Mairal (fc.).
10
Interestingly, this decision is fully consonant with Goldberg’s (2006, p. 25) claim that “each
argument structure pattern is best analyzed on its own terms, without relying on explicit or implicit
reference to possible alternative paraphrases” and with Ruiz de Mendoza and Mairal’s (2011) view of
the notion of alternation as epiphenomenal, i.e. an unsystematic side-effect of deeper processes
licensing or blocking lexical-constructional integration. For example, the causative/inchoative
alternation is grounded in the presence versus the absence of a ‘causal’ element (cf. The wind opened
the door ‘The wind caused the door to be opened’/The door opened ‘The door happened to open’). But
there are verbs that can be used in the alternation where the causal element is irrelevant. Take the verb
handle, which can be employed in the inchoative construction (This car won’t handle without the
proper shock package) but does not really have a causative alternate (People like to handle this car can
hardly be paraphrased as ‘People like to cause this car to be handled’). In its transitive use, the object
handle is not an affected object, but a relevant part of the scope of a controlled action. On a deeper
level of analysis, the inchoative use does not depend on the presence of a causative alternate, but on
whether it is possible to think of the object as allowing the agent to perform the action. This “enabling”
ability of the object, which has been frequently cited in the literature (cf. Heyvaert, 2003, p. 132), is an
act of conceptual re-construal of what actually happens (i.e. we see a car as willfully allowing the
action of which it is the object to be performed). In the LCM re-construal (usually in the form of highlevel metaphor and metonymy) is seen as a licensing factor for some cases of lexical-constructional
integration (see Ruiz de Mendoza and Mairal, 2008; Ruiz de Mendoza, 2013; and Ruiz de Mendoza
and Galera, 2014). The interested reader is further referred to Barðdal et al. (2011) who claim that
double object sentences and their corresponding to-variants (e.g. He gave her a sweater > He gave a
sweater to her) show differences in interpretation and are therefore not mutually exchangeable.
11
The interested reader is referred to van Trijp (2013) for a discussion of how approaches like Fluid
Construction Grammar and Sign-Based Construction Grammar understand the notion of construction.
41
Note that in Kay’s (2013) approach, for example, many configurations do not qualify as
constructions. Thus, a sharp distinction is sanctioned between constructions and ‘patterns of coinage’
(e.g. He sneezed the napkin off the table) on the grounds that the latter do not meet the condition of
productivity.
13
FunGramKB constructs can be found in the linguistic realization or input text and in the conceptual
representation (i.e. the COREL schema); the minimal constructs respectively being lexical units and
ontological concepts. For instance, Periñán (2013, p. 215) argues that He fried the egg in the pan only
consists of a Kernel-2 construction (“He fried the egg”), while the remaining components are
constructs. This, for example, contrasts with the Golderbgian (2006) approach to the notion of
construction according to which each of the words making up the above-mentioned sentence are
constructions (i.e. “egg”, “pan”, etc.). In FunGramKB, the meaning of these constructs is directly
derived from the meaning postulates of the concepts to which each of these lexical units is connected.
For instance, pan is related to the ontological concept +PAN_00, whose meaning postulate is given in
(i) together with a translation:
(i) +(e1: +BE_00 (x1: +PAN_00)Theme (x2: +UTENSIL_00 & +CONTAINER_00)Referent)
+(e2: +BE_01 (x1)Theme (x3: +ROUND_00)Attribute)
*(e3: +BE_01 (x1)Theme (x4: +METAL_00)Attribute)
*(e4: +COMPRISE_00 (x1)Theme (x5: +HANDLE_00)Referent)
*(e5: +USE_00 (x6: +HUMAN_00)Theme (x1)Referent (f1: (e6: +COOK_00 (x6)Theme
(x7)Referent))Purpose)
= ‘A pan is a round metal container which usually has a handle. Generally, humans use pans
to cook something’.
14
Periñán (2013, p. 215) further notes that “any construction is a construct itself, but not all constructs
can be deemed to be constructions”. This means that in order to attain constructional status, the
meaning of the whole sentence must exceed the sum of the parts, as in Anna ran the pavement thin (i.e.
‘Anna ran a lot and as a result, she caused the pavement to become thin’).
15
Asterisked examples inspired in the more exhaustive list provided by Wechsler (2005). See Boas
(2011) for an in-depth investigation of the factors that may influence language users to produce and
interpret unacceptable sentences like (4d) as acceptable, provided that the proper context is supplied.
16
As noted in Ruiz de Mendoza and Luzondo (2014), the origin of the resultative value of “to bits” is
to be found in the metaphor CHANGES OF STATE ARE CHANGES OF LOCATION (Lakoff, 1993).
This metaphor captures our tendency to correlate locations with how we feel in them (e.g. we feel cool
in the shade, warm under the sun, etc.). Expressing result by means of a PP is useful when there is no
clear adjectival form that captures the same meaning. Thus, in English the resultative element in The
blacksmith hammered the metal into different shapes would be impossible to express by means of an
adjective.
17
The intransitive resultative (e.g. The river froze solid) is also listed as a different construction.
However, this pattern will not be accounted for in the present proposal.
18
In Goldberg’s work, fusion processes are regulated by means of two broad constraints: (i) the
‘Semantic Coherence Principle’ (i.e. “only roles which are semantically compatible can be fused”), and
(ii) ‘Correspondence Principle’ (i.e. “each participant role that is lexically profiled and expressed must
be fused with a profiled argument role of the construction”) (see Goldberg, 1995, p. 50). Note that the
profiled participant roles of the verb and the profiled argument roles of the construction appear in bold
letters. Lexically profiled participants are those entities that are obligatorily accessed and thus function
as focal points. In turn, constructional profiling occurs as follows: “every argument role linked to a
direct grammatical relation (SUBJ, OBJ or OBJ2) is constructionally profiled” (Goldberg, 1995, p. 48).
19
These cognitive operations are termed high-level since they involve generic cognitive models thus
working at higher levels of abstraction (e.g. the notions of ‘action’, ‘process’, ‘effect’, ‘cause’, etc.).
12
42