Understanding Idioms
Nancy Chang, International Computer Science Institute, 1947 Center St., Suite 600, Berkeley, CA 94704-1198 USA,
Ingrid Fischer, Lehrstuhl für Programmiersprachen, Universität Erlangen-Nürnberg, Martensstr. 3, 91058 Erlangen, Germany,
Abstract
We describe a language understanding model in which idioms, like other units of linguistic knowledge, are treated as constructions linking elements of form with elements of meaning. They may also, however, invoke additional metaphorical mappings that can actively constrain both the interpretation and acceptability of idiomatic expressions. A wide variety of idiomatic phenomena, including compositional and non-compositional idioms, syntactic flexibility and modification can be explained within this framework.
1 Introduction
Any general-purpose language understanding system must accommodate a wide variety of idiomatic phenomena. The class of idioms resists simple definition; some are best considered large frozen word sequences, while others are quite flexible constructions and allow significant variation. We list here a few characteristic criteria to consider [4]:
Non-compositionality. The meaning of an idiom is typically not predictable from those of its parts, as in the often cited example kick the bucket (die). In contrast, spill the beans is decomposable into its parts, with beans modeling a secret and spill its revelation. Closely related is the presence of formal-semantic isomorphism between the surface form of an idiom and its semantics. This is the case for spill the beans but not for kick the bucket, which is lexically complex but has relatively simple semantics.
Fixed idiom constituents. Many idioms are lexically specific and do not permit substitution of even close synonyms. In contrast, German einen Korb bekommen/geben/austeilen/kriegen/erhalten? "to receive/get/give/obtain/catch a basket" describes a rejection in terms of the transfer of a basket (from the rejecter to the rejectee), and neither the manner nor direction of transfer is fixed by the idiom. Some of the fixed components may be so-called cranberry (or cran-) components that do not appear productively in other constructions, like the word Fettnäpfchenin ins Fettnäpfchen treten"step into the Fettnäpfchen" (drop a brick). Such idioms thus have no literal reading, a property also true of other examples, such as zwei linke Hände haben "to have two left hands" (be a bad craftsman).
Syntactic modifications. Even idioms with fixed lexical items often appear in syntactically diverse forms, as illustrated by these variations on the canonical active, unmodified form of spill the beans:
1.The beans were spilled by accident.
2.He spilled the political beans.
Example 1 is in passive form, with additional modification of the entire event; example 2 includes an additional modifier of the idiomatic interpretation of beans. Similarly, German einen Bock schiessen "shoot a bug" (make a mistake) has variations like:
3.Welchen Bock hat er geschossen?
"Which bug did he shoot?"
4.Der Bock, den er geschossen hat, ruiniert seine Karriere.
"The bug he shot ruined his career."
As illustrated here, some variations and modifications affect only one or several of the idiom's constituents, while others affect the idiomatic meaning as a whole.
Opacity. Idioms such as by and large and spick and span are often considered opaque: the semantic motivation or source for the idiom may be unclear, and no particular underlying image may come to mind. Such opaque idioms can be contrasted with motivated idioms [14]. In spill the beans, the dual mappings between spilling and revealing, and between beans and the secret, provide a framework for additional inferential mappings between the literal and idiomatic meanings; issues such as the location of the beans before and after being spilled, the size of the container, the nature of the spilling (purposeful or accidental, messy or neat) all influence the underlying picture and may have motivated metaphorical analogs.
Based on these considerations and following Dobrovols´kij [4,5], we can describe idioms as a radial category [14]. In their central (or prototypical) case, idioms are non-compositional, lexically simple and opaque; they have some fixed constituents and have some literal reading (apart from possible cranberry components); and allow only limited syntactic variations.
But none of these properties is absolutely required, and non-central examples that do not adhere to this strict definition appear regularly in natural language corpora and in idiomatic dictionaries, such as [6]. Our system must therefore be sufficiently expressive and flexible to handle the full range of combinations of these properties.
2Psycholinguistic assumptions
Psycholinguistic research provides clues about how humans handle the phenomena just described. This section summarizes the psycholinguistic theories that have influenced our model.
Literal versus figurative reading. A sentence like "He kicked the bucket" may evoke both a literal meaning based on a kicking event and a figurative meaning based on a dying event. Different psycholinguistic experiments support different conclusions about whether one reading precedes or triggers the other [10]. According to the simultaneous processing model [16], both readings are tried at the same time in a kind of race, and factors such as cranberry components, contextual cues and baseline frequency of the idiom's words determine which reading succeeds and is accessed first.
Storage of idioms. Three main proposals have been made for how idioms are stored in the semantic lexicon. First, in approaches based on lexical representation theory, an idiom is treated as a single, multiword lexical entry. This model is the simplest solution to the storage issue, but it cannot easily represent syntactic modifications of idioms. The configuration hypothesis [3] states that idioms are stored as configurations of words and recognized based on the so-called key. When thekey is detected, the literal reading is halted and the idiomatic reading is attempted. A refinement of this approach is the third possibility, described by Gibbs [9]. According to his decomposition hypothesis, idioms are not composed only of individual words, as in the configuration hypothesis, but instead of decomposable and non-decomposable phrases. The representation of idioms in our model allows both kinds of constituents.
Semantic motivation. There are two main approaches to handling semantically motivated idioms. The conceptual metaphor hypothesis [13, 14, 9] assumes that many motivated idioms are based on conceptual metaphors. That is, idioms are not merely dead metaphors; rather, metaphors actively participate in constraining both the interpretation and acceptability of idiomatic expressions, especially those that are analyzable and modifiable. The idiom spill the beans uses the metaphor Revealing a secret is Spilling the Beans, which itself draws on the more general metaphors The Mind is a Container and Ideas are Physical Entities. [12] includes an extensive list of idioms and the metaphors on which they are based.
Alternatively, the inference hypothesis by Cacciari, Ruminati and Glucksberg [2] assumes that idiom words always evoke their literal meaning, which in turn grounds the images or meaning associated with the idiom. Idiomatic meaning can then be inferred from the literal meaning of the words in the idiom. This approach has the computational advantage that it does not require separate literal and idiomatic interpretations of each word. The problem with this approach is that it is unclear how exactly the idiomatic meaning is to be inferred, especially in novel usages; in such cases the conceptual metaphor hypothesis seems more adequate.
As we shall see, the system we describe uses elements of both of the above approaches.
3Idioms as constructions in a simulation-based framework
Besides the psycholinguistic evidence just described, the approach to modeling idioms adopted here draws primarily on recent work on construction-based approaches to grammar [11, 1], which proposes that linguistic knowledge of all kinds consists of pairings of form and meaning constituents and constraints. Lexical and grammatical knowledge are represented uniformly in this framework: words and morphemes are simple constructions pairing particular form and meaning patterns, while phrasal and clausal constructions are defined somewhat more abstractly and may have constituent (sub)constructions. As a simple example, the lexical construction for bucket would specify that it has a formal constituent of type "bucket", corresponding to its phonological or orthographic representation, and a semantic (or meaning) constituent that refers to the bucket category in conceptual knowledge. General knowledge about buckets-that they are typically used as containers, that they are inanimate objects, their typical sizes and materials, etc.-is accessible through conceptual knowledge and useful for determining what larger constructions the word might appear in.
The more complicated Transitive construction, which licenses sentences like Mary spilled the beans and Harry kicked the bucket,has three constructional constituents, instantiated in the first example by Mary, spilled, and the beans; formal constraints (specifying the constituent word order, for instance); and a semantic constituent that evokes the Energy-transfer scene (or frame in the sense of [7]) from conceptual knowledge. (It should not be confused with a purely syntactic construction based on a syntactic notion of a transitive verb.) Finally, the construction's semantic constraints serve to bind the various entities designated by its constituents, for instance binding the referent of Mary as the source of the Energy-transfer (and, more specifically, as the spiller of the Spill-event frame evoked by spilled.
It is crucial to note that "meaning" in our model is not strictly compositional, since constructions may themselves contribute meaning separate from that of their constituents. Moreover, although constructions generally refer to conceptual knowledge structures, the meaning of an utterance arises from an active mental simulation in context and the set of additional entailments such a simulation produces. Constructions serve to map between linguistic forms and the parameters for the simulation, contained in a structure called the simulation specification. We adopt the formal representation for simulation-based constructions described in detail in [Bergen00]; relevant features will be highlighted through the examples in Section 4.
Given the basic architecture of a simulation-based language understanding system, it becomes relatively straightforward to incorporate idioms. In particular, the ease in Construction Grammar of expressing constraints of varying specificity and allowing both compositional and non-compositional approaches to meaning allows a simple treatment in which prototypical idioms are phrasal or clausal constructions with lexically specific constraints and a conventional meaning associated directly with the construction. Just as the Transitive construction can contribute an energy-transfer meaning, the Spill-the-Beans construction can evoke the revelation of a secret. Idioms may also introduce special-purpose metaphors as meaning constituents; see Section 4 for examples.
In summary, our approach to idioms is based largely on existing machinery for language understanding system based on constructions and simulation, which in turn relies on four major components:
1. The lexicon, grammar and idiom list are uniformly represented as a collection of grammatical constructions. Each construction represents well-formedness conditions across various domains of linguistic knowledge, including phonological, morphological, syntactic, semantic and pragmatic domains.
2. A rich, frame-based ontology of conceptual entities is needed to represent world knowledge. The network includes the standard sub- and supertype relations, part-whole relations. It is also organized using frames with sets of interdefined roles.
3. The construction analyzer uses surface cues to find possible matching constructions. Candidate constructions whose constraints can be satisfied are scored, and semantic and pragmatic information from the best-scoring construction cover of the original sentence is added to the simulation specification.
4. The simulation specification provides the parameters necessary to run the desired simulation. It draws on a dynamic action representation based on stochastic Petri nets; details of this representation are not necessary for the current work but are described elsewhere [1, 15].
This existing machinery suffices for many simple cases, particularly non-compositional idioms with no associated motivated metaphors; such cases can be most simply represented as individual constructions with associated conventionalized interpretations. To represent semantically motivated idioms and handle instances of syntactic modification, the underlying conceptual metaphors must also be represented. No additional system component is necessary, but idiomatic constructions must be able to contribute meanings in terms of local metaphor maps. Note that a large-scale language understanding system should include a repository of active metaphor maps, which should take the same form as the special-purpose ones we introduce here. For simplicity we will not show the relation between local metaphors and these general conceptual metaphors, but since the architecture described here allows a uniform representation for both kinds of metaphor, it is a simple matter to assert appropriate bindings between them.
4Modeling Idioms
In this section we illustrate the approach described with several idiomatic constructions, along with example sentences and their associated simulation specifications (or simspecs). We will concentrate on the general approach, highlighting the relevant portions of the formalism.
As discussed earlier, Kick-the-Bucket, shown in Figure 1, is a prime example of a non-compositional idiom. We will describe the construction formalism in some detail using this example. The construction begins with its name, followed by several blocks listing its constituents, formal constraints and semantic constraints. The final designates line singles out the particular meaning element that is the primary (or highlighted, profiled) meaning element for the entire construction.
construction Kick-the-Bucket
constituents:
form kick-form of type Kick-Verb
form the-form of type "the"
form bucket-form of type "bucket"
meaning dying-event construed as Dying-Event
construct kicker-form designates kicker
construct trans of construction Transitive
formal constraints:
the-form before bucket-form
trans.src-form = kicker-form
trans.sink-form = bucket-form
semantic constraints:
dying-event.dyer kicker
designates dying-event
Figure 1 Non-compositional idiom: Kick-the-Bucket
In the constituents block, each constituent is labeled form, meaning or construct and has an alias (name) for reference within the construction. Constituents may be constrained to be an instance of a specific (formal) type or construction, or construed as an instance of a specific (semantic) category. This idiom has form constituents corresponding to the and bucket, as well as kick-form with kick as its root (e.g., kicked); it also evokes (and designates) the Dying-Eventframe.
The kicker-form constituent is not constrained to be any particular construction type, but the designates tag assigns the alias kicker to the meaning associated with kicker-form. The other constructional constituent trans draws on the Transitive construction, the active clausal construction described in Section 3 for energy-transfer events but not shown here. The inclusion of the Transitive construction provides a concise way of adding various properties (e.g., the ordering of the various form constituents, and subject-verb agreement) that are not specific to Kick-the-Bucket.
The form constraints serve mainly to bind the Transitive construction's constituents (named src-form and sink-form) to the appropriate local constituents. In addition, the ordering constraint (the-form before bucket-form) asserts that the must precede bucket for the idiom to match the utterance being analyzed. Note that this construction is flat; it has no particular constituent associated with the unit the bucket.
The single semantic constraint establishes the crucial binding (symbolized with the arrow ) between the semantic element designated by the kicker-form constituent and the dyer of the Dying-Event, resulting in the basic simspec shown in Figure 2 (where the time of the event is separately established by the verb).
Dying-Eventdyer / Harry
action / die
Time / Past
Figure 2 Simspec for Harry kicked the bucket.
The Kick-the-Bucket construction is quite rigid in its requirements: according to this definition, it can appear only in an active transitive syntactic form. More flexible compositional idioms, as typified by spill the beans, require a slightly different notation, as shown in Figure 3.
construction Spill-the-Beans
constituents:
form the-form of type "the"
form beans-form of type "beans"
meaning revealing-event construed as Revealing-Event
meaning spilling-event construed as Spilling-Event
construct spiller-form designates spiller
construct spill-form of construction Spill-Verb
designates spill-action construed asSpill-Action
construct def-beans
of construction Definite-Determination
designates beans-thing construed asBeans
metaphor beanspilling-met linking source spilling-event
and target revealing-eventwith:
spilling-event.spiller revealing-event.revealer
spilling-event.spilled revealing-event.revealed
formal constraints:
def-beans.det-form = the-form
def-beans.thing-form = beans-form
semantic constraints:
spilling-event.action spill-action
spilling-event.spiller spiller
spilling-event.spilled beans-thing
designates revealing-event
Figure 3 Compositional idiom: spill the beans
We focus on three main changes that provide the richer interpretive framework and greater flexibility needed for this idiom. First, an additional Definite-Determination constructional constituent has been added to allow the phrase the beans to function as an instance of a construction in its own right. The Definite-Determination construction has as its meaning not a simple conceptual category but rather a specific contextual referent. This is important for allowing utterances with additional modifiers (drawing on the standard constructional machinery; see Section 5). As an added benefit, the order of the-form and beans-form needn't be explicitly mentioned here, since it is part of the Definite-Determination construction.