2.1 One frame element to one preposition structure mapping

Prepositions can be semantically vacuous or polysemous. In many cases such as in verb particle constructions, they have little semantic content. Because of the semantic emptiness of prepositions, many search engine indexing agents include prepositions in the bag of stop word lists. In other cases such as in locative expressions, the information content they contain is not low. For example, in ‘He drove down the road’ the semantic weight of the sentence lies in the preposition. Because prepositions are highly promiscuous distributionally, dictionaries usually give sets of examples demonstrating polysemous uses and various meanings of prepositions. Also, prepositions do not translate well to other languages, so they are one of the major sources of difficulties for students who are learning English as a second or foreign language (Boers and Demecheleer Reference Boers and Demecheleer1998; Celce-Murcia and Larsen-Freeman Reference Celce-Murcia and Larsen-Freeman1999).

The semantically polysemous nature of prepositions has been much studied in the past (Radden Reference Radden and Dirven1989; Vandeloise Reference Vandeloise1994; Pütz and Dirven Reference Pütz and Dirven1996; Cuyckens and Radden Reference Cuyckens and Radden2002; Tyler and Evans Reference Tyler and Evans2003). Linguists have been analyzing the literal senses of prepositions and the metaphorical meanings they enter into. In many cases, the polysemy of prepositions can be explained by radial categories (Lakoff Reference Lakoff1987) where many related senses of a preposition are connected to a semantic common core. Prepositions get their abstract senses from a primary proto-scene (Tyler and Evans Reference Tyler and Evans2003) or a semantic prototype (Wesche Reference Wesche1986) ‘by means of generalization or specialization of meaning or by metonymic or metaphoric transfer’ (Cuyckens and Radden Reference Cuyckens and Radden2002: xiii). For example, for the preposition at, the core meaning can be a simple point of location or reference as in at the front door. The following are some of the concepts that can be extended from the core meaning of at:

• • Extended locative concepts: the party at the beach, the meeting at Ottawa.

• • Extended intentional concepts: to wait at the bus stop, to be at school.

• • Extended temporal concepts: to wake up at midnight, to be busy at the moment.

• • Extended scale concepts: to work at 20 dollars an hour, to drive at 50 mph.

• • Extended causal concepts: to leave at my signal, to be happy at the arrival of the rescue ship.

• • Extended directional concepts: to throw the ball at him, to shout at him, to grab at the purse, to shoot at the rabbit.

To a certain extent, these conceptual extensions of at show that the semantic contents that a preposition can convey are not simply an arbitrary cluster of unrelated senses. They come from a highly schematic common core and constitute a unitary meaning. On the basis of this, we can assume that there can be a strong linking relationship between prepositional phrases and primary underlying meanings they represent. However, do we have corpus evidence to show that prepositions are associated with core meanings, and that some meanings can be represented by primary preposition? Are there correspondence patterns between frame elements and preposition structures? For grammar or syntax reasons, if we have to select a preposition structure that can be used with a specified lexical unit to represent a frame element concept, what prepositions are we most likely to choose? In other words, can we predict the selection of prepositions on the basis of the information about the frame element and the target lexical unit?

In addressing these questions, we analyzed the argument realization patterns for FrameNet sentences and observed some interesting cases where a few lexical words combined only with selected types of prepositions to represent some frame elements in the evoked frame. The prepositions are bound to target lexical words not in a way as strong as those traditionally called prepositional–verb constructions such as depend on, account for, and look at. The link between the preposition and the target lexical word seems to be much looser syntactically, but more widespread. In the prepositional–verb constructions, the connection between the preposition and the lexical verb is usually individually based, so that we say to listen to somebody, but not to hear to somebody. But in the one frame element mapping to one preposition structure case, the same preposition is selected not only for one individual lexical item, but also for a variety of target lexical units in the same frame. For instance, in the Seeking frame, the preposition for is used with almost all of the verbs such as comb.v, feel.v, forage.v, fumble.v, grope.v, hunt.v, listen.v, look.v, pan.v, ransack.v, rummage.v, scour.v, seek.v, and sift.v for the frame element Sought_entity. Here are more examples of such one-to-one mapping relations where one designated preposition is consistently used for the realization of a particular frame element:

• • Air: of 12 [breath.n, exhalation_act.n, inhalation.n]

• • Alterant: with 12 [stain.v, treat.v]

• • Arguer2: with 77 [altercation.n, argue.v, argument.n, bicker.v, disagreement.n, fight.v, quarrel.n, quarrel.v, quibble.v, row.n, row.v, spat.n, tiff.n, wrangle.v]

• • Artist: by 15 [direct.v, film.v, produce.v, shoot.v]

• • Assailant: by 16 [ambush.v, assault.v, attack.n, attack.v]

• • Authorities: by 25 [apprehend.v, arrest.v, bust.v, cop.v, detain.v, intern.v, internment.n, nab.v]

• • Beneficiary: for 18 [create.v, deal.v, fix.v, resolve.v, seek.v, solve.v]

• • Complaint: of 13 [complain.v, complaint.n]

• • Component_Sound: of 16 [babble.n, bellow.n, blare.n, boom.n, bray.n, burble.n, buzz.n, drone.n, peal.n, rap.n, titter.n, yammer.n]

• • Controlling_Act: in 12 [authoritarian.a, indulgent.a, lenient.a, liberal.a, strict.a, tolerant.a]

• • Conveyed_Emotion: of 19 [frown.n, grin.n, look.n, rictus.n, smile.n, smirk.n, sneer.n]

• • Crime_Jurisdiction: to 28 [extradite.v, extradition.n]

• • Employee: of 31 [dismissal.n, resignation.n, retirement.n, work.v]

• • Executed: of 14 [execution.n, guillotining.n]

• • Fact: by 11 [account.v, explain.v]

• • Hair: of 12 [braid.n, curl.n, cut.n, lock.n, plait.n, ponytail.n, topknot.n, tress.n]

• • Ingestor: by 17 [consume.v, devour.v, drink.v, eat.v, ingest.v, munch.v, nibble.v, sip.v]

• • Inspector: by 15 [double-check.v, examination.n, examine.v, inspect.v, inspection.n]

• • Interceptor: by 12 [intercept.v, interception.n]

• • Investigator: by 15 [investigate.v, investigation.n, probe.n, under investigation.prep]

• • Light: of 18 [flash.n, gleam.n, glimmer.n, glitter.n, glow.n]

• • Mass: of 357 [abundance.n, amount.n, avalanche.n, degree.n, deluge.n, flood.n, heap.n, load.n, mass.n, measure.n, mite.n, modicum.n, myriad.n, oodles.n, ounce.n, pile.n, pinch.n, plethora.n, quantity.n, raft.n, stream.n, ton.n, torrent.n, touch.n, trickle.n, wave.n]

• • Noisy_Event: of 76 [boom.n, burble.n, buzz.n, chatter.n, clack.n, clang.n, clatter.n, click.n, clink.n, creak.n, crunch.n, drone.n, gurgle.n, hiss.n, hum.n, patter.n, rap.n, rasp.n, roar.n, rustle.n, scrape.n, screech.n, slap.n, thump.n, tinkle.n, whine.n]

• • Object: of 25 [area.n, bless.v, confession.n, consecration.n, depth.n, length.n, rite.n, sacrifice.n, width.n]

• • Party_2: with 28 [agree.v, agreement.n, come together.v, meet.v, treaty.n]

• • Preventing_Cause: by 21 [derail.v, foil.v, nip in the bud.v, thwart.v]

• • Profiled_Item: by 20 [eclipse.v, equal.v, outdo.v, outshine.v, outstrip.v, rival.v, surpass.v]

• • Purpose_Of_Recipient: for 11 [provide.v, supply.v]

• • Sentence: to 15 [sentence.v]

• • Set: of 13 [proliferation.n]

• • Sought_Entity: for 86 [comb.v, feel.v, forage.v, fumble.v, grope.v, hunt.v, listen.v, look.v, pan.v, ransack.v, rummage.v, scour.v, seek.v, sift.v]

• • Themes: of 31 [clash.n, clatter.n, click.n, clink.n]

• • Transferors: between 11 [transfer.n]

• • Vantage_Point: from 35 [conceal.v, concealed.a, cover.v, hidden.a, hide.v, mask.v, obscure.v, obscured.a, screen.v, screened.a]

• • Weapon: with 12 [armed.a, assault.v, attack.v]

For each entry in the above list, a frame element name is followed by the primary preposition that is used to represent the frame element, the frequency (f) of this preposition in the realization of the FE concept, and the lexical units that are used with the primary preposition. The threshold for selection is f > 10. In these instances, there are outstanding patterns of correspondence between FEs and preposition structures. We observe, for example, that if we have to choose a preposition structure for the syntactic realization of the frame element Weapon, we are much more inclined to pick with to head the prepositional phrase. Thus, we say to be armed with a weapon and to attack somebody with a weapon. Similarly, we use for for the frame element Beneficiary. For the FE Arguer-2, the second party involved in an argument, we use the preposition with, no matter if the lexical unit is a noun such as argument.n and altercation.n, or a verb such as bicker.v and wrangle.v. For the two transferors involved in a transfer, we use between as in ‘It seems that a similar ambiguity is common in Britain in relation to [THEME money] [TARGET transfers] [TRANSFERORS between kin], although we do lack up-to-date evidence on this point.’ At the same time, evidence in this list shows that most of the by cases are not interesting in their own right. They are limited to passive uses of the verb, and it just happens to be the case that the FE that would have been Subject (Ext) in the active gets marked as by in the passive – but not at all because of the participation in a particular frame.

2.2 One frame element to multiple preposition structures mapping

Prepositions represent a closed class. On the one hand we have the richness of meaning; on the other hand we have the paucity of linguistic structures including prepositions to represent them. Thus, we can have a single preposition representing different conceptual relations, and a semantic role realized by multiple prepositions. While there are 214 frame element types that exhibit the one FE one preposition mapping relationship, there are almost twice as many (f = 426) FE types that manifest syntactic realization alternations when they use prepositional phrase representations. This means that most PP-FEs in FrameNet are represented by multiple preposition structures. Table 1 demonstrates the most frequent frame elements that have preposition phrases in their surface structures. The frequency threshold is f ≥ 500 for these PP-FEs.

Table 1.

Frame element and preposition distributions in the most frequent PP-FE structures in FrameNet (threshold f ≥ 500)

The FE column lists the frame element names. In FrameNet, target lexical units are classified on the basis of the semantic frames they belong to. In each semantic frame, there are participating FEs. The inventory and taxonomy of frames and FEs are created by the FrameNet group at Berkeley, and they are based on Fillmore's theory of frame semantics (Fillmore Reference Fillmore1982). Some FEs are general across frames such as Source and Goal, although definitions of the FE in those frames can be different to a certain degree. Others are unique names of thematic roles specific to frames. The FEs in Table 1 are the most frequent semantic roles that are represented by prepositional phrases. Prep_Tokens refers to the total number of phrases that are headed by prepositions; Prep_Types is the number of different prepositions that are used to head the prepositional phrases. For example, the FE Individual is used across nine frames (Aggregate, Congregating, Dispersal, Gathering_up, Make_acquaintance, Quantity, Relation_between_individuals, Social_connection, and Unemployment). In 1 449 (Prep_Token) cases, this FE is represented by prepositional phrases, and for these 1 449 occurrences, only 2 (Prep_Types) prepositions are used. Std. Dev. stands for standard deviation.

From the distribution of prepositions that head the preposition structures in argument realizations, we can see the diversity of syntactic configurations in selecting prepositions to represent frame element concepts. Goal in Table 1 is the frame element that is most frequently represented by prepositions (Prep_Tokens = 6 181). At the same time, it is also the frame element that can be represented by most of the preposition types (Prep_Types = 59). Direction is the frame element with the most even distribution (mean = 15.78 and Std. Dev. = 16.04) in the use of preposition structures in argument realizations. Although frame elements can have multiple argument realization possibilities, there is a strong tendency for the deep semantic arguments to consistently employ a few key prepositions much more than they do other prepositions in their realization expressions. We consider these key prepositions as central prepositions in representing the associated frame elements. For some frame elements in the one frame element to multiple preposition structure mapping scheme, the level of dependency on certain prepositions to represent the deep semantic arguments is almost as high as those listed in the one-to-one mapping conditions. In some cases, the centrality of prepositions in argument realizations is so notable that even though there are other prepositions structures in the realization patterns, the noncentral preposition structures are only few and far between. Now let us decompose the information in Table 1 and have a closer look at the preposition distribution information for each of the listed frame elements.

• • Goal: to 0.305, into 0.209, in 0.112, on 0.076, over 0.039, onto 0.027, at 0.024, on to 0.022, against 0.017, down 0.016.

• • Path: through 0.180, across 0.122, down 0.121, along 0.102, up 0.089, towards 0.080, over 0.063, past 0.041, round 0.026, in 0.025, around 0.023, by 0.018.

• • Topic: about 0.524, on 0.183, of 0.133, over 0.042, in 0.020, concerning 0.016.

• • Source: from 0.567, off 0.143, out 0.112, out of 0.084, of 0.050, away from 0.016.

• • Time: in 0.327, on 0.177, at 0.122, after 0.084, during 0.060, before 0.056, by 0.027, as 0.020, over 0.016.

• • Place: in 0.482, at 0.244, on 0.121.

• • Reason: for 0.630, as 0.114, of 0.070, with 0.033, on 0.027, from 0.021, because of 0.019, at 0.019, in 0.019.

• • Theme: with 0.509, of 0.338, by 0.077, in 0.052.

• • Individuals: of 0.999.

• • Manner: in 0.302, like 0.255, with 0.186, on 0.039, without 0.039, as 0.038, at 0.030, by 0.026.

• • Stimulus: by 0.368, at 0.255, with 0.133, of 0.085, about 0.056, in 0.038, for 0.025, over 0.015.

• • Area: around 0.281, about 0.190, in 0.145, round 0.086, on 0.064, over 0.045, of 0.042, among 0.027, at 0.015.

• • Addressee: to 0.621, at 0.192, with 0.042, of 0.033, from 0.031, in 0.022, on 0.019, for 0.017.

• • Means: by 0.511, with 0.226, through 0.066, in 0.060, on 0.058, of 0.034.

• • Whole: of 0.942, into 0.031.

• • Content: of 0.474, for 0.119, about 0.094, in 0.063, at 0.046, against 0.032, as to 0.030, from 0.025, to 0.022, with 0.022.

• • Ground: in 0.383, on 0.143, of 0.131, through 0.060, against 0.027, at 0.026, from 0.022, around 0.019, under 0.019.

• • Location: in 0.285, on 0.143, at 0.142, by 0.043, around 0.042, of 0.042, over 0.030, behind 0.025, under 0.024, outside 0.023, across 0.020, against 0.020, round 0.018, beside 0.016.

• • Stuff: of 0.999.

• • Substance: of 0.977, on 0.016.

• • Category: as 0.540, of 0.363, into 0.034, in 0.017, with 0.017.

• • Message: of 0.389, for 0.232, to 0.086, in 0.063, about 0.044, with 0.044, against 0.028, on 0.024, into 0.019, as 0.016.

• • Relative_location: in 0.359, on 0.162, at 0.140, of 0.082, between 0.036, near 0.031, around 0.026, above 0.023, outside 0.018.

• • Evaluee: of 0.464, for 0.159, on 0.136, at 0.044, in 0.040, about 0.037, to 0.034, against 0.027, with 0.015.

• • Speaker: by 0.578, from 0.288, of 0.125.

• • Direction: through 0.115, towards 0.109, down 0.079, in 0.073, over 0.073, into 0.071, around 0.053, across 0.051, round 0.044, up 0.040, out 0.034, to 0.034, about 0.030, from 0.028, along 0.022, behind 0.022, inside 0.022, past 0.018.

In the above list, each frame element name is followed by a head preposition and the probability of its occurrence. For the frame element names, they are taken independently of the frames to which they were assigned, which in our view, corresponds more to the newer view of these things in the current FrameNet database. The probability selection threshold is T ≥ 0.015. We can observe the obvious centrality in the use of prepositions for the realization of some FEs. For example, although there are five prepositions (as, of, into, in, with) for the frame element Category, the key prepositions are actually only the first two prepositions: as and of. These two prepositions subsume more than 90% of the uses of preposition structures in the realization patterns of this FE. For the frame element Reason, the probability of using the central preposition for is 0.630, much higher than the probability of using the second or the third most frequently used preposition (as 0.114, of 0.070). For the frame element Addressee, the central prepositions to, at, and of account for more than 80% of the uses. There is an obvious selection preference of of for the following four frame elements: Whole (of 0.942) in the frames of Part_whole, Part_orientational, Part_ordered_segment, Part_inner_outer, Part_edge, Calendric_unit; Stuff (of 0.999) in the frames of Measure_linear_extent, Measure_mass, Measure_volume; Substance (of 0.977) in the frames of Part_piece and Shapes; Individuals (of 0.999) in the frames of Aggregate and Quantity. The probabilities of using of for the frame elements Individuals and Stuff are so high (0.999) that they almost exclude the use of other prepositions.

Figure 3 shows the distribution of central prepositions and the degree of preposition centrality in the syntactic realization of semantic arguments. The figure includes the top fifteen frame elements that use preposition structures in their argument realizations. The four sections (p1, p2, p3, and p4) for each bar represent the first four most central prepositions for each frame element. For most of the frame elements, the first four prepositions cover more than 70% of the preposition uses for the associated semantic argument. The frame element Path is the only exception. In ten out of the fifteen frame elements, the first four central prepositions for each FE cover a PP-FE proportion that exceeds 80%. This means that in most cases, if we have information about the frame element and the lexical word, we would be able to predict the most likely prepositions to use.

Fig. 3.

Degree of preposition centrality for the fifteen top most frequent PP-FEs. p1, p2, p3, and p4 represent the first, second, third, and fourth most frequently used preposition types for the specified frame element.

It is worth noticing that some frame elements with a cross-frame occurrence nature such as Direction, Path, and Location have a rather noticeable tendency of using multiple realization configurations with different preposition structures. Here we use entropy to measure the uniformity of distribution of preposition frequencies to help us decide whether there is a tendency of using central prepositions for the realization of some semantic arguments.

Table 2 shows the entropy information about the degree of multiplicity and the density variance in the use of prepositions to represent FEs. It is for the PP-FEs that have an occurrence frequency of more than 500. The FE column is for frame element names. Given a preposition structure that is marked as a frame element in FrameNet, the probability of this preposition structure representing the specified FE is P(FE|Prep). Max_P is the maximum probability of a preposition type denoting the semantic argument of this specified FE. Prep refers to the preposition that bares the maximum probability. It can also be interpreted as the most frequently used preposition representing the FE concept. The Example column lists condensed examples from the preposition knowledge base, demonstrating how the most central preposition for each of the frame elements is used in FrameNet. Words such as army, mugs, and coils that are outside the square brackets in the Example column are frame-bearing target lexical units. They are the target lexical units that evoke the specific frame elements in the FE column and that are used with the prepositions in the Prep (preposition) column. The words included in the square brackets are the preposition structures that represent the corresponding frame element concept in the FE column. The table is sorted by the entropy values.

Table 2.

Preposition density variance in the most frequent PP-FE structures in FrameNet (threshold f ≥ 500)

A high entropy value coupled with a low maximum probability value means that the specified frame element has more varied preposition structures in its syntactic realization expressions. This is where the frame element lacks the centrality of preposition structures in representing the frame element concepts. Direction and Path are the two most obvious examples. Both of the frame elements indicate general spatial concepts. The spatial concepts can be associated with all directions in the real world such as through, across, towards, down, along, up etc. It is hard to say that one direction is more important than the others. That is why the uses of prepositions are more or less evenly distributed, and the prominence of using key prepositions for these frame elements is thus obscured.

2.3 Explaining multiple FE-preposition correspondence options

The argument realization options of both one-to-one mapping and one-to-many mapping clearly support the hypothesis that there is a link between the frame elements and the central prepositions in the surface syntactic realizations. The syntactic realization of arguments is predictable to a large extent from the frame element concept and from the target lexical units. In the FrameNet sentences analyzed, the tendency to assimilate the core meanings of individual prepositions with some basic semantic notions such as Agent, Manner, Means, Instrument, Theme, Reason, Topic etc. is rather marked. Now, let us focus on why at the same time noncentral prepositions are selected in some surface syntactic representations. Our analysis will be based on the factors that can have an impact on the selection of prepositions in the semantic argument realization process.

We isolated, from the preposition knowledge base, sentences that can demonstrate the use of multiple prepositions in representing the same frame element, and generated a mapping comparison list. This list contains 27 036 sentences. A close examination of the sentences in the list reveals some major factors that can affect the selection of different preposition structures as the surface manifestations of arguments.