Using translation corpora to explore synonymy and polysemy [1]

Abstract

1. Introduction

Polysemy and synonymy are sometimes viewed as two sides of the same coin; in both cases we are faced with one-to-many form-meaning relationships. The purpose of the paper is to show how translation corpora can be mined to shed light on such one-to-many relationships. With respect to synonymy, the greater the degree of semantic overlap there is between two lexemes or constructions in language A, the more difficult it should be to predict the original forms given their translations into language B. As for polysemy, it is hypothesised that putatively different senses of a lexeme or construction are more likely to be translated differently than similar senses (see Egan 2012).

The data for this paper are taken from the English-Norwegian Parallel Corpus (ENPC: see Johansson 2007). Compiled under the direction of Stig Johansson at the University of Oslo, the ENPC consists of 50 extracts from English texts of some 12,000 words in length, together with their translations into Norwegian, and 50 extracts from Norwegian texts of similar length, together with their translations into English. Both fictional and non-fictional texts are represented in the corpus. For the present paper only original English texts and their translations into Norwegian were consulted. The data comprise all tokens of the two near-synonymous verbs begin and start in both the fiction and non-fiction components of the corpus, and all tokens of the multi-polysemous preposition at in the fiction component alone. The reason for restricting the study of at to the tokens in the fiction component is its frequency. There are more than 2,500 tokens of the preposition in the fiction texts.

Section 2 contains a brief discussion of the notions of synonymy and polysemy and of the possible utility of translation corpus data for their study. Section 3 contains the first case study, of begin and start. The various senses of the preposition at and their translation equivalents, i.e. items in the translations corresponding semantically to items in original texts, are introduced and discussed in section 4. The results for at are compared to the results of previous studies of through (Egan 2012) and between (Egan forth.). Finally, section 5 contains a summary and conclusion.

2. Synonymy, polysemy and translation

As mentioned above, synonymy and polysemy both comprise one-to-many form-meaning relationships. This has led to their being viewed as two sides of the same coin, as it were. A moment’s reflection, however, will suffice to conclude that this metaphor is flawed, at best. Let us imagine that there are two lexemes, x and y, both of which are polysemous. For x and y to be defined as completely synonymous, they would have to have the exact same number of senses and each of these senses would have to be synonymous. It is unlikely, though not of course impossible, that two lexemes in a language would overlap to this extent. Moreover, if a language were to contain two such items, it is likely that over time one of them would fall into disuse. Given the improbability, then, of two polysemous lexemes overlapping completely in all of their meaning extensions, the reverse of the polysemy coin is much more likely to be partial, rather than complete, synonymy. The other side of the complete synonymy coin is monosemy. It is, in fact, doubtful if such a thing as complete synonymy actually exists, at least with respect to two expressions used in the same speech register over a period of time. As Murphy (2003) puts it:

One reason that synonymy is often defined as a sense relation is that synonyms usually involve a match between some, but not all, of a word’s senses. The sense relation description is convenient, then, because it only considers one sense of a word at a time. However, what is related here is not two senses (because in absolute synonymy a single sense is shared), but two lexical units, that is, instantiations of lexical items, associated with a particular sense. (Murphy 2003: 145)

Most studies of synonymy are studies of what Murphy calls sense synonyms, polysemous words that overlap in one, or possibly several, but not all of their senses. The challenge facing the researcher into putative synonymy is to distinguish between overlapping and non-overlapping senses of the lexemes being compared. One corpus-based method of establishing the extent of sense synonymy is the behavioural profile approach pioneered by Gries & Divjak (2009), whereby all tokens of the putative synonyms in a corpus are tagged for a wide range of morphological, syntactic and semantic features and then analysed using a hierarchical agglomerative cluster analysis. The results of this analysis show which lexemes are most similar to one another, both syntactically and semantically.

The present study makes use of what Krzeszowski (1990: 25) calls a 2-text translation corpus. The justification for the employment of cross-linguistic data in a study of synonyms in one language rests on the realisations that, on the one hand “Semantic relations are universal at both general and particular levels” (Murphy 2003: 40), while, on the other “all languages have different vocabularies, and so a map of relations in one language is not the same as that in another” (2003: 42). My hypothesis is that we can exploit the differences in coding of similar semantic relations in two languages to throw light on these relations in either language. For instance, it is axiomatic that if two lexemes or constructions in language A are completely synonymous, it should not be possible to predict the original form on the basis of translations into language B. In addition, the greater the degree of semantic overlap there is between two constructions in language A, the more difficult it should be to predict the original forms given their translations into language B. If there turn out to be significant differences between the translations of the two forms of putative synonymous lexemes or constructions, then we are forced to conclude that the translators are picking up on semantic differences between the two originals that were not apparent at first sight to advocates of their synonymy.

While a difference in translation equivalents should always alert us to the possibility of a difference in the semantics of the original items, the opposite is not necessarily the case. It may, for instance, turn out that a semantic distinction in the original language is not evidenced in the language of translation. In this sort of situation, one must take care not to be misled into postulating a greater degree of synonymy in the original language than is in fact the case. The opposite situation, where a semantic distinction in the language of translation is not evidenced in the original language, should not, however, carry the danger of leading to false conclusions.

Of the two phenomena explored in this paper, it is fair to say that polysemy has received a lot more attention in the literature in recent decades than has synonymy. The growth of interest in polysemy over the last thirty years or so was triggered by the exploration of the nature of linguistic categorisation that followed in the wake of linguists’ discovery of Rosch’s work on prototype theory (see, for instance, Geeraerts 2009: 183, Lakoff 1987: 15, Taylor 2003: 45). Of particular interest to linguists has been the interrelation of the various senses of a polysemous lexeme, as pointed out by Cuyckens, Sandra & Rice.

One of the major issues in cognitive lexical semantics over the past two decades has been the analysis of polysemous lexical items in terms of a family-resemblance network of multiple, interrelated senses of usage types. [...] The links between the different senses in a lexical network are manifold (conceptual/semantic overlap, metaphor, metonymy, image-schema transformation) and are supposed to represent the cognitive principles behind the processes of meaning extension. (Cuyckens, Sandra & Rice 1999)

One well-known example of the structure of a polysemous network is Langacker’s (1990) network for the verb run, reproduced here as Figure 1. The structure of the network in Figure 1 is informed by Langacker’s knowledge of his own mother tongue and his long training and experience as a linguist. The data on which it is based, however, are the product of his introspection. The introspective approach to the construction of networks may result in various scholars not only positing different nodes in a network, but also different numbers of nodes. In this connection, Tyler & Evans (2001) argue against what they term ‘rampant polysemy’. They maintain that the number of senses in a network should be limited to those spatial senses that cannot be understood given what they term the protoscene and contextual information as well as non-spatial senses. Nevertheless, their approach, like Langacker’s, is based on their own linguistically-informed analysis of data of their own devising.

If one wishes to avoid the dangers inherent in introspection as a source for the data to be analysed, one can make use of corpora. For instance, Gries (2006) makes use of a behavioural profile approach to investigate the polysemous network of the verb run. A hierarchical agglomerative cluster analysis showed that “the senses that amalgamated earliest are exactly those that are most strongly branching in the network-like representation of [a figure]” (Gries 2006: 83). The figure to which Gries refers here was drawn on the basis of semantic similarities between the various senses. The sense distinguished as the most central in the ‘run’ network is the ‘fast pedestrian motion’ sense which Langacker intuitively placed at the centre of the network reproduced above. Addressing the question of the links between the senses in the network, Gries computed correlations between the various senses and showed that some of them, including the ‘fast pedestrian motion’ sense and the ‘escape’ senses, are much more similar to one another than are other pairs of senses (Gries 2006: 80).

In my study of at in section 4 I again employ a 2-text translation corpus. The justification for the employment of translation equivalents in a study of polysemy rests on two common assumptions about linguistic networks. In the words of Tyler and Evans:

The lexicon constitutes an elaborate network of form-meaning associations (Langacker 1987, 1991a, 1991b), in which each form is paired with a semantic network or continuum (Brisard 1997). This follows from two basic assumptions, widely demonstrated within the framework of cognitive linguistics. First, semantic structure derives from and mirrors conceptual structure (see, for example, Fauconnier 1994, 1997; Heine 1997; Jackendoff 1983; Lakoff 1987). Second, the kinds of bodies and neural architecture human beings have – how we experience – and the nature of the spatio-physical world we happen to live in – what we experience – determine the conceptual structure we have. (Tyler & Evans 2001: 95)

If, as Tyler and Evans maintain, our conceptual structure is determined by the kinds of bodies and neural architecture human beings have and the nature of the spatio-physical world we happen to live in; and if, moreover, semantic structure derives from and mirrors conceptual structure: then one would expect that people with broadly similar cultural backgrounds might, given the similarity between their bodies and backgrounds, speak languages which contain similar (though, of course, not identical) semantic structures. In other words, one can hypothesise that:

If in language E the lexeme (or morpheme or construction) x is used with senses:
[a], [b], [c], [d], [e], [f], [g], [h], [i] ....
and in language N the lexeme y is used with senses:
[a], [b], [c], [e], [g], [j], [k], [l], [m] ....
and in language F the lexeme z is used with senses:
[a], [b], [d], [e], [i], [k], [l], [n], [i] ....

Then one can formulate a hypothesis that one or more of the senses [a], [b] and [e] represent(s) the prototype of each of the lexemes. This hypothesis presumes that one can identify senses cross-linguistically. But how can we be sure that the three lexemes, x, y and z code the same senses [a], [b] and [e]? In other words what evidence do we have for the (partial) equivalence of the semantic spaces in the three languages? In theory there are at least three sorts of evidence for cross-linguistic similarity. The first consists of introspection on the part of someone with a good knowledge of both languages. The second source of evidence might come from laboratory experiments, such as describing pictures or video snippets. The third consists of data from translation corpora. Translation corpora reveal which lexemes or constructions in one language are felt to correspond most closely to a given lexeme in another (see, for example, Dyvik 1998, 2004, Johansson 2007, Noël 2003).

Translation corpora have been mined extensively by Viberg (1998, 1999, 2002, 2003) in studies of the polsymies of certain very common verbs in various languages, including Swedish, English, Finnish and French. He demonstrates that translations can provide evidence of the internal structure of the polysemous network of a single lexeme. Moreover, different senses of a lexeme in language E which are usually translated into language N by one and the same lexeme (or construction) may be hypothesised, according to Garretson (2004), to be more closely related within the semantic network of the lexeme in language E than those translated by different lexemes. Should translations into other languages display a similar patterning, this would serve to strengthen the hypothesis. It goes without saying that any such hypothesis grounded on the testimony of translations should be evaluated in the light of other forms of (mono-linguistic) evidence.

3. Synonyms: begin and start?

Previous studies of begin and start, such as Freed (1979), Dixon (1991, 2005), Duffley (1999), Mair (2003) and Egan (2008), all agree that, while the two verbs both code ingressive aspect, in the sense that they both code a situation as (re)commencing, they are not completely synonymous. They also agree that it is difficult to tease out the difference between them. As Dixon puts it: “In many sentences start and begin may be substituted one for the other with little or no change in meaning [...]. But there do appear to be some semantic preferences for each verb, which motivates their use to a considerable extent” (Dixon 1991: 176 & 2005: 181). There is certainly a difference between the incidence of the two in spoken and written English, with start being significantly more common in spoken English (Egan 2008: 257).

One of the few clear semantic differences between the two matrix verbs posited in the literature is Freed’s (1979: 71) assertion that “[...] only from a sentence with begin does it necessarily follow that the nucleus (or characteristic activity) of the event has been initiated; a sentence with start followed by a to V complement can have as a consequence that only the onset of the event named in the complement has been initiated”. I will return to this assertion below.

There are 433 tokens of the lemma begin and 232 tokens of the lemma start in the English original texts in the ENPC. They both occur in four main construction types, listed here as (a)–(d). Raw figures for the distribution of each construction type are illustrated in Figure 2.

(a) to-infinitive complement constructions : i.e. begin/start to do
(b) -ing complement constructions: i.e. begin/start doing
(c) Intransitive constructions: i.e. begin/start #
(d) Transitive constructions with nominal objects: i.e. begin/start something

There are five main options for translators of expressions containing begin and start. These differ as to whether the ingressive aspect is encoded in the predication, and if this is the case, in the manner of its encoding. The five options and the number of times each is used are listed as (e)–(i):

(e) The Norwegian verb begynne: 481 tokens
(f) Other ingressive verbs (including starte ‘start’): 88 tokens
(g) Divergent forms (encoding ingression): 47 tokens [2]
(h) Ingressive aspect not translated: 34 tokens
(i) Whole phrase not translated: 15 tokens

There are thus four alternative ways of translating predications containing begin and start, given that option (i) consists of avoiding translating the predication altogether. In three of the remaining four options, (e) to (g), the ingressive aspect is encoded, albeit by different forms. In examples (1) and (2), both begin and start are translated by the Norwegian matrix verb begynne, which is cognate with English begin.

This is by far the most common strategy for translating both begin and start. In (3) and (4) the translators opt for alternative (f), coding the ingressive aspect by means of another verb.

Option (g), in which the ingressive aspect is coded by a non-verbal form is illustrated in (5) and (6).

Finally, the translator may choose to omit the ingressive aspect, in the case of the constructions with non-finite complements, (a) and (b), while still translating the predication in the complement clause. This option has been chosen in (7) and (8).

The extent to which the translators availed of the five strategies (e) to (f) is indicated in Figure 3.

The difference between the translations of begin and start shown in Figure 3 and Table 1 is significant at the p= 0.00001 level according to the Fisher Exact test with 4 df. However, the difference between the two forms in the construction with to-infinitive complements, type (a), is not significant at the p= 0.05 level (p= 0.264019) . Nor is there any significant difference between the forms with -ing complements, type (b) (p= 0.418827). It follows that the overall difference must be due to differences in either the translations of intransitive and transitive nominal constructions, types (c) and (d), or to both of these. This is indeed the case. The difference between the two forms in the intransitive construction is significant (p= 0.029692 according to the Exact test with 3 df) and the difference between the two forms in the transitive construction is also significant (p= 0.009382). [3]

Before proceeding to look in greater detail at the transitive nominal and intransitive constructions, I return to Freed’s (1979) assertion that by employing the begin to construction, as opposed to the start to construction, the speaker always guarantees that the action in the complement clause was actually carried out. Consider the evidence of examples (9)–(11).

The evidence of (9) confirms Freed’s conclusion that the complement of ‘start to’ may occasionally be interrupted before the subject has started to realise it. This element of non-realisation is brought out by the translator who employs the Norwegian equivalent of the ‘was about to’ construction. (10) is written by the same author and translated by the same translator as (9). It also contains the same complement predicate ‘move away’. However, the translation in (10) differs from that in (9) in that the subject is encoded as having taken the very first step towards the realisation of the complement event. The differences in the two translations may thus be seen as providing support for Freed’s contention. This cannot be said of example (11), in which the translation only serves to bring out the element of non-realisation already coded in the initial adverbial clause in the original English text. The translator has here availed of option (i), non-translation of the ingressive aspect. One way in which Freed’s thesis regarding ‘begin to’ could be saved is by considering the action of speaking as containing a prior act of thinking. Thus the (non-)speaker in (11) had certainly thought of what he was going to say. He was, nevertheless, interrupted before he actually did so. At the very least we can conclude that (11) constitutes a possible counter-example to the hypothesis that ‘begin to’ always guarantees the initiation of the complement event.

Turning now to the two types of construction that exhibit significant differences in translation equivalents, the transitive nominal object and intransitive constructions, these were analysed with respect to various parameters, including animacy and specificity of the subject (and object), TAM (Tense, Aspect, Mood) features of the verb and adverbial modification. Many of these parameters proved not to be significant, and no further mention will be made of them. Figure 4 contains raw figures for the translation options in the case of the transitive nominal constructions.

To begin with one area in which the translations of the two verbs resemble one another closely, those translated by begynne, we may consider the semantic type of the objects in these tokens. Table 2 contains details of the categorisation of these objects in five broad semantic fields.

As an illustration of the similarity between the translations of the two English constructions, consider (12) and (13), both of which contain an object from the semantic field ‘life’, and (14) and (15), both of which contain an object from the semantic field ‘work’.

One would not need to consult a translation corpus in order to discover that start and begin occur with the same semantic types of object in the transitive nominal construction. A monolingual English language corpus could serve the same purpose. On the other hand, the fact that different translators employ the exact same forms in (12) and (13) and in (14) and (15) does serve to underline the extent to which the two constructions are felt to be synonymous.

The raw figures in Figure 4 show that start transitive constructions with nominal objects are more likely to be translated into Norwegian by ingressives other than begynne. A total of 18 tokens of start are translated by 10 forms, while 6 tokens of begin are translated by 5 forms. Two forms are common to translations of both verbs, grunnlegge (‘found’) and påbegynne (‘begin on’). Both tokens of grunnlegge are from the same text and the object in both cases is the same: colony (of bees).

Of the ingressives exclusive to translations of start, two stand out in terms of number. Four tokens of start are translated by its Norwegian cognate starte, as in (16), and three by sette i gang (‘get going’, lit. set in motion), as in (17).

The occurrence of starte as a translation equivalent of start, but never of begin, may possibly be attributed, at least in part, to translation effects, the cognate form being the first to spring to the mind of the translator. On the other hand, it may also be the case that the English and Norwegian lemmas share a semantic component absent from the semantic make-up of both English begin and Norwegian begynne. Such a component might consist of the lack of any implication of the prior existence of a related situation, and would also motivate the employment of verbs like sette i gang in (17) and oppstå (‘merge’, lit. stand up) in (28).

There are three ingressives exclusive to begin. Of these, only one occurs more than once, innlede, which means open (lit. lead in), as in (18).

The other two forms just used for begin are legge ut på (‘set out on’) and ta fatt på (‘get a grip on’).

The second construction that exhibits significant differences in translation equivalents of begin and start is the intransitive construction. Figure 5 contains raw figures for both verbs.

Figure 5 resembles Figure 4 closely, at least with respect to the tokens translated by options (e) and (f), i.e. translation by begynne and by other ingressives. As was the case with the transitive nominal constructions there is little to distinguish the tokens of start and begin translated by begynne. There is one difference between them, nonetheless, that is perhaps worth noting. This concerns the tokens in which they occur with time adverbials. There are fourteen such tokens in the case of both verbs. The difference between them is related to the definiteness of the adverbial. 11 tokens of start (79%) occur with an indefinite time adverbial, as in (19), while 10 tokens of begin (71%) occur with a definite time adverbial, as in (20).

As was the case with the semantic types of object in the transitive nominal construction, one would not have needed to consult a translation corpus in order to distinguish between the two verbs when used intransitively in terms of their collocating adverbials. On the other hand, sometimes the translations can reveal differences not immediately apparent in data from a monolingual corpus. For instance, the Norwegian adverbial på ny(tt) (anew) is used in the translation of six of the start tokens and none of the begin tokens. Two of these are cited as (21) and (22).

Of the remaining four tokens translated by på ny(tt), three contain ‘start again’ in the original texts and one ‘start all over again’. There are also three tokens where ‘start again’ is rendered by begynne igjen (= begin again) as in (23), as well as three in which this translation is employed for ‘begin again’, as in (24).

(24) is typical of the three ‘begin again’ tokens insofar as they all encode the resumption of a situation after a comparatively short time interval, as indicated here by the adverbial soon. When something is said to start again, on the other hand, the interval in question is much longer, both in the cases where the adverbial igjen is used by the translator, as in (23), or where the translator prefers på ny(tt), as in (22). It would seem that when an interrupted situation is re-begun, the trajector can continue where he or she had left off, as it were. When an interrupted situation is restarted, one is more likely to begin at the very start.

As is the case with the constructions with transitive nominal objects, there are significant differences between the two verbs when it comes to translations using other ingressive verbs. 27 tokens of start are translated by 12 ingressives, while 11 tokens of begin are translated by eight ingressives. There are three forms in common, ta til (lit. take to), used by the same translator in (25) and (26), komme i gang (get going) and sette i (set in).

Of the ingressives exclusive to start, there are seven tokens translated by its Norwegian cognate starte as in (27) and 4 by oppstå (‘emerge’, lit. stand up), as in (28).

We have already seen that starte is also used to translate start in transitive nominal constructions, like (16). It is never used to translate begin. Two other translation equivalents used more than once for start are skulle til (‘be about to’) and ta fatt på (‘get going on’, lit. get a grip on).

Of five ingressives exclusive to begin, two occur more than once, åpne (‘open’), as in (29) and innlede, as in (30), which also means open (lit. lead in).

To conclude this section on begin and start, we have seen that if we look at the various constructions containing begin and start through the prism of Norwegian translations, we are unable to predict originals in the case of begin to and start to, begin -ing and start -ing. These two pairs would appear to be, to all intents and purposes, sense synonyms. This degree of synonymy is presumably also a prerequisite for the current expansion of start at the expense of begin, a development noted by Mair (2002) and Skandera (2003). In the case of the constructions with nominal objects and the intransitive constructions, differences in favoured translation options may reflect semantic differences in the English originals. For instance, in the case of interrupted situations in intransitive constructions we have seen that the longer the interruption, the more likely its recommencement is to be coded by start rather than begin. In addition, only tokens of begin are translated by Norwegian equivalents of ‘open’, while only tokens of start are translated by oppstå (= stand up). The use of a verb meaning ‘open’ implies the pre-existence of a participant (trajector or landmark) or process, since one cannot open something that does not already exist. On the other hand, the use of a verb like oppstå (= stand up), implies the creation of a new participant or process. The implications of these two translation equivalents may be taken as support for Freed’s contention that “start refers to the onset of an event while begin refers to the initial temporal segment of the nucleus of an event” (Freed 1979: 71).

4. The polysemous network structure of at

According to Herskovits (1986: 127) at is one of three basic topological prepositions in English (the other two are in and on). She states that: “The use types of at center around one ideal meaning: at: for a point to coincide with another” (Herskovits 1986: 128). There are many possible types of spatial coincidence, i.e. many possible locations of the trajector vis-à-vis the landmark, as pointed out by Lindstromberg (who uses the term ‘Subject’ for the more common ‘trajector’).

AT itself is often quite vague about whether the Subject is near the Landmark but not touching it; whether it is right by the Landmark and touching it; or whether indeed the Subject is on, in or among the Landmark. For this reason, even the spatial meaning of AT is barely depictable. (Lindstromberg 2010: 173)

This variety with respect to the exact locational relation of the trajector to the landmark in the spatial meaning of at is reflected in the translations in the ENPC. While there are close equivalents in Norwegian to the other two main topographical prepositions mentioned by Herskovits, in and on, this is not the case for at. In fact we often find tokens of locational at translated by the Norwegian equivalents of the other two prepositions, as in (31) and (32).

(31) and (32) may be said to be underspecified as to the exact location of the trajector with respect to the landmark. That is, in (31) it is of no importance exactly where at (or in) the National Gallery the listeners were actually located. In other cases, such as in (33)–(35), the area in question is more circumscribed.

Even in cases where the coincidence of the trajector and landmark do not admit of any great variation in terms of actual location, their relationship may be construed in quite different ways, as evidenced by (36) and (37).

The examples in (31)–(37) do not exhaust the list of Norwegian prepositions used to translate spatial at. Other prepositions used by at least two translators include langs (‘along’), rundt (‘around’) and hos (‘at the home of’, ‘chez’). The very wealth of translational options serves to underline the underspecification of mutual spatial orientation coded by English at, at least when it codes location. When used with a motion predicate to encode the end-point of a path, there is much less variation in prepositional use, the majority of translators employing til (‘to’), as in (38), although we also find multiple uses of mot (‘towards’), på (‘on’) and ved (‘by’). Goal predications closely resemble motion predications, differing from them in that no physical motion in the direction of the target actually takes place. They are also most commonly translated by til (‘to’) as in (39), followed in frequency by på (‘on’) and mot (‘towards’).

Temporal predications coded by at are translated into Norwegian by a broad range of prepositions, the most common of which are i (‘in’), as in (40), om (‘about’/‘around’), as in (41), på (‘on’), as in (42), til (‘to’), as in (43), and ved (‘by’), as in (44).

In addition to spatial and temporal predications, I have divided the predications coded in English by at into five more main types. These include predications of perception, as in (45), of degree, as in (46), of reaction, as in (47), of event, as in (48), as well as tokens in which the at phrase functions as the complement of an adjective, as in (49). These I have labelled ‘specifying’ since they serve to specify the exact nature of the quality coded by the adjective. [4] There is no space here to illustrate the variety of prepositional and divergent constructions used to encode all five. Instead, each will be exemplified with the most common prepositional translation.

Of all types of predications containing at, predications of perception are the most homogeneous with respect to their translations. Indeed perception verbs are homogenous in their predilection for collocating with at in English (see Perek & Lemmens 2010: 11). Just over 84% of the tokens of ‘look at’, ‘stare at’ etc. are translated into Norwegian by a perception verb followed by the preposition på (‘on’). It may be worth noting here that it was only in the Late Modern period that at replaced (up)on in English as the default preposition with these verbs. As for predications of degree, at occurs in several very common fixed expressions, such as ‘at least’ and ‘at all’. These are most often translated by fixed expressions, as in (46) for ‘at least’. Two prepositions account for most of the translations of predications of reaction. The most common of these is over (‘over’), as in (47). The other is av (‘of’). Event expressions containing at, as in (48), are most often translated by expressions containing the Norwegian equivalent of ‘by’. Finally, in all but two of the specifying tokens translated by prepositions, the preposition of choice is either til (‘to’), as in (49), or i (‘in’).

The nine broad categories of Location, Motion, Goal, Time, Perception, Degree, Reaction, Event and Specifying do not suffice to categorise all tokens of at in the ENPC. There remains a residue of over a hundred tokens of idiomatic expressions, three of which are exemplified by (50)–(52).

As was the case with perception predications, på (‘on’), as in (50), is the most frequent preposition found in translations of idiomatic expressions containing at. However, the majority of idioms are not translated using a preposition at all, but by a divergent construction, as in (51) and (52).

There are 2,570 tokens in all of at in the English original fictional texts in the ENPC. In the case of 142 of these the translator has omitted to translate the phrase containing the preposition. Of the remaining 2,428 tokens, 583 are translated by a divergent construction. The remaining 1,845 are translated by a congruent construction, i.e. one containing a preposition. By far the most common of these is på (‘on’), which accounts for some 31% the congruent translations. Since it is the most common translation equivalent, the tokens translated by på were separated from the remaining congruent translations. All of the 2,428 translated tokens were labelled according to whether they were translated by a prepositional construction containing på, a construction with another preposition, or a divergent construction. The translations of the various senses were then compared to one another, yielding the statistical results contained in Table 3. In the table comparisons which yielded non-significant results at the p= 0.05 level are in red.

Table 3 shows that there are significant differences in mode of translation between the perception, time, degree and idiom senses and all the other four senses. The location sense is translated in a similar fashion to both goal and event, being significantly different to the remaining seven senses. The specifying sense is translated similarly to motion and reaction and the goal sense similarly to location and event. Two of the senses, motion and reaction, are similar to three other senses, while the event sense is the most central in the network, as it is translated similarly to four other senses. Figure 6 is an attempt to convey in a network the resemblance between senses in Table 3.

Figure 6 illustrates a network of the various senses of English at as reflected in Norwegian translations. More accurately, it represents a network of similarities and differences between the way ‘at-ness’ is encoded in English and Norwegian. The degree, time and idiom senses are placed in proximity to one another because, although the translation options chosen in their cases are significantly different at p = 0.05, they are still less different to one another, as indicated by the statistical calculations, than they are to the other senses.

The structure of the network for the senses of at in Figure 6 resembles in one important respect similar networks drawn for through (in Egan 2012) and between (in Egan forth.) on the basis of their Norwegian translation equivalents. In the case of all three English prepositions, spatial meanings, which are not only the most frequent, but also often taken to represent the core or basic sense of the prepositions, do not assume a central position in the network of cross-linguistic correspondences. The most central meanings, in the sense of the meanings which are most similar to other meanings, including the spatial ones, are metaphorical extensions of the spatial senses, the channel sense in the case of through and the scalar sense in the case of between. These two senses occupy the same position in their respective networks as does the event sense in the case of at. This suggests that, at least for the prepositions in question, we should distinguish between the basic sense and the central sense, despite the fact that both basicness and centrality are often taken to be characteristics of a category prototype. The semantic network of a preposition may instead resemble the sort of structure portrayed in Figure 7.

Figure 7 is an extremely simplified sketch of a possible prepositional network. It does not attempt to illustrate any possible links between the various extended senses. What it does serve to emphasise, however, is the fact that the basic spatial sense, while it may be both the most frequent and the most psychologically salient, and thus a prime candidate for status as category prototype may well not be the most central sense in the network. (See Arppe et al. 2010 for a discussion of some of the problems related to operationalising the notions of frequency and salience in positing prototypicality.) Centrality is often also taken to a characteristic of the prototype (see Gries 2006: 76). Should further research result in our positing similar networks for other prepositions, we may need to further nuance our conception of prototypicality.

5. Summary and conclusion

In this paper I have explored how translation corpora may contribute to the exploration of synonymy and polysemy. In section 2 I mentioned some previous approaches to the study of synonymy and polysemy and proffered some arguments as to how translation corpora may be exploited in the study of these sense relations. Section 3 was devoted to a case study of the putatively synonymous verbs begin and start. The evidence of their Norwegian translation equivalents points to the two verbs being sense synonyms when they occur with to-infinitive and gerund complements. When used intransitively, or with a nominal object, on the other hand, there are significant differences in the translation options chosen by translators. These differences may point to differences in the semantics of the expressions in the original language.

Would the use of a translation corpus prove a useful addition to the methodological toolkit of the linguist studying synonymy? Insofar as translation corpora allow us access to the intuitive insights of a cross-section of competent speakers, they are obviously to be preferred to the intuitions of a single analyst. More particularly, translation corpora can contribute to studies of synonymy in two ways. Firstly, similarities/differences in translation equivalents can alert us to similarities/differences in the original constructions of which we may not have been aware. Secondly, the actual forms chosen in the case of different translation equivalents may highlight aspects of the semantics of the original forms (in the present case åpne/innlede for ‘begin’ and oppstå for ‘start’). It goes without saying that translation corpora containing more than one language may be even more useful as diagnostics for differences and similarities.

Section 4 was devoted to a study of the Norwegian translation equivalents of the preposition at. Nine senses of the preposition were distinguished and the translations of these various senses were analysed as either congruent or divergent in form. The congruent tokens were further divided into those containing the Norwegian preposition på (on) and those containing another preposition. The translation equivalents of the nine senses were compared to one another and a network was drawn on the basis of these comparisons. It transpired that the event sense is the most central in the network, insofar as it is the sense with the greatest number of links to the other senses.

I asked above whether the use of a translation corpus would prove a useful addition to the methodological toolkit of the linguist studying synonymy. The same question may also be raised with respect to polysemy. A similar answer may also be given. The point has been forcefully made in recent years (by, among other, Arppe et al. 2010 and Gilquin & Gries 2009) that we need to supplement corpus evidence with experimental evidence. Translation corpora partake of some of the advantages of elicitation experiments. They contain the intuitive linguistic responses of competent language users to a series of linguistic prompts. Moreover, since translators are not aware that their products will be subject to linguistic analysis, the dangers inherent in the observer’s paradox do not apply. Just as systematic differences in translation equivalents may serve to highlight differences in meaning between putative synonyms, they may also highlight differences in meaning between the various senses of polysemous items. Translation corpora provide a rich vein for future research into these semantic relations.

Notes

[1] I would like to thank the editors and two anonymous referees for insightful comments and suggestions for improving this paper. I would also like to thank Susan Nacey for her cooperation on a previous paper on begin and start, “The (near-)synonyms begin and start: evidence from translation corpora”, which we co-presented at the conference ”Rethinking Synonymy” at the University of Helsinki in October 2010.

[2] By ‘divergent forms’ I mean forms that differ syntactically from the original item, while at the same time conserving the meaning of the original.

[3] The reason why there are only three and not four degrees of freedom in the case of types (c) and (d) is that translation option (h), which involves dropping the ingressive element while retaining the complement predicate is not possible in constructions which lack a complement predicate.

[4] One reviewer makes the point that “it would seem that tokens “of reaction”, as in “horrified at” are simply a special case of the “specifying” class, as in “good at”. Certainly there is a generalization that can be made about the fact that we may say “horrified/surprised/astonished/shocked at”, but I see the real connection here as between the adjective and the preposition, not between the preposition and its own complement.” While I agree that the relationship between the adjective and the preposition is similar in the two types, I would still prefer to distinguish between them on the basis of the larger constructions in which they occur. Thus in the reaction type the trajector is an Experiencer and the landmark of the preposition codes the Stimulus. In the specifying type, on the other hand, the trajector is a Theme and the landmark of the preposition merely further specifies the quality predicated of this Theme.

References:

Arppe, Antti, Gaëtanelle Gilquin, Dylan Glynn, Martin Hilpert & Arne Zeschel. 2010. “Cognitive Corpus Linguistics: five points of debate on current theory and methodology”. Corpora 2010 Vol. 5(1): 1–27

Cuyckens, Hubert, Dominiek Sandra & Sally Rice. 1999. “Towards an empirical lexical semantics”. Human Contact Through Language and Linguistics, ed. by Birgit Smieja & Meike Tasch, 35–54. Wiesbaden: Peter Lang. Reprinted in The Cognitive Linguistics Reader. 2007, ed. by Vyvyan Evans, Benjamin K. Bergen, & Jörg Zinken. 57–74. London: Equinox.

Dixon, Robert M. W. 1991. A New Approach to English Grammar, on Semantic Principles. Oxford: Clarendon Press.

Dixon, Robert M. W. 2005. A Semantic Approach to English Grammar (2nd. ed.). Oxford: Oxford University Press.

Duffley, Patrick J. 1999. “The use of the infinitive and the -ing after verbs denoting the beginning, middle and end of an event”. Folia Linguistica, 33, 295–331.

Dyvik, Helge. 1998. “A translational basis for semantics”. Corpora and Cross-linguistic Research: Theory, Method, and Case Studies, ed. by Stig Johansson & Signe Oksefjell, 51–86. Amsterdam: Rodopi.

Dyvik, Helge. 2004. “Translations as semantic mirrors: from parallel corpus to wordnet.” Advances in Corpus Linguistics: Papers from the 23rd International Conference on English Language Research on Computerized Corpora (ICAME 23), Göteborg 22–26 May 2002, ed. by Karin Aijmer & Bengt Altenberg, 313–326. Amsterdam: Rodopi.

Egan, Thomas. 2008. Non-finite Complementation: A Usage-based Study of Infinitive and -ing Clauses in English. Amsterdam: Rodopi.

Egan, Thomas. 2012. “Through seen through the looking glass of translation equivalence: a proposed method for determining closeness of word senses”. Corpus Linguistics: Looking Back - Moving Forward, ed. by Sebastian Hoffman, Paul Rayson & Geoffrey N. Leech, 41–56. Amsterdam: Rodopi.

Egan, Thomas. forthcoming. “Between and through revisited”. Proceedings of ICAME31, ed. by Joybrato Mukherjee & Magnus Huber. (Studies in Variation, Contacts and Change in English). Helsinki: Research Unit for Variation, Contacts, and Change in English.

Freed, Alice F. 1979. The Semantics of English aspectual complementation. Dordrecht: Reidel.

Garretson, Gregory. 2004. The Meanings of English “of”: Uncovering Semantic Distinctions Via a Translation Corpus. Unpublished MA thesis, Boston University.

Geeraerts, Dirk. 2009. Theories of Lexical Semantics. Oxford: Oxford University Press.

Gilquin, Gaëtanelle & Stefan Th. Gries. 2009.“Corpora and experimental methods: a state-of-the-art review”. Corpus Linguistics and Linguistic Theory 5(1): 1–26.

Gries, Stefan Th. 2006. “Corpus-based methods and cognitive semantics: the many meanings of to run”. Corpora in Cognitive Linguistics: Corpus-based Approaches to Syntax and Lexis, ed. by Stefan Th. Gries & Anatol Stefanowitsch, 57–99. Berlin: Mouton

Gries, Stefan Th. & Dagmar S. Divjak. 2009. “Behavioral profiles: a corpus-based approach towards cognitive semantic analysis”. New Directions in Cognitive Linguistics, ed. by Vyvyan Evans & Stephanie S. Pourcel, 57&75. Amsterdam: Benjamins.

Johansson, Stig. 2007. Seeing through Multilingual Corpora: On the Use of Corpora in Contrastive Ctudies. Amsterdam: John Benjamins.

Herskovits, Annette. 1986. Language and Apatial Cognition: An Interdisciplinary Study of the Prepositions in English. Cambridge: Cambridge University Press

Krzeszowski, Tomasz P. 1990. Contrasting Languages: The Scope of Contrastive Linguistics. Berlin: Mouton.

Langacker, Ronald. W. 1990. Concept, Image, and Symbol: The Cognitive Basis of Grammar. Berlin: Mouton.

Lakoff, George. 1987. Fire, Women and Dangerous Things. Chicago, IL: University of Chicago Press.

Lindstromberg, Seth. 2010. English Prepositions Explained: Revised Edition. Amsterdam: Benjamins.

Mair, Christian. 2002. “The changing forms of complementation in late Modern English: a real-time study based on matching text corpora.” English Language and Linguistics, 6(1): 105–131.

Mair, Christian. 2003. “Gerundial complements after begin and start: Grammatical and sociolinguistic factors, and how they work against each other”. Determinants of Grammatical Variation in English, ed. by Britta Mondorf & Günter Rohdenburg, 329–343. Berlin: Mouton.

Murphy, M. Lynne. 2003. Semantic Relations and the Lexicon. Cambridge: Cambridge University Press

Noël, Dirk. 2003. “Translations as evidence for semantics: an illustration”. Linguistics 41(4):757–785.

Perek, Florent & Maarten Lemmens. 2010. “Getting at the meaning of the English at-constructions: the case of a constructional split”. CogniTextes [en ligne] 5: 1–23. http://cognitextes.revues.org/331

Skandera, Paul. 2003. “Start doing or start to do: Is the gerund spreading in American English?” English Core Linguistics: Essays in honour of D. J. Allerton, ed. by Cornelia Tschichold, 343–352. Bern: Peter Lang.

Taylor, John R. 2003. Linguistic Categorization [3rd edition]. Oxford: Oxford University Press

Tyler, Andrea & Vyvyan Evans. 2001. “Reconsidering propositional Polysemy Networks: The Case of Over”. Language 77(4): 724–765

Viberg, Åke. 1998. “Contrasts in polysemy and differentiation: Running and putting in English and Swedish.” Corpora and Cross-linguistic Research, ed. by Stig Johansson & Signe Oksefjell, 343–376. Amsterdam: Rodopi.

Viberg, Åke. 1999. “Polysemy and differentiation in the lexicon. Verbs of physical contact in Swedish.” Cognitive Semantics. Meaning and Cognition, ed. by Jens Allwood & Peter Gärdenfors, 87–129. Amsterdam: Benjamins.

Viberg, Åke. 2002. “Polysemy and disambiguation cues across languages. The case of Swedish få and English get.” Lexis in Contrast, ed. By Bengt Altenberg & Sylviane Granger, 119–150. Amsterdam: Benjamins.

Viberg, Åke. 2003. “The polysemy of the Swedish verb komma ‘come’: A view from translation corpora.” Meaning through Language Contrast. Vol. 2, ed. by Kasia M. Jaszczolt & Ken Turner, 75–105. Amsterdam: Benjamins.