We would like to thank Dorothy Kenny who first encouraged us in researching this topic, Roda P. Roberts for her very precious insight into the definitions of context and the history of lexicography, as well as for her advice on the methodology of the analysis, Anne Przewozny, Franck Sajous and François Maniez for their very sharp comments on the first drafts of the paper, and the reviewers for their helpful suggestions on how to enhance both the presentation and the contents of the study.

Introduction

1In a recent article (2023), Bartosz Ptasznik claims that “there is no doubt that example sentences enhance a dictionary’s value by providing dictionary users with the typical context of words.” [2023: 29]. Starting from the observation that dictionary entries of current English monolingual online learners’ dictionaries now contain a “plethora of examples” due to lexicographers “extracting detailed information from language corpora” [2023: 33], he wonders how many examples users really need and how many would lead to confusion. Using monolingual learners’ dictionaries and a translation task, he undertakes a study to investigate the usefulness to language learners of examples in language production and the number of examples needed for this purpose. He concludes that examples are beneficial to that target group in encoding tasks and that a large number of examples (up to eight) does not lead to confusion.

2The user group under study in this paper is seasoned or trainee translators, who cannot exactly be considered language learners. However, even though they are not targeted as the primary user group of such dictionaries, it is evidenced by surveys that, in conjunction with a wide range of other tools now at their disposal (online term banks, CAT tools, Machine Translation, etc.), this category of users does resort to them, as there are no monolingual dictionaries aimed exclusively at translators, and as this category of dictionaries provides richer context material and far more examples than most other types (Atkins & Rundell [2008: 222, 225], Dominguez Vázquez & Gouws [2023: 8]), something which is essential to translators both in understanding a source text and in producing a target text.

3Stemming from the oft-expressed desire on the part of translators as a distinct user group for more and better examples in dictionaries (e.g., Roberts [1994]), the aim of our study is to present a systematic review of the “additional example” feature in several online monolingual English learners’ dictionaries with the translator as the main user in mind. Both from a quantitative and a qualitative point of view, we aim to assess the usefulness and usability of the “additional example” feature for translators (trainees and/or seasoned) by critically reviewing the “additional examples” found for several words in various dictionaries.

4The present article is divided into four sections. Section 1. provides the general background of the study. It first discusses the theoretical notion of “context”. Then it describes the contextual needs of translators, in particular as evidenced by a number of surveys, and it lists the resources that they can turn to in order to glean contextual information. The next two subsections focus on those resources: first on dictionaries (their use by translators and the way contextual information is encoded in them through examples, and especially additional examples), and then on other technologized resources (online term banks, corpora, data-driven dictionaries, CAT tools and MT) with which dictionaries are now inextricably bound up. Section 2. outlines the methodology of the study, consisting in a literature-based justification of the focus of the study followed by a presentation of (i) the five online monolingual English learners’ dictionaries under study, (ii) the ten words under analysis and (iii) the variables to be analyzed. The analysis proper is found in Section 3. and explores the issues of general presentation, number, format, mode of selection and relevance to translators of additional examples based on the 1,306 extra examples provided in the five dictionaries under study. Lastly, Section 4. consists in a discussion of the results, which first puts forward a number of suggestions for improvement of the “additional example” feature before suggesting a number of avenues that could be investigated to refine the analysis and to better meet the needs of translators regarding contextual information in the age of e-lexicography.

1. Context, dictionaries and translation: an overview

5This section first discusses the notion of “context” from a theoretical point of view (1.1.) (definition, difference between context and example, types of contextual information, types of contexts/examples), then the attention shifts to the contextual needs of translators: the next subsection (1.2.) consists in the identification of their needs based on a number of parameters and a number of existing surveys, and of the resources that enable to meet those needs. The way in which, as assessed by surveys, dictionaries – especially monolingual ones – are used by translators, and the way they provide contextual information through (additional) examples is discussed next (1.3.). Finally, Section 1.4. contains a brief analysis of the way contextual information is provided in resources other than the monolingual general-purpose dictionary which are used by translators during the translation process.

1.1. Context

1.1.1. Definitions of “context”

6Context is a complex, multidimensional concept that can refer to various realities and can be approached in different ways based on the discipline (Dominguez Vázquez & Gouws [2023: 2]). It is particularly central to disciplines concerned with language use, such as pragmatics (House [2006: 340]), lexicography, and translation studies (House [2006: 338]). But even within the field of linguistics, it is not understood in the same way by all scholars. For example, Frédéric François in Martinet [1969: 65-66] defines it as follows:

Le contexte d’une unité d’une certaine nature sera alors défini comme l’ensemble des unités de même nature situées à proximité et qui, par leur présence, conditionnent la présence, la forme ou la fonction de l’unité considérée.

1.1.2. Context and examples

1.1.3. Contextual information

10(i) Semantic information, i.e., information about the meaning of a word

e.g., Three people have been killed in a bomb explosion in northwest Spain.

e.g., It was an explosion of anger against the practices of the occupying forces.

Killed and bomb illustrate one common sense of explosion, that of a violent burst of energy caused by a device such as a bomb, while anger in the second example shows how explosion can be followed by an abstract word to express a sudden violent expression of someone’s feelings.

11(ii) Collocational information, i.e., the words with which a linguistic element is commonly used

e.g., They warned him that a referendum might cause an explosion in the country.

e.g., ...the explosion of protest and violence sparked off by the killing of seven workers.

Cause an explosion or spark off an explosion are collocations.

12(iii) Syntactic information, i.e., how a linguistic element combines with others

e.g., The spread of the suburbs has triggered a population explosion among America’s deer.

This example illustrates in which syntactic pattern explosion is used: it is the direct object of the verb triggered and is followed by the preposition among.

1.1.4. Types of examples/contexts

13In the literature, both contexts and examples have been classified according to type depending on a) the purpose of the context/example, b) the type of contextual information presented and c) the amount of information provided. We will look briefly at three classifications: (i) Ptasznik’s for examples [2023], and those of (ii) Dubuc [1997] and (iii) Auger & Rousseau [1978] for contexts.

14Ptasznik [2023: 29] divides examples according to their purpose into decoding examples and encoding examples: the former, most necessary for lesser frequency items would contain more semantic information, whereas the latter would contain collocational or syntactic information.

15Both Dubuc and Auger & Rousseau have context categories that correspond to Ptasznik’s decoding example. Thus, Dubuc [1997: 72-73] proposes the following classification: defining, explanatory and associative “depending on the quality and quantity of semantic features”. A defining context contains the term and sufficient semantic features to provide a clear idea of its meaning. It does not necessarily constitute a formal definition. An explanatory context describes some characteristics of a concept without necessarily defining it (e.g., for the term shed, a context that explains how a shed is formed, without actually defining it). Finally, an associative context does not contain any semantic features; it simply enables the terminologist to link a term to the subject field through its association with the terms around it.

Dubuc’s first two classes of context above serve a decoding purpose. The third could provide some linguistic information, or could simply attest to the existence of the term.

16The Auger & Rousseau classification [1978: 34-36] also contains three categories: defining/descriptive, metalinguistic and attesting. The defining or descriptive context presents the semantic features of the term and shows the term in the context of other related terms. A metalinguistic context is a linguistic comment on the term and its usage. An attesting context provides little information other than attesting to the existence of the term. In this classification, the defining context corresponds to a decoding example. However, there is little support for encoding provided by any type of context enumerated.

1.2. Translators and context

1.2.1. What do translators need context for?4

1.2.2. Translators’ context-related needs as expressed through surveys

20All these findings justify the need for a study of contextual information as it presented in dictionaries (and other tools) used by translators.

1.2.3. Where do translators get contextual information?

21In a very recent paper, Dominguez Vázquez & Gouws [2023: 8-10] indicate that to obtain “at least some context information from a monolingual and/or bilingual perspective”, “a typical user (i.e., without (meta)lexicographic knowledge) can handle resources of different types”. They list the following resources: (i) a variety of (electronic) dictionaries, (ii) corpus-analysis tools, (iii) machine translation engines and tools, (iv) reading and writing assistants, (v) tools and programs that automatically generate various text types, literary sentences, texts from images and vice versa, and, in some cases, jokes, poems or stories. When it comes to a specific type of user, the translator, who does have some (meta)lexicographic knowledge and a rather advanced level of proficiency in both source and target language, this list has to be adapted. Categories (iv) and (v) are not relevant for our study as they refer to types of resources that are hardly – if ever – used by translators. Dictionaries, however, – and, among them, monolingual dictionaries – are essential tools found in the trainee or the professional translator’s toolbox, along with additional resources that are now an essential part of the translator’s increasingly technologized environment, such as online term banks, data-driven dictionaries (such as Linguee) and Computer-Assisted Translation (CAT) tools (also known as Translation Memories or TMs), and, to some extent, Machine Translation (MT) tools. We will first focus on (monolingual) dictionaries and the way contextual information is conveyed through examples, and particularly “additional examples” (Section 1.3.), but will also present other resources that translators can query for contextual information in Section 1.4. since, as will be shown in that section, the boundary between dictionaries and those other resources is now becoming blurred.

1.3. Dictionaries, translators and examples

1.3.1. Use of dictionaries by translators: findings of various studies

24Having stated the rationale for our study, we now focus on the way contextual information is encoded in dictionaries through examples, and in particular additional examples.

1.3.2. Examples as one way of encoding contextual information in a dictionary

26Another crucial function of examples is that they “provide models for language production” (Rundell [2015: 318]).

1.3.3. The “additional example” feature: what and why?

27Within the microstructure of monolingual dictionary entries, examples are traditionally nested at sense level right after a definition (Ptasznik [2023: 33]), and their number is usually limited, with an average of two to three.

28In addition to those “traditionally” presented examples, most monolingual dictionaries online now provide “additional examples” which, as will be thoroughly analyzed in Section 3., come under various names (“extra examples”, “more examples”, etc.), may be presented in different ways – e.g., they can be found within the dictionary entry or in an extra section –, may be automatically or optionally displayed, may be taken from specific corpora or from online sources, may have been edited or not, etc.

A passive dictionary could also be “activated” by giving the user the possibility to access relevant corpus data – e.g., concordance lines from target-language corpora for the potential translation equivalents given in the entry. [2002: 37]

31Cobb [2003] also advocated “a blend of dictionary and concordance” for language learners. About a decade later, when critically reviewing the terminological tools at the translators’ disposal, whose shortcomings regarding contextual information she highlighted, Bowker [2011: 215] made a similar suggestion. She showed how useful it could be for translators to generate concordances or frequency lists ‘on the fly’ [2011: 227-228]. In fact, a limited number of dictionaries, particularly learners’ dictionaries, started offering that option, which was one of the most innovative possibilities offered by the CD-ROM format [Nesi 1999: 62]. We will have a quick look at two of them, as the way they were devised and implemented has undoubtedly influenced the way the additional example feature is currently implemented in online dictionaries: (i) COBUILD and (ii) LDOCE.

33(ii) LDOCE
According to Asmussen [2014: 1082], it was the 4^th edition of LDOCE (2003) that first introduced the “Examples Bank” feature in the dictionary. The users could find additional examples of a “certain use” of a headword, which could be displayed both as KWIC concordances and as a listing of plain examples. A striking fact, though, is that the examples were not randomly sampled from the corpus, but “were chosen by applying ‘linguistic filters’ (which were not described). In that case the corpus was only used as “a repository of additional examples of how to use words and expressions”, but was not meant to be a stand-alone tool. Paquot [2012: 168-169] examined the next edition (LDOCE 5, 2009) and noticed that access data was offered in the microstructure of the dictionary. An example button on the right of the main entry allowed the users to access other dictionary examples, and also examples from the corpus (a sample of one million sentences taken from the Longman Corpus Network). Unlike COBUILD 5, the corpus was POS-tagged and there was no concordancing facility, which meant the corpus could not be used as a stand-alone tool.

35In 2023, the additional example feature, which started as a pioneering attempt by the COBUILD2, can be considered a “strong trend [in] English monolingual learners’ dictionaries which have been made available to us online” [Ptasznik 2023: 30], which justifies our study. However, dictionaries cannot be regarded as “stand-alone tools that should provide all the answers that the users need about language in use” (Varantola [2002: 34]), especially in the case of translators who rely on a combination of tools which tend to merge more and more, which is why we deem it necessary to devote a whole subsection to them.

1.4. Context in the translator’s toolbox: other resources

1.4.1. Specialized dictionaries and term banks

37In this paper, we have deliberately decided not to focus on specialized dictionaries and term banks. This choice may sound debatable: it could be argued that, as most of the professional translation takes place in specialized fields of knowledge (e.g., technical translation), translators make use of specialized terminographic resources such as term banks, much more often than general-purpose dictionaries. However, three main reasons account for our decision. First, various surveys, as we show in Section 1.3.1., have confirmed that translators still rely heavily on dictionaries and Bowker [2012: 382-384] has very convincingly described that what they actually use is in fact a combination of general and specialized tools, a behavior which is possibly more characteristic of trainee translators. In the experiments led by León-Araúz et al. [2019: 145] – a translation task carried out within a specific workstation, the participants, who were trainee translators, looked up specialized terms (estuary, storm-induced, detached breakwaters and soft cliffs) in general resources such as Cambridge dictionary. Second, not all professional translators translate heavily specialized texts. The information subtitlers need, for instance, has to do with everyday language (plus slang and colloquialisms, which are more likely to be found in (large) general dictionaries). Others include, for example, videogame localizers and people who translate biographies or general non-fiction. Third, and most importantly, there is so much information worth investigating that the topic is worth a separate paper (Josselin-Leray, underway). As a consequence, we will only briefly mention a few prominent elements here.

38There have been some attempts at so-called “contextual” specialized dictionaries – one can mention, in the field of earth sciences, the Dictionnaire contextuel de français pour les sciences de la Terre (Descamps [1973], which was not targeted at translators, or the Dictionnaire Contextuel anglais-français de la chromatographie, a bilingual dictionary explicitly aimed at translators (Serré [1981]), but these have remained isolated. Traditionally, contextual information is encoded in the “context” field of term records in online term banks, but it often fails to meet the translator’s needs for two main reasons: (i) the number of contexts provided – if any – is very limited and very often comes down to one single context (Bowker [2011: 214], Francœur [2015]), (ii) the type of contextual information included and the way the contexts have been picked is often unclear (Josselin-Leray [2018]).

1.4.2. Corpora

41Verplaetse & Lambrechts [2019] have very clearly summed up the findings of previous surveys in that respect. The MeLLANGE survey, in 2006, showed that only 41.8% of respondents used corpora when translating. More recent surveys, one in Spain, with 526 Spanish respondents (Gallego-Hernández [2015]) and one in Switzerland (Picton et al. [2015]), with 202 respondents, showed that the rate of translators using corpora has only slowly increased: “nearly 50% and even 70% of respondents […] use corpora sometimes, often or very often” (Verplaetse & Lambrechts [2019: 6]). The global survey conducted in 2015 by Zaretskaya et al. [2018] pointed to a very low adoption of corpora as translation aid by translators (the majority of participants – 85% of the 736 participants from 88 countries, had never worked with textual corpora). Finally, Verplaetse & Lambrechts’s own small-scale survey, conducted among 101 Belgian and Dutch professional translators in 2017-2018, only confirmed previous surveys: only 48 respondents (vs. 53) said they used corpora when translating. Another finding was that in-house translators seem to adopt corpus use more readily than freelance translators.

42Not only are translators not very familiar with corpora in general, but they are not familiar with corpus-analysis tools in particular. Zaretskaya et al.’s study showed that “one type of technologies, concordance systems, was completely unknown to a great number of translators who replied to [the] question” [2018: 45].

1.4.3. Data-driven dictionaries (Linguee)

Linguee […] invites its users, more than traditional (on or off line) dictionaries do, to look for the correct words and their correct form and combination in context: when one looks up a translation in the English-Spanish dictionary of Linguee, a drop down list pops up with different combinations of the word (collocations, prepositional phrases…) and after selecting one of these possibilities all occurrences of that combination in a large translation corpus are listed in a two-column format in context.

1.4.4. CAT tools or Translation Memories (TMs)

50Moreover, context can be considered as the “Achilles’ heel of translation technologies” and of translation memories, as stated by Killman [2015], a problem for which it is mostly segmentation which is to blame: these tools do not consider co-text beyond the sentence. Killman’s study has shown in detail that TM tools may suggest matches that are in some way semantically or lexically undesirable, even though they ‘match’ from a technical point of view (Killman [2015: 212]), and that TMs entail “contextual blind spots” [2015: 221]. The overall change of paradigm from a linear to a segmental non-linear approach in translation in TM tools has even led some scholars to theorize on the gradual disappearance of the source text as a complete textual unit (Bédard [2000]). Nevertheless, Jimenez-Crespo [2017: 162] qualifies this statement by saying that “nowadays it is clear that this lack of context and co-text can affect differently professionals and non-professionals”. He argues, referring to a previous study (Jimenez-Crespo [2013: 54]), that

from a cognitive perspective, professionals, through prior accumulated experience with similar texts or by constructing a mental model of what the texts might be, consciously or unconsciously can possess a model of the global text that compensates for this potential lack of communicative context or co-text. (Jimenez-Crespo [2017: 162])

1.4.5. Machine Translation (MT)

52Section 1. has provided the general background for our study by (i) discussing the theoretical underpinnings of context, (ii) presenting the translators’ needs regarding contextual information and the way this type of information has traditionally been found in dictionaries (iii), but also in other resources consulted by this distinct user group (iv). Despite the variety of resources at hand, and the fact that the distinction between those tools is not really clear-cut any more, the contextual information provided in the form of “additional examples” in online monolingual dictionaries can, to some extent, prove useful to translators as the analysis in Section 3. will show. But let us now introduce the methodology used to carry out that analysis.

2. Methodology

2.1. “Additional examples” in metalexicographical studies: a state of the art

Table 1. Number of extra-examples by dictionary in Szynalski’s findings

Cambridge 3	Collins 5	Longman 5	Oxford 7	Collins 6
58	180	131	32	0

55Heuberger’s study does mention the additional example feature, but since it depicts the then state of the art of monolingual online learners’ dictionaries “against the background of the unique opportunities of the electronic medium”, it only briefly examines this feature from the perspective of customization, hybridization and unlimited storage [2020: 409-412]. His study shows that (i) 4 out of the 6 dictionaries that he examines enable users to show or hide further examples, (ii) even though all publishers of those online dictionaries have their own corpora and would have the technical possibility of merging those corpora with the dictionary products, none has implemented a proper hybrid tool which would allow the user to have a direct access to the corpus, (iii) one area in which the extra value of online learners’ dictionaries appears in comparison with the corresponding print editions is the apparently tremendous increase in the number of example sentences and collocations – he cites the example of the entry goal (noun) in LDOCE, for which there is a 700-word difference in the size of the entry between the online and the print editions, and for which, in the online version, 24 further examples are added, in an extra section, to the 32 examples included in the regular entry. His straightforward suggestion that “The exact extent [of the apparently enormous increase in the number of examples and collocations] case would require – and be worth – a study of its own” caught our attention.

57Ptasznik’s study [2023] is an empirical study that did include a translation task, but which was conducted within the framework of language learning. 213 subjects participated in the study (143 university students, 70 high school students, all Polish learners of English at an upper-intermediate or advanced level). They had to translate 27 Polish sentences into English (which had been chosen because of the verb they contained) and were provided a bilingual dictionary and, depending on the experimental or control group, no examples, three or eight examples that were taken from the additional example feature of 7 online learners’ dictionaries. The study’s first aim was to investigate the general usefulness of examples in English monolingual learners’ dictionaries in language production, in other words, to explore the effectiveness of encoding examples. Its second aim was to determine whether the presence of eight examples in a dictionary entry confuses or benefits English learners. The overall conclusions that Ptasznik reached were that the presence of encoding example sentences in dictionaries had a “a significant influence on language production in the context of Polish learners of English” [2023: 43] and that “exposure to a multitude of encoding example sentences (eight-examples group) in a dictionary, mixed with target structure and non-target structure examples, does not confuse English learners when it comes to obtaining relevant information in the context of a language production task” [2023: 43]. Apart from the fact that the target users were language learners and not translators, the study shows three other limitations regarding what we want to study. First, it only focused on encoding needs, not decoding ones. Second, it mostly focused on lexico-grammatical patterns of use (and their correct use by participants), and did not delve into areas such as level of semantic equivalence, register, etc. Third, it focused extensively on quantitative, not qualitative, data. Fourth, the experiment was not carried out within an ecologically valid environment, since the participants had no access to other tools, Internet, dictionaries or books.

2.2. Dictionaries under study

59One methodological issue has to be mentioned here: our intention was to focus exclusively on learners’ dictionaries. However, we came to realize that the distinction is now blurred on several websites between dictionaries aimed at beginners, advanced learners and native speakers. For instance, the MW is not presented any more as a learners’ dictionary, and the Cambridge Learner’s Dictionary does not include any additional example because it is mostly aimed at beginners or intermediate learners. The Cambridge English Dictionary is described as being aimed at advanced learners of English, and includes additional examples, which is why we decided on that version of the dictionary. In fact, many dictionaries, such as the COBUILD, are now part of a portal, which leads to a rather confusing presentation. The COBUILD, for instance, is only accessible via the Collins English Dictionary.

2.3. Choice of entries to be analyzed

• Type of language (general language vs. specialized language): explosion vs. lava

• Morphology (simple noun vs. compound noun): house vs. greenhouse

• Register (unmarked word vs. informal word): money vs. dough

• Geographical variation (British word vs. American word): lorry vs. truck

• • 61 Since wokeness was not found in many dictionaries, we extended the criterion to the term awareness.

  Neologism: wokeness/awareness61

2.4. Choice of variables to be analyzed

61We looked up the ten words in the five dictionaries under study and looked at the following variables for the analysis of the example sentences: (i) general presentation and labelling, (ii) number, (iii) size and format, (iv) mode of selection and (v) relevance for translators. These are analyzed in Section 3.

3. Analysis

3.1. Labelling and general presentation

62We first looked at the way the five dictionaries under study named the additional example feature and how the additional examples were presented. The main information is summed up in Table 2.

Table 2. Various ways of labelling and presenting the additional example sentences

1	2	3	4
Dictionary	Label	Location	Show/hide further examples
CAMB	1. More examples	Within the dictionary entry	Yes
	2. Examples of [name of searchword]	Extra section	Yes
	3. All examples	Extra section	Yes
COBUILD	1. Examples of [name of searchword] in a sentence	Extra section	No
	2. Sentences (dictionary & corpus)	Specific tab outside of the dictionary entry	(Yes)
LDOCE	Examples from the Corpus	Extra section	No
OALD	Extra examples	Within the dictionary entry	Yes
MW	1. Example sentences	Extra section	(Yes)
	2. Recent examples on the Web		(Yes)

63Column 2 reveals a number of interesting facts. Not all dictionaries indicate explicitly that the examples provided are additional examples, which might be a bit misleading for users: in all cases, some examples are already included in the corresponding main dictionary entry. One dictionary (COBUILD, unsurprisingly) uses the word “sentence”, insisting on the fact that the headwords are presented in context, which might be more telling for advanced users such as translators. What seems a bit confusing is that, in three dictionaries out of five, there are in fact several ways of accessing additional examples: there are up to three different ways in the CAMB, and the COBUILD has both a list of examples at the very bottom of the page (Figure 2) and a separate tab under which both dictionary examples and corpus examples are included (Figure 3). Its usefulness might be questioned, all the more so as a quick look at the examples included in both options in COBUILD confirms some partial overlap (as will be seen in Section 3.2. about the number of examples). Another striking fact is that some dictionaries (LDOCE, MW) explicitly mention the source of the extra examples in the very label of the section (corpus, Web), which gives the user a hint about the fact that examples might not have been edited by lexicographers: on the one hand, they might be more authentic but, on the other hand, the reliability of the web-sourced examples might be questionable.

64Column 3 in Table 2 shows that there are two different ways of presenting the additional example feature: either within the dictionary entry in a distinct box, or as an extra section found after the dictionary entry or after several dictionary entries. When the extra examples are presented within the entry, they are necessarily found next to the relevant sense division, after “regular” examples, as can be seen in Figure 4 (CAMB) and Figure 5 (OALD). This suggests that the examples have been analyzed by lexicographers for sense mapping, which is extremely beneficial for users (as underlined by Rundell [2015]) but time-consuming for lexicographers, who have to do it manually. Only two dictionaries (CAMB, OALD) out of five have decided on that option, but only some sense divisions are illustrated by extra examples (e.g., in CAMB there are extra examples only for the first meaning (“the fact of something such as a bomb exploding”), but none for the other senses). The remaining three dictionaries, and the CAMB as well, have an extra section that is found after the entry. In the case of the LDOCE, the section is found at the very end of the entry, after the collocation boxes – when there are any, as can be seen in Figure 6. In the case of CAMB and COBUILD, the section appears after a potentially very long list of dictionary entries, since these learners’ dictionaries, as explained in Section 2.2., are in fact part of a dictionary portal. This, coupled with the fact that there is no sense mapping, might be considered a major drawback for users such as translators, who work under tight time constraints, since scrolling down the long list of dictionaries takes some time.

65Ease of access is an important criterion for this kind of user, and it goes hand in hand with customization. What the data in Column 4 implies is that for only half the dictionaries, users have the choice of hiding or displaying all or part of the additional examples. By default, extra examples in OALD are hidden, and users have to click on the “plus” sign in the title of the section to have them displayed. Conversely, in LDOCE for instance, extra examples are displayed permanently. CAMB is half-way: by default, all extra examples (the ones that are found within the dictionary entry) are displayed, and users can click on “fewer examples” if they want to see less. In MW (see Figure 7), a limited number (usually around three) are permanently displayed, and users can decide to see more. All in all, it seems that there is still rather limited customization of dictionary content in that respect and room for improvement, as Heuberger [2020: 409] has already noted. He concludes that “The more data electronic dictionaries offer, the more important it becomes for users to be able to determine how much and which information they require.” [2020: 412].

3.2. Number

66Table 3 below sums up the various numbers of additional examples found in the five dictionaries under study for the ten words we analyzed.

67Before listing the various tendencies that this analysis has unearthed, a few methodological issues we were faced with when counting the examples need to be addressed.
First, as mentioned in Section 3.1., many dictionaries include extra examples in various locations. We therefore had to add up all those examples, and the sum is presented within a “total” column for four dictionaries out of five.
Second, in some cases the examples that are provided do not correspond to the right part of speech; in some other cases the examples actually illustrate a compound (cmp) in which the headword is found, not the headword by itself (e.g., truck driver instead of truck). Only the LDOCE makes a clear distinction between the two, so that we decided to include all additional examples notwithstanding the possible confusion.

68Here are the more significant findings that our analysis of the data reveals.

3.2.1. Overall number of additional examples

69All in all, the total number of extra examples provided in all five dictionaries for just ten words is huge: 1,306. This seems to confirm Heuberger’s assumption that the general increase in the number of examples in online dictionaries is enormous [2020: 411].

3.2.2. Number of additional examples depending on the type of word

71The very last column of Table 3 displays the total number of examples per word. Here is the list of words in decreasing order of the number of extra examples: money (259), house (252), explosion (180), greenhouse (100), awareness (95), truck (88), lorry (85), dough (76), lava (74), wokeness (11). The general tendency seems to be that it is the most frequent and general words that get the highest number of examples (money, house, explosion) while more specialized (lava) or recent (wokeness) words get fewer examples. Some words do not get any additional examples at all, in particular, rather surprisingly, dough in CAMB. MW does not provide any additional examples at all for five dictionaries out of ten (dough, lorry, truck, wokeness, awareness) in its “more examples” section (but there are examples taken from the Web).

73What Table 4 reveals is that the number of examples per word seems to strongly correlate with the frequency of the word in the general language and thus to confirm our hypothesis: the most frequent words (money, house, explosion) get the most examples. Greenhouse, which gets a relatively high number of examples, is considered as a rather high-frequency word in the Oxford lists, which are more recent than the Longman 3000 (and which might rely on a corpus when environmental issues are more present). Words that are more specialized (lava) or recent (wokeness) are less frequent and thus get fewer examples. The case of dough is more complex: its low frequency might account for the rather low mean number of examples, but further investigation could reveal whether this is related to a register issue (informal) or a specialization (the baking field) issue.

74Such a correlation might be relevant for learners of English, who are the primary target groups of the dictionaries under study, and who need to master the core of the English language in order to communicate effectively in both speech and writing, but for translators, numerous examples for words that are less frequent would be particularly useful.

3.2.3. Consistency and customization

75On the whole there seems to be no precise policy regarding the number of examples to be provided in each dictionary for each word. For instance, in the “all examples” section of CAMB there are 84 examples for money vs. seven examples for lorry (which is slightly more than for truck, even though the former is supposed to be more frequent in British English); in OALD there are 67 extra examples for money and none for lava, greenhouse and dough. For all the dictionaries that are corpus-based, it probably depends on the contents of the corpus being used.

76The only aspect for which there seems to be some consistency is the number of examples that are first displayed before the user chooses to get more examples in CAMB (five examples, except for lava, greenhouse, wokeness, i.e., specialized or recent words), and in COBUILD (ten examples, except for explosion and wokeness). As for MW, it consistently provides eight examples in the Web-sourced examples section, three of which are initially displayed (Figure 7).

77Finally, there does not seem to be any particular criteria in the ordering of the additional examples, which might be randomly sampled and ordered in the case where they are automatically retrieved.

3.3. Size and format

78We analyzed the way the additional examples are visually presented to the user. Table 5 sums up whether the search word is highlighted or not.

Table 5. Highlighting of the search word in the five dictionaries under study

Name of dictionary	Label	Highlighted Search word
CAMB	1. More examples	No
	2. Examples of [name of searchword]	Yes (italics)
	3. All examples	Yes (italics)
COBUILD	1. Examples of [name of searchword] in a sentence	No
	2. Sentences (dictionary, corpus)	No
LDOCE	Examples from the Corpus	Yes (bold)
OALD	Extra examples	No
MW	1. Example Sentences	Yes (italics)
	2. Recent examples on the Web	Yes (italics)

79What stands out is that not all dictionaries have highlighted the search word (either by adding italics or bold type), which makes it harder to spot. The lack of this usability feature may be considered a drawback for translators. Indeed, one of our previous studies has shown that a context that was considered useful by translators was one that highlighted the search word: a majority of the translators interviewed “stressed the importance of clearly displaying the search word in bold or colour font in the context, for easier consultation” (Picton et al. [2018: 126]). The lack of visibility also makes the additional examples different from a KWIC concordance format where the search word is easy to observe, as we will discuss below.

80The length of the additional examples varies greatly from a very short noun phrase to a rather long and complex sentence, as exemplified in Table 6 for explosion and occasionally for lava and awareness. The web-based example for explosion in MW is 38 words long.

Table 6. Length of the additional examples

Name of dictionary	Label	Noun phrases	Full sentences
CAMB	1. More examples	No	Yes: Eight people, including two children, were injured in the explosion
	2. Examples of [name of searchword]	No	Yes: Section 4 contains necessary conditions for heteroclinic points to be chain explosions.
	3. All examples	No	Yes: The average velocity of the glass fragments was 295 ft./sec in the 18 g air-sphere explosions, and 215ft./sec in the 16g helium-sphere explosions.
COBUILD	1. Examples of [name of searchword] in a sentence	No	Yes: The force of the explosion could have blown him out into the sea.
	2. Sentences (dictionaries)	No	Yes: Three people have been killed in a bomb explosion in northwest Spain.
	3. Sentences (corpus)	No	Yes: The second guard was kneeling on the floor, his hands clasped over his eyes, temporarily blinded by the explosion.
LDOCE	Examples from the Corpus	Yes: An explosion of laughter A nuclear explosion Political awareness	Yes: Rabbits and ducks have been contributing to a population explosion in the park.
OALD	Extra examples	Yes: a great explosion of creativity a sudden explosion in the number of students	Yes: A huge explosion rocked the entire building.
MW	1. Example sentences	Yes: a flow of molten lava	Yes: The filmmakers staged the car’s explosion.
	2. Recent examples on the Web	No	Yes: And the Kremlin’s image took another battering this week with the apparent death of the mercurial mercenary leader Yevgeny V. Prigozhin in a crash following what U.S. and other Western officials say was an explosion aboard his private jet.

81Most of the examples take the form of full sentences. In examples that are automatically retrieved, sentences are defined as a group of characters starting with a capital letter and ending with a period. This kind of segmentation, which is the same as the one used in CAT tools (see Section 1.4.4.) might yield some problematic results. For instance, one can find the following truncated sentence in COBUILD for lava:

I was watching the lava flow out of Mt.
Christianity Today (2000)

where the end of the sentence has been omitted due to the fact that the volcano’s name already contains a period, a shortcoming that is well-known to people who use TMs and corpora. The length and both lexical and syntactic complexity of the sentences vary rather markedly depending on the dictionary and, quite obviously, on whether the sentences have been made-up or edited by lexicographers, or have not been edited at all (e.g., examples from the Web in MW). Examples that are shorter than a sentence seem to be taking the form of Multi-Word Expressions, in particular noun phrases: this is the case in three dictionaries: mostly in LDOCE, but also in OALD and occasionally in MW, but there are no explicit criteria.

82The phrase or the sentence format differs from the format allowed by KWIC concordances in that the search word is not instantly visible, unless in bold or italics, but, most importantly, because the searchword can be found at any place within the sentence, including the very beginning (e.g., explosion in OALD) or at the very end (which happens in most cases for explosion in Table 5). This implies that the user does not have access to the context to the right and/or to the left of the searchword, unlike a KWIC concordance format that uses a wide layout to display maximum in-context information – something that a real direct access to the underlying corpora would allow.

83The choice of noun phrases and full sentences can be accounted for by the fact that it is the format that most users, including translators, are familiar with and we can gather that the noun phrases have been selected because they correspond to collocations, but a KWIC format (which used to be available in CD-ROM versions of some dictionaries, see Section 1.3.3.) would allow for more context, and, in the case of a large number of examples provided, would “bring out patterns that are “immediately obvious” but which otherwise may pass unnoticed (Clear [1987: 41]).

3.4. Mode of selection

3.4.1. Manually or automatically retrieved examples

84Based on the information that was available on the various websites, we identified which dictionaries provided additional examples that had been automatically retrieved. The information is summarized in Table 7 below.

Table 7. Mode of selection of the additional examples

Dictionary	Label	Automatic retrieval
CAMB	1. More examples	No
	2. Examples of [name of searchword]	Yes
	3. All examples	Yes
COBUILD	1. Examples of [name of searchword] in a sentence	Yes
	2. Sentences (dictionary)	Yes?
	3. Sentences (corpus)	Yes
LDOCE	Examples from the Corpus	No
OALD	Extra examples	No
MW	1. Example sentences	No
	2. Recent examples on the Web	Yes

87The fact that examples are automatically retrieved has pros and cons. Obviously, this allows for a very large amount of data, and, as we will see below, of potentially up-to-date data. On the other hand, what may be considered a drawback is that, in some cases, the examples retrieved are irrelevant for various reasons:

1. the part of speech might be wrong. This happens when the corpus that is used has not been part-of-speech tagged. For example, in CAMB, the first two lines of the extra examples of house show first an example of the verb (“However, this is not because of the objects it houses”) and then of the noun (“Their relationship with the houses is beyond doubt, as they do not occur further away from the plan”). This also happens in the first two lines of the extra examples provided in the COBUILD for house (“Would they give one up to house a refugee family, The Sun, 2016”), and is also often the case in MW, with examples taken from the Web. While it is understandable that algorithms might fail to work on Web data which are not necessarily POS-tagged, it is rather surprising that both the Collins and the Cambridge corpora seem not to be POS-tagged. Another explanation could be that they are just instances of POS-tagging errors, or that the use of the POS-tagged version of the corpora is restricted to the lexicographers.

2. • 68 See for example http://termostat.ling.umontreal.ca/, accessed on August 21, 2023, which has been de (...)

   the word has not been identified as being part of a compound, which should appear in a separate dictionary entry. Greenhouse and lava can be considered as specialized words or terms, which implies that they are often used to build other, more complex terms, typically greenhouse gas and lava flow. The algorithms used to automatically retrieve the examples do not make a distinction between the stand-alone term and the term within a compound. For instance, five out of the first six additional examples provided for greenhouse in COBUILD yields do not illustrate the use of greenhouse, but the use of greenhouse gas. The same can be said of the examples provided by LDOCE for greenhouse, and greenhouse gases, while MW yields greenhouse plants and greenhouse farming. Among the 8 additional examples provided by LDOCE for lava, only half illustrate lava, while the other half illustrate lava flow (1), submarine lava flow (1), molten lava (1), river of lava (1). COBUILD provides at least one context for lava which, in fact, illustrates lava lamp, which might be totally irrelevant for the translator of a text dealing with volcanoes. This is an issue for translators who are often faced with very precise terms, and a well-known problem in terminology where it is often hard to determine whether a lexical unit is a term per se or just part of a compound term (L’Homme [2004: 166-200]). This problem, however, has been addressed (but not solved) by automatic term acquisition tools which can extract both simple and complex terms from corpora68.

3. • 69 https://www.lerobert.com/faq-dictionnaire-gratuit.html#exemples, accessed on August 21, 2023.

   there might be homographs – this issue is raised by the lexicographers about the free online version of the French Robert Dico en ligne69, but there were none in the corpus of entries we examined. This is probably the issue that lexicographers at CAMB have in mind when they ask users to report example sentences where “the word in the example sentence does not match the entry word”.

4. the same examples have been retrieved several times. Duplication of examples (which often happens with corpus tools) is an issue that lexicographers are aware of: they now allow users to send feedback about the data found in the dictionary, in particular in that respect. MW users are expected to report “duplicate example sentences”. Unlike some other issues such as homographs (iii) or absence of sense mapping (v), this problem could easily be solved automatically.

5. finally, which seems rather self-evident, there is no sense mapping, and there are many other semantic issues which will be discussed below. Lexicographers at MW are aware of the semantic weaknesses since users are invited to report the extra examples that fall into the following two categories: “error in the example sentence”, and “example sentence does not match the definition”. However, those two categories are too vaguely phrased (for instance, one may wonder which definition we are talking about when the word is polysemous), and might in fact better correspond to neology-related issues.

88To conclude on that aspect, what is particularly striking is that the main drawback that seems to matter to the publishing houses regarding automatic retrieval is the appropriateness of the content. COBUILD, MW and CAMB all warn against “opinions” or “sensitive content” that do not reflect the respective publisher’s opinions. Both CAMB and MW expect users to report example sentences that contain “offensive/harmful/abusive content” but all other issues, in particular the semantic one, are not mentioned.

3.4.2. Type of source

89Translators need to make sure the information they use to make translation choices is reliable, so they need to be provided with as much information as possible about the sources where the examples are taken from, which is why we now investigate this aspect.

3.4.2.1. Other dictionaries

90Only one dictionary, COBUILD, provides additional examples taken from other dictionaries from the same publishing house. It may be beneficial for users, but there is no mention of the dictionaries they are taken from and no formal indication as to whether they have been coined by lexicographers or taken from an authentic corpus. Given the size of the examples and the type of vocabulary they contain, it could be both. “The club is planning a public debate” or “This house believes that journalism has not gained from the introduction of new technology” seem to come straight from a corpus, while “she has moved to a smaller house” could have been either made up or corpus-based, but simplified. Providing metadata on the source of the examples would make the feature even more useful.

3.4.2.2. Identified “in-house” dictionary corpora

91Three dictionaries out of five explicitly mention that the additional examples are taken from a corpus: CAMB, COBUILD, LONG.

94Unlike the CAMB, the COBUILD does provide some metadata: name of the source (in most cases, the title of the newspaper or magazine) and year of publication. In the case of literary quotations, the name of the author is provided, but it is not always the case for the title of the book or the year of publication.

95All of the extra examples taken from LDOCE are very explicitly taken from “the” Corpus as explained in Section 3.1. There is hardly any information found on the website regarding the contents of the corpus, so we might assume that the corpus from which the examples are retrieved is the “Longman Citation corpus”, as it is referred to in the print versions, which has traditionally been used by lexicographers at Longman, originally as the Longman/Lancaster English Language Corpus which contained over 30 million words in 1993 (Summers [1993]).
Finally, OALD does not explicitly mention on the website that the examples are taken from a corpus.

3.4.2.3. “The Web”

97Only one dictionary out of the five under study includes additional examples that are automatically retrieved from “the Web”, namely MW (which is also one of the reasons why we included it in our study). According to the website, the extra examples “are programmatically compiled from various online sources”. However, the sources are not detailed. From the examples we analyzed, we gathered that the main websites that were being used for data collection belonged to various genres: newspapers and magazines from all over the United States (New York Times, Washington Post, Los Angeles Times, the Arizona Republic, USA Today…), American news websites (CNN, NBC news), nonfiction storytelling (Longreads), but also more specialized resources. The types of sources included here are, in fact, very similar to the ones found in the “in-house” corpora that we mentioned in the previous subsection, but we felt it necessary to make “the Web” a different subsection because the name was used as such by MW – which is why we use inverted commas.

98What is striking is that, for the three words under analysis that could be considered as belonging to a specialized field (greenhouse, lava and dough), the examples that are provided come from very relevant specialized or semi-specialized sources: e.g., for greenhouse: Better Homes and Gardens, Popular Mechanics, Scientific American, House Beautiful, Ars Technica; for lava: Smithsonian Magazine and Travel+Leisure; for dough; Southern Living and Better Homes & Gardens. Table 8 presents some of the examples that were provided at some point for those three words between early June and late August 2023.

Table 8. Example sentences taken from specialized sources

dough	Drizzle in water, 1 tablespoon at a time, and pulse until dough begins to clump together. —Pam Lolley, Southern Living, 25 Aug. 2023
	When the cycle is complete, remove dough from machine. —Bhg Test Kitchen, Better Homes & Gardens, 22 Aug. 2023
greenhouse	During power outages, close the curtains during the day to keep your home from heating like a greenhouse. —Lucy Tu, Scientific American, 5 Aug. 2023
	Unlike other small greenhouse kits on the market, this greenhouse is not tall, measuring only three feet in height. —Rachel Center, Better Homes & Gardens, 29 June 2023
lava	Flowing lava permanently altered the surrounding landscape and destroyed more than 700 homes, Brigit Katz reported for Smithsonian magazine at the time. —Nora McGreevy, Smithsonian Magazine, 20 May 2020
	During the initial eruption, fountains of lava spewed from fissures in the Earth. —Evie Carrick, Travel + Leisure, 9 Aug. 2023

99Another major advantage of using web-crawled data is that the linguistic evidence is up to date – something traditional corpora such as the ones used by COBUILD, for instance, cannot really compete with: we found very few examples from 2021 and 2022, and none from 2023. Even though the website does not give any indication about the frequency with which all examples are updated – what is mentioned is just the fact that examples are programmatically compiled “to illustrate current usage of [the looked-up word]”, regular look-ups on the website have shown that the web-based examples changed approximately every other week. This provides up to date material, but can also seem a bit frustrating for a user who consults the entry several days in a row and does not always find the same information.

3.5. Relevance of the various types of examples for translators

101Carrying out a fine-grained semantic analysis of the 1,306 additional examples provided for the ten words would have been worth another paper, so we will restrict the analysis here to some general tendencies that we were able to identify.

3.5.1. Sorting, ordering and sense mapping

102As shown above, only the examples provided in the “Examples Sentences” section of the MW (and potentially some examples found in the other COBUILD dictionaries) seem to have been made-up by lexicographers. The bulk of the examples are thus either corpus-based but edited by lexicographers (LDOCE, OALD) or raw extractions from corpora (CAMB, COBUILD, web-based examples in MW). While OALD provides extra examples that have been edited but, most importantly analyzed for sense mapping, LDOCE only provides long lists of unsorted examples, and the raw extractions provided by the remaining three dictionaries have not been filtered at all. In the latter case, there might be irrelevant examples because of duplication, wrong POS identification or wrong headword identification (part of a compound; e.g., lava lamp instead of just lava), as we showed earlier.

103The ordering of extra examples in all dictionaries, especially the ones that are automatically retrieved, does not seem to match any specific criteria, which is an issue for all types of users. The examples that come up first do not necessarily correspond to very frequent uses of the searchword, which might be an issue for language learners. For instance, the first extra example that is given for truck in COBUILD is “People have little truck with that kind of things, anyway” (Times, Sunday Times 2016). It does show an interesting pattern (to have (…) truck with sb/sth), but it does not help regarding the core meaning of the word or its most frequent use as a vehicle. Such an example, however, could be more relevant for a translator, who is more interested in examples related to less frequent uses of the word or of specific patterns. But he/she has no guarantee that the less frequent uses will come up first.

104Another consequence of the absence of sorting of the data extracted from corpora is the under- or over-representation of a given meaning, which is most probably very dependent on the contents of the corpora that are being used. This is particularly striking in the case of dough, where the informal money-related meaning is overwhelmingly underrepresented in the extra examples and the baking-related meaning overrepresented. Out of the ten extra examples provided in LDOCE, only two are related to the informal meaning (“He only married her for her dough”, “I’d go on vacation three times a year too, if I had his dough!”). The ten contexts that are automatically displayed for dough in COBUILD only evidence the baking-related meaning, but a quick look at the other contexts accessible via the “Sentence” feature allowed us to reach the same conclusion. The same kind of problem has been observed, for instance, with explosion and house in MW, where the more figurative meanings (e.g., an explosion of anger, the whole house) are underrepresented. Extra examples that correspond to the societal meaning of awareness cannot be found in any of the dictionaries of our study. This balance issue might be representative of the real difference in frequency between the two meanings, but a user who is not familiar with this kind of corpus-related issues might be left with no clue.

3.5.2. Types of contexts

105At the very beginning of the paper (Section 1.1.3.), we mentioned that several scholars or terminologists make a distinction between various types of contexts. Sorting the 1,306 examples under analysis based on specific types of contexts would have been a very tedious task, so here we just analyze a few that caught our attention as we were browsing through them. We picked 20 which are presented in Table 9, and for each example we tried to roughly assess the type of contextual information they contained, based on the classification we presented in Section 1.1.4. Such an analysis may be considered as subjective as it is sometimes hard to make clear-cut decisions, which is why there are some question marks here and there in the table.

Table 9. Classification of 20 extra examples based on the type of contextual information contained

106Here are the general trends that can be observed based on this selection of examples:

1. Some examples do not seem to bring any useful information at all.
   This is the case for examples (1), (2) and (3), and the only function that could be assigned to them is mere attestation (cf. the attesting contexts defined by Dubuc and Auger & Rousseau and presented in Section 1.1.4.). Typically, they are very vague examples (2, 3), or examples with problematic anaphoric references (1).

2. Some examples bring very limited information or only one kind of information.
   The only kind of semantic information that example (4) could bring, even if this may sound a bit far-fetched, is that lava fields is a place in which some events can take place. Example (5) only shows the collocation between lamp and glow (but this does not provide any specific information regarding lava).

3. Some examples might be wrongly classified.
   Example (6) can in fact be considered an idiom and should be classified as such, preferably with an explanation or definition.

4. Some examples lack intelligibility.
   Coupled with the fact that there is no access to the surrounding context, the vocabulary in some examples is so complex that the example might not prove very useful to users. This is the case for example (7), which is very highly specialized but with no defining elements, but also, to a lesser extent, for example (8) where the combination of informal vocabulary (top notch), and rather specialized vocabulary (wireless charging) makes the context slightly more difficult for non-native speakers, who might not spot immediately the fixed expression “for the money”.

5. Some examples do provide several types of information at the same time.
   A number of contexts provide both collocational patterns and some kind of semantic information: this is the case for examples (9) to (12), where we can see, for instance, that truck collocates (as an object) with the verb load, or (as a subject) with the verb block, and that it is used in the semantic field of transportation and vehicles (headed west, bridge, road, traffic…). Some contexts can bring both collocational pattern and pragmatic information regarding geographical variation (lorry is used in two contexts, (13) and (14), that relate to British life and British English – M62 and motorway). Some can bring some kind of semantic information and some syntactic information: example (15) shows the cause/consequence link between warming and greenhouse gases and the syntactic structure to be released into, example (16) shows the conceptual link between pollution and trucks and the preposition in pollution from trucks. Example (17), which relates to the specialized field of linguistics, provides both semantic and collocational information.

6. Finally, some examples can be considered to be much more informative than others.
   This is the case of examples (18) to (20), which are all specialized contexts. Example (18) is particularly rich conceptually speaking as it provides information on what volcanic eruptions consist in (expulsion of lava and emission of sulfur dioxide by volcanoes) and specialized collocational patterns (a volcano expels lava; gases such as SO₂ are emitted). The addition of “researchers said” is also a reliability indicator for the user. All three types of information can prove particularly useful for translators. Examples (19) and (20) also provide some conceptual information related to the field of geology and volcanology, but also show a very interesting pattern that is only found in specialized discourse: the transitive use of the verb to erupt (Josselin-Leray [2005: 453-455]; Josselin-Leray [2010]).

107What is rather striking is that very few metalinguistic contexts and very few defining contexts (two types of contexts that we identified in 1.1.4.) are found within the additional examples available in the five dictionaries under study, unlike what is recorded, for instance, in term banks. These types of examples might have been used directly by lexicographers when writing the definitions (Krishnamurthy [1987: 66]), so they may indirectly benefit users. But instead of trying to fit each of the examples into one particular category of context, another fruitful way of looking at the various types of examples might be to analyze their ‘degree of richness’ by resorting to the notion of “Knowledge Rich Context” which has been extensively used in the field of terminology and translation (Josselin-Leray et al. [2014], Picton et al. [2018]) – the examples analyzed above range from “poor” contexts (ex. (1) to (3)) to “very rich” contexts (ex. (18) to (20)).

108To briefly conclude on Section 3., it seems that more choice is not necessarily synonymous with improved choice, and one may wonder to what extent such a large number of extra examples might be useful to users, and especially translators. The lack of sense mapping in particular does not help for text comprehension or decoding, and the fact that collocational patterns are not always clearly highlighted within the examples does not provide immediate help for production or encoding purposes.

4. Discussion

109This section first puts forward a number of suggestions for improvement of the “additional example” feature with regard to the variables analyzed in Section 3 (Section 4.1.), then it presents some avenues to be explored for further analysis (Section 4.2.), before concluding on the “translator’s dream tool” (Section 4.3.).

4.1. Suggestions for improvement of the feature

4.1.1. General presentation

110One of the simplest suggestions for improvement regarding the presentation of additional examples would be, for dictionaries such as CAMB or COBUILD which are in fact portals, to group together in one single section all the examples that are scattered in several locations. The presentation in COBUILD is particularly confusing because the same ten examples that are found in the first section of the “example in a sentence” feature are also found in the “sentence” tab, either at the very beginning or sometimes mingled with all the other sentences.

111The labelling could also be improved in some cases: the label “examples from the corpus” for additional examples in LDOCE is misleading as it gives users the impression that examples provided within the dictionary entry are not taken from corpora while the very same examples are sometimes provided both within in the entry and in the extra section (e.g., “an explosion of laughter”).

4.1.2. Number

112Our analysis has shown that the number of additional examples available for a given word varies greatly, from simply none (with no obvious explanation, except, maybe, for a neologism such as wokeness) to above 80 (money in CAMB). Given the current opportunities provided by huge corpora and the Web, it seems to us that “none” should not be an option, and that some extra examples could and should be systematically provided.

113On the other hand, the presentation mode consisting in a very long list of examples that need to be scrolled down (as can be found in CAMB and COBUILD) might lead to a situation of information overload, and we may wonder how much is too much. Gouws & Tarp [2017: 394] warn against the “nature and extent of the data provided in online dictionaries – where overloading of data can be an inviting trap” and insist that there is an important balance to be reached between the storage space and the presentation space. As underlined by Ptasznik [2023: 45], further research by metalexicographers on the phenomenon of confusion possibly induced by this type of presentation is needed, especially in an experiment carried out in real-life conditions as we suggested above.

114All in all, the ideal number of sentences is hard to establish and, again, it might depend a lot on the user profile but also on a number of parameters such as the nature of the translation task itself, the conditions within which the task is performed, the stage of the translation process, etc. Ptasznik’s study, which tested whether eight examples helped or confused learners of English (as opposed to three examples only), concluded that

the upper-intermediate and advanced language users did not seem to be troubled by the excess of information offered to them in the translation tasks, and the availability of either three or eight examples led to approximately identical translation accuracy. […] [T]he availability of eight examples makes it marginally easier for the students to get necessary information, as opposed to the availability of only three examples [2023: 45].

116But even more than a precise number, what really matters is whether the examples have been sorted or not, as will be discussed below.

4.1.3. Size and format

117Our analysis has shown that the choice of the noun phrase or, in most cases, of the sentence as the unit of segmentation is far from always being satisfactory, something which has also been evidenced for TMs and MT (see Section 1.4.), since translators aim to achieve equivalence “at text level” (Bowker [2006: 176]).

119A middle course could be to resort to the paragraph as the basic unit of segmentation, as suggested by Bowker [2006: 184], but this might be harder to implement in the end than access to the full text.

4.1.4. Source and metadata

4.1.5. The need for sorted (or sortable?) information

121Back in the 1980s and 1990s, both Hanks and Varantola were very assertive about the role of the lexicographer:

Should electronic dictionaries then give an unlimited number of usage examples to suit every possible user need? It can certainly be argued that there is no point in giving the user an unsorted collection of examples. It is the lexicographer’s duty to crystallize the corpus evidence into well-chosen examples and to give the user a good breakdown of the lexical item’s behaviour, both its prototypical behaviour and its range, as well as its semantic mobility. (Varantola [1994: 608])

A vast citation file is no substitute for the judgment of the lexicographer. (Hanks [1979: xxxv])

123Systematic sense-mapping as is done in OALD is probably what is most relevant for users. However, it seems hard to implement on a large scale. In 2012, Hanks stated that “analysing meaning in context is a skill that is still in its infancy” [2012: 64] – so the full automation of sense disambiguation or sense mapping is still far from being a reality, as underlined by Bothma & Gouws [2022: 76].

124It might be more fruitful to investigate other ways of filtering information. For instance, the Cristal project aimed at providing translators with automatically retrieved corpus contexts that contained terminological relation markers, as these Knowledge-Rich Contexts were deemed very useful for translators (Josselin-Leray et al. [2014]). Those contexts were available within a CAT tool. Let us mention here that in 2016, a Sketch Engine plugin for CAT tools was created, but it was discontinued a few years later, probably due to the fact that translators did not feel comfortable with it.

4.2. Possible avenues to be explored

4.2.1. A larger corpus of examples

126Our study had a rather limited scope with only ten words, and only focused on one part of speech, nouns. Further studies could cover more words and other parts of speech such as verbs and adjectives, whose lexicographical treatment is often more problematic for translators. Studying series of words belonging to the same type (i.e., words belonging to a given semantic field, words with a formal or an informal register) might also reveal some interesting tendencies. A comparison of the treatment of the additional example feature by monolingual dictionaries in other languages (e.g., the French Robert Dico en ligne) could also provide some valuable insight.

4.2.2. More specific translator profiles

127As highlighted by Zetzsche, translators are “very, very diverse and versatile” [2023: 126]. The analysis we carried out here could be refined depending on the translator’s profile. Three distinctions at least seem particularly relevant. First, a more clear-cut distinction between seasoned translators and trainee translators would be beneficial when analyzing the usefulness and usability of the additional example feature. Indeed, seasoned translators who are more familiar with the terminology and concepts in their field of specialization, and who might use dictionaries less than other tools such as TMs, are more likely to be interested in saving time when they look up words in dictionaries; whereas trainee translators might make a more extensive use of ‘traditional’ online dictionaries and be more interested in the information they contain than on their time-saving potential. Second, a distinction between in-house translators and freelance translators could also prove pertinent. While the former might be more familiar with a number of translation technologies and tools, as suggested by the findings of Picton et al. [2015], the latter might need more user-friendly material. The relevance of introducing a proper concordance feature in dictionaries might be questioned and seriously investigated based on the user profile, as Rundell ([2015: 319]) had already pointed out. Finally, another criterion that needs to be taken into account is the directionality of the translation: the usefulness of the additional example feature might greatly vary depending on whether translators are translating into their first or second language. Varantola [2000] showed how specialized corpora – where most extra examples are taken from – have high “reassurance value”, particularly where the target text is in the translator’s L2, insofar as they illustrate similar contexts to those the translator is working on. The notion of context similarity is also critical for translators, as will be explained below (Section 4.3.).

4.2.3. The need for a workplace study

128Most importantly, our discussion of the usefulness of the various characteristics of the additional examples for translators is based on a number of assumptions regarding their needs, as they have been described in translation studies or as evidenced by surveys (see Section 1.2.). It would be worth confronting our findings with an empirical study. Translators with various profiles (e.g., trainee vs. professional, in-house vs. freelance) could be asked to perform a real translation task in an ecologically valid, or nearly ecologically valid, environment, i.e., with all the tools they normally have at their disposal (including other dictionaries, term banks, CAT tools, MT…) – unlike what Ptasznik [2023] did, where the experimental settings were rather artificial, as he himself acknowledges [2023: 147]. In that context, they could assess the usefulness and the usability of the extra example feature, in complementarity with the contextual information retrieved from other tools and the features found in those tools (e.g., WordSketch), in the same way as the usefulness of Knowledge-Rich Contexts automatically retrieved from a specialized corpus was assessed within the framework of the Cristal Project (Picton et al. [2018]), using logs, screen recordings, online surveys and retrospective verbalization tasks.

4.3. The translator’s dream tool: where to start from?

129Several lexicographers and translators have written extensively about the translator’s dream tool (Atkins [1996], De Schryver [2003], Orozco-Jutorán [2017]), which very often includes access to a corpus and corpus-analysis tools such as concordancers. Since translators still seem rather reluctant to use these tools (see Section 1.4.2.), it might be a better idea to start from what they are mostly familiar with (CAT tools) or becoming more and more familiar with (MT). As mentioned earlier, contexts that are considered most useful by translators are contexts that are similar to the texts they are translating. The technology that allows context matching and fuzzy matching in CAT tools is based on the very notion of similarity, so it might be interesting to see how this technology could be used to match the text to be translated with “resembling” contexts in the Source Language but also in the Target Language in large corpora available within the CAT tools. Bothma & Gouws [2022: 58-70] thus suggest “mapping […] a word in a text and an item in an e-dictionary article”, emphasizing that the presence of metadata could be especially helpful [2022: 76].

Conclusion

131The aim of this paper was to investigate the usefulness and usability for translators of the additional example feature found in online monolingual learner’s English dictionaries. This feature was first introduced by lexicographers to address the “typical complaint about dictionaries […] that [dictionaries] do not give enough usage examples” (Varantola [1994: 607]), a complaint that was voiced specifically by translators. The possibilities opened up by the availability of huge amounts of data from carefully designed corpora or even the Web need to be closely monitored if lexicographers do not want this feature to be reduced to just smoke and mirrors. What a rather detailed study of the 1,306 additional examples provided in five dictionaries has shown is that more is not always better, and that the translator’s needs regarding contextual information are still only partially met in the age of e-lexicography. We agree with Tarp & Gouws [2019: 264] that “full contextualization is still a challenge to modern lexicography”.

132These findings should be corroborated or invalidated by evidence from additional empirical studies where translators would deal with this feature in the context of a real translation task, at various stages of the translation process, and in conjunction with all the other tools at their disposal. Some findings of the Cristal project highlighted in Picton et al. [2018: 126] have indeed revealed that corpus-extracted contexts are “only really useful when they complement or are completed by other resources that should be integrated in the translator’s environment”.

133We believe, as Atkins already stressed in 1996, that users’ needs are paramount in dictionary-making. Producing a multi-purpose customizable tool that better meets translators’ needs might start with listening to what they have to say about the tools they use daily, especially translation technologies, and observe the way they interact with these. Translation workplace-based research (see Ehrensberger-Dow & Massey [2020]) could probably cast even more light on the subject.