Using corpora to investigate antonym acquisition

Transcription

1 Using corpora to investigate antonym acquisition Steven Jones University of Manchester M. Lynne Murphy University of Sussex In this study, a purpose-built corpus, containing both child-produced and child-directed speech, is used to conduct a longitudinal examination of antonym use among children from the age of two to five years old. Previous research has tended to approach antonym acquisition using either elicitation techniques or corpora of printed adult language. In contrast, this research focuses on the speech of preschool children in naturally-occurring interactions. The discourse functions of antonymy in child-produced and child-directed speech are quantified and compared with those identified in adult, written English (Jones 2002). Despite its complexity, Ancillary Antonymy is found to be most common in child-produced speech, even from the age of two, perhaps because of its particular usefulness in structuring ideas and discourse. This study presents a detailed inter-corpus comparison, assesses the discourse functions of antonymy at different stages of childhood, and discusses the correlation between antonym use in child-directed and child-produced speech. Keywords: antonymy, child language acquisition, speech, CHILDES 1. Introduction Antonyms 1 are widespread in language. Using a 280-million-word corpus of broadsheet newspaper text, Jones (2002) estimates that as many as one in fifty sentences includes a pair of antonyms. This ubiquity in printed, adult English raises important questions about how, when and at what rate we learn International Journal of Corpus Linguistics 10:3 (2005), ISSN E-ISSN John Benjamins Publishing Company

2 402 Steven Jones and M. Lynne Murphy antonyms and their appropriate discourse contexts. Lyons talks about a general human tendency to categorize experience in terms of dichotomous contrast (1977: 277), and this study uses corpus data to determine whether children are indeed predisposed towards using antonyms as a mechanism for organising their discourse. That antonymy plays an important part in young children s lives is clear: stories, rhymes and songs make heavy use of lexical contrast; parenting handbooks (e.g. Eisenberg et al. 1989) recommend that binary opposition be emphasised in order to encourage cognitive development; and, in later childhood, the UK National Literacy Strategy promotes antonym awareness throughout primary education, as a means for building the metalinguistic skills required for reading (DfES 1998). Furthermore, developmental psychologists have found that children appreciate antonymy from very young ages and are able to grasp the notion without difficulty (e.g. Kreezer & Dallenbach 1929; Clark 1972). However, analyses of how children use antonyms in naturally occurring speech remain rare. 2 Most previous empirical studies of antonymy in discourse have focussed on adult English and made exclusive use of written language corpora (e.g. Justeson & Katz 1991; Mettinger 1994; Fellbaum 1995; Jones 2002). These studies have shown that antonyms co-occur at higher-than-expected rates and tend to favour a relatively small number of grammatical frameworks. This could be cited as evidence to support the view of some commentators (e.g. Ervin 1963; Charles & Miller 1989; Justeson&Katz 1991) that pairs of words come to be learnt as antonyms because of repeated exposure in particular antonymous constructions. However, aside from question marks about the circularity of this position, the problem here is that inappropriate data have been examined. Surely safe conclusions about child language acquisition cannot be based on corpora of formal, written, adult-produced data. It is for this reason that our research into antonym acquisition makes use of a purpose-built, longitudinal corpus of child-produced speech, and that spoken input received from adult speakers (usually the parents of each child) is included. We believe that by identifying and quantifying antonym use in a corpus of child-adult interaction, a clearer picture will emerge of how antonymy is acquired. This methodology differs from those used in earlier studies of antonym acquisition, most of which relied on elicitation techniques. For example, previous research in this area includes a study by Akiyama (1985), who asked 32 Japanese-speaking and 32 English-speaking children to state the opposite of sentences such as the ship is large in order to ascertain whether their prefer-

3 Using corpora to investigate antonym acquisition 403 ence was for semantic denial (the ship is small) orsyntacticdenial(the ship is not large). Similarly, O Dowd (1980) investigated semantic differentials by asking 48 children to grade ten pairs of antonyms, and Clark, Carpenter and Deutsch (1995) looked at how English and German-speaking children used unto create reversal verbs in contexts such as will you unpeel the banana? Elicitation techniques are useful for providing a snapshot of a child s familiarity with the notion of antonymy, but a corpus-based approach allows for usage, rather than metalinguistic awareness of antonyms, to be examined. 3 As our corpus includes child-directed as well as child-produced speech, the picture that emerges developed over a period of three years for each child allows for any correlation between antonym input and output to be assessed, and a profile of how antonym use changes throughout childhood to be created. 2. Data and methodology The CHILDES database, 4 as incorporated in the TalkBank project, 5 includes child-adult interactions gathered from three suitable longitudinal studies. 6 The data take the form of transcripts of recorded family interactions which took place when the child was between the ages of two and five. 7 The children A(be), N(aomi), and R(oss) are all speakers of American English. Indication of the size of the resulting ANR corpus is given in Table 1, whichshowsword count, turn count and mean length of turn (MLT) for language spoken by and spoken to each child at the three ages investigated. In total, the ANR corpus contains over half a million words, 280,000 of which are spoken by children. Though very small in comparison with written corpora 8 and fairly small in comparison with some adult spoken corpora, 9 this corpus remains large enough for us to make some initial observations about antonym use in childhood which can contribute usefully to the wider debate about antonym acquisition. However, future investigations of this issue will, we hope, be able to make use of denser longitudinal corpora (see Tomasello &Stahl2004). We searched for within-turn co-occurrences of 43 antonym pairs 10 in the ANR corpus, using the CLAN software developed for the CHILDES database. Although previous studies of antonym co-occurrence in adult language have measured antonym co-occurrence within sentences, we chose instead to look at co-occurrence within turns because (a) sentence punctuation is not reliable within the corpus, (b) sentence boundaries are not always clear in child lan-

4 404 Steven Jones and M. Lynne Murphy Table 1. Details of the ANR corpus child child-produced speech child-directed speech age types tokens turns MLT types tokens turns MLT Abe Naomi Ross Total 2;4.24 2; ;0 3; ;0 4; ;0.2 2; ;2 3; ;7 4; ;6 2; ;0 3; ;1 4; (or weighted (2476) 281,736 35,891 (7.850) (2342) 247,076 37,999 (6.502) average) guage, and (c) young children s limited syntactic abilities may make them particularly likely to contrast antonyms across sentence boundaries (e.g., This is big. That is little.). Child language studies typically concentrate on syntacticallydefined utterances, rather than turns, but again we felt that the short Mean Length of Utterance for children at the youngest ages would limit our ability to tell whether they were using antonyms contrastively. Furthermore, within the CHAT transcription system in CHILDES, sentential punctuation (.?!) is used to delimit utterances, but it is unclear that all transcribers used these symbols consistently. Therefore, the turn unit was deemed more reliable across the different data sets. Our aim was to cast the net as widely as possible, so we concentrated the search on morphologically simple, high-frequency nouns, verbs, adjectives and prepositions that were likely to be found in child-produced and -directed speech. In addition to canonical pairings, like big/little and large/small, some variant pairings, such as big/tiny and large/little, were investigated in order to determine whether they were found in the same types of antonymous constructions as the more established pairs. In total, 43 pairs were searched for. 11 Our search was designed to locate antonyms in inflected forms (e.g. lose, lost,

5 Using corpora to investigate antonym acquisition 405 Table 2. Antonym pairs found to co-occur in the ANR corpus bad/good before/after big/little big/small big/tiny black/white bottom/top boy/girl clean/dirty close(d)/open close/far cold/hot dark/light day/night different/same down/up dry/wet easy/hard empty/full far/near fast/slow father/mother forget/remember full/hungry happy/sad hard/soft hate/love heavy/light in/out large/small long/short lose/win loud/quiet loud/soft low/high new/old off/on losing)andderivedforms(loser, loss)aswellasintheirbaseform,and37pairs were found to co-occur at least once 12 intheanrcorpus.thesearelisted in Table 2. Morphological variations of antonym pairs were located using wildcards in the searches; for example the search term happ* captured instances of happy as well as happier and happiest. (Of course, non-target words, such as happen, were sometimes captured too, but these were manually excluded.) We also accepted any antonymous compounds and phrasal verbs which came to light (day-time/night-time, inside/outside, put in/take out etc.). Therefore, the real number of antonym pairs searched for and found exceeds that indicated by Table 2. Co-occurrences of antonyms were not analysed further if they proved not to be in any specific semantic relation. For example, the utterance it comes out early in the spring includes both in and out, but these words do not carry contrastive significance. Table 3 summarises the types of co-occurrences that were counted and not counted. In total, the ANR corpus was found to contain 415 (270 child-produced; 145 child-directed) instances of antonym co-occurrence that satisfied our criteria. Table 4 provides a breakdown of each antonym co-occurrence according to whether the pair was child-produced or child-directed, which child was involved in the exchange, and how old the child was when the exchange took place. The next step was to classify these 415 co-occurrences according to the categories of antonym use introduced in Jones (2002), a study that quantified the discourse functions of 3,000 intra-sententially co-occurring antonym pairs in a 280-million-word corpus of text from The Independent newspaper. In order to minimise the danger of inconsistent or idiosyncratic assignment, the authors separately coded the ANR data and then compared results. In the vast majority of cases the data were coded identically. In the remaining instances,

6 406 Steven Jones and M. Lynne Murphy Table 3. Summary of criteria used in search for antonym co-occurrence Types counted symmetrical inflection asymmetrical inflection parallel compounds antonymous non-compound and compound phrasal verbs with antonymous particles We re littler, yeah, and they re bigger I m bigger and I m not little They ride inside,theydon trideoutside Daniel goes in my house, I went outside This hippo got up and this hippo laid down Types counted only once iterative uses verbatim repetition within a turn up and down and up and down and up the sponge has to be in,notout,daddy, the sponge has to be in,notout Types not counted different parts of speech non-antonymous compounds lexicalised compounds with internal antonyms non-contrasting co-occurrences uninterpretable/ambiguous utterances turn the light off and make it dark in here It s K s birthday, soweshouldcallhim tonight. it s upside down it comes out early in the spring I m too far closer Table 4. Raw frequency of antonym co-occurrence in the ANR corpus source 2years 3years 4years allages tokens types tokens types tokens types tokens types Abe: child-produced Abe: child-directed Naomi:child-produced Naomi:child-directed Ross: child-produced Ross: child-directed All: child-produced All: child-directed either both authors agreed that one of the incongruent codings was a mistake, or, more rarely, discounted the item as being uninterpretable without further contextual information.

7 Using corpora to investigate antonym acquisition Results Data were examined from three different perspectives. Firstly, an inter-corpus approach was taken to determine whether the classification system used to describe antonym use in adult, written English could account for antonym use in child-produced speech, and, if so, how the distribution of categories differed across corpora. Secondly, a longitudinal approach to the data was taken to identify which discourse functions of antonymy were favoured at each stage of childhood. And thirdly, drawing on evidence from the previous two approaches and including frequency information about antonym input as well as output, a broader analysis of antonym acquisition was undertaken. 3.1 Comparing antonym use in The Independent and ANR corpora In this section, the way in which antonyms are used in a corpus of adult, written English is compared with the way in which antonyms are used in a corpus of child-produced speech. Intuitively, one would predict that most linguistic phenomena would function very differently across corpora with such dissimilar compositions. Infants aged two to five would not be expected to use antonyms in a fashion comparable to that of journalists writing for The Independent newspaper. However, in terms of the categories of antonym use introduced in Jones (2002), we found surprising similarities, as the distribution frequencies in Table 5 show. The most common discourse function served by antonyms in both corpora is that of Ancillary Antonymy. 13 In Ancillary Antonymy, two oppositions appear and the inherent contrast of the more established antonym pair is used to affirm or generate a second, parallel contrast. So, in the example given from The Independent in Table 5, the familiar opposition between dead and alive is used to emphasise a second contrast between communism and fascism. Likewise, in the ANR corpus example in Table 5, the contrast between stand up and sit down is closely linked to the contrast between mommies and daddies.inthe Independent corpus, 38.7% of antonym co-occurrences sampled contained an ancillary contrast. In the corpus of child-produced speech, the proportion of Ancillary Antonymy contexts was found to be 45.6%. Because of the syntactic and semantic acumen involved in creating such parallel constructions, our original prediction was that Ancillary Antonymy would be among those functions acquired in later childhood. In fact, we found it to be present in discourse from the age of two.

8 408 Steven Jones and M. Lynne Murphy Table 5. Discourse functions of antonymy in The Independent corpus, compared with the ANR corpus Ancillary Antonymy Coordinated Antonymy Transitional Antonymy Negated Antonymy Other usage Written English example from The Independent corpus 38.7% Communism may be dead,butfascism is most assuredly alive. 38.4% Whitehall was yesterday unable to confirm or deny other devolutions. 3.0% Inflation redistributes wealth from the sophisticated to the unsophisticated. 2.1% However, the citizen pays for services to work well,notbadly. Childproduced Speech 17.8% 8.6% example from the ANR corpus 45.6% Mommies sit down when they tee-tee and daddies have to stand up. 22.9% Yeah daddy at school they have big blocks and little blocks. 12.1% I went upstairs and got us a bath and then and went downstairs into the kitchen. 10.8% It won t hurt me it s just a little pear tree not a big pear tree. The reasons why children favour Ancillary Antonymy can only be speculated about at this stage, but it is possible that the human tendency to dichotomise which Lyons (1977:277) speaks of is being mobilised from the moment language acquisition begins. In that case, Ancillary Antonymy is helpful for children because it enables them to impose structure and order on their immediate environment. Using one antonym pair in tandem with another contrast is an efficient way for children to create or reinforce analogies that are helpful to them in organising and expressing their lexical knowledge. In doing this, the child is able to broaden their conceptual knowledge of the world around them. The corpus examples in (1) and (2) show how children use Ancillary Antonymy to dichotomise in this way. In the former, a contrast between two humans is emphasised; in the latter, the secondary contrast is non-human. (1) a. Ross (aged 2) He s a girl and you re a boy. b. Ross (aged 3) I m the big brother and Marky s the little brother. c. Ross (aged 4) Billy s inside is dirty and my inside is clean.

9 Using corpora to investigate antonym acquisition 409 (2) a. Ross (aged 3) Milk is good for you but gum is bad for you. b. Abe (aged 3) Time to go home the school is closed time to come back the school is open. c. Ross (aged 3) On this one the boys are different and the hats are the same. Ancillary Antonymy is frequently used to draw attention to differences between two people, as seen in (1), and one of those people is often the speaker of the utterance. In these examples, the antonymy of the more established pair (girl/boy, big/little and dirty/clean) is being exploited, and a contrast relating to people (he/you, I/Marky and Billy s/my) is created. Self-versus-other comparisonisparticularlycommoninyear2(accountingforhalfoftheancillarydata) but tapers off in the later years, perhaps as the child s sense of self-awareness and identity becomes more pronounced and such distinctions, therefore, become less important. The corpus also shows that young children invoke Ancillary Antonymy to create oppositions between non-human items, as seen in (2). For instance, the contrast set up between milk and gum is hardly a textbook opposition, but the child has used the basic, known opposition between good and bad to express the relative merit of these items with respect to dental health. In this way, a nonce opposition has been created between milk and gum. Ancillary Antonymy is the most frequent category of antonym use and, in both corpora, Coordinated Antonymy is ranked second. In Coordinated Antonymy, the inherent semantic opposition between the two antonyms is not emphasised, and their coordination signals inclusiveness or exhaustiveness, as in (3). (3) a. Abe (aged 2) I wanna play with my big cars and my little cars. b. Ross (aged 3) And you can either take it out or put it in. c. Abe (aged 3) hunhunh I m getting cold and hot. No dichotomy is set up in these utterances. Playing with my big cars and my little cars is synonymous with playing with all my cars. The child knows that antonyms express opposition, but also knows that this opposition can serve an inclusive, as well as an exclusive, function. The difference between big cars and little cars is understood by Abe; his comment signals an awareness that, in this coordinated context, it is the similarity between antonyms that is more noteworthy than the dissimilarity. So why is the relative frequency of Coordinated Antonymy only half that of Ancillary Antonymy in the ANR corpus when these two categories occur in nearly identical proportions in The Independent? Given the greater complexity

10 410 Steven Jones and M. Lynne Murphy of ancillary usage, this inter-corpus difference may seem counter-intuitive, but one possible explanation revolves around the greater cognitive sophistication of Coordinated Antonymy. 14 In order to use antonyms in this way, one must be able to conceive of unity between two things or properties that are logically too distinct to both hold for the same referent at the same time under normal circumstances. For example, the ANR corpus includes an utterance in which Abe (aged3)saysthatatschoolthey have big blocks and little blocks. Here, the child is noting that the general category of blocks can be divided into contrasting categories, but he is also demonstrating an awareness that the inherent difference between big blocks and little blocks does not demand that the two categories be treated as mutually exclusive in all contexts. Thus, even though Coordinated Antonymy is usually grammatically simpler than Ancillary Antonymy, it remains more cognitively complex. The child must understand the contrastive potential of antonyms before realising that this potential need not always be activated, and that antonyms can also be coordinated as a means to emphasise inclusiveness. In both corpora, the distribution of categories is dominated by Ancillary and Coordinated Antonymy. However, among lower-frequency categories, differences arise. For example, Transitional Antonymy and Negated Antonymy arebothmoreprevalentintheanrcorpusthaninthe Independent corpus. This is partly because young children tend not to use the other four minor categories 15 identified by Jones (2002), but it is also the case that these two classes share something in common with Ancillary Antonymy: they both employ antonyms to emphasise a contrast and are therefore about dichotomising rather than uniting. Transitional Antonymy describes a shift from having, being or doing one thing to having, being or doing its opposite. Its use reflects young children s natural interest in change and their need to differentiate temporally related things, states, and events, as in it used to be day-time, now it s getting night-time. Negated Antonymy involves the assertion of one term and the denial of its antonym, as in that s not little, that s big. Itsusebychildren indicates an awareness that the meaning of a word or phrase can be reinforced through explicit rejection of its opposite. Therefore, both discourse functions may be more suited to the informal rhetoric of spoken interaction 16 than to journalistic monologue. However, it should be noted that while Negated and Transitional Antonymy are relatively common in child-produced speech compared to the written corpus, the frequency of these classes is still only 12.1% and 10.8% respectively, while Ancillary Antonymy alone weighs in at 45.6%.

11 Using corpora to investigate antonym acquisition Tracking changes in antonym use over time So far, the discourse functions of antonymy in the ANR corpus have been shown to be broadly similar to those in a corpus of adult, written English. In this section, the children s antonym use is analysed longitudinally to determine the extent to which it changes throughout childhood. Table 6 shows the proportion of all antonym use assigned to each class for children at the ages of two, three and four. The proportional use of both Ancillary Antonymy and Coordinated Antonymy remains fairly constant as the child ages (always between 40% and 50% for Ancillary and between 20% and 25% for Coordinated). However, changes are more noticeable among other classes. For example, almost one quarter of all antonym use by two-year-olds involves Negated Antonymy. When children reach the age of three, this proportion falls to one eighth, and, in our data, Negated Antonymy does not occur at all among four-year-olds. Therefore, a pattern emerges here: very young children tend to use antonyms in a negated framework, perhaps because this discourse function is the most explicitly contrastive; then, as their linguistic competence improves, they increasingly use antonyms in more rhetorically sophisticated contexts and the proportion of negating XnotYframeworks decreases accordingly (as one might expect given the syntactic redundancy of the not Y component). Complementary explanations for the high frequency of Negated Antonymy among younger children involve behavioural and communicative factors. First, on the behavioural side, the age profile of Negated Antonymy use is consistent with the negative and rebellious behaviour patterns often associated with children during their so-called terrible twos. (Of course, whether ability to use Negated Antonyms is a symptom of negative attitudes or a cause of the terrible twos stereotype is up for debate.) Second, on the communicative side, rejection may be particularly important for young children because their ability to express themselves is Table 6. Proportional use of antonym discourse functions in the ANR corpus (by age) Age of Ancillary Coordinated Transitional Negated Other child antonymy antonymy antonymy antonymy usage all ages

12 412 Steven Jones and M. Lynne Murphy limited, and their need to emphatically correct misapprehensions is therefore greaterthanitisinlaterchildhood. Conversely, ANR corpus data show that the frequency of Transitional Antonymy increases with age, perhaps because children become progressively more concerned with how the world around them is changing and more able to describe these changes. Mastery of transitional frameworks (e.g. from X to Y) andtransitionalconjunctions (e.g. and then) provide the linguistic means by which such contrasts can be expressed. Aside from comments about natural progressions, such as that from day to night, manyofthecorpusexamples refer to things being lost then found, orgoingup then down. Bythe age of four, children use Transitional Antonymy almost as much as they use Coordinated Antonymy. The two examples below are typical of how Negated Antonymy and Transitional Antonymy occur in the ANR corpus. The former (4a) is an utterance from a two-year-old child and features relatively simple antonyms being used in a very basic grammatical construction. The latter (4b) is an utterance from a four-year-old child and makes use of a more complex contrast (involving comparative morphology) and a more sophisticated grammatical construction. (4) a. Naomi (aged 2) not out # in. b. Ross (aged 4) The first part is the hard part and then once we get this all mixed in here # then it s easier Thus, Negated Antonymy and Transitional Antonymy have an inverse correlation in the corpus: the former is most prevalent in early childhood and acts as a prototype for antonym use; the latter is most prevalent in later childhood and reflects a more advanced application of antonymy. Use of both Ancillary and Coordinated Antonymy remains relatively constant during childhood in the ANR corpus, as does antonym use serving other minor discourse functions. 3.3 Examining the correlation between antonym input and output Having established how children use antonyms and how these patterns of use change during childhood, we now compare child-produced and child-directed speech in the ANR corpus with the aim of identifying whether the distributions of antonym discourse functions are similar or different. Similarities between distributions would suggest that children parrot (to a degree) the discourse functions they hear around them; differences would suggest that other motivating factors are driving children s antonym use. Table 7 compares child-

13 Using corpora to investigate antonym acquisition 413 Table 7. Discourse functions of antonymy in child-produced speech and child-directed speech. Childproducedirected Child- Example from Child-directed speech speech speech Ancillary Antonymy 45.6% 39.1% Oh sometimes I m little and you re big and other times I m big and you re little? Coordinated Antonymy 22.9% 18.4% They re winter shoes that you can wear outside or inside. Interrogative 2.2% 15.5% Is she a good mommy or a bad mommy? Antonymy Negated 10.8% 12.1% That s not making it clean, that s making it dirty. Antonymy Transitional 12.1% 5.6% Tonight hasn t happened yet tonight follows today. Antonymy Other usage 6.4% 9.3% produced speech and child-directed speech, with an example of each discourse function supplied from the latter. In terms of the relative frequency of Ancillary Antonymy and Coordinated Antonymy, child-directed speech is similar to child-produced speech. When adults talk to children, their use of antonyms resembles that found in childproduced speech more closely than it resembles that found in adult, written English. For example, Coordinated Antonymy occurs at a rate of 18.4% in child-directed speech and 22.9% in child-produced speech, but 38.4% in adult, written English. This trend is also observable among some lower frequency classes. For instance, Negated Antonymy occurs in similar proportions in both child-produced and child-directed speech, but is only a fraction as prevalent in The Independent corpus. The most notable difference between child-produced and child-directed speech is the proportionally high use of Interrogative Antonymy in the latter. This is a new class of antonymy, not common enough to warrant separate treatment in the Jones (2002) investigation. However, the ANR corpus includes so many antonym pairs in an XorYquestion framework that they merit closer attention here. Note that while these questions have a superficial resemblance to Coordinated Antonymy, in that the antonyms are joined by a conjunction, the Interrogative has a very different semantic role. As mentioned already, Coordinated Antonymy indicates unification of opposed items. So, in the example in Table 7, shoes that you can wear outside or inside, the speaker indicates that there is no important difference between inside and outside for the wearer of the

14 414 Steven Jones and M. Lynne Murphy shoes. In contrast, the Interrogative framework is truly disjunctive, in that the answer to the question posed must be one or the other of the antonyms. 17 In a sense, Interrogative Antonymy is more closely related to Negated Antonymy, because the reply to Is she a good mommy or a bad mommy? could very well be She s a good mommy, not a bad mommy. This similarity between Interrogative and Negated Antonymy is interesting because both discourse functions are more common in the language to and of children than they are in adult, written English. The presence of Interrogative Antonymy in child-directed speech accounts for the fact that the other major types occur in smaller proportions, 18 though Ancillary Antonymy remains marginally more frequent than in adult written English, which is further testimony of the cross-corpus ubiquity of the class. The differences between children s and adults usage of Interrogative Antonymy is a function of the difference in status between child and adult speakers. Interrogatives are more widespread in child-directed speech because adults have more reasons to ask such questions. The ANR corpus shows that children are being given choices (do you want to do letters down here or up there?), being asked to supply further information (is this the new Billy or the old Billy?), or being quizzed for pedagogical purposes about a wrong antonym choice that they have made (Is that a big table or a little table?,saidafterthechild had erroneously referred to a little table as a big table). However, use of Interrogative Antonymy in child-produced speech is not motivated by pedagogical intent, but tends to reflect genuine requests for information (for example, three year old Ross s question should I turn it off or on?). Even though Interrogative Antonymy is much more common in childdirected speech than child-produced speech, it remains the case that antonym use overall is more prevalent in children s linguistic output than it is in their input. Table 8 shows the frequency of antonym co-occurrence per 100,000 words in the corpus for both adults (when talking to children) and children. At all ages, antonym use is proportionally greater in the speech of children than adults. Indeed, this differential actually widens throughout infancy: data show Table 8. Frequency of antonym co-occurrence per 100,000 words of in child-produced and child-directed speech Age 2 Age 3 Age 4 Child-produced speech Child-directed speech

15 Using corpora to investigate antonym acquisition 415 that two-year-olds produce slightly more antonyms than are directed at them, and that this gap increases at the age of three, then again at the age of four, by which time the proportion of antonyms in child-directed speech has actually begun to fall. Further longitudinal study is necessary to discover at what age children s antonym use becomes more adult-like, and a comparative study of child-adult interaction and adult-adult interaction is required to show the extent to which the adult use of antonyms here is normal or a special feature of child-directed speech. Nevertheless, these figures are interesting because they show that children s use of antonyms exceeds adults from the earliest ages, indicating that the process of learning antonyms and how to use them is not particularly gradual or input-dependent. As soon as children have the vocabulary and grammar required to do so, they use antonyms profusely. Table 9 provides a longitudinal breakdown of antonym use in the ANR corpus according to discourse function and speaker. This allows us to assess the degree to which children s output mirrors their input for each discourse function of antonymy at each of the ages examined. Some evidence that antonym use is input-driven is provided in Table 9. Transitional Antonymy, for example, is used with increasing frequency both by children and the adults who speak to them. This class of antonymy is favoured slightly more by children than adults at all three ages, but the rate of increase for input and output is very similar. However, the input-output correlation is not as close for all classes of antonymy. For example, the significant increase in adult use of Interrogative Antonymy as the child ages is not mirrored by output, nor is the steady fall in Ancillary Antonymy in child-directed speech. Conversely, the rapid decrease in children s use of Negated Antonymy is accompanied by a much steadier fall in input. In other words, longitudinal inspection Table 9. Proportions of antonym discourse functions in child-directed speech (CDS) and child-produced speech (CPS) by age Ancillary Coordinated Interrogative Negated Transitional other age Antonymy Antonymy Antonymy Antonymy Antonymy CDS CPS CDS CPS CDS CPS CDS CPS CDS CPS CDS CPS all

16 416 Steven Jones and M. Lynne Murphy of the corpus shows that similarity between aggregate input and output level for a particular discourse function (as shown in Table 7) neednot necessarily be a reflection of close year-by-year correspondence. Some discourse functions of antonymy are favoured in child-directed speech at certain stages of childhood (e.g. Interrogative Antonymy when the child is four) and others are favoured in child-produced speech (e.g. Negated Antonymy when the child is two), at least as far as this limited corpus is concerned. Thus, it seems to be the changing discourse needs of children and adults talking to children (as well as children s ability to master the necessary syntactic and semantic skills) that determine how antonyms are used. 4. Conclusions Though a larger corpus of child-produced and child-directed speech would allow firmer conclusions to be drawn, this paper demonstrates the value of a corpus-based approach to language acquisition, and sheds new empirical light on the ways in which antonymy a ubiquitous and important feature of any language is learnt and used. The corpus of child-adult interaction investigated in this study contains over half a million words, over 70,000 turns and, most significantly, over 400 criteria-fitting, analysable examples of antonym co-occurrence. This is only a fraction of the sample size used by Jones in his 2002 study, but spoken corpora especially spoken corpora which incorporate children s language present more problems than written corpora in terms of the resources available and analysis required. Therefore, the conclusions drawn from this investigation represent a valuable contribution towards the debate about the ways in which children learn and use antonyms. The first conclusion reached involves inter-corpus similarity. In terms of the discourse functions developed by Jones (2002), antonym use in the spoken language of children is, broadly speaking, comparable with that found in the written language of adults. Though some of the minor categories found in The Independent corpus did not occur in the ANR corpus, the two major uses of antonymy (Ancillary and Coordinated) recorded the highest frequencies in both corpora. In other words, when antonyms are used, they tend to be used in two particular ways, regardless of who they are being used by. Clark argues that even when children have only words in their active vocabulary, they use some two-thirds of them much as an adult would (1993:37). Our research

17 Using corpora to investigate antonym acquisition 417 suggests that this is true of antonym use, which, according to the ANR corpus, tends to conform to adult-established discourse norms throughout childhood The second conclusion is based on longitudinal examination of the ANR corpus. By examining the frequency of each discourse function among children aged two, three and four years old, developmental trends in antonym use were identified. For example, the data show that Transitional Antonymy becomes more common as the child ages, but Negated Antonymy becomes less common. However, at all three ages investigated, Ancillary Antonymy was found to be more frequent than Coordinated Antonymy. This was an unexpected finding because the double-contrast complexity of the former (a cat sawoman s best friend and a dog s a man s best friend) would seem to require more syntactic sophistication than the latter (it s upstairs or downstairs). The explanation could lie in the fact that, cognitively speaking, it is easier to dichotomise (Ancillary Antonymy) than it is to unify (Coordinated Antonymy). By using a corpus of data collected over three-year time spans, we have been able to show that children s antonym use is an evolving, dynamic phenomenon, with changing discourse functions being used to express changing discourse aims. Our third and final conclusion offered more tentatively than the previous two is that childhood antonym output is not inextricably linked to input. In other words, no evidence of a firm correlation between how antonyms are heard by children and how antonyms are used by children was found. Within some individual classes of antonymy (e.g. Transitional Antonymy), a correlation between input and output is detectable, both globally and within individual years of childhood. However, other classes remain more prevalent in the speech of either children or adults (e.g. Interrogative Antonymy, which is frequent and rising in child-directed speech but almost non-existent in childproduced speech). Furthermore, some classes differ in frequency between early and late childhood despite having a similar proportion of aggregate input and output (e.g. Negated Antonymy, which is more common in child-produced speech when the child is two years old, but in child-directed speech when the child reaches four). The ANR corpus also shows that not only do children use proportionately more antonyms than they hear, but that the difference between frequency of antonym use in child-produced speech and child-directed speech actually increases throughout childhood. The fact that antonyms are used at earlier stages of childhood than previous, non-corpus-based studies have suggested leads to the supposition that antonymy is a key experienceorganising mechanism, exploited more in childhood than adulthood perhaps because of its useful properties for cognitive and linguistic development. This

18 418 Steven Jones and M. Lynne Murphy is reflected in the way that, once familiar with the basic grammatical frameworks associated with antonymy, children are able to be creative in their choice of antonyms. For example, when two-year-old Abe says I m not a tiny infant, I m a big infant, he is exploiting a standard Negating Antonymy framework but inserting a non-canonical antonym pair. Children s frequent use of Ancillary Antonymy also reflects an urge to experiment with new oppositions, as in two-year-old Ross s observation that Cookie Monster drinks it down and I drink it up.corpusexamplesofthisnatureshowhowchildrenuseantonymyto impose structure on the world around them and point towards antonym use in childhood not being entirely input-driven. This is consistent with Kagan s (1984) view that antonymy is not the product of painstaking instruction and adds weight to Lyons s (1977) claim that the human mind is drawn towards dichotomising. So, the findings of this paper support the contention of Murphy (1998)that children use antonymy in robust and creative ways from the earliest ages. They furthermore support Jones s (2002) contention that there are two major uses of antonyms in discourse: Ancillary Antonymy and Coordinated Antonymy. In using corpora to support these contentions, we have demonstrated the value of an empirical, quantitative methodology in the study of how semantic relations are acquired by children. However, this is no more than a first step towards a better understanding of antonym acquisition. Clearly, a corpus study of adult speech would provide a more realistic picture of the end-point of antonym acquisition than a written, journalistic corpus. Also, a longitudinal study based on a longer time-frame would allow the stage at which child antonym usage slows its development and becomes adult-like to be identified. Furthermore, a broader corpus that includes other input types (such as children s literature and other media, formal education and inter-child discourse) would indicate whether our findings for child-directed speech truly reflect the ways in which children are exposed to antonyms. Research of this nature is currently being undertaken by the authors. 19 Notes 1. Following Jones (2002) and Murphy(2003), the term antonymy is here used in its broadest sense to refer to all kinds of lexical-semantic opposition, rather than the narrower sense (covering gradable pairs such as cold/hot but not non-gradable pairs such as female/male) favoured by Lyons (1977) andcruse(1986).

19 Using corpora to investigate antonym acquisition An exception is Murphy (1998, also reported in Murphy and Jones, forthcoming), which shows that children from the age of two to five not only use antonyms contrastively in their utterances, they often do so at higher type and token rates than the adults speaking to them. 3. Previous research shows that until the age of five children do not know what an opposite is and do not exploit paradigmatic semantic relations in metalinguistic tasks such as free word association (McNeill 1966). After 5, children master the notion of oppositeness in a flashof insight (Kreezer/Dallenbach 1929) and show an early preference for antonymy over synonymy in various experiments (e.g., Heidenheimer 1978). Therefore, investigations which rely on elicitation techniques can provide only a limited explanation of antonym acquisition. 4. The CHILDES (Child Language Data Exchange System), administered by Brian MacWhinney, provides tools for studying child-adult interactions ( 5. TalkBank is an interdisciplinary research project funded by the National Science Foundation and undertaken by Carnegie Mellon University and the University of Pennsylvania ( 6. The Abe data were originally collected for Kuczaj (1976); the Naomi data for Sachs (1983); the Ross data is from the MacWhinney files in CHILDES (MacWhinney 1995). 7. Data for Naomi begin at 2;0, but data for Abe and Ross begin at 2;4 and 2;6, respectively. Some of Ross s speech is only partially transcribed in CHILDES, often consisting of short vignettes taken from longer conversations. 8. For example, The Independent newspaper corpus (used by Jones 2002, and against which findings are compared) consists of 280 million words. 9. For example, the spoken component of the British National Corpus consists of ten million words. 10. The index of pairs searched for was designed specifically to meet the stated research aims of this paper. Lists of antonyms created for other studies (e.g., Deese 1964; Lehrer 1985; Jones2002) were considered, but as these were designed to be representative of adult written language, their suitability for this project was limited. 11. As well as belonging to several different word classes, the pairs reflect a wide range of antonym types, including gradable pairs (fast/slow), non-gradable pairs (off/on), reversives (forget/remember) and gender-based antonyms (boy/girl). 12. Six of the pairs searched for did not co-occur in the corpus: empty/fill [verb], large/little, noisy/quiet, old/young, short/tall and yucky/yummy. 13. A detailed analysis of these classes, and how they were formulated, can be found in Jones The relatively low frequency of Coordinated Antonymy in the ANR corpus (relative to The Independent corpus) may also be explained partly by its tendency towards intrasentential use. As this study includes antonym co-occurrence which crosses the boundaries of sentences/clauses (unlike Jones 2002), those categories which allow for such usage, most notably Ancillary Antonymy, may appear over-represented.

20 420 Steven Jones and M. Lynne Murphy 15. The other four classes (together with an example from The Independent and frequency information for adult, written English) are: Comparative Antonymy (one has to be more optimistic than pessimistic; 6.8%), Distinguished Antonymy (the difference between right and wrong; 5.4%), Extreme Antonymy (conditions are either too dry or too wet; 1.3%) and Idiomatic Antonymy (teach this old dog a few new tricks; 0.8%). 16. Jones (forthcoming) finds the frequency of both Negated and Transitional Antonymy in the spoken component of the BNC to be lower than that found in the ANR corpus but higher than that of The Independent corpus. 17. Note that Coordinated Antonyms can be used in questions, but that this does not constitute Interrogative Antonymy as described here. An example from the ANR corpus of a Coordinated use in a question is: Does it go up and down? In this case, the answer to the question is not one member of the antonym pair, but rather the acceptance or rejection of both. Questions in which or linksthetwoantonymscanbeambiguousastowhethertheyare a Coordinated or Interrogative, although their prosodic contours will disambiguate them in speech. For example, Can you wear them outside or inside? is Coordinated, while Can you wear them outside, or [just] inside? is Interrogative. 18. If examples of Interrogative Antonymy are excluded (on the grounds that this discourse function is almost exclusively the preserve of child-directed speech) then the distribution similarity of major classes is even more striking: Ancillary Antonymy occurs at a rate of 46.3% in child-directed speech (compared to 45.6% in child-produced speech) and Coordinated Antonymy occurs at a rate of 21.8% (compared to 22.9%). 19. In particular, we are engaged in a study of the discourse functions of antonymy in the BNC spoken (Jones, forthcoming) and a more wide-ranging investigation of antonym input/output in childhood (Murphy & Jones, forthcoming). References Akiyama, M. M. (1985). Denials in young children from a cross-linguistic perspective. Child Development, 56, Charles, W. G., & Miller, G. A. (1989). Contexts of antonymous adjectives. Applied Psycholinguistics, 10, Clark, E. V. (1972). On the child s acquisition of antonyms in two semantic fields. Journal of Verbal Learning and Verbal Behavior, 11, Clark, E. V. (1993). The lexicon in acquisition. Cambridge: Cambridge University Press. Clark, E. V., Carpenter K. L., &. Deutsch, W. (1995). Reference states and reversals: Undoing actions with verbs. Journal of Child Language, 22, Cruse, D. A. (1986). Lexical Semantics. Cambridge: Cambridge University Press. Department for Education and Skills (DfES) (1998). The National Literacy Strategy framework for teaching YR to Y6. Accessed at literacy/publications/?pub_id=135&top_id=327&art_id=2100 (Mar 2003).

21 Using corpora to investigate antonym acquisition 421 Deese, J. (1964). The associative structure of some common English adjectives. Journal of Verbal Learning and Verbal Behavior, 3, Eisenberg, A., Murkoff, H., & Hathaway, S. E. (1989). What to expect the first year.newyork: Workman. Ervin, S. M. (1963). Correlates of associative frequency. Journal of Verbal Learning and Verbal Behavior, 1, Fellbaum, C. (1995). Co-occurrence and antonymy. International Journal of Lexicography, 8, Heidenheimer, P. (1978). Logical relations in the semantic processing of children between six and ten. Child Development, 49, Jones, S. (2002). Antonymy: A corpus-based approach. London: Routledge. Jones, S. (Forthcoming). A lexico-syntactic analysis of antonym co-occurrence in spoken English. Justeson, J. S. & Katz, S. M. (1991). Co-occurrences of antonymous adjectives and their contexts. Computational Linguistics, 17, Kagan, J. (1984). The nature of the child. New York: Basic Books. Kreezer, G. & Dallenbach, K. M. (1929). Learning the relation of opposition. American Journal of Philosophy, 41, Kuczaj, S. (1976). -ing, -s, and -ed: A study of the acquisition of certain verb inflections. Doctoral thesis, University of Minnesota. Lehrer, A. (1985). Markedness and antonymy. Linguistics, 21, Lyons, J. (1977). Semantics (2 vols.). Cambridge: Cambridge University Press. MacWhinney, B. (1995).The CHILDES Project: Tools for analyzing talk (2nd edn). Hillsdale, NJ: Erlbaum. McNeill, D. (1966). A study of word association. Journal of Verbal Learning and Verbal Behavior, 5, Mettinger, A. (1994). Aspects of semantic opposition in text. Oxford: Oxford University Press. Murphy, M. L. (1998). Acquisition of antonymy. Presented at the Linguistic Society of America meeting, New York, 8 11 Jan. Murphy, M. L. (2003). Semantic relations and the lexicon.cambridge:cambridgeuniversity Press. Murphy, M. L., & Jones, S. (Forthcoming). Antonymy in children s and child-directed speech. O Dowd, S. C. (1980). Semantic differential studies of children s language development: Some methodological problems. Journal of Experimental Child Psychology, 29, Sachs, J. (1983). Talking about the there and then: The emergence of displaced reference in parent-child discourse. In K. E. Nelson (Ed.), Children slanguage,vol.4(pp ). Hillsdale, NJ: Erlbaum. Tomasello, M. & Stahl, D. (2004). Sampling children s spontaneous speech: How much is enough? Journal of Child Language, 31,