Subject Omission In Children’s Language: The Case For .
Subject Omission in Children’s Language:The Case for Performance Limitations in LearningDaniel Freudenthal (DF@Psychology.Nottingham.Ac.Uk)Julian Pine (JP@Psychology.Nottingham.Ac.Uk)Fernand Gobet (FRG@Psychology.Nottingham.Ac.Uk)School of Psychology, University of NottinghamUniversity Park, Nottingham, NG7 2RD UKAbstractSeveral theories have been put forward to explain thephenomenon that children who are learning to speaktheir native language tend to omit the subject of thesentence. According to the pro-drop hypothesis, childrenrepresent the wrong grammar. According to theperformance limitations view, children represent the fullgrammar, but omit subjects due to performancelimitations in production. This paper proposes a thirdexplanation and presents a model which simulates thedata relevant to subject omission. The model consists ofa simple learning mechanism that carries out adistributional analysis of naturalistic input. It does nothave any overt representation of grammatical categories,and its performance limitations reside mainly in itslearning mechanism. The model clearly simulates thedata at hand, without the need to assume large amountsof innate knowledge in the child, and can be consideredmore parsimonious on these grounds alone. Importantly,it employs a unified and objective measure of processingload, namely the length of the utterance, which interactswith frequency in the input. The standard performancelimitations view assumes that processing load isdependent on a phrase’s syntactic role, but does notspecify a unifying underlying principle.Subject OmissionChildren who are acquiring English often producesentences with missing subjects, like those shownbelow.Hug MummyPlay BedWriting BookSee RunningWhile these examples clearly do not adhere to adultEnglish grammar, many contemporary theories of childlanguage assume that children produce their sentenceson the basis of an abstract grammar. Theories differwith respect to how much the hypothesized grammardiffers from the adult grammar. According to the prodrop hypothesis (Hyams, 1986; Hyams & Wexler,1993), children represent a grammar that is differentfrom the adult grammar in that it allows null subjects.In this respect, children’s grammar resembles that ofadult Italian and Spanish speakers. Other authors haveargued that children actually possess the correct adultgrammar, but drop subjects because they have difficultyexpressing the (correct) underlying form due to somekind of processing bottleneck (L. Bloom, 1970; L.Bloom, Miller & Hood, 1975; Pinker, 1984; P. Bloom1990; Valian, 1991). Thus, a child producing anutterance is thought to represent a grammaticallycorrect underlying structure, but, due to performancelimitations, some elements have a lower probability ofbeing expressed than others.A number of phenomena have been cited as evidencefor the performance limitations view. P. Bloom (1990)showed that, in utterances with a subject, the length ofthe Verb Phrase (VP) is shorter than it is in utteranceswithout a subject. The load associated with theprovision of a subject is thought to decrease thelikelihood of expressing a longer verb phrase. Alongsimilar lines, the length of the VP is greater when thesubject is a pronoun, than when it is a noun. This isthought to result from the fact that pronouns arephonetically shorter, and the fact that non-pronominalsubjects may be longer than pronominal ones. L. Bloom(1970) has also found that subject omission is morelikely in negated sentences or in sentences withrelatively new (unfamiliar) verbs. Presumably, the loadassociated with negation and novel verbs is such that itinduces subject omission.While the performance limitations view makes sensefrom an information-processing point of view, it is notvery precise in its predictions (Theakston, Lieven, Pine& Rowland, 2001). Performance limitations accountsalso tend to be rather ad hoc in nature. Given theimprecise nature of performance limitations, it becomesall too easy to posit a greater processing load wheneverthe provision of a certain element leads to a greaterlikelihood of the omission of another, especially whenthere is an interaction with frequency. Furthermore, it isnot clear whether an explanation of the patterns in thedata requires a limitation in production coupled withfull knowledge of a language’s grammar (as theperformance limitations view typically has it). In fact,as Theakston et al. point out, performance limitedlearning of lexical items (independent of syntacticcomplexity) may well give rise to the same pattern of
results without the need to assume a full representationof the grammar, and a different processing load forvarious types of grammatical roles. The present paperaims to test these claims by seeing to what extent aperformance limited distributional analysis ofnaturalistic input can account for the pattern ofomission and provision of grammatical categories thatis found in children’s speech. To this end, we aim tosimulate the effects that P. Bloom (1990) attributes toperformance limited production. We will now introducethe model we have used for these simulations.development of the net through the three presentationsof the sentence.MOSAICMOSAIC (Model of Syntax Acquisition In Children) isan instance of the CHREST architecture, which in turnis a member of the EPAM family of models. CHRESTmodels have successfully been used to modelphenomena such as novice-expert differences in chessand computer programming. In language acquisition,MOSAIC has been applied to the modelling of the useof optional infinitives in English and Dutch, thelearning of sound patterns and the Verb Islandphenomenon. Due to space limitations, we refer thereader to another paper in this volume for the relevantreferences (Freudenthal, Pine & Gobet, 2002).The basis of the model is a discrimination net, whichcan be seen as an index to Long-Term Memory. Thenetwork is an n-ary tree, headed by a root node.Training of the model takes place by feeding utterancesto the network, and sorting them (see Figure 1).Utterances are processed word by word. When thenetwork is empty, and the first utterance is fed to it, theroot node contains no test links. When the model ispresented with the utterance He walked home, it willcreate on its first pass three test links from the root. Thetest links hold a key (the test) and a node. The key holdsthe actual feature (word or phrase) being processed,while the node contains the sequence of all the keysfrom the root to the present node. Thus, on its first pass,the model just learns the words in the utterance. Whenthe model is presented with the same sentence a secondtime, it will traverse the net, and find it has already seenthe word he. When it encounters the word walked it willalso recognize it has seen this word before, and willthen create a new link under the he node. This link willhave walked as its key, and he walked in the node. In asimilar way, it will create a walked home node underthe primitive walked node. On a third pass, the modelwill add a he walked home node under the he walkedchain of nodes. The model thus needs three passes toencode a three-word phrase with all new words. (Forexpository purposes, here we assume that a node iscreated with a probability of 1. As is explained underlearning rate, this probability is actually lower anddependent on a number of factors). Figure 1 shows theFigure 1: MOSAIC learning an input.As the model sees more input, it will thus encodelonger and longer phrases. Apart from the standard testlinks between words that have followed each other inutterances previously encountered, MOSAIC employsgenerative links that connect nodes that have a similarcontext. Generative links can be created on every cycle.Whether a generative link is created depends on theamount of overlap that exists between nodes. Theoverlap is calculated by assessing to what extent twonodes have the same nodes directly above and belowthem (two nodes need to share 10% of both the nodesbelow and above them in order to be linked). This isequivalent to assessing how likely it is that the twowords are preceded and followed by the same words inan utterance. Since words that are followed andpreceded by the same words are likely to be of the sameword class (for instance Nouns or Verbs), thegenerative links that develop end up linking clusters ofnodes that represent different word classes. Theinduction of word classes on the basis of their positionin the sentence relative to other words is the onlymechanism that MOSAIC uses for representingsyntactic classes.The main importance of generative links lies in therole they play when utterances are generated from thenetwork. When the model generates utterances, it willoutput all the utterances it can by traversing the networkuntil it encounters a terminal node. When the modeltraverses standard links only, it produces utterances orparts of utterances that were present in the input. Inother words, it does rote generation. During generation,however, the model can also traverse generative links.When the model traverses a generative link, it can
supplement the utterance up to that point with a phrasethat follows the node that the current node is linked to.As a result, the model is able to generate utterances thatwere not present in the input. Figure 2 gives an exampleof the generation of an utterance using a generativelink.has been provided by Naigles & Hoff-Ginsberg (1998)and has been attributed to prosodic highlighting of thesentence final position (Shady & Gerken, 1999). Incontrast to the standard performance limitations view,processing load in MOSAIC does not vary as a functionof grammatical role. Also note that the version ofMOSAIC used for these simulations is identical to thatwhich Freudenthal, Pine & Gobet (2002) used for thesimulation of the optional infinitive phenomenon inDutch. No free parameters were fitted to obtain theseresults.Subject Omission in MOSAICFigure 2: Generating an utterance. Because she and hehave a generative link, the model can output the novelutterance she sings. (For simplicity, preceding nodes areignored in this figure.)Learning RateMOSAIC does not simply learn all the utterances itencounters. The probability of the creation of a node isdependent on the size of the net and the length of theutterance it encodes. This has the effect of making thelearning process frequency sensitive. If an utterance isseen more often, it has a higher probability of beingcreated. Finally, phrases that occur in an utterance finalposition in the input (have an end marker) have a higherprobability of being encoded. The precise formulagoverning learning rate is given elsewhere in thisvolume (Freudenthal, Pine & Gobet, 2002).Performance Limitations in MOSAICThe only performance limitations in MOSAIC are thefollowing: Frequency: high frequency items have a higherlikelihood of being encoded, and thus feature inlonger utterancesShort phrases have a higher likelihood of beingencoded than long phrasesUtterance final phrases have a higher likelihood ofbeing encoded.An utterance will only be produced (generated) ifits final phrase has occurred in sentence finalposition in the input.It may be appropriate to point out that theseperformance limitations are plausible from generaltheorizing in the cognitive psychology and learningliterature. Huttenlocher et al. (1991) provide evidencefor the effect of frequency on vocabulary learning.Evidence for the importance of sentence final positionMOSAIC creates utterances without subjects becausethe model can output partial utterances, provided thatthe utterance final element has occurred in a sentencefinal position in the input. As a consequence,constituents that take a position early in the sentence,have a higher probability of being omitted than thosethat take a position further downstream. Since thesubject takes first position in English, it has the highestlikelihood of being omitted. However, this prediction isnot tied to the English language. MOSAIC wouldgenerate utterances with omitted subjects in alllanguages that have the subject as the first element intheir underlying word order.MethodIn order to simulate the data presented by P. Bloom(1990), two MOSAIC models were trained usingcorpora of maternal speech available in the CHILDESdatabase (MacWhinney & Snow, 1990). We used thefiles of Anne and Becky. The mean length of utterance(MLU) in the output generated from the models was2.87 for Anne’s model, and 3.41 for Becky’s model. Inline with Bloom’s analysis, we limited our analysis toutterances which could not be interpreted asimperatives. This is necessary as subjectless sentencesin English are grammatical as imperatives (e.g. Put itdown). Bloom selected a list of nonimperative verbsand past tense verbs for his analysis. Since these verbscannot be used in an imperative form, sentences whichcontain a verb from these lists, and do not contain asubject, are true examples of subject omission. Tables 1and 2 give the lists of verbs that were used for theseanalyses.Table 1: Nonimperative verbs used for sLiveSneezeForgetLivesWantGrowLoveWantsKnowLoves
In line with Bloom’s analysis, we removed from oursamples all questions, all utterances that contained thewords not or don’t, all utterances where the verb wasnot used in a productive way, and all utterances wherethe target verb was part of an embedded clause.Table 2: Past tense verbs used for dWentFellSatWroteTable 3 gives the data for three children that Bloomreports and the two simulations (Anne’s and Becky’smodel). It can be seen that for the children, the VerbPhrase length in utterances with a subject is shorter thanin utterances without a subject. It can also be seen thatMOSAIC readily simulates this result, and the size ofthe effect is quite comparable to that in the children.The difference in verb phrase length is statisticallysignificant for both Anne’s model (t(330) 4.82, p .001), and Becky’s model (t(314) 4.64, p .001).Table 3: Mean length of Verb Phrases insentences with and without Eve2.022.72Sarah1.802.46Anne’s ModelBecky’s Model2.142.582.763.31MOSAIC obtains this result because the probabilityof learning an item in MOSAIC is dependent only on itsfrequency and length, and not on its grammatical role.There is thus no reason (apart from differences infrequency), why sentences with subjects should, onaverage, be longer (or shorter) than those without. Thefact that verb phrases in utterances with subjects shouldbe longer than verb phrases in utterances without asubject is a straightforward consequence of this fact.A second analysis performed by Bloom was to lookat the length of the verb phrase as a function of the typeof subject (no subject, pronoun or non-pronoun). Thereasoning was that, since the processing load of asubject is higher than that of a missing subject, and theprocessing load of a non-pronoun subject is higher thanthat of a pronoun (since the pronoun is bothphonetically shorter as well as shorter in word length),this should again result in length effects on the VerbPhrase. The results of this analysis are shown in table 4.Table 4: Mean length of Verb Phrase as afunction of subject sizeNo Subject rah2.451.901.50Anne’s ModelBecky’s Model2.763.312.452.931.601.67Again, it is clear that MOSAIC has no difficulty insimulating these results (though the size of the effect inMOSAIC appears to be slightly larger than in thechildren that Bloom analysed). The difference in verbphrase length between utterances with a pronounsubject and those with a non-pronominal subject isstatistically significant for both Anne’s model (t(64) 3.45, p .001) and Becky’s model (t(104) 4.40, p .001). There are two possible reasons why MOSAICmight simulate this result. Firstly, non-pronoun subjectsare on average slightly longer than pronoun subjects.Pronouns are by definition one word long, while nonpronoun NP’s can contain determiners and adjectives.In fact, Bloom indicates that the average non-pronounsubject for the children he analysed was 1.16 wordslong. Secondly, pronouns have a higher frequency ofoccurrence than non-pronominal subjects. In MOSAIC,this increases the likelihood that they will be learnt, andthe likelihood that they will feature in longer utterances.We decided to test these two explanations in MOSAICby performing the analysis on non-pronominal subjectsof length one and greater separately. As it turned out,only a small proportion of the non-pronominal subjectshad a length greater than one. For non-pronominalsubjects of length one, the size of the VP was 1.58 forAnne’s model, and 1.88 for Becky’s model1. Bothvalues are smaller than the VP length for pronoun1One would expect the length of the verb phrase to increasewhen limiting this analysis to subjects of length 1. This is thecase for Becky’s model, but not for Anne’s model. This is dueto the fact that, for Anne’s model, there were relatively fewlong non-pronominal subjects, but one of those that did occurhad a particularly long verb phrase.
subjects. Given the low incidence of long nonpronominal subject in both these and Bloom’s data, thisclearly indicates that the lower complexity effect thatBloom attributes to the fact that pronouns arephonetically shorter, can be explained by frequency inthe input. Note that MOSAIC does not employ aphonetic component. Phonetic differences can thereforenot have contributed to MOSAIC’s simulation of theeffect.The importance of frequency in the input as anexplanation for the difference between pronouns andnon-pronouns is also highlighted by a point made byHyams & Wexler (1993). Though pronouns may bephonetically shorter, the process of assigning thereferent to a (potentially ambiguous) pronoun mayactually result in its processing load being higher, ratherthan lower. This would predict a shorter Verb Phraselength for pronominal than for non-pronominalsubjects.Subject versus Object OmissionIt has often been shown that subjects are omitted moreoften than objects. In order to test how often objects areomitted, Bloom selected utterances which contain verbsthat require an object, and calculated the proportion ofobject omission from these obligatory contexts. Table 5shows this list of verbs.Table 5: Verbs that take obligatory veSaidWashedTable 6 compares the proportion of omitted subjectsand objects from obligatory contexts (verbs from tables1 and 2 for subjects, verbs from table 5 for objects). Itcan be seen that the proportion of subject omission isconsiderably higher than the proportion of objectomission. The subject-object asymmetry was significantfor both Anne’s model (X2 (1, N 560) 98.83, p .001), and Becky’s model (X 2(1, N 548) 125.97, p .001). Bloom suggests several possible causes for thisasymmetry. Firstly, it may be due to pragmatic factors.Since subjects typically convey given information,while objects convey new information, it may be morepragmatically appropriate to omit subjects whenprocessing capacity is limited. A second possible causemight be that there is a ‘save the heaviest for last’ bias.This would result in subjects having a higher processingload than objects, and as a result, in them being omittedmore often.Table 6: Omission from obligatory %Anne’s ModelBecky’s Model64%60%21%14%The explanation for the effect in MOSAIC is simple.As a result of MOSAIC’s performance limitations, aconstituent is less likel
phenomenon that children who are learning to speak their native language tend to omit the subject of the sentence. According to the pro-drop hypothesis, children represent the wrong grammar. According to the performance limitations view, children represent the full grammar, but omit subjects due to performance limitations in production. This paper proposes a third explanation and presents a .
General Standard Disclosures Page number (or link) Identified omission(s) Reason(s) for omission(s) Explanation for omission(s) External assurance
Mark Scheme 128 Unit 2 Paper 02 129 Mark Scheme 144 CAPE Food and Nutrition Subject Reports 2004 Subject Report 2005 Subject Report 2006 Subject Report 2007 Subject Report 2008 Subject Report (Trinidad and Tobago) 2008 Subject Report (Rest of the Caribbean) 2009 Subject Report 2010 Subject Report 2011 Subject Report 2012 Subject Report 2013 .
Independent Personal Pronouns Personal Pronouns in Hebrew Person, Gender, Number Singular Person, Gender, Number Plural 3ms (he, it) א ִוה 3mp (they) Sֵה ,הַָּ֫ ֵה 3fs (she, it) א O ה 3fp (they) Uֵה , הַָּ֫ ֵה 2ms (you) הָּ תַא2mp (you all) Sֶּ תַא 2fs (you) ְ תַא 2fp (you
Portfolio Media, Inc. 648 Broadway, Suite 200 New York, NY 10012 www.law360.com Phone: 1 212 537 6331 Fax: 1 212 537 6371 customerservice@portfoliomedia.com Is There A Hidden Omission In Your E&O Policy? Law360, New York (December 16, 2008
act and omission to act in Chapter 4 "The Elements of a Crime". For now, it is important to understand that criminal act, omission to act, . criminal law or criminal procedure? 2. Read Payton v. New York, 44
Coordinated Omission in NoSQL Database Benchmarking Steffen Friedrich 1, Wolfram Wingerath 2, Norbert Ritter 3 Abstract: Database system benchmark frameworks like the Yahoo! Cloud Serving Benchmark (YCSB) play a crucial role in the process of developing and tuning data management systems. YCSB
Benchmarking Cloud Serving Systems with YCSB, SoCC‘10, ACM, 2010. Friedrich, Wingerath, Ritter Coordinated Omission in NoSQL Database Benchmarking 3 YCSBs load generation _targetOpsTickNanos (long) (1 000 000 000 / target) long overallStartTime System.nanoTime();
Russell, S. and P. Norvig Artificial Intelligence: A Modern Approach. (Upper Saddle River, NJ: Prentice Hall, c2010) third edition [ISBN 9780132071482 (pbk); 9780136042594 (hbk)]. Russell and Norvig is one of the standard AI textbooks and covers a great deal of material; although you may enjoy reading all of it, you do not need to. The chapters that you should read are identified in the .
