Lexical Configuration and Lexical Engagement: When Adults Learn New Words

Laura Leach; Arthur G Samuel

doi:10.1016/j.cogpsych.2007.01.001

. Author manuscript; available in PMC: 2008 Dec 1.

Published in final edited form as: Cogn Psychol. 2007 Mar 23;55(4):306–353. doi: 10.1016/j.cogpsych.2007.01.001

Lexical Configuration and Lexical Engagement: When Adults Learn New Words

Laura Leach ¹, Arthur G Samuel ¹

PMCID: PMC2134982 NIHMSID: NIHMS34989 PMID: 17367775

Abstract

People know thousands of words in their native language, and each of these words must be learned at some time in the person's lifetime. A large number of these words will be learned when the person is an adult, reflecting the fact that the mental lexicon is continuously changing. We explore how new words get added to the mental lexicon, and provide empirical support for a theoretical distinction between what we call lexical configuration and lexical engagement. Lexical configuration is the set of factual knowledge associated with a word (e.g., the word's sound, spelling, meaning, or syntactic role). Almost all previous research on word learning has focused on this aspect. However it is also critical to understand the process by which a word becomes capable of lexical engagement – the ways in which a lexical entry dynamically interacts with other lexical entries, and with sublexical representations. For example, lexical entries compete with each other during word recognition (inhibition within the lexical level), and they also support the activation of their constituents (top-down lexical-phonemic facilitation, and lexically-based perceptual learning). We systematically vary the learning conditions for new words, and use separate measures of lexical configuration and engagement. Several surprising dissociations in behavior demonstrate the importance of the theoretical distinction between configuration and engagement.

Keywords: Lexical configuration, lexical engagement, learning new words

According to Herodotus, the experimental study of lexical development can be traced to an Egyptian Pharaoh, Psammetichus. In an effort to discover which language was innate, the Pharaoh arranged for two newborns to be raised without any spoken language input. When their first spoken word was reported to be “bekos”, the Pharaoh concluded that Phrygian was in fact the pre-packaged language of humans, because “bekos” referred to a kind of bread in Phrygian.

Although one must admire Psammetichus for his willingness to report an experimental result that ran counter to his hypothesis (he of course had predicted that Egyptian was the natural language of humans), a more plausible view is that each human acquires the words of whatever language(s) the environment presents. There is now a substantial literature that tracks the first words learned by very young children, and the development of increasingly long utterances. In the current study, we are concerned with much later lexical development: The addition of new words to the lexicon of an adolescent or adult language user. According to Nation and Waring (1997), this is a surprisingly frequent event. They estimate that people add about a thousand word forms (a word and its close morphological relatives) per year, up to an asymptote of about 20,000 word forms. This translates to about three word forms per day, every day of the year, which entails more lexical change than one might expect.

Adding a word to the mental lexicon is presumably at least partially incremental, given all that is potentially involved in a fully specified lexical entry. At a minimum, such an entry should include a word's input form (phonological and orthographic), its meaning (which will typically be multidimensional and often context-specific), and its syntactic role(s). In most cases, the language user will get some of this information initially, and more of it over time. For our current purposes, we will refer to all of this information as the “Configuration” of a lexical entry. The lexical configuration is the set of factual information that one knows about a word. As we have noted, this information develops over a rather long time course, perhaps measured in weeks, months, or even years. From this perspective, at any particular moment, lexical configuration can be thought of as relatively static.

We can distinguish lexical configuration from what we will call lexical “Engagement”. By lexical engagement, we mean how the lexical representation behaves dynamically, over a much shorter time scale. For example, most current theories of word recognition talk about the activation of lexical entries by input signals, and the effect that such activation can have on other lexical entries, and on other types of representations. Perhaps the most established example of one type of lexical engagement is semantic priming: Activating the representation for “doctor” leads to an increase in the activation of another lexical representation, the one for “nurse” (Meyer & Schvaneveldt, 1971). There are also hypothesized inhibitory or competitive functions of lexical representations (e.g., if “cat” is activated, it can compete with and possibly inhibit “cap”) in most theories, beginning with Cohort theory (e.g., Marslen-Wilson, 1990), and including models such as Trace (McClelland & Elman, 1986), Shortlist (Norris, 1994), and NAM (Luce & Pisoni, 1998). Interactive models such as Trace also posit lexical influences on sublexical representations, with activation of a lexical entry (e.g., “cat”) producing increased activation of its component phonemes (e.g., /k/).

We believe that in order to understand the lexicon and its development, it is important to keep the distinction between lexical configuration and lexical engagement clear. In examining the existing literature on how people add words to the lexicon, we found that it was difficult to discern a coherent story, because different studies seemed to tap into different lexical properties, using different experimental techniques. Thus, the goal of the current study is to take a more systematic approach, both theoretically and experimentally. We will first attempt to establish plausible measures of both lexical configuration and of lexical engagement, so that in the following experiments we can see how each of these aspects of lexical representations may develop. With these measures in hand, we will then examine three different learning procedures for new words. Finally, we then bring together language production and language perception, to see the effect of their interaction on new word learning. This approach allows us to provide a much more cohesive understanding of how new lexical representations get established than has been available previously.

Lexical Configuration

It is instructive to examine some of the prior studies of new word learning to see how each has given us bits of the big picture. For example, Salasoo, Shiffrin, and Feustal (1985) conducted an interesting and thorough examination of new word learning in the visual modality. Each training trial consisted of a word (e.g., “pelican”) or nonword (e.g., “shargat”) briefly presented on a computer monitor, immediately followed by a mask; each stimulus was presented thirty times during the course of training. The task was simply to report what was presented. In several experiments, participants either typed in their response or orally identified the word. In such a test of perceptual identification under challenging sensory conditions, Salasoo et al. predicted words should be named more quickly and more accurately than nonwords. Initially this was in fact the case, but after just six presentations, participants' responses to nonwords matched their speed and accuracy of the real words, suggesting that lexical representations were developed rather quickly. These representations were apparently quite well developed, because they produced comparable results in a test conducted a year later, without intervening training. From our perspective, this is an excellent example of lexical configuration development, because the orthographic information was learned and available for recognizing the words; there is no information provided by this study about the possible development of lexical engagement, as there was no measure of more dynamic effects of these representations on any other representations (i.e., we have no evidence that the activation of “shargat” had any effect on anything other than its own lexical representation).

There are a number of studies of spoken word acquisition that share many properties with the Salasoo et al. approach, with a number of differences as well. For example, Storkel (2001) examined how children (ages three to six years old) acquired new lexical entries as a function of repetition and semantic reference. Children heard new words in the context of a paragraph, along with a drawing depicting the story. The new words were repeated in the story from one to seven times. Storkel asked the children to identify the new word by selecting from three recorded nonwords, and by naming the word when a picture of it was presented. The participants' performance increased through such training, though not as much as in the Salasoo et al. study. Note that Storkel's study differed in two important ways. First, the new words were given a semantic referent (both through the story context, and the drawings). Second, the representations were probed with two different but related tasks (naming, and forced choice recognition). Nonetheless, this study is still only assessing lexical configuration, as both tasks call for retrieval of an individual lexical item's phonological codes.

Gupta's (2003; Gupta, Lipinski, & Abbs, 2004) examination of new word learning also falls into this category. Gupta's primary interest was in the relationship between learning new words and measures of memory span, since learning a new word can be viewed as learning an ordered series of sounds or letters. For the current purpose, the similarities to Storkel's (2001) study are most relevant: The (college-age) participants were taught names for novel animal pictures (Experiment 1) or drawings of “aliens” (Experiment 2), by pairing the names and the pictures. For the aliens, participants also learned various semantic properties (e.g., each alien's favorite food). Later testing showed that the participants could both produce the name, given a picture, and could provide the relevant semantic information. These are the sorts of factual properties that we have categorized as lexical configuration information.

Lexical Engagement

The Salasoo et al. (1985), Storkel (2001), and Gupta (2003) studies are all well designed investigations of the acquisition of new lexical representations. They are also typical of most studies in this literature, in that they assess lexical development by relatively direct measures of “factual” information, such as the ability to report a word given a brief orthographic input or a picture, or the ability to retrieve information associated with a newly learned word. These studies were not designed to assess the development of more dynamic lexical engagement.

There are two recent papers that have begun the investigation of lexical engagement. Magnuson, Tanenhaus, Aslin, and Dahan (2003) trained participants on sets of either 16 disyllabic new words (Experiments 1 and 2), or 20 monosyllabic ones (Experiment 3). Each word was paired with a meaningless visual pattern formed by randomly filling eight contiguous cells in a 4×4 grid. Participants were taught to associate each word with a particular pattern by presenting two (or later, four) of the patterns visually, and providing feedback after the participant selected the pattern that matched a given word. In the experiments of primary interest, the new words came in sets of four phonologically similar items (e.g., /pibo/, /pibu/, /dibo/, and /dibu/). In the later phases of the experiment, Magnuson et al. used an eyetracker to determine which visual patterns were being considered as a participant heard a word. The central theoretical question was whether the newly learned words would show a form of lexical competition typically found with normal words: Would listeners look at the “pibu” relatively often, when they were hearing “pibo”? Note that lexical competition is an example of lexical engagement, as it potentially involves the activation of one lexical item affecting the activation of another one (competitively). Magnuson et al. did in fact find that newly learned words showed looking patterns comparable to those for more developed lexical entries.

However, it is not clear that these results tell us about “normal” lexical development. A central goal for Magnuson et al. (2003) was the establishment of an encapsulated artificial lexicon, because such a small-scale system could be used to test various hypotheses without the immense messiness of a real lexicon. They demonstrated that their procedures had succeeded in this respect, as the new words did not seem to be affected by properties of the normal English lexicon (e.g. the presence or absence of many similar words in English). Presumably, the use of organized item sets, and the general experimental procedures, helped to seal off the newly learned words from the existing lexicon. Although this was methodologically attractive, it potentially limits the conclusions one can draw about lexical engagement, as the latter is by definition based on interactions among representations in the lexicon.

Gaskell and Dumay (2003) took an approach to studying new word learning that was essentially the opposite of Magnuson et al.'s (2003): Rather than try to design conditions that would create an encapsulated artificial lexicon, Gaskell and Dumay created conditions that maximally linked new words to existing ones. For example, for the real word “cathedral”, Gaskell and Dumay created the new word “cathedruke”. Over the course of five days, participants were repeatedly exposed to such nonwords in the context of a phoneme monitoring task. In this task, on each trial participants were given a target (e.g., “p”) and then heard the nonword, and judged whether the nonword contained the sound. This task was purely a vehicle for exposure to the items. Each day, the participants also completed a lexical decision task that included real words (e.g., “cathedral”) that were similar to the newly learned nonwords (e.g., “cathedruke”). If a functional lexical entry for “cathedruke” has been created, it should compete with the entry for “cathedral” in a lexical decision task, slowing responses to such similar words (compared to controls without new competitors). By the third day of training (i.e., after approximately 30 learning trials for each new word), Gaskell and Dumay found exactly this pattern. This is the first clear evidence for the emergence of lexical engagement. Although there were of course many differences between the stimuli and procedures of Gaskell and Dumay's study and those of Salasoo et al. (1985), it may be instructive to note the much longer learning period (30 exposures) needed to see evidence of lexical engagement than that needed for the emergence of lexical configuration (6 exposures).

The Current Study

The existing literature thus provides some information about the constraints and time course of learning the configuration of a new word (Gupta et al., 2003; Salasoo et al. 1985; Storkel, 2001), and some preliminary results about lexical engagement (Magnuson et al., 2003; Gaskell & Dumay, 2003). However, each study provides a rather isolated look at some aspect of lexical development, without a systematic comparison of configuration versus engagement, or of the effects of different word learning conditions. The current study is designed to provide a much more comprehensive and systematic analysis of these two factors. Across experiments, we examine different learning procedures. Within each experiment, we compare measures of configuration learning to measures of lexical engagement. We begin by adopting Gaskell and Dumay's training regime (exposure through phoneme monitoring). In Experiment 1, we assess lexical development with four tasks: Two were chosen to be natural measures of lexical configuration, and two were appropriate measures of lexical engagement. In addition to providing the first comparison of the development of lexical configuration versus lexical engagement, Experiment 1 also serves a methodological purpose. The results will be used to select one good measure for each aspect of lexical representations, which will allow us to then compare how different learning conditions affect the development of lexical configuration and lexical engagement (Experiments 2-5).

EXPERIMENT 1

Gaskell and Dumay's (2003) stimuli were constrained by their measure of lexical engagement – lexical competition. This measure required them to study new words that were close variations of existing words (e.g., cathedruke-cathedral). Magnuson et al.'s (2003) stimulus set was also strongly affected by their desire to study shared onsets (e.g., /pibo/-/pibu/) and rimes (e.g., /pibo/-/dibo/). In the current study, our measurements did not place such strong constraints on our stimulus selection. We therefore created new words that were not tied to particular existing words, and that were not systematically related to each other, which is probably more typical of normal lexical additions. Because previous research has shown that relatively long words generally produce stronger lexical effects in various tasks (e.g., Davis, Marslen-Wilson, & Gaskell, 2002; Pitt & Samuel, 2006), our stimulus set consisted of three- and four-syllable words. Table 1 shows the twelve critical items. As the table shows, six of the words included an /s/ (e.g., “gatersy”), and six included an /∫/ (e.g., “wickoshah”). This contrast was needed for one of the measures of lexical engagement (see below).

Table 1.

The twelve critical new words

Words with /s/	Words with /∫/
lifrisen	wickoshah
gatersy	bibershack
galasod	marfashick
figondalis	benemshalow
naronesay	penivasher
nomemsoly	komalsheum

Open in a new tab

Participants in Experiment 1 each learned all twelve new words. The learning took place in the context of making many phoneme monitoring judgments for each word over the course of five days, just as in Gaskell and Dumay's (2003) study. On each day, participants also did two other tasks, one of which was intended to measure the current state of lexical configuration, and the other of which was designed to assess lexical engagement. By using both types of measurement each day, we can see whether the two aspects of lexical representations develop together, or separately.

As we have noted, a methodological goal of Experiment 1 is to help us select a good measure of lexical configuration, and a good measure of lexical engagement. We chose two tasks that by their nature seem to primarily tap lexical configuration, and two which should depend on lexical engagement:

Lexical Configuration Measures

One task used to assess the state of lexical configuration was a three alternative recognition judgment (Storkel, 2001; see also Gaskell & Dumay, 2003, for a two-alternative version). For each of the critical words, we recorded two very similar items, and we tested how well a word was learned by a participant's ability to identify which word was the one that had been heard during training. For example, for the critical word “penivasher”, participants would hear “penivasher”, “penikasher”, and “penivawer”, and they would have to indicate which one of these three stimuli was the one that they had heard previously. This task clearly depends on the establishment of the kind of factual information (in this case, primarily phonological form) that defines lexical configuration.

The second measure of lexical configuration was a “word-in-noise” task, a type of perceptual identification test (Salasoo et al., 1985). The critical words (and many filler words and nonwords) were initially presented to participants buried under a high level of white noise. The stimulus was then played repeatedly, with each iteration including lower and lower amplitude noise. The participant was told to hit a button to stop the presentations when the speech could be identified (the participant was then required to write down what the word or nonword was). Pilot work shows that participants can identify the speech sooner (i.e., at higher noise levels) when a real word is presented than when a nonword is the target. To the extent that a word has developed a strong lexical configuration, it should be recognized at higher noise levels, much as Salasoo et al. showed for visually-presented stimuli.

Lexical Engagement Measures

The defining feature of lexical engagement is the ability of a lexical representation to affect the activation of another lexical representation, or the activation of a sublexical representation. The phonemic restoration effect seems to be a good example of such lexical functioning. Phonemic restoration occurs when part of an utterance is removed, and is replaced by an extraneous sound. In the initial report of this phenomenon, Warren (1970) recorded a sentence, and then spliced out a phoneme from one of the words, replacing that segment with the sound of a cough. When the result was played to listeners, they consistently reported that none of the speech was missing; they heard the cough as co-occurring with an intact utterance. In a series of experiments, Samuel (1981a, 1996) has shown that the strength of this effect depends on the amount of lexical support: When the replacement occurs within an utterance that should provide strong lexical activation, the restoration of missing phonemic information is strongest. Phonemic restoration thus appears to involve a lexical representation's activation of a sublexical phonemic representation. As such, it is a good example of lexical engagement. We therefore tested the ability of each of the trained words to support phonemic restoration. To the extent that the training enables lexical engagement, the words should be better able to support restoration of missing phonemes.

Norris, McQueen, and Cutler (2003) recently reported another perceptual effect that also seems to depend on lexical engagement – lexically-driven perceptual learning. Participants in the Norris et al. study first did a lexical decision task in which they categorized 100 words and 100 pseudowords. Among the 100 words were 20 critical items. For half of the participants, the critical items were words that end in /f/ (e.g., sheriff); for the other half, the critical items ended in /s/ (e.g., Paris). Norris et al. created a sound that was midway between /f/ and /s/ by mixing the two sounds together. This ambiguous sound replaced all of the final /f/ or final /s/ sounds in the critical items. After the lexical decision task, participants identified items from an /Ef/-/Es/ continuum, labeling each syllable as either /f/ or /s/. Norris et al. found that exposure to the ambiguous sound produced substantial changes in how listeners identified the test syllables. Participants who heard the sound in a lexical context that called for /s/ expanded their categorization of /s/; the reverse was true for those who had /f/ context. The results show that lexical items can cause substantial changes in the definition of sublexical categories, a clear case of lexical engagement (no such perceptual learning occurred when the ambiguous sound was placed in nonwords). Thus, we can use the ability of our trained words to produce perceptual learning as an index of the development of lexical engagement.

METHOD

Participants

Forty Stony Brook University students participated. Eight participants were replaced after Day 1, based on their failure to reach criterion performance on the phoneme monitoring test (85% correct). All were native speakers of American English and reported no hearing problems. They were either given credit toward a requirement in a psychology course or paid for their participation. No individual participated in more than one experiment in this study.

Apparatus

Participants were tested in a sound-shielded booth that accommodated three individual testing stations. Responses were entered on a 4-button response board, with appropriate labels on the buttons used in each task. The experiment was controlled by a Pentium PC, on which the responses were stored for later analysis.

Stimulus recording was done in a sound shielded booth, using a high quality microphone. The speech was digitized on a Pentium PC (12 bit analog-to-digital conversion, 16 kHz sample rate). Individual tokens were edited from the original digital recordings using Goldwave sound processing software. Stimuli were output through a digital to analog converter (12 bit; 16 kHz rate), low-pass filtered (7.8 kHz), amplified, and presented binaurally over high quality stereo headphones.

Design

Each participant was given a phoneme monitoring training task every day for the first four days. On all five days, each participant was given one test of configuration (either words-in-noise or three-alternative recognition) and one test of engagement (either phonemic restoration or perceptual learning). There were four groups of participants, incorporating the 2×2 combination of the two configuration tasks with the two engagement tasks. Within each group, order of the two tasks was counterbalanced across participants. On Day 1, the phoneme monitoring training preceded the two test tasks. On Days 2-4, the training came after the two tests. This provided a way to look for any change in performance between Days 1 and 2 due to consolidation without any further training (cf. Dumay & Gaskell, 2006; Gaskell & Dumay, 2003; Storkel, 2001). However, because there were no interesting consolidation effects found here, this manipulation will not be considered further. On Day 5, there was only testing, not training.

In the phoneme monitoring training task, the participants received six words (e.g., marfeshick) with /∫/ sounds, and six words (e.g. “figondalis”) with /s/ sounds, as shown in Table 1. In the testing phase, the twelve critical items were divided such that each participant heard six of the items in a configuration testing task, and the other six items in a task measuring lexical engagement. The tasks and the stimuli used in them are described below.

Phoneme Monitoring (training)

In the phoneme monitoring training task, listeners responded to the nonword items presented one at a time. Each day the twelve critical items were presented a total of 24 times each, with a new randomization on each pass. A target letter was displayed on a computer screen for 1500 msec before the playing of each nonword. The participant was given three seconds to respond. Participants were asked to press a button labeled “yes” if the target sound was present anywhere in the stimulus they heard, or to press a button labeled “no” if the target was not present. We instructed them to respond as quickly and as accurately as possible. Latencies were recorded at button push, measured from the onset of the stimulus. The order of the trials was randomized for each participant.

Target letters were chosen so that each nonword had twelve target present and twelve target absent trials. For each, three target present target letters were repeated four times, and three target absent target letters were repeated four times. Target letters that were target present trials for some nonwords were target absent trials for other nonwords. All targets appeared in three different nonwords in both target present and target absent cases.

Three-Alternative Recognition

For each critical nonword we recorded two foils that differed from the original nonword by one phoneme. Table 2 presents the stimuli used in this task. The three items (critical nonword plus two foils) were played in a random order. Buttons on a response board were labeled with “1”, “2”, and “3”. Participants were instructed to identify which of the three items they had heard in the training task by pressing the corresponding button.

Table 2.

The targets and foils used in the three-alternative forced choice task

Critical Item	Foil #1	Foil #2
marfashick	marfashil	marpashick
bibershack	bizershack	bibershav
gatersy	gabersy	gaterfee
galasod	gapasod	galasot
lifrisen	lifriken	lifrisib
wikoshah	wifoshah	wikogah
nomemsoly	nomeksoly	nomempoly
figondalis	figonbalis	figondalit
benemshalow	benezshelow	benemvelow
komalsheum	komarsheum	komalsheul
penivasher	penikasher	penivawer
naronesay	narobesay	naronetay

Open in a new tab

Threshold Discrimination: Words-in-Noise

Sixty test items were presented in a random order: six of the trained nonwords, thirty word fillers, and twenty four nonword fillers. Each stimulus was played with a mask of white noise completely covering the item. The item was then played repeatedly; with each repetition we reduced the noise level by ten percent. Participants were instructed to press a button at the earliest point that they could recognize the word or nonword. They then wrote down the item on an answer sheet. We recorded the level of noise at which the participant responded, and scored the written answers for accuracy.

A response was coded as correct if it was within one phoneme of the original stimulus. For example, if a participant responded with “fibershack” and the correct word was “bibershack”, this was coded as correct, whereas a response of “fiberdack” would be incorrect.

Phonemic Restoration

The stimuli included six of the trained nonwords, thirty six filler words and thirty six filler nonwords. The test followed the signal detection methodology used by Samuel (1981a). Critical phonemes (primarily fricatives) were selected in each word or nonword, chosen to be near the middle or end of the item. Half of the items were played with noise replacing the selected phoneme, and half of the items were played with noise merely added on top of it. On each trial, the participant first heard a clear version of the item and then a version containing (added or replacing) noise. The task was to discriminate between the added and the replaced versions by selecting “added” or “replaced” on the button board. Strong lexical support for a missing phoneme produces strong perceptual restoration, and in such cases the “replaced” stimulus should be perceived as intact, yielding poor discriminability.

Perceptual Learning

As in Norris et al.'s (2003) perceptual learning study, a categorization task followed an initial exposure task. Their exposure task was lexical decision, which we did not use here because we did not want to ask our participants to make any explicit judgments about the lexical status of the words they were learning. Instead, we used an old-new recognition task. Participants first heard a list of twenty items – ten words and ten nonwords, and were told that a second list would follow. Participants were to call any item on the second list “old” if it had been played in the first list; otherwise, the correct response was “new”.

The first list included six of the critical new words, but with one sound slightly mispronounced (see below). For half of the participants, the six critical items were new words that had been presented during training with an /s/; for the other half, the six critical items were based on new words that had had /∫/. Each list was balanced in the sense that if it had six critical /s/ words, there were also six (real, untrained) words that had an /∫/; if the critical items were based on /∫/, the balancing words had /s/.

We used the second list to expose participants to slightly mispronounced versions of six of the words that they were being taught. For half of the participants, the mispronunciations were of /s/; for the other half, the mispronunciations were of /∫/. For example, one group of participants heard the critical nonword “gatersy” during the training phase, and heard an ambiguous version of that nonword (“gater?y”) in the second list. All mispronunciations were generated by mixing an /s/ and an /∫/ taken from recordings of the critical item (e.g., “gatersy” and “gatershy”), and splicing the ambiguous mixture in place of the original /s/ or /∫/. As noted above, each list was balanced by including an equal number of (real) words with the opposite sound.

Participants were exposed to the six critical items (and the six balancing real word items) three times each, for a total of eighteen mispronunciations of either /s/ or /∫/. This total was designed to be similar to the twenty critical items used by Norris et al. (2003). The critical items were randomly mixed in with sixty filler items (thirty words and thirty nonwords) that did not contain either an /s/ or an /∫/. All filler items were presented twice over the course of the second list.

After completing the old/new task, participants categorized members of a seven-step /asa-a∫a/ continuum. Participants identified each sound as either an /s/ or an /∫/ by pressing one of two labeled keys on a button board. Ten different randomizations of the seven sounds were presented. The /s/-/∫/ continuum was created by first recording clear tokens of /asa/ and /a∫a/, and then creating various mixtures (a weighted average of the waveforms) of the /s/ and /∫/ sounds. The sounds were digitally mixed in 5% increments, e.g., 5% /s/ mixed with 95% /∫/, 10% /s/ mixed with 90% /∫/, etc. Based on pilot testing, five steps were selected: 85% /s/ mixed with 15% /∫/, 75% /s/+ 25% /∫/, 65% /s/ + 35% /∫/, 55% /s/ + 45 % /∫/, and 45% /s/ + 55% /∫/. The original /asa/ and the original /a∫a/ were included as the endpoints of this seven-step series.

RESULTS

Training Task

Phoneme Monitoring

Figure 1 shows average reaction times and accuracy rates across the four days of training. An ANOVA revealed that there was a significant difference in response times across the four days, F(3,117) = 135.90, p < 0.001. Accuracy also improved across the four days, F(3,117) = 20.07, p < 0.001. The training data indicate that participants were learning the nonwords well enough to know the sounds within them.

Lexical Configuration Measures

Recognition

The responses for the recognition task showed that participants were moderately accurate on Day 1 (56%) and reached 71% accuracy by Day 5 (see Figure 2). This trend was marginally significant, F(4,76)=2.25, p=.07. The recognition foils were very similar to the newly-learned words, and the developing representations are apparently only moderately detailed: They are sufficient to generate performance well above chance (33%) but not yet precise enough for extremely fine distinctions.

Threshold Discrimination

The other configuration test, threshold discrimination, produced more systematic results. Figure 3 shows average accuracy rates and noise thresholds (given in the percentage of the signal that was noise, rather than speech). Both the accuracy and noise threshold improved systematically across the five days. Accuracy on Day 1 averaged 69%, and by Day 5 performance reached 94%, F(4,76) = 16.02, p<.001. Listeners were also able to recognize the trained items in increasingly louder noise across the five days, F(4,76) = 20.01, p< 0.001. In conjunction with the recognition and phoneme monitoring results, these data show that lexical configuration can be established through the type of training used in Experiment 1.

Experiment 1: Threshold task, average noise level and average accuracy for recognizing items in noise

Lexical Engagement Measures

Phoneme Restoration

We conducted a one-way ANOVA, examining the differences in the participants' ability to distinguish between added and replaced versions of the critical nonwords. Participants were not able to successfully tell the difference between added and replaced stimuli for these critical nonwords, F < 1. The task was clearly not useful in measuring lexical engagement, due to the essentially chance level of performance. In retrospect, it is clear that using mostly fricatives as the critical phonemes made the task much too difficult (Samuel, 1981b).

Perceptual Learning

For each participant, we calculated the average choice of “sh” for each of the seven items on the /asa/-/a∫a/ test series. Figure 4 shows the group average each day, broken down by whether the Old-New exposure task should have increase “sh” report (solid lines) or “s” report (dashed lines). As Figure 4 shows, on each day there was a small trend towards a perceptual learning effect, with the solid curve consistently lying above the dashed one.

For statistical analysis, we focused on the middle three members of the test continuum (Samuel, 1986; Bertelson, Vroomen, & de Gelder, 2003), where any effects would be most evident. For each subject, we calculated the average percentage of “sh” choices for these three items, and then compared these averages in an ANOVA with Day of testing as one factor, and Exposure condition (favoring “sh” vs “s” perceptual learning) as the other. Despite the consistent 4-5% shift in the direction favored by perceptual learning, Exposure condition did not produce a reliable effect (F<1), and there was no interaction of the two factors (F<1). The consistency of the pattern suggests that there might be a small amount of perceptual learning going on, but clearly any such effect is too weak to imply any significant engagement by the representations for the words being learned.

DISCUSSION

During the training phase, phoneme monitoring accuracy was close to ceiling, and the participants grew faster across the four days of training. The steady increase in speed and accuracy suggests that the training regime accomplished the goal of giving listeners many attended-to exposures to the novel words.

The impact of such exposure could clearly be seen on the tasks that measured lexical configuration as a function of training. The results of the recognition task showed that participants were able to recognize the words at levels well above chance, improving from Day 1 (56%) to Day 5 (71%). The results of the threshold discrimination task were more systematic: There were large and significant performance gains across days of training. The difference in sensitivity between the two tasks may be related to the coding criterion used for accuracy in each task. For the recognition task, the foils generally only differed from the targets by only one phoneme. The participant had to get 100% of the sounds correct to make a correct response. In contrast, in the threshold discrimination task, responses were coded as accurate if they were within one phoneme of the original word. Taken together, the results of these two tasks showed that with phonological training, participants developed a good (though not perfect) configuration for each new word. Of course, the configuration was limited to form information, because no higher level (semantic, syntactic, pragmatic) information was available.

The development of lexical engagement was quite different than what we observed for lexical configuration. We will limit our discussion of lexical engagement to the ability of newly learned words to sustain perceptual learning, as the phonemic restoration data were uninformative because of our use of mostly fricatives as the critical phonemes.

If the training had produced fully functional lexical representations, then the new words should have supported perceptual learning: Hearing ambiguous versions of /s/ or /∫/ during the Old-New exposure task should have caused listeners to expand their category for the target phoneme. Although there was a 4-5% trend in the appropriate direction on each of the five days, the effect was very small and showed no hint of growing across days of training. Overall, the responses were not statistically different for the /∫/-trained group than the /s/-trained group.

Taken together, the results of the Threshold task and the Perceptual Learning task provide a striking contrast. The training regime produced very strong evidence for lexical configuration, but very weak evidence of lexical engagement. This dissociation underscores the importance of the theoretical distinction between knowing the “facts” for a word, and for that word's representation actually behaving as a true lexical representation.

This dissociation suggests that certain information about each new word was being represented, but that the representations had not achieved some criterion needed for behaving as “true” lexical items. This lack of full lexical engagement is especially interesting in light of the fact that we used the same training task as Gaskell and Dumay (2003) – phoneme monitoring – and they found evidence for lexical competition, which is one type of lexical engagement.

One important difference between Experiment 1 and the Gaskell and Dumay (2003) study is the nature of the new words to be learned. As we pointed out previously, the words to be learned in Gaskell and Dumay's study were necessarily variants of existing words (e.g., “cathedral-cathedruke”), whereas our new words were not closely linked to any members of the lexicon. One interpretation of the different results is that it may be easier to grow an offshoot of an existing lexical representation than to establish a fully functional one with no such links. That is, it is possible that “cathedruke” inherited some of its lexical functioning by virtue of its obvious ties to “cathedral”.

The other important difference between Experiment 1 and the Gaskell and Dumay study is the measure of lexical engagement. It is possible to devise models of lexical competition in which there is actually no real lexical engagement – the lexical representations do not directly affect each other's activation levels. For example, simply having two similar lexical items could cause a slowdown in any decision process, if that process needed to consider both entries, even if the entries themselves had no communication with each other. In contrast, for perceptual learning to occur, it is absolutely necessary for the lexical representation to inform the sublexical representation.

We noted that a methodological goal of Experiment 1 was to establish one good measure of lexical configuration, and one good measure of lexical engagement. Based on the patterns that we observed, we believe that the Threshold task provides a very orderly measure of lexical configuration; both accuracy and the level of noise at which a listener responded improved very systematically over training. The Perceptual Learning task seems promising as a measure of lexical engagement. It seems to be sensitive to lexical engagement (there were very consistent, though small, trends), and it seems to absolutely require the type of lexical function we wish to understand. Thus, in the following experiments, we will use these two tasks as our measures.

In the Introduction, we reported that several recent studies of the establishment of lexical representations have used a training technique in which participants learn to associate each new word with some visual stimulus (e.g., Gupta et al., 2003; Magnuson et al., 2003; Storkel, 2001). Although these studies only assessed lexical configuration, the training technique seems to be one that could potentially lead to both lexical configuration and engagement. In Experiment 2, we will use a variation of this training method, and examine the growth of lexical configuration and engagement.

EXPERIMENT 2

In Experiment 2 (and in the experiments that follow) we will use the same basic design that was used in Experiment 1: Participants will learn the same twelve new words over the course of five days, and we will track changes in the words' lexical configuration and engagement. In the current experiment, instead of using phoneme monitoring to present the new words, we will associate each new word with the picture of an unusual object (see Figure 5, below, for the pictures). Participants will be given the same number of training trials as before, and we will use the Threshold and Perceptual Learning tasks to assess changes in lexical configuration and lexical engagement, respectively.

METHOD

Participants

Twenty Stony Brook University graduate and undergraduate students participated in the five day experiment; four were replaced after Day 1 due to their failure to reach the criterion of 75% accuracy on the picture choice task (see below). All were native speakers of American English. They were either given credit toward a requirement in a psychology course or paid for their participation.

Design

Each participant received picture identification training every day for the first four days. Each participant was given one test of lexical configuration (threshold) and one test of lexical engagement (perceptual learning) every day for five days. The two tests were identical to the tests in Experiment 1.

Stimuli

The same twelve critical items used in Experiment 1 were used here.

Training Procedure

Each training session consisted of hearing the nonwords repeatedly in a picture identification task. We arbitrarily associated each nonword with the picture of an unusual object. Pictures were shown in color, and were collected by searching the Internet for hard to recognize objects. The twelve pictures are shown in Figure 5.

On each training trial, pictures of two objects were presented side by side on a computer screen. Each picture was presented as approximately the size of a normal playing card, making each very easy to see. At the same time that the pictures appeared, one of the critical words was presented over headphones. The participant was asked to identify which of the two pictures represented the word by pushing one of two buttons (left button = left picture, right button = right picture). As soon as the participant responded (correctly or incorrectly), the picture that was NOT associated with the word disappeared, leaving only the correct picture. This provided feedback to the participant about the correct association of words and pictures. Before participants had learned the associations, they simply guessed. In each training session (i.e., each day), each word was played 24 times with the picture that represented that word, each time paired with a randomly chosen picture from the remaining 11; the foil picture was randomly selected each time. Responses were coded for accuracy for the last 12 presentations on Day 1 (i.e., after the initial guessing period), and for all 24 repetitions on Days 2 through 4.

Testing Procedure

In the testing phase, each participant did the threshold task and the perceptual learning task. The twelve critical items were divided such that each participant heard six of the items in the threshold task, and the other six items in the perceptual learning task. The procedures were identical to those in Experiment 1, and the order of the two tasks was counterbalanced across participants.

RESULTS

Picture Identification

For the training task, responses were recorded for analyses of reaction times and accuracy. We conducted a one-way ANOVA on the reaction times, and found a significant effect for Day of training, F(3,57) = 46.45, p< 0.001. As Figure 6 shows, this effect reflects the much slower responses on Day 1 than on the other days. For accuracy, the second half of the responses on Day 1, and the complete set of responses on Days 2 through 4 were near ceiling, as can be seen in Figure 6. The 2% improvement over days (95% to 97%) did reach significance, F(3,57)=2.91, p<.05.

Experiment 2: Response times and accuracy for selecting pictures