Artificial Grammar Learning by 1-year-olds

Rebecca L. Gomez and LouAnn Gerken (1999) conducted four experiments using head-turn preference procedures with infants to study their ability to remember and extract information from auditory strings of an artificial grammar. Their findings in these experiments showed that infants can discriminate new grammatical and ungrammatical strings after less than 2 minutes of exposure to the stimuli. These experiments showed that the infants were acquiring new and specific information about the grammar. Previous research has shown that by 10 months of age infants are able to acquire specific information about their native grammar. However, the researcher felt that more research into what types of information infants acquire at different ages of development was necessary due to the limitations in understanding the actual mechanisms involved in language acquisition.

The researchers highlight that past studies did not account for these limitations or allow for controlling of prior learning, specific structural factors or the learning environment. To control for these aspects, Gomez and Gerken used artificial language stimuli. Artificial grammars had been used in similar studies with adults, but are more often being used in research with infants as well to discover the mechanisms used to segment, process, and understand speech. Past studies using artificial grammars to test infants suggest that infants can discriminate transitional probabilities within words from ones between words and that infants show a listening preference for familiar strings of sounds from the grammar, even once they are reordered.

Although previous studies have answered many questions about the acquisition of syntax by infants, Gomez and Gerken felt there were more questions to be answered. Namely, they are interested in whether infants will generalize grammatical strings they have been trained on to new grammatical strings, if they learn sequential dependencies to a grammar that allows variable word orders, if infants can detect relations among units at low probabilities and if infants would show the ability to abstract beyond utterances to which they were previously exposed. To answer these questions, the researchers used strings created by a finite-state grammar. These grammars are complex, but limited in their output ability. The use of a finite-state grammar allowed them to train the infants on a subset of the outputs, and then test for generalization to new strings generated by the grammar.

The research was conducted with participants 11- and 12-months-old since they would have been at a more mature linguistic ability after the developmental language changes occurring at 10 months. These 11- and 12-month-old infants are also more likely to pick up on grammatical morphemes at this age, which are thought to cue categories in language (Shady, 1996; Shafer et a., 1998). These infants participated in two experimental phases during the research; acquisition and test. During acquisition the head-turn preference procedure was used to measure them as they were exposed to a subset of the artificial grammatical strings. During test, they were presented with new grammatical and ungrammatical strings.

The first experiment was designed to test whether the infants could discriminate new grammatical strings from ungrammatical strings. The second experiment focused on the infants’ abilities to discriminate new grammatical strings from ungrammatical strings based on pairwise violations of word order. (Experiments 3 and 4 looked at the generalizability of the learning accomplished in Experiment 2 and assessing if the infants could abstract beyond word order. Please refer to the study for more information on these experiments.) The researchers expected to find that infants would listen longer to (prefer) grammatical strings than to ungrammatical strings (Gomez and Gerken, 1999).

Experiment 1

Participants:
Infants were recruited for the research by surveying birth announcements in local newspapers of each location (New Mexico State University and University of Arizona). 16 American infants with an average age of 361 days were tested.

Stimuli:
Grammatical strings were created from the finite-grammar (G1) with a limit of two loop representations. The strings were 3 to 6 words in length; a total of 23 strings were created. For acquisition, ten strings from G1 were used and chosen to represent all possible paths through the grammar. Three random orders were generated for the two sets of five strings created. Ten new grammatical and ten ungrammatical strings generated by G1 and not in the acquisition set were used for testing. Ungrammatical strings were created by reversing the order of words of the grammatical strings. This allowed them to match the strings in terms of word frequency and length to avoid confounding variables in the data.

The twenty new strings were grouped into eight test samples, four grammatical and four ungrammatical. The samples were read aloud by a female college professor who was blind to the experiment’s conditions and who was given random strings to say. The words were to be spoken in an animated way to give each word the same stress, except the last word which had a higher pitch and more stress than the other words. These strings were recorded onto tape and the strings were recorded back-to-back to create longer strings. The resulting samples were similar in average number of words per second and average frequency.

Apparatus:
The apparatus in Location 1, University of New Mexico, was manual. The apparatus in Location 2, the University of Arizona, was automated. The mother was in the sound booth with the infant at Location 2.

Location 1:
The experiment was conducted in a 4×6 ft, three sided plywood test booth. A hole was cut in the center board above a green light mounted at the infant’s eye level. A video camera was aligned behind the hole; loud speakers with a red light at the infant’s eye level were mounted on each of the side walls. Samples were played from a cassette player and the infant’s behavior was monitored through a cellophane covered hole in the center panel. The observer recorded data on the duration of head-turns with a response box and the duration of looking times on a coding sheet. The second observer monitored the video, controlled the lights and used a puppet to refocus the infant’s attention on the center after each trial. Both observers wore head phones and were blind to the samples being played.

Location 2:
The tape from Location 1 was recorded onto a computer which controlled the presentation of the samples and recorded the infant’s responses. The booth for this location was 9×9 ft and soundproof. Sessions were videotaped and the computer recorded heard-turn and looking times.

Procedure:
The head-turn preference procedure was used. The infant’s caregiver was seated in the middle of the room with the infant held on their lap. Infants were exposed to the acquisition samples twice and ended after 12 trials or when the infant no longer showed interest in the task. They were then given a five minute break during which the caregiver and experimenter played with the infant. They next participated in the eight-trial test phase. The trials began by blinking the center light until the infant focused on the light, the center light was turned off and one of the side lights was turned on to flashing while the sample played. The caregivers wore headphones with background noise and were unaware of the sample being played. Reliability coding was performed on the videotapes to ensure there was no observer bias by an observer who did not participate in the experiment.

Results:
There was no interaction found between the looking times for grammatical versus ungrammatical trial sin the first or second blocks to indicate the infant was learning. This experiment determined that the infants distinguished new strings from those that were ungrammatical. There were significantly longer looking times towards grammatical strings than ungrammatical strings even after a 5 minute delay. This seems to be linked due to training rather than learning during the experiment.

Experiment 2

Participants:
The participants from Experiment 1 were used. The average age of the infants in Experiment 2 was 360 days.

Stimuli:
The same acquisition strings from Experiment 1 were used and grouped in the same way, generating three random orders for each set of 5 strings. The test strings were 10 grammatical and 10 ungrammatical strings, also from Experiment 1.

Apparatus:
The infants were tested with the same apparatus in each location as in Experiment 1.

Procedure:
The acquisition and testing procedures were the same as in Experiment 1. The same reliability tests were performed on the videotapes form Experiment 2 as in Experiment 1.

Results:
The results form Experiment 2 showed the infants had longer average looking times for grammatical over ungrammatical strings. The infants’ preferences for grammatical strings is not though to be due to prosodic differences in the strings because of the reliability checks of having the roughly the same word frequency and average in grammatical and ungrammatical strings. As a further check on this assumption, the experiment was conducted with adults with low-pass filtered versions of the strings.

The low-pass filter eliminates the phonetic and phonotactic properties of the strings while maintaining the prosody. The adults’ results showed that prosodic cues were not a factor in discriminating grammatical and ungrammatical strings. The results from this experiment reinforced the results from Experiment 1 in showing that infants were showing longer looking times for grammatical versus ungrammatical strings. The retention of the learning after a 5 minute delay was not limited to familiar strings as discrimination was shown for new strings as well.

Experiment Results Discussion:
This research attempted to prove that infants could discriminate between grammatical and ungrammatical strings of an artificial grammar. In each of the experiments, the infants were able to do so after less than 2 minutes exposure to the new grammar and retained what they had learned after a 5 minute delay. Gomez and Gerken conducted these experiments to expand on previous research on language acquisition and to focus on the acquisition of syntax in particular. To do this, they tested the infants’ abilities to generalize rather than memorize the grammars to which they were exposed.

Infants exhibited flexibility on their learning by generalizing the grammars and distinguishing between grammatical and ungrammatical strings based on word pairs and word order. This research was unique because it looked at whether infants could learn these rules in ranges more like natural language versus unnatural fixed-order units like previous research. They proved their hypothesis by showing that infants could generalize to new strings form the grammar as well as discriminate between new grammatical strings and ungrammatical strings with pairwise violations.

This research supports two explanations for these phenomena: on that states learners abstract frames of a grammar and then use them to identify grammatical strings and another that states learners may use a complex association mechanism to abstract sequential elements of the grammar. More research with infants will need to be conducted to further examine these theories. A limitation on this research as stated by Gomez and Gerken is the use of an unnatural artificial grammar which may not properly reflect how learners acquire speech and language. They would like to see additional research conducted to compare the developmental trends of language to developmental cognitive trends to develop a better understanding of these acquisition processes.

Research Discussion:
This research with artificial grammars has significance in other areas of research on language, and specifically syntax, acquisition amongst infants. The use of finite-state grammars like the one used in this study can help establish a simple novel grammar which other researchers can use to study similar aspects of language acquisition. This same novel grammar** is being used in a word order study by Dr. Plante, Associate Professor Vance, Shari Moore and others in the Speech & Hearing Sciences Department at the University of Arizona. This novel grammar allows these researchers to present 4-6 -year-olds with simple strings to test their ability to distinguish between grammatical and ungrammatical strings at their age. To conduct this experiment, these researchers created a simple computer game that children can play.

Simple grammatical strings from the grammar are presented in the context of the game in 10 test trials before the child’s responses are collected during 40 trials. The object of the game is to identify the times when the visiting alien, Zork, speaks in the same way as the home aliens, the Snarkys. Using response time software, the game is run on a laptop. Responses are collected by having the child press the happy (grammatical) or sad (ungrammatical) face on the integrated mouse pad after Zork tries to talk to the Snarkys. The children then play a second game with a fish theme and a different grammar to ensure that the results are not valid simply for one grammar.

This research is currently in beta-testing, but preliminary results are already showing that these children can determine, above chance, the grammatical from the ungrammatical strings. It is interesting to note that this research is also showing trends relating to speech and language disorders in the children and their ability to complete the task. This research is helpful in identifying ways to determine causes for speech and language acquisition problems in children as well as to study how children acquire syntax rules for their language. It is the hope that through understanding what children can and cannot complete these tasks, and the correlation of children with varying speech and language disorders and their ability to complete the tasks a better understanding of how their disorder arises and can be treated will be found.