Presentations

Abstract

Infant-directed speech (IDS) is typically slower, higher-pitched with greater pitch
modulation and larger vowel space than adult-directed speech (ADS) (Saint-Georges et
al., 2013). IDS may aid development of infant attention (Senju & Csibra, 2008), emotion
(Fernald, 1992) and language (Golinkoff et al., 2015), though IDS quantity (Cristia et al.,
2019) and acoustic features vary across languages and cultures (Moser et al., 2020).
One proposed source of cross-cultural variability is the time that caregivers have infant
body-contact (Falk, 2004). However, most studies involve small samples from WEIRD
populations, so cultural variability is poorly estimated. We focused on free play and
mother-infant interactions in Uganda and the UK to assess cross-cultural differences in
IDS quantity and acoustic features and test the body-contact hypothesis. In Study 1,
we calculated the proportion of free play mothers spent producing IDS and/or were in
body contact with their infant (3-9 months). In Study 2 we recorded mothers speaking
to their infant (3-6 months) and an adult experimenter, including naming objects to
elicit the corner vowels /i, u, a/. We extracted mean pitch, pitch modulation, speech
rate and vowel space measures. In contrast to the body-contact hypothesis, mothers
in Uganda and the UK produced comparable amounts of IDS, despite Ugandan mothers
spending significantly more time in body contact with their infant. Study 2 showed that
IDS was higher in mean pitch and pitch modulation than ADS in both Uganda and the
UK, but this difference was more pronounced in the UK. Speech rate for IDS was
significantly slower than ADS in Uganda, but not the UK. We found no evidence of
group level vowel-hyper articulation in either population. We discuss possible drivers
of this cultural variation in acoustic features of IDS and highlight the importance of
future work probing downstream effects of this variation on infant behaviour.

Abstract

High-frequency words have been found to benefit speech segmentation (Bortfeld, Morgan, Golinkoff, & Rathbun, 2005) and grammatical categorisation (Monaghan, Christiansen, & Chater, 2007) - possibly simultaneously (Frost, Monaghan & Christiansen, 2016). Previous studies have tested the effect of high-frequency words on language acquisition by presenting them reliably within the experimental language, however natural language contains noise and variability that may provide further opportunities for robust learning (Monaghan, 2017). We tested the effect of variability on learning by familiarising adults with continuous speech comprising repetitions of target words, which were preceded by one of two high-frequency marker words 100%, 67%, or 33% of the time, with marker words distinguishing targets into two categories. Participants completed speech segmentation and categorisation tests, followed by a cross-situational word learning task where the same target words present in the continuous speech were mapped onto referents from two different grammatical categories. There was a clear advantage of variability: the 67% group performed best on measures of segmentation, categorisation, and learning the meaning of words in the cross-situational learning task. The data indicate that variability can help learners to use the same high-frequency words to inform both speech segmentation and grammatical categorisation.

Abstract

Learners can extract transitional information from speech and use it to infer word boundaries and linguistic regularities. Critically, studies suggest that statistical language learning may benefit from the presence of high-frequency marker words (Bortfeld et al., 2005) that may act as anchors around which speech segmentation can occur, while also assisting with grammatical categorisation (Monaghan & Christiansen, 2010). To address these claims, we familiarised adults with a continuous stream of language comprising sequences of 8 bisyllabic target words, and compared learning to the same language but with high-frequency monosyllabic marker words preceding target words, and distinguishing them into two distributionally-defined categories. Participants completed a 2AFC testof segmentation, and a similarity judgement categorisation test containing word-pairs from the same versus different categories. We then tested transfer to a word-action/object learning task, where target word categories were either consistent or inconsistent with the action/object distinction. Participants segmented the speech stream better than chance, but only the marker word condition resulted in effective categorisation for the similarity judgement task. The advantage of marker words extended to the early stages of the word-learning task. Findings indicate that high-frequency marker words may assist grammatical categorisation even when they are not required for speech segmentation.

Abstract

The existence of sentences containing hierarchical dependencies is taken as evidence that language is not a finite state system. Whilst phrase structure is necessary in producing these sentences, their comprehension may draw on linear processing. Testing processing of language structures can be controlled in laboratory studies by constructing artificial language fragments and examining participants’ learning of the structures therein. However, hierarchical artificial language structures are difficult to acquire, meaning that measuring their processing is difficult to accomplish. Prior studies have employed nonsense word sequences to determine the learning of hierarchically generated grammatical structures. These may derive from prosodic information: rhythmic cues can indicate structural dependencies at phrasal boundaries, and pitch information can also highlight hierarchical structure, with lower pitch variance within phrases than between phrases. Further, information about similarity within dependencies (e.g., knowing dogs chase cats, and cats meow constrains the meaning of “the cat the dog chases meows”) can promote processing of hierarchically-generated structures. More importantly, the type of process required to comprehend the sentence may then not require phrase structure, but instead may be consistent with finite state operations. Detection of similarity and grouping of material using prosodic information may instead bypass the need for more complex constructions. In the present study, we tested whether participants were able to use prosodic and similarity cues to process the grammatical structure of hierarchical centre embedded structures in an artificial language. Participants listened to sequences of a hierarchical centre embedded language, of the AnBn form (e.g. A1A2B2B1, where A1 and B1 always occur as a dependent pair). 80 participants took part in one of five conditions manipulating the presence of several cues: a baseline condition with no additional cues to the structure; a similarity condition, where dependent pairs shared phonological information; a pitch condition where the pitch change within dependent pairs was smaller than that between pairs; a rhythmic condition, where short pauses intervened at phrasal boundaries; and a condition that combined these cues. In the baseline condition, we found no evidence of learning, consistent with prior literature. Individual cues – particularly the phonological similarity cue – promoted learning in the later stages of training. The combined cues condition, however, processing was facilitated by multiple sources of information in the initial stages of training, before the advantage of individual cues was observed: to the early learner, a wealth of information is beneficial, though with greater expertise, individual cues can be used equally effectively.

Abstract

The existence of sentences containing hierarchical dependencies is taken as evidence that language is not a finite state system. Whilst phrase structure is necessary in producing these sentences, their comprehension may draw on linear processing. Testing processing of language structures can be controlled in laboratory studies by constructing artificial language fragments and examining participants’ learning of the structures therein. However, hierarchical artificial language structures are difficult to acquire, meaning that measuring their processing is difficult to accomplish. Prior studies have employed nonsense word sequences to determine the learning of hierarchically generated grammatical structures. These may derive from prosodic information: rhythmic cues can indicate structural dependencies at phrasal boundaries, and pitch information can also highlight hierarchical structure, with lower pitch variance within phrases than between phrases. Further, information about similarity within dependencies (e.g., knowing dogs chase cats, and cats meow constrains the meaning of “the cat the dog chases meows”) can promote processing of hierarchically-generated structures. More importantly, the type of process required to comprehend the sentence may then not require phrase structure, but instead may be consistent with finite state operations. Detection of similarity and grouping of material using prosodic information may instead bypass the need for more complex constructions. In the present study, we tested whether participants were able to use prosodic and similarity cues to process the grammatical structure of hierarchical centre embedded structures in an artificial language. Participants listened to sequences of a hierarchical centre embedded language, of the AnBn form (e.g. A1A2B2B1, where A1 and B1 always occur as a dependent pair). 80 participants took part in one of five conditions manipulating the presence of several cues: a baseline condition with no additional cues to the structure; a similarity condition, where dependent pairs shared phonological information; a pitch condition where the pitch change within dependent pairs was smaller than that between pairs; a rhythmic condition, where short pauses intervened at phrasal boundaries; and a condition that combined these cues. In the baseline condition, we found no evidence of learning, consistent with prior literature. Individual cues – particularly the phonological similarity cue – promoted learning in the later stages of training. The combined cues condition, however, processing was facilitated by multiple sources of information in the initial stages of training, before the advantage of individual cues was observed: to the early learner, a wealth of information is beneficial, though with greater expertise, individual cues can be used equally effectively.

Abstract

Language is composed of complex grammatical structures that learners must make sense of in order to achieve linguistic proficiency. Key questions concern how learners come to realise that these structures are present in the language input, and how they grow to understand the purpose they serve. Much research has tested learners’ processing of language structures in laboratory studies by training participants on artificial languages, and examining their learning of the structures they contain. For example, prior work has shown that learners can extract transitional information from language input, and use it to identify word boundaries (e.g. Saffran et al., 1996) and grammatical regularities, such as non-adjacent dependencies (e.g. Frost & Monaghan, 2016). However, successful acquisition of statistically defined recursive structures has proven notoriously difficult to achieve, making measuring the way that learners process them even more difficult to accomplish. Previous studies have tended to use sequences of nonsense words to determine learning of recursive structures. Natural language, in comparison, contains multiple other sources of information to support the processing of dependencies within recursive sentences. Key structural information is provided by; prosody (learners have been found to use voice pitch directional changes to indicate phrase boundaries), rhythm (pauses occur in speech at phrase boundaries), and information about semantic relationships within dependencies (e.g., knowing that dogs chase cats, and that cats meow, helps us to determine the meaning of “the cat the dog chases meows”). In a series of experimental studies, we show that the addition of explicit phonological cues - reflecting similarity across dependent elements in recursive structures - facilitates learning. We found no evidence of prosodic pitch cues promoting learning in isolation. However, rhythmic cues did seem to support learning, and combining all three cues resulted in the best performance of all. Our results indicate that learners are able to use phonological and rhythmic information in isolation to support learning of recursive structures. In addition, we show that learners are able to combine phonological, rhythmic, and prosodic information to assist in acquisition of recursive structures.