Semantic and syntactic bootstrapping posit that children use their prior knowledge of one linguistic domain, say syntactic relations, to help later acquire another, such as the meanings of new words. Empirical results supporting both theories may tempt us to believe that these are different independent learning strategies. Here, we argue for a unified approach, where instead they are both contingent on a more general learning strategy for language acquisition: joint learning. Using a series of neural visually-grounded grammar induction models, we demonstrate that both syntactic and semantic bootstrapping effects are strongest when syntax and semantics are learnt simultaneously via joint learning. This more general learning strategy results in better grammar induction, realistic lexical category learning, and better interpretations of novel sentence and verb meanings. Joint learning makes language acquisition easier for learners by mutually constraining the hypotheses spaces for both syntax and semantics. Studying the dynamics of joint inference over many input sources and modalities represents an important new direction for language modeling and learning research in both cognitive sciences and AI, as it may help us explain how language can be acquired in more constrained learning settings.

• In aphasic repetition we documented rarely noted strictly increasing performance with position along with U-shaped and decreasing patterns. • Formal statistical modelling showed that patterns were dependent on prior context, not serial position. • Previous errors controlled decreasing functions in participants with phonological and apraxic impairments. • Previous items correct controlled increasing functions only found in participants with apraxia of speech. • A single mechanism for serial production in language and STM was not supported and chaining-like mechanisms may operate at some levels. Organising and producing a sequence of events is a basic human cognitive capacity. It occurs across a wide variety of domains including speech, writing, memory, planning and almost every type of skilled action. Errors involving sequences have been widely studied and often present two kinds of profiles: performance either declines across positions or it declines and then improves in the final positions (a U-shaped pattern). Studies of errors in aphasia have also reported these patterns with letters (in spelling) or phonemes (in speech). Another pattern, with more difficulty initiating speech, has been reported in apraxia of speech. Contrasting declines and increases in performance, however, have not been described in studies using the same methodology and evidence of performance linearly improving is very limited. We document all three patterns using statistical models in a case series of 23 people with aphasia (PwA) who make speech errors when repeating single words. We found that the declining pattern and the U-shape patterns occurred across patients, independent of whether their main impairment was a phonological impairment or apraxia of speech. Only people with apraxia of speech, however, showed the inverse pattern of linearly improving performance. Upward and downward patterns were not the consequence of a general factor like severity. Importantly, further exploration with statistical models revealed that phoneme position in the word was not, in fact, the dominant factor determining the visual patterns. Instead, performance was determined by either the number of previous errors (for declining performance) or the number of previous phonemes correct (for improving performance). Errors were almost never governed by serial position or word length per se . Our results support an important role for evolving context in the serial production mechanisms supporting single word production and we discuss implications for current models of speech production and, more generally, for models of serial performance. We suggest that temporary retention of novel sequences may rely more on an explicit representation of position, while stored articulatory representations may benefit from a contextual format (of the chaining type) where activation of previous units helps to support retrieval of units further along in the sequence.

• Speakers lengthen words that share initial phonological segments. • Segments may be planned by abstract representations or acoustic features. • Participants heard speakers of two dialects, which manipulated acoustic input. • Participants still lengthened words regardless of prime dialect. • Results suggest abstract phonological representations affect speech planning. Speakers often lengthen the duration of a word when it shares initial phonological segments with a previously uttered word (e.g., candy and candle ). One explanation for this is that words with initial similarity affect phonological encoding during sequence planning, yet it is unclear whether this similarity is phonetic or phonological. We manipulated phonetic differences by using a dialect variant: the pin-pen merger in American English. Participants completed an event description task in three experiments. We manipulated whether the participant’s target vowel ([ɪ] or [ɛ]) either phonetically matched or mismatched the vowel of the prime speaker, depending on the participant’s dialect. In the second experiment, we introduced a control vowel in the prime word ([æ] vs. [ɛ]). Participants in both dialect groups lengthened target words when they shared an initial phoneme, even when the vowel of the overlapping prime word was not shared across dialects. In the third experiment, we replicated this finding in a larger cohort of non-merger participants. All three experiments showed word lengthening despite the phonetic realization of phonemes, suggesting this effect is driven by phonological representations rather than surface-level pronunciations.