The ability to detect abstract grammatical rules, i.e., principles that govern speech sound streams, is essential for learning a language. To investigate the infants' ability to extract abstract algebraic rules, Marcus et al. (1999) familiarized infants to sequences of syllables (or sentences) that followed a particular “grammatical” rule (e.g., “ga ti ga” for ABA). During the test, infants were observed to be more attentive to sequences that were grammatically inconsistent (e.g., “wo fe fe,” which is ABB) than to those sequences that were consistent with grammatical rules (e.g., “wo fe wo”). Because the test sentences were different to those used in the training phase, the authors concluded that infants can extract an abstract rule and generalize it to novel instances. Also, detection of ABB and AAB structures were compared, and it was found that even if both structures have a reduplication element, the infants paid more attention to the inconsistent patterns.

It is not known, however, whether the ability to extract grammatical rules from speech sounds only applies to human linguistic cognition or whether this cognitive element has originally evolved for other, more general purposes. In the latter case, these skills could also be found in non-human animal species.

It is known that non-human animal species can process speech up to a certain level of cognitive complexity. Speech sound discrimination has been demonstrated in various animal species neurophysiologically (e.g., Dooling and Brown, 1990 in birds; Kraus et al., 1994 in guinea pigs, Ahmed et al., 2011 in rats), and on a behavioral level (e.g., Engineer et al., 2008 in rats; Sinnott et al., 1976 in monkeys; Sinnott and Mosteller, 2001 in gerbils). Also, word segmentation based on transitional probabilities has been demonstrated, on a behavioral level, in rats (Toro and Trobalon, 2005) as well as in cotton-top tamarins (Hauser et al., 2001). Extraction of grammatical rules (i.e., structural patterns) from speech sounds in non-human species has been studied in tamarin-monkeys and rats with similar stimulus conditions as applied originally by Marcus et al. (1999). The report concerning tamarin-monkeys (Hauser et al., 2002) was later retracted (Retraction notice, 2010). In rats, no evidence of pattern extraction was found (Toro and Trobalon, 2005).

It might be, however, too early to conclude that rats are not able to extract structural patterns from three-syllabic speech sequences, as were applied in a classic study by Marcus et al. (1999) in infants. Since there is evidence in rats of representing abstract rules from pure tones (Murphy et al., 2008), this issue should be further explored. To the present study we applied a neurophysiological mismatch response (MMR), a measure of automatic cognition, which is the equivalent of the human electrophysiological response called mismatch negativity (MMN; Näätänen et al., 1978, 1997, 2010). MMR can reflect auditory cognition before its behavioral manifestation (e.g., Tremblay et al., 1998). Based on this method, we have previously demonstrated that the rat's brain is able to detect changes in abstract auditory features, such as melodic patterns in tone-pairs (Ruusuvirta et al., 2007) and in combinatory rules between frequency and intensity of the sound objects (Astikainen et al., 2006, 2014). Rats also make representations of spectro-temporally complex sounds such as speech sounds in their brains, and they can detect changes in these sounds based on the content of the transient memory (Ahmed et al., 2011). Rats, anesthetized with urethane have been used in these studies as urethane is known to largely preserve the awake-like function of the brain (Maggi and Meli, 1986).

In the present study, capitalizing on the above mentioned studies, we recorded local-field potentials (LFPs) from the dura, above the auditory cortex in urethane-anesthetized rats. We presented the animals with a series of synthesized speech sounds. The stimulus series (modified from Marcus et al., 1999) consisted of several different sequences consisting of three pseudowords (called sentences here). Ninety percent of the sentences followed a specific pattern structure (“standards”). Acoustically novel sentences were introduced (“deviants”) rarely (10% of the sentences) and randomly in the sequences. Deviant sentences were of two different types: 1) “pattern-obeying deviants” that shared the pattern structure of the standard sentences but deviated from them physically, and 2) “pattern-violating deviants” that differed from the standards physically but also presented a different pattern structure. We expected to observe an early MMR to be triggered by the first pseudoword for both types of deviant sentences due to their acoustical differences from the standard pseudowords. We also expected to observe a later MMR to be triggered by the second word in the pattern-violating deviant sentences. This would indicate that the syntax-like rule, carried by the standard patterns, was extracted by the animals' brains.

The first MMR, i.e., an amplitude difference in local-field potentials, between the standard and the deviant sentences, was found for both the pattern-violating deviant sentences (Figure 1, left) and the pattern-obeying deviant sentences (Figure 1, right). This first MMR for the pattern-violating deviant sentences, was significant at 194–213 ms after the sentence onset, [t₍₁₃₎ = 2.2–2.7, p = 0.020–0.047], and at 231.5–251 ms after the sentence onset, [t₍₁₃₎ = 2.2–2.3, p = 0.039–0.050]. For the pattern-obeying deviant sentences the corresponding latency ranges were 187.5–206.5 ms after the sentence onset, [t₍₁₃₎ = 2.155–2.379, p = 0.033–0.050], and 228–261 ms after the sentence onset, [t₍₁₃₎ = 2.2–2.8, p = 0.016–0.048].

The second MMR was found only for the pattern-violating deviant sentences, in which the second word at a low probability (probability 0.05) violated the pattern that the rest of the sentences followed (probability 0.95). The latency for this MMR second was 217.5–316.5 ms from the onset of the second word, [t₍₁₃₎ = 2.2–3.6, p = 0.003–0.050] (Figure 1, left).

Next, an ANOVA comparing the responses to the pattern-violating and pattern-obeying deviants and their consecutive standards in the time window in which the second MMR was found (i.e., 217.5–316.5 ms from the onset of the second word) was conducted. Significant interaction effect of stimulus type × deviant type was found, [F_{(1, 13)} = 8.7, p = 0.011, η²_p = 0.401]. Main effects were non-significant. Responses to pattern-violating deviant sequences and those to the preceding standard sequences differed significantly, [t₍₁₃₎ = 3.5, p = 0.004, d = 1.02]. The corresponding difference was non-significant for the pattern-obeying deviants and preceding standards, [t₍₁₃₎ = 0.7, p = 0.525, d = 0.23]. Figure 2 depicts the mean amplitude values, standard deviation, and individual subjects' amplitude values for the differential responses.

Both types of deviant sentences, pattern-obeying and pattern-violating deviants, were detected from among the repeated standard sentences in the rat brain as indexed by the electrophysiological mismatch response. The earlier difference starting at 187.5 ms, after the sentence onset, was most probably elicited by the physical novelty of the deviant sounds; since the probability for the each standard variant was 22.5% and that of the deviant variants was 5%. An additional mismatch response, starting at 217.5 ms from the onset of the pattern change, was specifically found for the deviant sound sequences that were different in pattern structure from the frequently presented standard sequences. This finding suggests that anesthetized rats are able to extract structural patterns from speech stream that is carried out at a fast pace, and generalize this information to new items (since the deviant sentences differed physically from the standard sentences). Namely, in order to detect the pattern-violating deviant sequences, the brains of the animals needed to make a representation of the structure in the frequently presented “standard” sequences (Näätänen et al., 2001, 2010).

There is previous evidence of non-human animals' ability to extract grammatical rules from speech sounds. Common marmosets (New World monkeys) detected the grammatical differences based on simpler learning strategies than Rhesus monkeys (Old Wold monkeys) (Wilson et al., 2013). Similar ability for rule extraction has been previously reported from sinusoidal sounds in rats (Murphy et al., 2008) and from speech-specific calls in song birds (e.g., Gentner et al., 2006). In human infants, there is evidence that they learn more easily rule-like regularities from speech than from other auditory material (Marcus et al., 2007). It is not known whether this preference is related to linguistic potential in an infant's brain, familiarity of the speech sounds, or some other factors. Future studies in non-human animals could enlighten this issue.

In the present study we tested the rats' ability to detect pattern violation is speech sound sequences that all included a repetition of an element. Therefore, they were not possible to differentiate by detecting only this property of the stimulus. On the other hand, the generalization of the present results may be restricted to stimuli in which the pattern is defined as a repetition of an element and only the position of the repetition in the three-syllabic sequence is varied. Humans are particularly sensitive to rules that are expressed as a repetition of an element at the edges of a sequence (Endress et al., 2005). In our experiment, repetitions were always at the edge of the sequence. It is thus unclear as to what extent the present results in rats can be generalized to other types of rules. Furthermore, the types of rules applied to the previous studies on rule extraction have been under debate (Gentner et al., 2010; ten Cate et al., 2010). Thus, far studies in song birds have been progressive in solving this problem (e.g., van Heijningen et al., 2013), but there are still open questions (ten Cate and Okanoya, 2012). Electrophysiological methods which provide accurate information on the timing of neural activity (recorded in animals and humans) would be a feasible addition when studying different levels of cognitive complexity required in rule extraction. In humans, event-related potentials to study processing of non-adjacent dependencies, i.e., AXC structure in which the first and the last element are dependent (De Diego Balaguer et al., 2007; Mueller et al., 2009) and structural rules (ABB vs. ABA, Sun et al., 2012) in speech sounds have been utilized.

Previous behavioral research has failed to find evidence for rule extraction from speech sounds in rats (Toro and Trobalon, 2005). In this study, rats were presented with similar three-syllabic sequences of speech sounds, as in Marcus et al. (the third experiment, 1999). Our stimuli were nearly identical and the variability in the “standard” and “deviant” sequences was also the same (16 standard variants and 2 deviant variants of both deviant types). In the study by Toro and Trobalon (2005), rats indicated the detection of the pattern violation by pressing a lever. The present positive finding may be related to the methodology used. Namely, the mismatch response is known to be capable of probing into auditory cognition regardless of its behavioral manifestations (Tremblay et al., 1998). This method can bypass a wide range of factors related to behavior, for example, motivation, attention, or requirements of overt behavior. However, the constraints of such non-behavioral measures should also be acknowledged. Namely, it is unclear whether this ability can support behavioral adaptation in rats or not. Nevertheless, its existence in an animal species, which do not use complex sequences of calls in intra-species communication, (as compared to human speech or birdsong, e.g., Doupe and Kuhl, 1999; Gentner et al., 2006) supports the notion of its non-linguistic origin. Moreover, these findings endorse the view that even the most complex functions, quintessentially considered inherent to the human brain only, may in fact, also be represented in a primitive form (Näätänen et al., 2010) in brains thus far considered evolutionarily incapable of such procedures. Since extraction of rule-like patterns, in serially presented spectro-temporally complex sounds, is one of the mechanisms utilized by humans in receptive language learning the results might imply that some of the mechanisms supporting human language learning may not have evolved solely for human language during evolution.

In conclusion, the present results demonstrate the ability of the anesthetized rat brain to detect and represent the common abstract rule or pattern obeyed by a sequence of speech-like sound stimuli with a wide acoustic variation. Hence, these results appear to give a major contribution to the evidence suggesting the presence of the automatic sensory-cognitive core of cognitive function that is shared by humans and different other, at least higher species, at different developmental stages, and even in different states of consciousness, as proposed by Näätänen et al. (2001, 2010).

All authors contributed substantially to the conception and design of the work; Tanel Mällo, Timo Ruusuvirta, and Piia Astikainen contributed to the acquisition and analysis of the data, and all authors contributed to the interpretation of data for the work; Tanel Mällo and Piia Astikainen drafted the work, and Timo Ruusuvirta and Risto Näätänen contributed to revising it critically for its important intellectual content. Final approval of the version to be published was attained from all authors who also agreed to be accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.