Sensory Linguistics presents a timely examination of how perceptual experience maps onto linguistic patterns. The first half of the book builds the theoretical background critical to the study of sensory linguistics. Afterwards, Winter indulges us in a rich investigation of English sensory adjectives, combining data from human ratings and corpora. Winter not only presents elaborate analyses that reveal the relations between language and the senses, but simultaneously attends to the methodological steps, highlighting important considerations relating to reproducible and open science.

The theoretical introduction (Part I: Chapters 2–9) outlines three key themes: ineffability, embodiment, and metaphor and polysemy. Ineffability describes the extent to which a language has the linguistic means to verbalise perceptual experience (Levinson & Majid, Reference Levinson and Majid2014). A review of the literature reveals clear “differential ineffability” across the senses: vision is most codable (least ineffable), with a high number of type and token frequencies of vision words, and the proximal senses, particularly smell, are least codable (most ineffable). Ineffability is not only about codability across the senses, but extends broadly to whether or not subjective experiences are linguistically amenable.

Winter describes possible causes of ineffability (Chapter 4). Firstly, it may be difficult to encode perceptual experience in language because of cognitive–architectural constraints. For example, the cognitive representation of odour may not be sufficiently detailed to connect with language (Olofsson & Gottfried, Reference Olofsson and Gottfried2015). Secondly, language may be the limiting factor: the digital nature of language and restrictions on vocabulary size make it difficult to depict fine-grained perceptual experience. The vast potential of perceptual phenomenology mismatches the number of sensory-related words available. The final relevant factor is communicative need, which can have perceptual and cultural roots. Humans may have a communicative need to talk about visual aspects of the world because vision is the dominant sense. At the same time there may be little pressure to talk about smell because it is culturally appropriate to disregard smell (at least in Western societies).

In Chapter 5, Winter lays out the Embodied Lexicon Hypothesis, a hypothesis critical to the empirical chapters. This hypothesis builds on embodied approaches to language (Barsalou, Reference Barsalou1999, Reference Barsalou2008) which propose that language comprehension engages mental simulation in the perception and action systems. This means that representations of word meanings are the result of sensory and motor experience. Moving beyond experimental and neuropsychological work, Winter extends the study of embodiment to the composition of the lexicon and statistical patterns of word use. The central proposal of the Embodied Lexicon Hypothesis is that, since language and perception are tightly connected, perceptual patterns and the inter-relations between modalities should be seen in the structure and use of language: embodiment leaks into language. The Embodied Lexicon Hypothesis also sees emotional valence as a key component of lexical representation. The relevance of valence may differ across the senses, and could be more important for the lower senses, i.e., smell, taste, and touch (see also Speed & Majid, Reference Speed and Majid2020). This is another strength of the approach of Sensory Linguistics: differences across sensory modalities are brought to the fore.

A common use of sensory adjectives is in so-called “synaesthetic metaphors” – where sensory adjectives from one modality describe experiences in another sensory modality (e.g., sweet melody, sharp sound). In Chapters 6 and 7, Winter takes issue with the concept of synaesthetic metaphors and concludes that they are not, after all, synaesthetic. Canonical synaesthesia is a neurological phenomenon where individuals experience automatic and vivid associations between the senses, for example experiencing colour sensations when smelling an odour. Comparing cases of synaesthetic metaphor such as sweet fragrance to this unique perceptual experience is problematic for several reasons: canonical synaesthesia is rare whereas synaesthetic metaphors are prevalent, canonical synaesthetic associations are idiosyncratic and do not overlap with associations in synaesthetic metaphor, canonical synaesthetic associations occur automatically and synaesthetic metaphors are consciously accessed, and there is no evidence that links the occurrence of canonical synaesthesia with the use of synaesthetic metaphors.

Not only are synaesthetic metaphors not synaesthetic, according to Winter, they are also not metaphorical. Synaesthetic metaphors are not conceptual metaphors because source and target are sensory, whereas conceptual metaphors map from concrete to abstract concepts (e.g., Lakoff & Johnson, Reference Lakoff and Johnson1980). Neither are they primary metaphors (e.g., Lakoff & Johnson, Reference Lakoff and Johnson1999), where the target metaphor must be non-sensory. Metonymy is also not an accurate characterisation of synaesthetic metaphor because it assumes two distinct but correlated qualities within the same domain (e.g., Gibbs, Reference Gibbs1994). Adjectives used in synaesthetic metaphor are said instead to be supramodal and to have a broad denotational range. Rather than mappings across distinct senses, synaesthetic metaphors can be seen as literal descriptions of the same perceptual event. For example, because smell and taste are coupled in flavour, sweet smell literally describes an aspect of the same experience that sweet taste does. The evaluative content of sensory adjectives may drive other uses of synaesthetic metaphors (also described in Chapter 16).

The final chapters of the theoretical introduction (Chapters 7–9) are devoted to the hierarchy of the senses – how frequently each sense is used to talk about the other senses. First proposed by Ullman (Reference Ullman1945, Reference Ullmann1959), various versions of the hierarchy depict which sensory modalities are most likely to be mapped onto the others. In general, mappings from touch are considered most frequent (e.g., smooth taste) and mappings from sight least frequent (e.g., bright sound). Perception-based accounts explain the hierarchy in terms of differences across the senses, as does the Embodied Lexicon Hypothesis. For example, Shen and colleagues’ (e.g., Shen & Cohen, Reference Shen and Cohen1998) view is that the hierarchy reflects differences in accessibility across the senses. Touch, taste, and smell allow greater access to objects and are more subjective, whereas vision and sound are more distal. However, when considering wider evidence of asymmetries across the senses from psychophysical and neuropsychological studies, it is unclear if such explanations hold. The hierarchy could also be explained in terms of ineffability, as discussed in Chapter 4. If a sense is less codable in language, it may borrow terms from other senses to help convey perceptual detail. Initial evidence supports this idea (e.g., Ronga, Bazzanella, Rossi, & Iannetti, Reference Ronga, Bazzanella, Rossi and Iannetti2012; Strik Lievers, Reference Strik Lievers2015), and Winter contributes to this in Chapter 17.

The evaluative nature of cross-modal metaphors may also be important in the hierarchy. Certain modalities afford more affective descriptions than others. For example, in Chapter 16 smell and taste words are shown to be particularly emotion-laden. The gradability of sensory terms may also influence the hierarchy. Vision terms are not typically gradable, but touch and temperature terms are: an object cannot be more blue or round (vision), but it could differ in intensity of heat (temperature) or roughness (touch), for example. Terms that are more scalar are therefore useful when describing gradability of perceptual experience. A final proposal is that iconicity constrains the hierarchy of the senses. Words that are highly iconic, such as hissing and booming, may not be applicable to the other senses because their iconic auditory meanings are fixed. Words with more flexible meanings are instead more likely to be used cross-modally. In sum, a monolithic account of the hierarchy is unlikely, instead multiple causes are expected to be at play. Chapter 17 provides evidence for three factors: composition of the sensory vocabulary, emotional valence, and iconicity.

In Part II (Chapters 10–17), the focus turns to English sensory adjectives. Chapter 10 details the methodological approach for the empirical chapters. Following Lakoff and Johnson (Reference Lakoff and Johnson1999), sensory linguistics should connect closely with other domains such as cognitive psychology and neurophysiology. Theories of sensory linguistics should also be supported by converging evidence across methods and datasets, and they should be generalizable across a range of phenomena. Commendably, Winter makes a strong argument for reproducible science. Principles of reproducible science include making data and code accessible, and comprehensive reporting of statistical tests and results. In general, all components of the research should be accessible so that another researcher could exactly reproduce the research.

The empirical investigation focuses on sensory norms, described in Chapter 11. Pioneered by Lynott and Connell (Reference Lynott and Connell2009), sensory norms are ratings of how much a word is associated with the dominant sensory modalities (sight, touch, sound, taste, and smell). Based on the ratings, a word’s mean perceptual strength in each modality is calculated, a word is assigned a dominant modality, and a word’s modality exclusivity – how multimodal its referent is – is calculated. By distinguishing between modalities, the importance of multimodal representations in word meaning is demonstrated. The validity of the ratings is supported across studies, with sensory ratings significant predictors of behavioural responses to words (e.g., Connell & Lynott, Reference Connell and Lynott2014; Lynott & Connell, Reference Lynott and Connell2009). The norm-based approach is valuable to cognitive linguistics because it goes beyond the intuition of individual linguists, averaging over individual participants’ intuitions. Using existing norms also exemplifies the value of reproducible science.

In Chapter 12, Winter analyses the sensory dominance and multisensoriality of the adjectives of Lynott and Connell (Reference Lynott and Connell2009). Vision is the dominant modality of English adjectives: vision receives the highest ratings of perceptual strength, more adjectives are classified as vision-dominant, and all adjectives are somewhat visual. In contrast, ratings of sound, taste, and smell are low across adjectives, and few words are taste- or smell-dominant. In terms of distribution of the ratings, aside from vision all modalities appear to be two-peaked, with small clusters of words associated with high perceptual strength and a larger peak of words with low perceptual strength. Adjectives are less multisensory than a randomly permuted distribution of modality exclusivity and are therefore somewhat specialized in terms of sensory modality. This is in line with the proposal that adjectives are “compression devices” (p. 39), highlighting specific perceptual qualities whilst supressing others.

The validity of the five senses model of perception is tested in Chapter 13. There are numerous ways in which the senses can be categorized (O’Meara, Speed, San Roque, & Majid, Reference O’Meara, Speed, San Roque, Majid, Speed, O’Meara, Roque and Majid2019), and Winter tests which model best fits the sensory norms. Principal components analysis revealed larger subgroups of modalities than the standard five senses model: taste and smell, and touch and sight are highly correlated, with sound forming its own group. In contrast, cluster analyses divided the data into smaller subgroups: twelve distinct subgroups were identified based on similar profiles of modality strength. Some subgroups are more strongly associated with a single modality, such as those heuristically labelled “pure sight”, “smell”, and “taste”, whereas other clusters reflect abstract or multimodal properties, such as “gross surface properties” or “impression-related”. The number of words within each cluster also speaks to the ineffability of certain perceptual experiences. For example, the “pure sight” cluster includes 41 words whereas the “smell” cluster includes only 12, supporting the ineffability of smell in English. Importantly, the analyses show that the constraints of the five-category model can be overcome to reveal both a macro- and micro-perspective on sensory experience.

By integrating natural language data into the analyses, Chapter 14 assesses the “modality affinity” of sensory words. Modality affinity describes the semantic preference of sensory words: which senses tend to co-occur together. Analyses compared adjective–noun pairs containing adjectives from Lynott and Connell (Reference Lynott and Connell2009) and nouns from Lynott and Connell (Reference Lynott and Connell2013) taken from the Corpus of Contemporary American English (COCA; Davies, Reference Davies2008). Cosine similarity between adjective sensory modality vectors and noun sensory modality vectors is greater than the baseline similarity of all possible adjective–noun comparisons (0.81). This shows that sensory adjectives tend to modify nouns with similar modality profiles – sensory adjectives have a semantic preference for nouns with similar perceptual features. When looking at differences across the senses, there are bi-directional correlations between vision and touch. Touch adjectives tend to modulate visual nouns (e.g., rough blanket) and vice versa (e.g., shiny belt). This is similar to taste and smell, with smell adjectives frequently used with taste nouns (e.g., smoky sauce) and taste adjectives used for smell nouns (e.g., sweet cologne). Such mappings fit the Embodied Lexicon Hypothesis, reflecting correlations found in real-world perception. Touch and vision work together in manual actions and show multisensory integration in the brain. Similarly, smell and taste are integrated in flavour and are processed similarly in the brain. The correlations observed in the corpus support the larger subgroups of modalities that emerged from the principal components analysis in Chapter 13. A strength of this approach is that it shows that human intuitions (ratings) correspond to patterns in real-world data (corpus), confirming the findings with different methodological tools.

In Chapter 15 the sensory adjectives are related to other linguistic measures: word frequency, semantic complexity, and iconicity. As expected based on existing work with perception verbs (San Roque et al., Reference San Roque, Kendrick, Norcliffe, Brown, Defina, Dingemanse and Majid2015, Viberg, Reference Viberg, Hylenstam and Viberg1993) vision adjectives are most frequent, reflecting the visual dominance seen in perception. Touch adjectives are the most complex, with the highest average number of dictionary meanings. This reflects the frequent use of touch adjectives in metaphor. Sound is most iconic, which is to be expected considering the existence of onomatopoeia (e.g., meow, rustle). Winter assesses whether the linguistic measures are best predicted by the cluster model, the categorical model, or a model including continuous sensory ratings. Across the three measures all models reliably fit the data, but the model with continuous ratings best explains word frequency and semantic complexity, and the traditional five-category model best explains iconicity. Therefore, taking into account perceptual information across the senses is informative about broader patterns in language.

Chapter 16 focuses on the evaluative dimension of sensory adjectives – whether they have positively or negatively valenced connotations. Taste and smell are strongly associated with emotion: odour and taste are well connected with emotion systems in the brain (Rolls, Reference Rolls2004; Royet, Plailly, Delon-Martin, Kareken, & Segebarth, Reference Royet, Plailly, Delon-Martin, Kareken and Segebarth2003), odours are thought to be processed primarily in terms of valence (e.g., Khan et al., Reference Khan, Luk, Flinker, Aggarwal, Lapid, Haddad and Sobel2007; Yeshurun & Sobel, Reference Yeshurun and Sobel2010; Zarzo, Reference Zarzo2008), and odours can cue emotional autobiographical memories (e.g., Chu & Downes, Reference Chu and Downes2002; Willander & Larsson, Reference Willander and Larsson2006). According to the Embodied Lexicon Hypothesis, if odour and taste are more emotional, odour and taste language should be too. This was confirmed with ratings of absolute valence from Warriner, Kuperman, and Brysbaert (Reference Warriner, Kuperman and Brysbaert2013). In addition, the USENET corpus (Shaoul & Westbury, Reference Shaoul and Westbury2013) was used to assess the valence of the context in which sensory adjectives occur – the valence of the five words preceding and following each adjective. Smell and taste again emerged as having the highest absolute valence. Interestingly, the sensory models explained more variance in the context-based measure than the explicit ratings of valence. This suggests that the valence of sensory adjectives can be more directly inferred through context. When looking at valence direction a different pattern emerges. In explicit ratings, taste and sight are most positive, whereas smell is most negative. Taste adjectives also appear in the most positive contexts and smell adjectives appear in the most negative contexts. Overall however, sensory words can be best distinguished in terms of their absolute valence when considering the amount of variance explained in the models. In addition, the continuous model again outperforms the cluster and categorical models in explaining valence. Considering fine-grained differences in sensory associations appears to best explain differences in emotional association.

After the thorough analysis of the sensory norms, Chapter 17 re-evaluates the hierarchy of the senses. Analyses using the dominant sensory modality of the adjective–noun pairs in Chapter 14 support the hierarchy: after controlling for lexical differentiation, touch and taste are likely to be sources, smell is equally likely to be a source and a target, and sound and sight are more likely to be targets. However, on closer inspection, the patterns in the data do not support a monolithic account of sensory language. Firstly, there is a disproportionate representation of certain mappings counted as “hierarchy-consistent”. Of the possible eleven cross-modal mappings, three explain two-thirds of the hierarchy-consistent mappings: touch-to-sound, sight-to-sound, and taste-to-smell. Secondly, across studies, different mappings are used as evidence supporting the hierarchy. Finally, Winter proposes that dominant patterns of the hierarchy can be explained by the composition of the sensory lexicon rather than by more general principles of cross-modal language. That is, the differential ineffability of the senses in English is mirrored in the patterns of the hierarchy. Even when adjectives and nouns are randomly combined, the resulting mappings support the sensory hierarchy. In general, a single monolithic account of cross-modal language is unlikely, but rather local explanatory accounts are useful for explaining specific cross-modal mappings.

Two factors that can explain specific patterns of cross-modal language are emotional valence and iconicity. Ratings of absolute valence (Chapter 15) and ratings of iconicity (Chapter 16) significantly predict whether an adjective–noun pair is a cross-modal or within-modality pair. The more valenced an adjective–noun pair is, the more likely it is to be cross-modal. In contrast, the more iconic an adjective–noun pair is, the more likely it is to be within-modality. The same pattern is observed in the cosine similarity between adjectives and nouns (as used in Chapter 14). Contextual valence (and a trend for absolute valence) negatively predicts cosine similarity: the more valenced the adjective–noun pair, the lower the similarity in perceptual profile of the adjectives and nouns. On the other hand, iconicity positively predicts cosine similarity: the higher the iconicity of the adjective and noun, the greater the similarity in the perceptual profiles of the pair. These findings support the evaluative function of cross-modal metaphors and the constraints that iconicity imposes on cross-modal language use.

The theoretical background and empirical evidence is brought into synergy in Chapter 18 – the final chapter of Sensory Linguistics. Several conclusions about the English sensory lexicon should be taken away from this book. We have learned that although a five-senses model of the lexicon can help make research more tractable, language and cognition are overall better explained as a continuous phenomenon. Related to this, multisensoriality is prevalent in language, implying that cross-modal language is more literal than metaphorical. It is clear from the empirical chapters that the codability of perceptual experiences differs across the senses, with, for example, smell being most ineffable. Patterns that are observed in the English sensory lexicon closely match patterns seen in perception, supporting the Embodied Lexicon Hypothesis. Finally, although across decades the hierarchy of the senses has been supported, evidence presented here suggests that it cannot be explained by a single hypothesis, and instead multiple factors, including valence and iconicity, play a role in how sensory language combines together.

As well as the advancement of knowledge of the English lexicon, further evidence for the success of this book is the future research it will inspire. The work presented is restricted to English adjectives and nouns, which means there are many directions in which sensory linguistics can still be explored. The link between language and perception can be examined in other contexts: across languages, cultures, discourses, words classes, and time, for instance. If perception is mirrored in language, we should expect to find linguistic differences in people with different perceptual experiences. One can wonder what cross-modal language looks like in individuals who lack one of the senses (e.g., blind or deaf individuals) or those with an exceptionally developed sense, such as sensory experts. Language is clearly a powerful tool for shedding light on our inner perceptual experiences, and I expect in the near future the burgeoning field of sensory linguistics will teach us a lot more about this rich, subjective experience and how it can be communicated.