2.1. Manual selection of appropriate seed words

The first step in producing a list of adjectives reflecting a product's brand identity is to establish the initial field for exploration. An FMCG product can be described in terms of the triad of functional qualities of the product itself, its pack, and the brand equity. For instance, in terms of a product's functional qualities, people buying a washing powder consider characteristics such as cleaning clothes, perfume, retaining shape and color, and so forth. In terms of the pack, washing powder comes in boxes (usually cardboard), which have different important qualities, such as its size and shape and, how easy it is to grasp, and so forth. A branded washing powder can be described through the system of values that brand owners attribute to potential buyers, such as caring for the family, or in terms of the product's “personality,” such as engaging, genuine.

Original lists of functional qualities and packaging words can come from general knowledge about the domain or from market research, which identifies a list of benefits or characteristics that consumers treat as being important qualities of a product when buying and using it. The source for building the list of brand-specific concepts for a branded product is a strategic branding statement that outlines the company's vision of how the product can be made to appeal to consumers, and how they perceive tangible and intangible characteristics of the product.

The resulting set of principles that are used to evaluate products therefore derive from a triad of characteristics, deriving from the pack, the overall function of the product, and the desired brand position. These characteristics may be captured in a single word, a phrase, or even a paragraph (from “soft” to “ensures no static cling”). From these characteristics, “seed words” [a word that is used as a basis for the remainder of the process and forms the key input into the British National Corpus (BNC)] must be derived to begin the process of developing a set of adjectives for evaluation of the product.

Each member of the function–package–brand triad produces a list of several (usually 5–15) concepts, not all of which are immediately useful for a list of seed words, as some of them can be specialized terminology, descriptive phrases, or unusual words, which are not frequent in the general language. The next step involves exploration of each concept to find words touching several aspects in a domain identified by the concept. For instance, the functional property of “perfume” can give rise to such words as scent, odor, aromatic, and deodorized. The properties of the pack can give rise to handle, hold, steady, and so forth, whereas the brand values can produce care, engaging, love, and so forth.

The resulting list of seed words consists of about 50–70 words, and captures the company's perception of qualities that people consider as a basic requirement of the total product that they buy and the differentiator that makes one product better than the other for their purposes. However, the list of seed words is not ready for affective engineering experiments, as it suffers from problems both in precision and recall. From the viewpoint of recall, there can be many other expressions, some of which are more suitable. From the viewpoint of precision, the list of seed words can include concepts with inherent ambiguity, which may not immediately be clear to the participant.

2.2. Automated extension of seed words list

In this step we improve both recall and precision of the seed word list by comparing it against examples of real language use. The source for language data in the current study is the BNC, which is a 100 million word collection of texts from a wide range of sources, designed to represent a comprehensive picture of how British English is used (Aston & Burnard, 1998). Recall can be improved by extending the list with other words that are used in similar contexts, whereas precision can be improved by taking words that stand out in the original list as they do not share many contexts with others. The procedure for improving recall is completely automatic, whereas precision can be improved by manually pruning the list of candidates following a set of rules.

Automatic detection of similar words is based on the distributional similarity hypothesis (Harris, 1985), according to which two words have similar meaning, if they share a significant number of other words occurring in their context. There have been several proposals for computing lists of words with similar meanings. They can be classified into two groups. One approach involves detection of typical patterns that can relate words, such as AND/OR, IS-A, and so forth (Pantel & Ravichandran, 2004): if words frequently co-occur in similar patterns (e.g., frustration and anger), they have similar meanings. The second approach involves representing the context of each word by a context vector (also known as feature vector):

Features in this vector are weights, which can represent collocates (other content words co-occurring in the window of N words), syntactic features, such as subject, object, or modifier relationships, or the strength of association of word w with documents in which it occurs. In our case, the feature space was represented by a list of collocates ranked according to the log-likelihood score (Manning & Schütze, 1999), which in comparison to other measures (such as chi-square or mutual information) ensures that both common expressions (e.g., enjoy life or spend one's life) and less frequent terminological constructions (life jacket or life imprisonment) receive high weights. Then the procedure involves computing the distance between vectors representing respective words D(C(w₁), C(w₂)) and finding words most similar in the feature space. The distance can be measured in several metrics, such as cosine, Jackard, Dice, and so forth (see Manning & Schütze, 1999, chap. 8.5).

The advantage of the first (pattern-based) approach is that it can be efficiently applied to large-scale dynamically changing corpora, because it is based on a relatively small number of topical patterns. However, it requires either development of elaborate parsers to apply patterns to running text or an additional step for extraction and evaluation of suitable patterns. The second (context-vector) approach produces high-quality lists automatically, but it is computationally expensive, because most typically the dimensionality of the vector space is huge. For instance, in our case we wanted to collect statistics on the most typical environment for every sufficiently frequent content word, so we computed collocates for words where frequency is above 50 occurrences in 100 million words of the BNC. This leaves us with the feature space of about 35,000 items (examples of words at the bottom of our frequency list are imprudent, tarragon, and uprating).

Given that computing the distance between vectors of 35,000 dimensions is computationally expensive, the next step in the automatic procedure was to reduce the dimensionality of the feature space by means of singular value decomposition (SVD) of the resulting matrix of collocates following the procedure designed by (Rapp, 2004). SVD-based transforms were also used in latent semantic analysis (LSA) to compute the similarity between terms in documents (Landauer et al., 1998). However, in our case we do not deal with terminology, so our matrix is based not on the co-occurrence of words per document, but on the strength of collocations between words in their contexts.

SVD transform involves decomposition of the initial matrix of word co-occurrence M = UΣV^T, in which Σ is a diagonal matrix containing singular values of the original matrix M (Berry et al., 1999). The selection of k largest singular values of Σ gives a reduced dimensionality space, in which the original n × n matrix of correlation between content words can be approximated by n × k matrix, in which vectors for n words (35,000 in our case) contain only k features (300 in our experiment). This reduces the dimensionality substantially without loosing information about the relationship between words.

For each word in the original seed list we produce its simclass consisting of about 20 words, which SVD-reduced vectors have the closest distance to the vector of the source word according to the cosine metrics. For instance, love produces the following simclass:

Not all words in the list are suitable for affective engineering, but recall here is more important than precision, as the initial investigation could have missed words like beautiful or passionate.

At this stage we deal with the issue of precision by first using collocate lists to remove seed words, which frequently co-occur with words outside of the desired domain, and second, by checking the ambiguity of respective simclasses. For instance, even if care is suitable for the original list of brand keys, the set of its collocates reveals that its most typical contexts are intensive care, residential care, and home care services, which can cause problems with interpretation of these words by participants in affective experiments. This suggests that care is not suitable for an affective engineering test, as it is likely that it in the mind of some subjects it will be linked to health care, and hence, result in misinterpretations.

WordNet (Fellbaum, 1998) is a lexical database that also captures the relationship of synonymy between lexical items and is frequently used for research in computational linguistics; however, its coverage of affective words is not good enough. For instance, its output for care included words such as attention, aid, tending, caution, and guardianship, but did not highlight the subtlety of the health-care analogy, an aspect critical for affective engineering studies. This is because it uses a thorough but thesaurus-based approach rather than considering how a word can be used in natural language.

The ambiguity from the BNC output can produce two unrelated strands in their simclasses, such as the simclass of spirit (because of corpus processing, words are in lowercase):

At the same time simclasses of other words, which belong to the same semantic field as spirit, but are less ambiguous, such as resilient, can produce interesting simclasses:

In our experiments we also found that packaging words, for example, bottle, typically referred to the content of the pack and not to its inherent qualities. The simclass for bottle in the BNC is the following:

The most frequent collocates of bottle also refer to the content (most typically beverages): drink half a bottle of brandy, buy a bottle of Scotch, an empty vodka bottle, with occasional milk or hot-water bottles. In 100 million words of the BNC there are only seven examples of bottle combined with shampoo, none of which really evaluates qualities of the pack. This confirms the position of branding agencies that “the pack communicates the product and the brand,” so no evaluative words could be generated from the packaging description as such.

In further studies we concentrate on product and brand seed words exclusively. In both cases the aim is to generate an exhaustive range of language: the wider the range of language the better. The range of words that are used to describe the product creates its semantic space (see Fig. 2). However, as examples 1–4 show, the semantic space for a product contains various words, many of which are not suitable for an affective engineering test, for example, lover or never for love, whereas others are relevant, but are nouns or verbs, whereas in our tests we use adjectives. What is more, we also need gradable adjectives for which one can apply a judgment on a scale: a plastic box cannot be judged as more or less plastic. Thus, in the next step we convert the semantic space to produce gradable adjectives that can occur in constructions like to be M X, where M stands for a modifier, for example, more, very, not so. For instance, we can convert pleasure from example 1 into pleasant, hate into hateful, friend into friendly, but we cannot produce suitable adjectives for kiss or mother.

An example of semantic space.

3.1. Manual application of expert rules

As yet, no automatic process is available to reduce the burden of the task to remove unsuitable words from the generated list. This manual process uses the application of a set of linguistically and grammatically informed rules, examples of which are given in Table 1. This is a relatively easy process whereby each rule in turn is applied to the list of words and any adjective that violates the rule is removed.

Example of rules used to reduce adjective set

After these rules have been applied, the adjective list is considerably shorter, and many of the words that are not central, or that are likely to produce confusion or error, have been removed.

The next step reduces the list still further by reapplying the remaining words back to the set of keys resulting from the brand, product, and pack description. In this step, we find that some of the words satisfy more than one such characteristic, which makes them strong candidates for inclusion in the test. This step may, in some cases, be the final one, as the words that are selected may be only those that satisfy more than one quality or criterion. In practice, however, it is also necessary to ensure that all the characteristics have one or more words that relate to them included in the test. Thus, the selection process should include not only words that relate to more than one criterion, but also some words that may relate to only one, but that are its sole representative in the test.

3.2. Adjectival candidates for consumer survey

All adjectives that claim more than one “parent” quality are good candidates for use in the tests. However, further procedures are sometimes necessary to reduce a still overlarge word list down to the 10 or 20 adjectives that are manageable for experimental participants. This step is done in collaboration with the brand owner, who has ultimate oversight of the final list. For example, although in our case study (see Section 4) the adjectives “tender,” “luxurious,” and “conventional” occurred in the context of more than one brand quality, the brand owner did not favor them and felt that other candidates were better. Although this is a subjective step, it also relates to the fact that brand descriptions are not themselves perfect and do not always represent fully the character of the brand. In practice, the more specific and distinctive a brand description is, the more distinctive and representative the words that this process generates will be. A more general brand description is likely to generate less interesting and incisive words and less valuable results for the brand owner.

4.1. Identification and reduction of relevant adjective set

The first stage in the process was to identify a set of appropriate seed words for input into the system. The client, a linguistic expert, and an affective engineering expert defined a set of seed words under the headings of functional benefits, brand equity, and packaging attributes. In this situation, packaging attributes did not result in any suitable adjectives, so it will not be discussed any further within this case study. These seed words were manually extended using the artificial intelligence (AI)-supported linguistic process described in Section 2.2 to about 70 words (see Table 2). The list of extended seed words were in agreement with the client to ensure they accurately reflected the product and brand essence.

Extending seed words

Each seed word was then input into the BNC as described in Section 2.2 to automatically identify other adjectives that have similar lexical behavior in naturally occurring language. A list was produced of the most significant collocations for each sufficiently frequent word from 100 million words of the BNC. The SVD method was used to group words that occur in similar lexical contexts.

Each seed word resulted in 10–20 significantly related words from the SVD method, representing the saturation of the semantic exploration for the product and brand attributes. Table 3 shows part of the relevant adjective set placed into columns under the heading of their original seed word. The numbers indicate the cosine semantic distance between the vectors of respective words in the resulted SVD matrix; this indicates the similarity between the derivative and original word, but is not used to discriminate in this methodology.

Part of the relevant adjective set

The next stage was to reduce the size of the relevant adjective set into the range of adjectives suitable for the affective engineering consumer study. This is achieved using the rule set hierarchy and accompanying grammatical tests as described in Section 3. Table 4 shows a representative selection of adjectives that were removed along with the rule violation reasons.

Examples of removed adjectives

At this point there were still a few hundred adjectives representative of attributes of the product/brand/pack that met the basic criteria for evaluating objects. The adjectives that occurred across several seed words were arranged into a matrix, a sample of which is shown in Figure 3. This shows the relevant adjective set after the reduction process and relates these to the original seed words from which they were extracted. This facilitates the selection of appropriate adjectives for the affective engineering survey. Adjectives that have multiple roots test more than one concept, and can reduce the number of questions required. The client must be confident that all the relevant brand, product, and pack attributes are covered by the adjectives, and this matrix can be used to ensure this occurs.

A matrix of the relevant adjective set versus the original seed words.

The final adjective list for the consumer survey is highlighted in Figure 3 and included words such as tender, conventional, fun, luxurious, showy, everyday, slender, cosy, and bold.

Three additional adjectives were added to test whether the selection process was discriminatory. These were chosen because they represented concepts that the client was keen to include:

4.2. Consumer survey

To validate the process described, the selected adjectives were carried forward into a full consumer survey. Fifty female participants took part in the experiment between the ages of 35 and 50 who were all frequent users of premium products within the market sector and who all considered added value aspects important (e.g., brand promise, smell, and other detailing).

A set of 10 prototype samples was used for the study with all graphics, branding, and color removed to constrain the evaluations to pure shape. Figure 4 shows an exemplar sample used for the study.

An example sample stimulus used in the consumer survey.

Each participant was asked to complete a semantic differential experiment. Figure 5 shows the semantic differential questionnaire as used in this survey, illustrating how each adjective was presented with its corresponding negative. Each participant was asked to evaluate each sample, in turn, according to each adjectival pair and rate it on a 7-point scale. Each participant completed one questionnaire per sample and was allowed to do the experiment at his or her own pace.

An example survey questionnaire.

4.3. Results and discussion

The participants' scores were aggregated and cleaned. The raw evaluation data and preference scores were analyzed primarily with regard to the following points:

1. evaluation and preference score distribution,
2. principal component analysis of the subject's perceptual space, and
3. contribution of adjectival selection method.

4.3.1. Preference score distribution

Preference score distribution plots show the randomness of the participants' scores. If the AI-supported method described in this paper has not resulted in suitable adjectives, and there is little/no significant correlations between the evaluative adjective and the object then one of the following scenarios will be observed for the adjectives (see Fig. 6).

a. Flat distribution: Responses are equally distributed across the scale showing no overall consensus by the test group of participants.

b. Central peak distribution: A cluster of responses at the central point showing that most participants had no significant opinion of the interaction between the adjective pair and sample.

c. Double peak distribution¹

This may not necessarily be an issue, as it evidences two separate populations of opinion and thus another potential market opportunity.

: Two polarizing opinions of the group suggesting the evaluative adjective is suitable for differentiating the properties of the shape, but participants have opposing opinions of the correlation.

Examples of response distributions. [A color version of this figure can be viewed online at www.journals.cambridge.org]

These distributions are frequently found in such studies and require that the data be removed. However, that was not the case in this experiment, as all the responses for the identified good adjectives showed the following:

1. useful correlations for evaluating objects, and differing degrees of consensus were observed across different adjective and sample combinations, which evidences the ability of participants to use a set of evaluative adjectives to communicate differences in perceptual properties between the samples; and
2. a lowered level of ambiguity in comparison to previous studies.

4.3.2. Principal component analysis of the subject's perceptual space

Although the samples can be represented in a perceptual space where each evaluation adjective has an independent dimension, it is too complicated to determine the relative location of each sample in such dimensional semantic space. Moreover, it is not known whether there exists any interaction among the adjectival words.

Therefore, a principal component analysis (with a varimax rotation and using significance as eigenvalues > 1), can define the underlying similarities and relationships between the individual words and samples.

Two components were extracted, which means that there are two orthogonal sets of perceived similarities between the adjectival relationships to the samples: principal component 1 (PC1) accounts for 68% of total variance, and PC2 for 21% of total variance. The two adjectives that load highest onto PC1 are luxurious and showy; cosy and friendly load very highly onto PC2. Figure 7 shows diagrammatically the semantic space with all sample scores plotted. From this one can see that sample 2 loads highly on both PC1 and PC2, making it the ideal sample for the brand if it wants to be seen as both stylish and natural. If one wants a sample that is only perceived as natural then samples 3 or 9 are appropriate. However, if the target is to have a sample that is only stylish, then on should select sample 5. Sample 4 scores poorly against both factors, so is unlikely to represent the right design.

A semantic map showing sample ratings versus principal components. [A color version of this figure can be viewed online at www.journals.cambridge.org]