Data cleaning and item combining were manually applied to individual cases. Typos were identified and corrected; for example, “債卷” (typo) was changed to “債券”(corrected, “zaai3 hyun3,” “bond,” as a response to “a kind of investment tool”). The mixed usage of simplified Chinese characters was adjusted; for instance, “生气” (Chinese-Simplified) was changed to “生氣” (Chinese-Traditional, “sang1 hei3,” “angry,” in response to “a mood state”). Allographs were unified; for example, “雞” and “鷄” (allographs for “chicken,” “gai1”) were merged to yield “雞.” Variations of words used to describe a very similar or identical concept were merged into a common form and treated as identical, as in the case for “乳牛”(“jyu5 ngau4”, milk cow) and “奶牛” (“naai5 ngau4”, “milk cow”), which were deemed to describe the identical concept of “milk cow” as a response to “a kind of farm animal.” These variations are due to differences between formal and informal speech rather than to conceptual differences (there is no analogous example available in English).

Understanding whether the familiarity of a word impacts its categorical typicality is an essential step toward understanding how semantic knowledge is organized. This experiment assessed the familiarity of each concept in the general context of participants’ daily lives. Participants were instructed to rate how often they encountered a target word (an exemplar from Experiment 1) in all the life scenarios, instead of being under a specific category. Note that here the categorical information was not given to the word to be rated.

The database included 1298 items in 72 categories. The results from the two experiments were integrated and presented in tables, one for each category. A representative example is shown in Table 1. A word code (Word Code) was assigned to every exemplar under a specific category using the format HKC (the acronym for Hong Kong Category) followed by a 3-digit category code (e.g., “001” for “a kind of farm animal”) and a 2-digit exemplar code. The exemplar (Word) was numbered to indicate the descending rank of total probabilities within the category. “Slot 1,” “Slot 2,” and “Slot 3” indicate the probabilities of a given exemplar being allocated on the best/second-best/third-best slot. The probability on a slot is the number of mentions in the given slot divided by the total number of eligible entries for that slot, rounded to three decimal places. “Total” is the instance probability of an exemplar, regardless of the slot. “Accumulative” is the summed instance probability of the exemplar and that of all its precedents. Logically, this value increases with each exemplar in the category until it reaches 1.000 at the last exemplar. “Familiarity” is the average of all rating scores given by all participants who viewed that exemplar. Possible English Equivalents are listed as the possible corresponding concepts in the English, for the convenience of the researchers who are interested in further studies on cross-language comparisons.

At the end of each table, Invalid xx is a designed virtual exemplar indicating sum of the probabilities of all invalid responses, where “xx” is the category code. Invalid responses may be due to mistyping, misunderstanding of the category name, or lack of knowledge of the category (as mentioned in the data compiling and cleaning section). In some cases, when the participants were unable to think of a word, they repeated a response or rephrased it to an interchangeable lexical item. For example, both “石屎” (“sek6 si2”) and “英泥” (“jing1 nai4”) refer to “cement” in the category of “construction materials”; only that “石屎” (“sek6 si2”) is more colloquial. In other cases, participants generated non-referring items, such as “square” or “round” in the category of “natural geographical feature,” indicating unfamiliarity of geographical terminology; while participants in other studies were able to produce more relevant and referring terms such as “mountain” and “lake,” as the cases in other norm studies (Van Overschelde et al., 2004). Such non-referring items were considered as invalid responses. The probabilities of the invalid responses for the individual slots were not considered informative and thus were omitted from the table by designating the values as “N/A.” No familiarity scores were associated with the invalid responses, and hence this field was also marked as “N/A.” When there was no invalid response in a given category, the Total of the Invalid xx was 0.000, and when the Total of Invalid xx reached 0.500, the category would be excluded from the final table. Invalid responses may have occurred because the category and its related information were unfamiliar or unavailable to Hong Kong Cantonese speakers; therefore, categories with more than 50% invalid responses were discarded because they were not able to represent the consensus of semantic knowledge in the population. Twelve categories (e.g., a kind of natural geographical feature) were discarded from the final list (see the Appendix). A total of 72 categories were included in the following analyses.

Previous studies have suggested that the most frequently generated exemplar within a category is the most typical and hence the central member of that category (Barsalou, 1985). The more central an exemplar, the faster and more frequently it is recalled as a category exemplar, as the search process follows a fixed order (Rosch, 1973). Given this logic, the exemplars that are recalled most frequently (higher Total) should also be recalled as the first responses (best-fit) more frequently (in Slot1). To examine this hypothesis, the Pearson’s correlation coefficient of the two sets of probabilities (Slot1 and Total) of the exemplars was calculated for each category as shown in Table 2. Note that only the exemplars that had been mentioned in Slot 1 at least once (i.e., the value of Slot1 was >0.000) were included in the correlation analysis. Besides the number of included exemplars in the correlation analysis n, the total number of exemplars listed in a category N was also presented. Significant positive correlations were observed for 62 out of 72 categories and marked in the table, confirming that most frequently recalled exemplars are also likely to be mentioned first for the majority of the categories.

Here, familiarity is defined as the average score of participants’ subjective ratings of the frequency of encountering an exemplar across all daily contexts and scenarios. The participants were not given categorical information about the target words in the familiarity experiment (except in cases of ambiguity), which differs from the procedures in some studies (e.g., Hampton and Gardiner, 1983). Familiarity measures how often the target word (the written form of a concept) is experienced in a general context, among other words which are not necessarily from the same category. In experiment 2, we asked participants to rate how often they experienced (by hearing, reading or using, etc.) the word “狗” (“gau2,” dog) in their daily lives, instead of asking them to rate how often they had experienced it as “a kind of domestic pet.” This approach of avoiding the co-occurrence of the exemplar and its category limited the potential interaction between general familiarity with the concept itself, as well as familiarity with the concept cued by a certain category name. To further examine the relationship between the probability the production probability of an exemplar under a given category (Total) and the familiarity of the concept in general (Familiarity), the correlations between Total and Familiarity were calculated within each category (see Table 3). Each category contains different numbers of exemplars in this analysis, and for each category there is a correlation r and a corresponding p-value. No significant correlations were identified for the majority of categories (51 of 72), indicating that more frequently experienced concepts were not necessarily produced more frequently in response to a category cue. As mentioned earlier, instance probability is a legitimate a measurement of exemplar typicality, and the familiarity of a word is highly correlated with its frequency. Thus, the results of the current study are in line with those of previous studies (Mervis et al., 1976; Rosch et al., 1976).

The familiarity-based explanation of faster and more accurate processing for typical exemplars could be due to the generally high degree of familiarity of the concepts (Ashcraft, 1978), and familiarity confounds to the pattern of experimental results (e.g., processing efficiency) that argued for a semantic memory model (McCloskey, 1980). However, even if word familiarity is an important determinant of typicality, it cannot account for all of the variance in typicality ratings (Rosch et al., 1976).

In this study, all combinations of “Familiarity” and “Total” were observed (high-F and low-T; high-F and high-T; low-F and high-T; low-F and low-T) for the exemplars. For example, dog (“狗,” gau2) was a highly familiar concept among the participants (6.30 out of 7.00, higher than the category average of 5.30), but was retrieved as a low typical member in the category “a kind of farm animal” (Slot1 = 0.000, Slot2 = 0.025, Slot3 = 0.025, Total = 0.017). In contrast, solar eclipse (“日蝕,” jat6 sik6) was a far less familiar concept (4.25 of 7.00, lower than the category average of 4.81), but was the top-mentioned exemplar in the category of “an astronomical phenomenon” (Slot1 = 0.425, Slot2 = 0.150, Slot3 = 0.05, Total = 0.208).

Previous norming studies have often found familiarity (or word frequency) to be correlated with indices of typicality, such as overall frequency, first-occurrence, and mean rank (Montefinese et al., 2012), because typicality and familiarity are both associated with the ease of production of an exemplar. The non-correlation discrepancy may be due to the experimental designs used in the current study. In Experiment 1, the number of category responses was restricted to three, so all three responses were more likely to be highly familiar items, though their instance probabilities would still differ. In Experiment 2, the familiarity ratings were provided without the category context or other category items, though familiarity ratings were done within a category in some other studies (e.g., Hampton and Gardiner, 1983). The familiarity ratings in Hampton’s work are of more comparative and relative results among category members.

The comparative statistical indices of the categories are shown in Table 4, where “Valid Exemplars” represents the number of valid exemplars listed in the category. “Exemplars to 0.90 Coverage” is defined as the number of exemplars covering 90% of the occurrences of all valid entries. “0.90 Coverage%” is calculated as Exemplars to 0.90 Coverage divided by Valid Exemplars. “Invalid Exemplars%” is the proportion of invalid responses in a category, the same as Total of Invalid xx in Table 1. “First Exemplar Total%” is the Total probability of the top-ranked exemplar in that category, which indicates the degree of dominance of that exemplar and how typicality congregates in that category. “Average Familiarity” is the average familiarity score of all of the valid exemplars in a category, along with its standard deviation.

Over the years, the Battig and Montague (1969) English norms have been constantly updated and expanded, while researchers have compiled norms in other languages by adapting the category list and using similar methodologies (e.g., Storms, 2001 in Flemish; Marful et al., 2015 in Spanish, and many others). Among them, the norm of Van Overschelde et al. (2004) as an updated English norm, reflected contemporary category membership knowledge and captured the recent cultural changes based on Battig and Montague (1969). It has also been used as a comparable work to many other norm studies (Bueno and Megherbi, 2009 in French). A cross-norm comparison to Van Overschelde et al. (2004) should be representative as the comparison between the current study and the general body of norm studies.

In addition to the current representation of the categories, the exemplars, and the descriptive statistics, the database could provide primary training data for a more complex model with additional variables explored. The data presented in the current study is rather straight forward, as categories independent of each other and the exemplars are associated by their mutual categorical information. To further examine an interconnected semantic knowledge structure, more variables such as semantic relatedness of the exemplars and categorical feature analysis would be necessary, such that both intra-category exemplar relations and inter-category relations would be captured. This approach would provide a more sophisticated analysis of the semantic network of Hong Kong Cantonese, with an exploration of the concept clustering and interconnections between categories and concepts.

The processing efficiency of the highly typical exemplars suggests that categorical typicality imposes a spontaneous contextual prime on an exemplar, which can be considered as the stored semantic information about an exemplar. If we accept the hypothesis that Slot1 measures a kind of instant typicality, then this time-sensitive quality may be exploited in psychophysiological experiments to investigate the online processing of exemplars with congruent and incongruent categorial information primes. This type of investigation could be achieved by monitoring brain activity using technologies such as electroencephalography and event related potentials (Stuss et al., 1988; Kounios and Holcomb, 1992; Kutas and Iragui, 1998; Federmeier and Kutas, 1999). Furthermore, the data collected from young, healthy adults can serve as a benchmark for studies of other age groups, namely older adults and children, and of patients with cognitive deficits. As the semantic knowledge is generally preserved in the elder population (e.g., Park et al., 2002), this database provides resources in examining neural mechanisms of word retrieval for Cantonese-speaking elderlies. On the other hand, comparisons between the responses provided by neurologically impaired subjects and the database may reveal the domain-specific degeneration of semantic knowledge. The database may also benefit developmental studies examining how children establish lexical inventories by observing category and exemplar learning.