A corpus-based study of the ADJ-looking adjectivization in English

: This article aims to conduct a corpus-based study of the diachronic and synchronic distributions of a special type of participle adjectivization, the ADJ-looking adjectivization. The study based on the Corpus of Historical American English (COHA) ﬁ nds that this process of adjectivization consists of two phases: (1) The downward rank-shift from the look ADJ construction to the ADJ looking adjectivization is a process of metaphorization; (2) The transcategorization from the ADJ looking adjectivization to the ADJ - looking adjectivization is a process of lexicalization. The study based on the Corpus of Contemporary American English (COCA) ﬁ nds that ADJ-looking adjectivizations are mode and register sensitive but not discipline sensitive. The modi ﬁ er use prefers to occur more in hard science texts to increase the complexity of nominal groups and the predicative use prefers to occur more in soft science texts to increase the grammatical intricacy of sen-tences. The reason for the non-sensitivity across disciplines is that evaluative adjectives tend to occur in neither soft nor hard science texts.


Introduction
By what grammatical steps might a clause complex that straightforwardly realizes a semantic content as in (1a) be shifted to a simple clause realizing the same semantic content as in (1b)?
(1) a. They shredded the documents before they departed for the airport. b. Their shredding of the documents preceded their departure for the airport. (Halliday and Matthiessen 1999: 343) The two clauses in the clause complex in (1a) are shifted to two nominal groups in (1b). This shift is referred to as nominalization in the Hallidayan sense (Halliday 1985(Halliday , 1994Halliday and Matthiessen 1999, 2004, 2014. In the process of nominalization, the conjunction group before realizing the logico-semantic relation between the two clauses is correspondingly shifted to the verbal group preceded realizing process in the simple clause in (1b). This shift is referred to as verbalization (He and Yang 2018). With the nominalization of the verbal groups, the personal pronoun they has been shifted to the processive pronoun their to function as the modifiers of the head nouns shredding and departure. This process of downward rank-shift is referred to as a type of adjectivization of nominal groups identified by He (2019b). It is from nominalization, verbalization, and adjectivization that the clause complex in (1a) has been shifted to a simple clause in (1b). Besides nominal groups, participles also have the potential to be adjectivized. When a participle modifying the head noun (e.g. the moving blackboard, the sleeping lion) can be intensified or accept degrees of comparison (e.g. the very moving story, the most touching poem) (Biber et al. 1999), adjectivization occurs. This type of adjectivization can have predicative use. For example: (2) a. It was entertaining, instructive and often very touching. (COHA_1972) b. I haven't even mentioned the most touching part. (COHA_2005) According to , participles functioning as modifiers can themselves be further modified, resulting in complex modifiers (e.g. the fast-moving car, the well-boiled water). In the fast-moving car, for example, fast is the modifier of moving which can itself function as the modifier of the head noun car (e.g. the moving car).
Similarly, in the well-boiled water, boiled can itself function as the modifier of the head noun water. However, not all complex modifiers can be regarded as being further modified. Some participles can only function in complex modifiers (e.g. a good-looking boy, a fund-raising activity). In a good-looking boy, for example, looking cannot be considered as the modifier of boy, and in a fund-raising activity, fund has no potential to function as the modifier of raising. These complex modifiers violate the principle of right-headedness (see Williams 1981). Some complex modifiers can be intensified or accept degrees of comparison and have predicative use. For example, fast-moving can be intensified as in the very fastmoving car, and the car is very fast-moving is also acceptable. However, fund-raising in a fund-raising activity cannot be intensified by very or have predicative use. According to He (2019a), adjectivizations function as Classifier or experiential Epithet (Epithet 2 ) in nominal groups. In the fast-moving car, fast-moving functions as experiential Epithet and in a fund-raising activity, fund-raising functions as Classifier. Classifier adjectives cannot be intensified or accept degrees of comparison.
According to He and Guo (2021), a participle cannot be adjectivized by itself. Rather, it is the auxiliary verb be and the present or past participle together that have the potential to be adjectivized. For example, the verbal groups is pleased in (3a) and is worrying in (3b) can be adjectivized as pleased (e.g. the pleased city) and worrying (e.g. the worrying advertising industry) respectively. Seen from this perspective, the adjectivization of participles is also a process of downward rankshift as the adjectivization of nominal groups.
(3) a. The city is pleased by what it has seen so far. (COCA_ACAD) b. The enormous cost of sports programming is worrying the advertising industry. (COCA_NEWS) The ADJ-V-ing adjectivization is also a type of participle adjectivization that is getting increasingly prevalently used in the Present-Day English (Sun and Baayen 2021;Vartiainen 2016). Other popular sense-perception ascriptive verbs are sound, smell, feel and taste (Halliday and Matthiessen 2014: 269). In the present research, we will take the ADJ-looking adjectivization as our research target because the frequency of looking in this type of adjectivization is much higher than those of the other four sense-perception verbs in the corpus, accounting for 86.37% of the total five ADJ-V-ing adjectivizations. 1 The ADJ-looking adjectivization should have originated from the look ADJ construction. According to Vartiainen (2016), the ADJ looking complex came into being in the 18th century mainly functioning as noun modifier. From the early 19th century, the hyphenated ADJ-looking and its predicative use started to develop. A hyphen is added in between to reduce the possibility of parsing the entire construction into two separated units (Sanchez-Stockhammer 2018;Sun and Baayen 2021 It should be noted that in the look ADJ construction, the ADJ is always evaluative because look construes a relational process and evaluative adjectives are always gradable. After the look ADJ construction has been adjectivized as ADJ-looking, this 1 We retrieved 16,934 ADJ-looking adjectivizations, 927 ADJ-sounding adjectivizations, 987 ADJ-smelling adjectivizations, 527 ADJ-feeling adjectivizations and 231 ADJ-tasting adjectivizations in the COHA. The ADJ-looking adjectivization in English adjectivization is still evaluative. It, therefore, functions as the interpersonal Epithet in the nominal group. For example, the adjectives good-looking and young in the goodlooking young boy realize the interpersonal Epithet and the experiential Epithet respectively. Nominalization in the Hallidayan sense is a characteristic feature of scientific writing (Biber 2006;Biber et al. 1999;Galve 1998;Halliday 1998, because it gives scientific writing "an appearance of solidity, stability, and fixed factuality" (Banks 2005: 350). According to He and Yang (2018), nominalizations are more prevalent in written texts than in spoken texts with regard to mode distribution, and more prevalent in academic texts than in fiction texts with regard to register distribution. However, it is verbalization rather than nominalization that is discipline sensitive. Verbalizations are more prevalent in hard science texts than in soft science texts. Since nominalization, verbalization, and adjectivization are the three resources for creating ideational grammatical metaphor 2 (He 2019b), adjectivizations should have a similar distribution pattern to nominalizations and verbalizations.
However, of the two types of adjectivization (i.e. nominal group adjectivization and participle adjectivization), nominal group adjectivizations like nominalizations and verbalizations tend to occur more in formal academic texts with regard to register distribution. Participle adjectivizations, however, are oppositely distributed because participle adjectivizations function mostly as interpersonal Epithet, and "the speaker involvement is not a characteristic feature of academic writing" (He and Guo 2021: 13). As a special type of participle adjectivization, ADJ-looking adjectivizations should have a similar distribution pattern to present or past participle adjectivizations. That is, they should occur more in the relatively nonformal fiction texts rather than in the formal academic texts.
In this study, we will investigate the diachronic and synchronic distributions of the ADJ-looking adjectivizations. For this purpose, we will employ real language data to conduct a corpus-based study of the look ADJ constructions, the ADJ looking adjectivizations, and the ADJ-looking adjectivizations. We can hereby work on the following two hypotheses: (1) The look ADJ constructions have the potential to shift to the ADJ looking adjectivizations which will be further lexicalized as the ADJ-looking adjectivizations, and the modifier use of the ADJ-looking adjectivizations has the potential to shift to the predicative use; (2) ADJ-looking adjectivizations are more prevalent in spoken texts than in written texts with regard to mode distribution, more prevalent in fiction texts than in academic texts with regard to register distribution, and more prevalent in soft science texts than in hard science texts with regard to discipline distribution.
We will introduce the corpora and data collection in Section 2. Then we will present the diachronic and synchronic distributions of the ADJ-looking adjectivizations in Sections 3 and 4 respectively. The findings of this study will be discussed in Section 5.

Corpora
The corpora we used in the present research are the COHA and the COCA. The COHA is the largest structured corpus (406 million words) of historical English, covering a time span of 200 years from 1810 to 2009. This permits the "analysis of historical change on a scale not imaginable with earlier corpus designs" (Biber and Gray 2013: 108). Regarding the study of the ADJ-looking adjectivization, we can use this corpus to explore the diachronic distribution of this construction across decades over the past 200 years.
We use the COCA because this corpus is the only large, register-balanced corpus of American English. The COCA has been expanded by 20 million words each year since the early 1990s. It has reached one billion words till 2020 and new registers have been added. Regarding the present study, we will only use 600 million words distributed over five sections, i.e. Spoken, Fiction, Popular Magazines, Newspaper, and Academic (around 20% in each). The academic register is further organized with regard to disciplines (e.g. History, Education, Geography/ Social Science, Law/Political Science, Humanities, Philosophy/Religion, Science/ Technology, Medicine, Miscellaneous, and Business). This can facilitate the investigation of the synchronic distribution of the ADJ-looking adjectivizations across modes, registers, and disciplines.
Both the COHA and the COCA are available online. 3 They also provide convenient search tools. We can write search queries following the query syntaxes to retrieve the ADJ-looking adjectivizations.

Data collection
According to the query syntaxes provided by the corpora, we wrote the following search queries: This first search query can be described as any construction consisting of look and an adjective not following any adjective. The second to the fourth search query can be described as any noun preceded by an adjective or a comparative or superlative adjective and looking. The fifth search query can be described as any construction consisting of an adjective and looking following a linking verb and followed by a punctuation mark. Using these five search queries we can retrieve concordance lines as in (6): (6) a. We will have to look sharp. Note that the occurrences retrieved using the second search query subtract those retrieved using the fourth search query will result in the modifier ADJ looking adjectivizations without a degree modifier. We then wrote the following search queries to retrieve the ADJ-looking adjectivizations. The sixth search query can be described as any ADJ-looking adjectivization followed by a noun. The seventh search query can be described as any ADJ-looking adjectivization with intensifier or degree modifier followed by a noun. The eighth search query can be described as any ADJ-looking adjectivization following any form of the relational verb be. For example: (7) a. His deputy was an aimless-looking goof with fresh acne scars.
(COHA_2006) b. It was a very thin, very flat-looking weapon. (COHA_2005) c. Mr. Geller is soft-looking and bottom-heavy like an old pear.
Similarly, subtracting the occurrences retrieved using the seventh search query from the occurrences retrieved using the sixth search query will result in the modifier ADJ-looking adjectivizations without intensifier or degree modifier. The data we retrieved from the two corpora are shown in Tables 1 and 2.

Diachronic distribution of the ADJ-looking adjectivizations
In this section, we investigated the diachronic distribution of the ADJ-looking adjectivizations.

Comparing the ADJ looking and the ADJ-looking adjectivizations
We added up the occurrences retrieved using the second and the fifth search queries for the frequency of the ADJ looking adjectivizations. Similarly, we added up the occurrences retrieved using the sixth and eighth search queries for the total frequency of the ADJ-looking adjectivizations. The size of the COHA is not evenly divided across decades. To facilitate comparison, we converted the raw frequencies in each sub-corpora to the normalized frequencies of per million words 5 and then further converted the normalized total frequency of the ADJ looking adjectivizations (111.81) and that of the ADJ-looking adjectivizations (451.65) to that of the look ADJ constructions (1,678.34) (see Figure 1). Figure 1 shows that the diachronic distribution of the look ADJ constructions maintains an overall trend of increase, whereas that of the ADJ looking adjectivizations shows a general trend of decrease. The diachronic distribution of the look ADJ constructions and that of the ADJ looking adjectivizations are not necessarily correlated. This is not in agreement with our hypothesis that the look ADJ constructions will shift to the ADJ looking adjectivizations. It is also shown that the diachronic distribution of the ADJ-looking adjectivizations is leveling off since the 1850s and this distribution trend is significantly negatively correlated with that of the ADJ looking adjectivizations at the 0.05 level (r = −0.461; p = 0.041 < 0.05), indicating that the ADJ looking adjectivizations have the potential to shift to the ADJ-looking adjectivizations.
The look ADJ construction functions as process and attribute in relational clauses, and so the ADJ looking adjectivization and the ADJ-looking adjectivization maintain the same nature and can be intensified or accept degrees of comparison. The question is whether the comparative ADJ-looking adjectivization originated directly from the comparative ADJ looking adjectivization (e.g. good looking → better looking → better-looking) or from a further word-formation process (e.g. good looking → good-looking → more good-looking).  In the following, we took the adjective good as an example to compare the diachronic distribution of the better-looking adjectivization and that of the more good-looking adjectivization. The reason for choosing good is that the frequency of the good-looking adjectivization is the highest (3,106) of all the ADJ-looking adjectivizations (16,934), accounting for over 18.34%. Another reason is that good has the comparative form better (e.g. a better-looking boy).
The data we retrieved from the COHA regarding the comparative good-looking adjectivization are shown in Table 3, and the diachronic distributions of the normalized frequencies of the data are shown in Figure 2.
It can be seen from Table 3 and Figure 2 that of all the five groups of data goodlooking adjectivizations count the most. This type of adjectivization keeps a general trend of increase over the past 200 years. The diachronic distribution of the good looking adjectivization, however, does not show an obvious trend of increase or decrease. Both the better looking and better-looking adjectivizations show a general trend of increase. Bivariate correlation test shows that the diachronic distribution of the better-looking adjectivization and that of the better looking adjectivization are significantly correlated at the 0.01 level (r = 0.610; p = 0.004), indicating that the better looking adjectivization does not have the potential to shift to the betterlooking adjectivization; rather, they are of the same group of data. There are only a very small number of more good-looking constructions as in (8). However, of the eight more good-looking constructions, there is still one in which more functions as Numerative in the nominal group as in (9). It can be concluded that the better-looking adjectivization does not occur later than the good-looking adjectivization; rather it occurs while the good looking adjectivization is shifting to the good-looking adjectivization. It can also be seen that, in recent years, the good-looking adjectivization has begun to accept degrees of comparison, indicating the full lexicalization.

Comparing modifier and predicative uses of ADJ-looking adjectivizations
In this section, we compared the diachronic distribution of the modifier ADJ-looking adjectivizations and that of the predicative ADJ-looking adjectivizations. To facilitate comparison, we converted the normalized total frequencies of The ADJ-looking adjectivization in English the modifier ADJ looking adjectivizations (108.2), the predicative ADJ looking adjectivizations (3.61) and the predicative ADJ-looking adjectivizations (22.97) to that of the modifier ADJ-looking adjectivizations (428.68) (see Figure 3). Figure 3 shows that, in general, the modifier use of the ADJ looking adjectivizations is decreasing while the predicative use is fluctuating. Bivariate correlation test shows that the two groups of data are negatively but not significantly correlated (r = −0.399; p = 0.08 > 0.05), indicating that some of the modifier ADJ looking adjectivizations do not shift to the predicative use.
As shown in Table 1, the raw frequencies of the modifier and the predicative ADJ looking adjectivizations are 1,796 and 86 respectively, the predicative use accounting for 4.57% of the total, whereas the raw frequencies of the modifier and the predicative ADJ-looking adjectivizations are 9,253 and 534 respectively, the predicative use accounting for 5.46% of the total. This also indicates that some modifier ADJ looking adjectivizations have not yet been ready to shift to predicative use.
The diachronic distribution of the modifier ADJ-looking adjectivizations increases quickly over the first five decades, then begins to decrease slowly from the 1850s to the 1900s, and levels off from the 1900s to the 2000s, whereas that of the predicative use, though sometimes fluctuates, maintains an overall trend of increase. The two groups of data are significantly correlated at the 0.05 level (r = 0.492; p = 0.028 < 0.05), indicating that the ADJ-looking adjectivizations have both the modifier use and the predicative use.
The diachronic distribution of the modifier ADJ looking adjectivizations and that of the modifier ADJ-looking adjectivizations are negatively correlated (r = −0.426; p = 0.061 > 0.05), although not significantly. The distributions of the predicative ADJ looking adjectivizations and the predicative ADJ-looking adjectivizations are even less correlated (r = 0.379; p = 0.133 > 0.05). This indicates that the modifier ADJ looking adjectivizations have the potential to shift to the modifier ADJ-looking adjectivizations. However, the distribution of the predicative ADJ-looking adjectivizations is significantly negatively correlated with that of the modifier ADJ looking adjectivizations at the 0.01 level (r = −0.772; p = 0.000 < 0.01), indicating that the predicative ADJ-looking adjectivizations are shifted from the modifier ADJ looking adjectivizations. The predicative use of the ADJ-looking adjectivizations is significantly correlated with the predicative use of the ADJ looking adjectivizations at the 0.01 level (r = 0.719; p = 0.000 < 0.01), indicating that they are of the same group of data, both being shifted from the modifier ADJ looking adjectivizations.
From the above analysis, we can conclude that the ADJ looking adjectivizations tend less and less to occur at the modifier place in nominal groups. A small number of modifier ADJ looking adjectivizations have been shifted to the predicative use and a larger number to either the modifier use or the predicative use of the ADJ-looking adjectivizations.

Synchronic distribution of ADJ-looking adjectivizations
In this section, we investigated the synchronic distributions of the ADJ-looking adjectivizations across modes, registers, and disciplines.

Mode distribution of ADJ-looking adjectivizations
As shown in Table 2, we collected 12,823 and 56,138 occurrences of the total three types of constructions in the spoken sub-corpus and the written sub-corpus. To facilitate comparison, we converted the total normalized frequency of the ADJ looking adjectivizations (4.69) and that of the ADJ-looking adjectivizations (23.83) to that of the look ADJ constructions (186.52) in the COCA (see Figure 4). Figure 4 shows that under equal total frequencies, the distributions of the three groups of data are significantly different in spoken texts and written texts (x 2 = 45.051, df = 2, p = 0.000 < 0.01) and the correlation strength is modest (F = 0.284). The look ADJ constructions are not obviously mode sensitive. The ADJ looking adjectivizations are more prevalent in spoken texts than in written texts, The ADJ-looking adjectivization in English whereas the ADJ-looking adjectivizations are more prevalent in written texts. This is in agreement with our hypothesis that the ADJ-looking adjectivizations tend to occur more in spoken texts than in written texts with regard to mode distribution.
In the following, we compared the mode distributions of the modifier and predicative ADJ-looking adjectivizations. To facilitate comparison, we converted the normalized total frequencies of the modifier ADJ looking adjectivizations (4.04), the predicative ADJ looking adjectivizations (0.65) and the predicative ADJ-looking adjectivizations (1.84) to that of the modifier ADJ-looking adjectivizations (21.99) (see Figure 5). Figure 5 shows that both the modifier use and the predicative use of the two types of adjectivizations are mode sensitive. The distributions of the four groups of data are significantly different in spoken texts and in written texts (x 2 = 12.416, df = 3, p = 0.008 < 0.01) and the correlation strength is moderate (F = 0.371). Generally, both  the modifier and the predicative ADJ looking adjectivizations tend to occur more in spoken texts, whereas both the modifier and the predicative ADJ-looking adjectivizations are more prevalent in written texts. Both the modifier ADJ looking adjectivizations and the modifier ADJ-looking adjectivizations prefer to occur more in written texts, whereas both the predicative uses are more prevalent in spoken texts. The reason is that noun modifiers increase the complexity of nominal groups and hence contribute to the lexical density of written texts (Halliday 1989).

Register distribution of ADJ-looking adjectivizations
Of the 56,138 occurrences of the three types of constructions in the written subcorpus, there are 31,889, 14,241, 7,920, and 2,088 occurrences in the Fiction, Magazine, Newspaper, and Academic sub-corpora of the COCA respectively. In this section, we compared the register distribution of the ADJ-looking adjectivizations. To facilitate comparison, we converted the normalized frequencies of the ADJ looking adjectivizations (6.95) and the ADJ-looking adjectivizations (68.02) to that of the look ADJ constructions (385.37) in written texts of the COCA (see Figure 6). Figure 6 shows that all three types of constructions tend to occur mostly in non-academic texts, especially in fiction texts, but tend least to occur in academic texts. This is in agreement with our hypothesis that the ADJ-looking adjectivizations are more prevalent in fiction texts with regard to register distribution. The reason is that the ADJ looking adjectivizations and the ADJ-looking adjectivizations are evaluative and the speaker's involvement is not a characteristic feature of academic writing (He and Guo 2021). According to , academic language is characteristic of objectivity and impersonality. The ADJ looking and the ADJ-looking adjectivizations are similarly distributed in the fiction and the academic registers (x 2 = 0.061, df = 1, p = 0.805 > 0.05) but the correlation strength is very weak (F = 0.003). However, the distribution of the look ADJ constructions and that of the ADJ looking and the ADJ-looking adjectivizations are significantly different in the two registers (x 2 = 238.920, df = 1, p = 0.000 < 0.01), but the correlation strength is also weak (F = 0.084). Comparatively, the ADJ looking and ADJ-looking adjectivizations tend less to occur in fiction texts.
It can also be seen from the data shown in Table 2 that the raw frequency of the look ADJ constructions accounts for 86.03% of the total frequency of the three types of constructions in fiction texts, ranking the highest among all the four written registers, 80.56, 82.27 and 73.66% in the magazine, newspaper, and academic texts respectively. According to Copeck et al. (1997), fiction is the least technical, and academic is the most technical. The reason for the highest frequency of the look ADJ constructions in fiction texts is that relational clauses of the attributive type are mostly evaluative, which is not a characteristic feature of academic writing.
Adjectivizations can package information to increase the semantic density in academic texts, but our data also show a lower frequency of adjectivizations in academic texts. The reason is that the ADJ-looking adjectivizations are all evaluative. They function as the interpersonal Epithet in nominal groups. However, it can also be seen that in the academic register the ADJ looking adjectivizations count the most in the three groups of data, and their frequency is much higher than that of the look ADJ constructions.
In the following, we compared the register distributions of the modifier and the predicative ADJ-looking adjectivizations. We converted the nominalized frequencies of the modifier ADJ looking adjectivizations (6.31), the predicative ADJ looking adjectivizations (0.63) and the predicative ADJ-looking adjectivizations (4.21) to that of the modifier ADJ-looking adjectivizations (63.82) (see Figure 7).  Figure 7 shows that the modifier and the predicative ADJ looking adjectivizations and the modifier and the predicative ADJ-looking adjectivizations, on the whole, tend to occur more in non-academic texts, especially in fiction texts (x 2 = 84.665, df = 9, p = 0.000 < 0.01). In fiction texts, both the predicative ADJ looking adjectivizations and the predicative ADJ-looking adjectivizations are more prevalent than the modifier uses. This distribution pattern is the opposite of that in academic texts. This is in agreement with our hypothesis. The reason is that fiction texts are the least technical and tend to represent experience as processes (Halliday 1989). The predicative use can explicitly express the speaker's judgement and tends to appear in relational clauses to construe quality. This fits into the feature of non-academic language.
One reason for the higher frequency of the modifier use than the predicative use in academic texts is that modifier adjectivizations can increase the complexity of nominal groups, which is a characteristic feature of academic texts. The other reason is that the modifier use can avoid the author's involvement and allows little negotiation between speaker and hearer in the text, hence increasing the objectivity of academic texts. The predicative use of the ADJ-looking adjectivizations will again increase the non-technicality of fiction texts. This explains the relatively lower frequency of the predicative use of the ADJ-looking adjectivizations than that of the modifier used in academic texts.

Discipline distribution of ADJ-looking adjectivizations
As shown in Table 2, we collected only a very small number of ADJ-looking adjectivizations from the academic sub-corpus of the COCA. Of the eight disciplines in the academic sub-corpus of the COCA, Science/Technology and Medicine are categorized as hard sciences and History, Education, Geography/Social Science, Law/Political Science, Humanities, and Philosophy/Religion as soft sciences. The data we retrieved from the eight disciplines of the academic sub-corpus are shown in Table 4.
For comparison, we converted the normalized total frequencies of the ADJ looking adjectivizations (0.97) and the ADJ-looking adjectivizations (6.78) to that of the look ADJ constructions (20.59) (see Figure 8). Figure 8 shows that the look ADJ constructions are more prevalent in soft science texts than in hard science texts, but the distribution of the look ADJ constructions is not significantly different from that of the ADJ-looking and the ADJ-looking adjectivizations (x 2 = 2.529, df = 1, p = 0.112 > 0.05) and the correlation The ADJ-looking adjectivization in English  strength is very weak (F = 0.038). However, it is too early to conclude that the ADJ-looking adjectivizations are not discipline sensitive because the ADJ-looking adjectivizations have modifier and predicative uses.
As shown in Table 4, we retrieved 476 occurrences of the ADJ looking and the ADJ-looking adjectivizations from the soft science texts and the hard science texts in the COCA, including 53 modifier ADJ looking adjectivizations, 4 predicative ADJ looking adjectivizations, 397 modifier ADJ-looking adjectivizations, and 22 predicative ADJ-looking adjectivizations. For comparison, we converted the nominalized total frequencies of the modifier ADJ looking adjectivizations (0.92), the predicative ADJ looking adjectivizations (0.045) and the predicative ADJ-looking adjectivizations (0.307) to that of the modifier ADJ-looking adjectivizations (6.47) (see Figure 9).  The ADJ-looking adjectivization in English Figure 9 shows that neither the modifier nor the predicative ADJ-looking adjectivizations tend to occur in hard science texts. Comparatively, under equal total frequencies, the predicative uses are more prevalent than the modifier uses in soft sciences texts, whereas the modifier uses are more prevalent than the predicative uses in hard science texts. This is in agreement with our hypothesis. The reason for the prevalence of modifier uses in hard science texts is that hard science text writers seek to downplay the personal role and distance from interpretations while soft science text writers value personal credibility and tend to express their conviction and emphasize their contribution (Hyland 2008). The modifier use enables speakers to package the quality of nouns without revealing who is responsible for the evaluation. Conversely, the predicative use presents a straightforward comment or judgement (Kress and Hodge 1979).

Discussion
The COHA-based study finds that the look ADJ constructions have the potential to shift to the ADJ looking adjectivizations and the transcategorization from the ADJ looking adjectivizations to the ADJ-looking adjectivizations increases the degree of lexicalization. The COCA-based study finds that the ADJ-looking adjectivizations are mode and register sensitive but not discipline sensitive and that the modifier use prefers to occur in hard science texts and the predicative use prefers to occur in soft science texts. This section is devoted to the discussion of these research findings.

On the diachronic distribution of ADJ-looking adjectivizations
The process of adjectivization can be considered as a process of metaphorization. Metaphorization in the Hallidayan sense (Halliday and Matthiessen 1999) can be interpreted as the semantic blending or semantic junction. The semantic categories of sequence, figure, and element are realized congruently as clause complex, clause, and group/phrase respectively at the lexicogrammatical stratum. The logico-semantic relation between the two clauses in a clause complex is construed by the functional element relator, which "forms figures into sequences" (Halliday and Matthiessen 1999: 177). The relator is realized as a conjunction group. The elements of a figure include participant, process, and circumstance which are congruently realized as nominal group, verbal group, and adverbial group or prepositional phrase. The mismatch between semantics and lexicogrammar creates ideational grammatical metaphor. According to Halliday and Matthiessen (1999), adjectivization joins process with quality and hence is a sub-type of grammatical metaphor (He 2019b;Taverniers 2004). For example: (10) a. One day he came upon a funny looking little cap. (COHA_1921) b. One day he came upon a little cap. The cap looked funny.
The second clause in (10b) is rank-shifted to a nominal group in (10a). In the process of nominalization, the look ADJ construction looked funny is adjectivized to funny looking, joining process with quality, and hence creating ideational grammatical metaphor. The process of adjectivization can also be considered as a process of lexicalization. Lexicalization is the process by which new linguistic entities become conventionalized on the level of the lexicon (Blank 2001). It involves fusion concerning alternative ways of expressing the same concept. Lexicalization is a gradual historical change which results in new lexical forms of any complexity. It should be noted that lexicalization is not a word formation process such as clipping, conversion, compounding and borrowing (Brinton and Traugott 2005).
The shift from the ADJ looking adjectivization to the ADJ-looking adjectivization is a process of lexicalization. According to Vartiainen (2016), the ADJ looking adjectivization has to be lexicalized because when it is modified by a degree adverb the whole construction will be grammatically ambiguous. The hyphenated ADJ-looking adjectivization will reduce the ambiguity. For example: (11) a. In all, Mr. Tripp was a most impressive-looking man. (COCA_FIC) b. He spoke for 10 minutes to the ever more skeptical-looking official.
Our research also shows that the ADJ-looking adjectivizations (e.g. good-looking) and their comparative forms (e.g. better-looking) are of the same group of data. This indicates that better looking does not necessarily have the potential to shift to better-looking; rather, they are also of the same group of data. In recent years, the fully lexicalized good-looking adjectivization has begun to accept degrees of comparison.
If an adjective is semantically non-gradable, it always functions as modifier instead of predicative (Bauer and Huddleston 2002). According to Biber et al. (1999), being predicative and accepting degrees of comparison are two defining characteristics of adjectives. The more characteristics an adjective has, the more central it is as an adjective. The ADJ-looking adjectivizations are semantically gradable, so they have both the modifier use and the predicative use since their formation.
The ADJ-looking adjectivization in English

On the synchronic distribution of ADJ-looking adjectivizations
The ADJ-looking adjectivization is a resource for creating ideational grammatical metaphor, it should have a similar distribution trend as nominalization and verbalization. However, our research shows that the ADJ-looking adjectivizations tend to occur more in spoken texts than in written texts with regard to mode distribution, more in fiction texts than in academic texts with regard to register distribution and more in soft science texts than in hard science texts with regard to discipline distribution. This is the polar opposite of the synchronic distribution of nominalization or that of verbalization. There might be two reasons: (1) Nominalization shifts clause to group and verbalization shifts clause complex to clause. The ADJlooking adjectivization functions as interpersonal Epithet in nominal groups or as Attribute in relational clauses. Both uses highlight the author involvement, and the author involvement is not a characteristic feature of academic writing. (2) The target semantic domain of the ADJ-looking adjectivizations is Quality which is regarded to be less stable and represents values on a single dimension (Halliday and Matthiessen 1999). Academic language tends to move towards thingness because the Thing is regarded as the most stable domain to be typologized. The modifier use of the ADJ-looking adjectivizations helps increase the lexical density of clauses or the complexity of nominal groups. This is a characteristic feature of written texts (Halliday 1994). The predicative use, however, helps increase the grammatical intricacy. Seen from this perspective, nominalization, verbalization, and adjectivization are not necessarily the characteristic feature of academic language. They are the characteristic feature of academic language only when they increase the lexical density of language. This can explain why the modifier ADJ-looking adjectivizations are more prevalent in academic texts than in fiction texts with regard to register distribution and more prevalent in hard science texts than in soft science texts with regard to discipline distribution despite the fact that they function as interpersonal Epithet.

Conclusions
In this study, we investigated the diachronic and synchronic distributions of the ADJ-looking adjectivizations. The COHA-based diachronic research finds that the process of ADJ-looking adjectivization includes two phases. The first phase is the shift from the look ADJ construction to the modifier ADJ looking adjectivization. This is a process of metaphorization. The second phase is the shift from the ADJ looking adjectivization to the ADJ-looking adjectivization. This is a process of lexicalization. During the second phase, the shift from the modifier use to the predicative use reinforces the adjective features of the ADJ-looking adjectivization.
The COCA-based synchronic research finds that the ADJ-looking adjectivizations are mode and register sensitive but not discipline sensitive and that the modifier use prefers to occur more in hard science texts and the predicative use prefers to occur more in soft science texts. The modifier ADJ-looking adjectivizations are more prevalent in written texts with regard to mode distribution, more prevalent in academic texts with regard to register distribution, and more prevalent in hard science texts with regard to discipline distribution. The predicative ADJ-looking adjectivizations, however, are oppositely distributed. The reason is that the modifier use of the ADJ-looking adjectivizations increases the lexical density and the predicative uses of the ADJ-looking adjectivizations increases the grammatical intricacy.