A collostructional approach to Japanese noun-modifying clause construction use and acquisition: a learner corpus study

2.1 Overview of Japanese clausal noun modification

The Japanese NMCC (hereafter NMCC) has been discussed in the literature (e.g., Matsumoto 2017, 2018) as a single construction with two main constituents: a modifying clause and a head noun (HN), see Figure 1.

Figure 1:

The NMCC construction (adapted from Matsumoto 2018: 464).

The construction is described as having a “wide range of interpretations” (Matsumoto 2018: 464) due to relations between the HN and the modifying clause that are grammatically and semantically broad (Matsumoto 2017). Structurally, the modifying clause predicate is a well-formed independent clause (Matsumoto 2018: 465) that consists of a noun combined with a copula, a verb, or an adjective in the finite or adnominal form. Three simplified examples of typical varieties of NMCCs are provided in Table 1.

As can be seen in the examples in Table 1, the modifying clause predicate can share the same verb (e.g., kat-ta, ‘buy’) in the past prenominal/finite form and contain unexpressed arguments (e.g., subject, direct object) and adjuncts, which are interpreted as having an association with the HN. However, the array of arguments in the modifying clause will result in different interpretations of the NMCC (e.g., argument NMCC, adjunct NMCC). For example, the HN can play the role of the subject or direct object of the verb as seen in the instances where the argument is internal to the clause. The HN can also indicate the location where an action or event occurred, as in the second argument structure in Table 1, where the HN is the adjunct of the modifying verb. Matsumoto (2018) notes that language users’ acceptability judgements concerning interpretations of the construction are not only constrained by syntactic, semantic, and pragmatic knowledge but also shared knowledge of the world. An example demonstrating a selection of possible interpretations for a single instance of the NMCC as determined by the interaction between linguistic and contextual factors is provided in Table 2.

It is important to note that, while it is possible for Japanese NMCCs to have a modifying clause that is identical in structure to an independent clause in terms of expressed arguments, there are instances where the HN cannot be associated with a missing argument in the modifying clause. There are also instances where the HN cannot be interpreted as an argument or adjunct of the modifying clause (see Table 3 for examples of these instances). Matsumoto (2017) observed that these characteristics of Japanese NMCCs can lead to the creation of instances with potentially ambiguous interpretations.

2.2 Descriptive approaches to analyzing Japanese NMCCs

The descriptive approach to analyzing NMCCs can be attributed to the in-depth analyses of the construction in early typological research, most notably Teramura (1969). Teramura (1969) investigated the syntactic and semantic relationships between the HN and the modifying clause predicate within an NMCC. The author identified inner- and outer-relationship constructions as two broad types of clausal noun modification. Inner-relationship constructions (uchi no kankei) embed into the modifying clause predicate a HN with one of the five grammatical functions: Subject (SU), Direct Object (DO), Indirect Object (IO), Oblique (OBL), and Genitive (GEN). Conversely, outer-relationship constructions (soto no kankei) have HNs that cannot be embedded into the modifying clause predicate due to syntactic restrictions related to the semantic properties of the HNs. Teramura’s (1969) bipartite classification of NMCCs based on the relationship between the semantically loaded HN and the modifying clause predicate has informed the design of subsequent typological studies and empirical investigations into the acquisition of the construction by L1 Japanese speakers and JFL learners, the latter of which will be the focus of Section 2.3.

2.3 Empirical studies of NMCC acquisition by L2 Japanese learners

We are only aware of a small number of empirical studies written for an English-speaking readership that have investigated the acquisition of NMCCs by JFL learners (Ozeki and Shirai 2007; Yabuki-Soh 2007). Ozeki and Shirai’s (2007) corpus study tested the applicability of Keenan and Comrie’s (1977) noun phrase accessibility hierarchy (NPAH). The authors hypothesized that JFL learners are capable of producing more marked NMCCs even at lower levels of proficiency. Using the KY (Kamada Yamaguchi) Corpus (Kamada 1999, 2006), a learner corpus of transcribed ACTFL Oral Proficiency Interviews, instances of L1 and L2 NMCCs were analyzed. NMCCs were operationalized as nouns modified by verbs, adjectives with complements, and adjectives in the past tense (see Appendix A for examples of each of these NMCC types). The study found that with the exception of L1 Korean-speaking learners, NMCC use was contingent on the animacy status of the HN rather than its markedness as predicted by the NPAH. JFL learner productions from the KY Corpus revealed an initial mapping of SU relatives with animate-head NPs and DO and OBL relatives with inanimate-head NPs. Similar animacy effects were found in the analysis of the L1 Japanese data.

The quasi-experimental study by Yabuki-Soh (2007) explored whether JFL learners’ acquisition of a marked NMCC, particularly the OBL relativization type, would facilitate learning NMCC types higher in the NPAH. Adult learners of first-year Japanese were randomly assigned to a form-based, meaning-based, or form-plus meaning-based instructional group and were exposed to only the OBL-type NMCC in three treatment sessions. Comprehension and sentence completion tasks were used to measure participants’ progress across the three groups. Highly explicit instruction was found to play an important role in participants’ ability to recognize and use NMCCs. Moreover, instruction on only the OBL-type was found to facilitate participants’ ability to generalize its structure to other NMCC types. The study demonstrated that the acquisition of Japanese NMCCs is a complex and nonlinear process, which cannot be encapsulated by the NPAH predictions correlating less marked constructions and learner accessibility, suggesting a need to examine NMCCs using alternative theoretical frameworks and analytical approaches.

Departing from research testing the NPAH, Yabuki-Soh (2013) analyzed NMCCs in seven postsecondary JFL textbooks to determine the extent to which instructional materials reflect the characteristics of Japanese clausal noun modification. The author found that textbooks tend to focus on inner-relationship NMCCs, namely, the SU and DO types, and neglect outer-relationship NMCCs. The author makes several suggestions to improve the existing coverage of NMCCs in JFL textbooks, notably, the provision of grammar explanations that extend the use of NMCCs from providing additional information about people and things (i.e., inner-relationship NMCCs) to expressing conceptual content (i.e., outer-relationship NMCCs) and the inclusion of more varied HNs in discussions of outer-relationship NMCCs.

Taken together, the above studies show changes in the trajectory of research on NMCCs, which shifted from a focus on the learnability of RC constructions by JFL learners to the quality of JFL textbooks regarding their inclusion and treatment of noun-modifying clauses of different types. There remains a need, however, to better understand JFL learners’ knowledge of NMCCs as reflected in their use of the construction. Our study seeks to address this research gap by drawing on learner corpus data to analyze NMCCs in JFL learner and L1 Japanese speaker production and to identify potential discrepancies in their usage patterns.

2.4 The present study

According to Mori et al. (2020), the development of quality instructional materials for JFL teaching is contingent on “the integration of fine-grained qualitative linguistic analysis and a systematic analysis of a large database enabled by technology” (109). Answering this call for more corpus-based research investigating L2 Japanese development, our study uses quantitative methods to complement existing qualitative linguistic analyses of the structure of and lexical items in NMCCs. The study aims to provide insights that can inform the treatment of NMCCs in pedagogical materials and in classroom instruction. Three research questions guide this inquiry:

1. What NMCC types are part of JFL learners’ spoken repertoire?

2. How does JFL learners’ use of NMCCs in spoken Japanese compare to that of L1 Japanese speakers?

3. Are there any significant associations between NMCCs and specific head nouns in L1 and L2 usage?

The remaining sections will outline the methodology of the study, present findings on the three research questions, and conclude with a discussion of results, limitations, and future directions.

3.1 The International Corpus of Japanese as a Second Language (I-JAS)

The International Corpus of Japanese as a Second Language (I-JAS; Sakoda et al. 2016, 2020) is a large-scale (8,076,969 words) learner corpus in the National Institute for Japanese Language and Linguistics (NINJAL) family of corpora. The corpus is composed of written and spoken language data from three groups of adult learners (n = 1,000) ranging from 17 to 46 years in age: (1) learners of Japanese in a foreign language context (n = 850); (2) learners of Japanese in a classroom context in Japan (n = 100); and (3) learners of Japanese in naturalistic contexts (n = 50) (Sakoda et al. 2016, 2020). It also includes comparison data from L1 Japanese speakers (n = 50) ranging from 20 to 50 years in age. A key characteristic of the I-JAS is its inclusion of rich metadata pertaining to learners’ backgrounds and learning environments, which was collected through a 20-item questionnaire administered online in the participants’ respective L1 (see Appendix B for a breakdown of the metadata collected in the questionnaire). Additionally, the corpus includes information on various aspects of learners’ Japanese language proficiency, such as their linguistic knowledge and automatic processing ability, as measured by the Japanese Computerized Adaptive Test (Imai et al. 2009) and the Simple Performance-Oriented Test, respectively.

3.2 Data collection and analysis

The present study draws from the dialogue task (taiwa tasuku; ‘dialogue task’) in the I-JAS, a 5,952,601-word subcorpus of spoken interview data, in which the participant data isolated from interviewer data totals 3,168,198 words. The task is a 30-min semi-structured conversation facilitated by a L1 Japanese-speaking interlocutor on 15 predetermined topics largely pertaining to daily life. Appendix C presents the overall structure of the dialogue task and provides sample topics and questions. Instances of NMCCs produced by 850 learners of Japanese in a foreign language context, ranging in proficiency from beginner to advanced,^[1] with a majority of intermediate level learners (see Appendix D), and 50 Japanese L1 speakers, referred to as JJJ by the corpus compilers, were extracted in two separate key-word-in-context (KWIC) searches using the online corpus concordance system Chunagon (version 2.7.0; National Institute for Japanese Language and Linguistics n.d.). The graphical user interface-based query builder uses dependency parsing to retrieve mono- and multi-morphemic lexical items, more commonly referred to as short-unit words (SUWs) and long-unit words (Koiso et al. 2020; Sakoda et al. 2016, 2020), from individual NINJAL corpora based on annotated morphological information. Adopting an SUW search for lexemes tagged as nouns preceded by lexemes in the attributive form, a total of 25,738 KWIC concordance lines were retrieved from the corpus of L2 Japanese learner utterances (2,466,125 words), hereafter referred to as the Japanese as a Foreign Language (JFL) corpus. See Table 4 for an overview of the composition of the JFL corpus. A total of 4,441 KWIC concordance lines with 661 SUW HN lemma types were retrieved from the corpus of L1 Japanese speaker utterances or JJJ corpus (228,818 words).

A combination of Microsoft Excel and R scripts was used to prepare and process the data from the two corpora for analysis. The raw word counts were normalized to frequencies per 1,000,000 words to allow for comparisons between speaker groups. A simple collexeme analysis (Stefanowitsch and Gries 2003) was performed, focusing on HNs in the JFL and L1 Japanese corpora.

3.3 Collostructional analysis

Collostructional analysis (CA) is described as a “family of methods” (Gries 2019: 388) – collexeme analysis, distinctive collexeme analysis, and co-varying collexeme analysis – developed by Stefanowitsch and Gries (2003) that measures the strength of attraction or repulsion of lexical items (collexemes) to slots in constructions (Goldberg 1995: 1–6, 2013). Advantages of CAs include increased precision in grammatical descriptions over traditional collocation analysis, which have practical implications for fields such as language pedagogy wherein knowing the attraction between collexemes and constructions can inform materials design and classroom practices.

The collostructional analysis selected for the present study was a simple collexeme analysis (e.g., Stefanowitsch and Gries 2003), which focused on the HN slot of the NMCC. Chunagon was first used to extract all instances of nouns preceded by attributive forms in the JJJ and the JFL corpora. Using the downloadable search results, pivot tables were then generated in Microsoft Excel to identify the types and the tokens of nouns present in the corpora. To determine the observed frequency of each noun, all instances of nouns in the JJJ and the JFL corpora were also extracted. Pivot tables were again created to obtain all noun types and their frequencies in the corpora. Due to the larger size of the JFL corpus and limitations of the concordancer regarding downloadable search results (i.e., a maximum of 100,000 concordance lines), nouns in the JFL corpus had to be retrieved separately for each L1 group and were later combined into a single spreadsheet, mirroring that of the JJJ corpus. The resulting information was compiled into two data frames containing three columns – types of nouns in the HN slots, their frequency in the NMCC, and their frequency in the corpus – which served as input for the simple collexeme analysis performed using R (version 4.2.2) (R Core Team 2022). The package {collostructions} (Flach 2021) was used to compute the collostructional strength, measured as log-likelihood (Dunning 1993), for the HNs in the corpora, and a Bonferroni correction was applied. The alpha level was set at p < 0.05. In the calculation, the size of the complete JJJ corpus (in words) was used for the “corpsize” argument of the function.

4.1 Distribution of NMCCs in JFL learner speech

Two searches – nouns preceded by attributive forms and attributive forms followed by nouns – were performed to isolate SUWs in the HN and the modifying clause predicate slots of the NMCC in the JFL corpus. Analyses of the modifying predicate slot revealed that JFL learners use a variety of NMCC types categorizable by POS. Four POS categories, in particular, were identified (see left side of Table 5):^[2] auxiliary verbs (jodōshi; 10,068 instances), verbs (dōshi; 9,130 instances), adjectives (keiyōshi; 6,506 instances), and suffixes (setsubiji; 34 instances). With the exception of auxiliary verbs, these categories are further subdivided as indicated in Table 5. Analyses of the HN slot in the NMCC revealed four categories of nouns (see left side of Table 6). The most frequent category was common nouns (meishi-fūtsūmeishi; 25,456 instances) with six subtypes as detailed in Table 6.^[3] The second most frequent noun types were proper nouns (meishi-koyūmeishi; 196 instances) of six subtypes (see Table 6). Less frequent noun types were auxiliary nouns (meishi-jodōshigokan; 60 instances) and numeral nouns (meishi-sūshi; 38 instances).

4.2 Contrastive analysis of NMCCs in L1 and L2 Japanese

Comparisons of the NMCC predicates in the L1 and L2 data showed complete overlap in the POS categories and subcategories of SUW directly to the left of the HN (see Table 5 for an overview). The most frequent predicate type in both corpora were auxiliary verbs (39.12 % of all NMCC tokens in JFL and 45 % in JJJ data), followed by verbs (35.47 % in JFL; 27.98 % in JJJ) and adjectives (25.38 % in JFL; 16.59 % in JJJ). Suffixes were the least frequent type of predicate in NMCCs in both corpora (0.13 % in JFL; 0.20 % in JJJ).

An examination of the lexemes in the HN slots of L1 and L2 NMCCs also revealed overlap in POS categories. In the NMCCs produced by both JFL learners and L1 Japanese speakers, the HN slot of the construction was filled predominantly by common nouns (98.86 % in JFL; 99.27 % in JJJ) of the same five subcategories: general common nouns, adverbial nouns, verbal nouns, counter nouns, and adjectival nouns (see Table 6). Similarities were also observed in JFL learners’ and L1 Japanese speakers’ use of other noun subcategories, such as country names, auxiliary nouns, counter nouns, general places, verbal adjectival nouns, general names, and first names. NMCCs with general proper nouns and last names in the HN slot of the construction are the only point of difference between the two corpora, occurring exclusively in the JFL data while being the only noun subcategories absent from L1 Japanese speaker productions. These findings can be attributed to self-referential utterances typical of lower-proficiency speakers of a language, which would include proper nouns in the form of country of origin and general place and location names relating demographically heterogeneous JFL speakers. Table 6 summarizes the HN tokens and the top three lexemes within each subcategory.

The next step in the analysis involved identifying and comparing realizations of the construction within the three most frequent modifying clause predicate types (auxiliary verbs, verbs, adjectives). Suffixes were omitted given their low frequency. In the auxiliary_HN realization, the top three most frequently used auxiliary lexemes in the JJJ corpus were -ta, -da, and -te iru/teru (continuation), which contrasted with the lexemes -da, -ta, and -nai found in the JFL corpus. The general nouns koto and toki were the two most frequent HNs in NMCCs in both corpora. The third most frequent HN, however, differed between the L1 and the L2 Japanese speakers with the former demonstrating a preference for the general noun kanji over the general noun hito, which was preferred by the latter. In verb_HN NMCC realizations, both JFL learners and L1 Japanese speakers showed a strong preference for the general verb iu/yū as the modifying predicate. This lexeme, however, was more frequent in the JJJ corpus (73.72 %) than in the JFL corpus (48.53 %). Realizations of i-Adjective_general_HN and i-Adjective_bound_HN NMCCs types were more frequent in the JJJ corpus (3,198.91 instances pmw) than in the JFL corpus (2,638.15 instances pmw). There was minimal overlap in the top three modifying clause types in the i-Adjective_generalHN NMCC. The most frequent types in the JJJ corpus were sugoi (463.23 pmw), chīsai (463.23 pmw), and kowai (135.47 pmw), whereas chīsai (352.27 pmw), warui (137.05 pmw), and ōki (114.35 pmw) were most frequent in the JFL corpus. As for the types of HNs in the i-Adjective_generalHN NMCC, there was little variation in lexical preferences between speaker groups. The adverbial nouns koro (231.62 pmw) and toki (201.02 pmw) followed by the general noun koto (113.62 pmw) were the three most frequent HNs used by L1 Japanese speakers in this construction. In contrast, the top three HNs in the JFL corpus were koto (195.85 pmw), toki (169.09 pmw), and hito (123.68 pmw). In i-Adjective_{bound_}HNs NMCC realizations, the modifying clause predicates ii, nai, and hoshī were the top three most frequently used across both the JJJ (493.82, 244.73, and 69.92 instances pmw) and the JFL corpora (508.08, 111.11, and 20.27 instances pmw). The HN types were also similar. The top three most frequent HNs were mono (91.77 pmw), sensei (61.18 pmw), and koto (52.44 pmw) in the JJJ corpus and sensei (63.66 pmw), hito (62.85 pmw), and koto (49.47 pmw) in the JFL corpus.

4.3 Simple collexeme analysis of head nouns in NMCCs

The collostructional strength measured as log-likelihood (G²) was computed for each collexeme in the HN slot of the NMCC. In the JJJ corpus, 101 out of the 650 collexemes used by L1 Japanese speakers were found to be significantly attracted to the construction after the Bonferroni correction. Of the 1,246 collexemes used by JFL learners in the JFL corpus, 222 noun collexemes were found to be significantly attracted to the construction. Tables 7 and 8 present the top 20 collexemes in each corpus organized by collostructional strength, that is, their level of attraction to the NMCC. In the L2 NMCC productions, the top 20 significantly attracted collexemes in descending order were koto, toki, tokoro, mono, hito, kanji, sensei, machi, tame, kata, ato, basho, koro, tabemono, tsumori, ko, kodomo, shigoto, hanashi, and keiken. The top 20 significantly attracted collexemes in the L1 Japanese NMCC productions were koto, tokoro, mono, kanji, toki, wake, hito, fū, kata, sensei, ki, koro, hanshi, yastsu, omoide, imi, basho, tame, kioku. Comparison between the corpora indicated substantial overlap (70 %) among these top-20 HNs. Both speaker groups show strong preferences for general common nouns that broadly refer to concrete or abstract entities. Notable differences include the HNs kodomo, machi, tsumori, tabemono, ato, shigoto, and keiken which are among the most significant collexemes in the JFL but not in the JJJ corpus, and the HNs hazu, omoide, ki, wake, fū, imi, and kioku for which the opposite is the case. These differences do not pertain to POS subcategorization but rather the semantic weight of the HNs.

The simple collexeme analysis also identified HNs that are significantly repelled by the NMCC. Only those in the JFL corpus, however, reached statistical significance. Table 9 displays the most significantly repelled collexemes. Of the 131 repelled collexemes in the JFL corpus, 15 reached statistical significance, namely, -go, ni, jū, kane, haha, jibun, issho, mina/minna, ichi, toshi, go, yon/shi, ima, chichi, kinō, tsuki. The analysis also revealed 16 repelled collexemes in the JJJ corpus after the Bonferroni correction, only one of which came close to reaching statistical significance: ima ‘now.’