skip to main content
research-article
Open Access

Analyzing Variations of Everyday Japanese Conversations Based on Semantic Labels of Functional Expressions

Published:10 March 2023Publication History

Skip Abstract Section

Abstract

To achieve effective dialogue processing, the kinds of daily conversations people have must be clarified. Unfortunately, the characteristics of everyday conversations remain insufficiently investigated. In recent years, the Corpus of Everyday Japanese Conversation (CEJC) was developed, which is a large-scale corpus constructed by recording everyday Japanese conversations. In this article, we investigate the linguistic variations of everyday conversations in a multitude of situations using CEJC. We conducted factor analysis of it using the semantic categories of functional expressions that represent such subjective information as modality, thoughts, and communicative intention in addition to various tenses and facts. Our analysis identified seven factors that characterize everyday conversations and suggests that they are expressed by a combination of a dialogue’s purpose (e.g., “Explanation” and “Suggestion”) and its manners (e.g., “Politeness” and “Involvement”). We also analyzed the BTSJ–Japanese natural conversation corpus with transcripts and recordings and the Nagoya University conversational corpus and confirmed the generalizability of these factors.

Skip 1INTRODUCTION Section

1 INTRODUCTION

Dialogue systems require the ability to adapt to conversational situations to effectively engage in everyday conversations. Although recent advances in natural language processing based on deep learning have rapidly improved the naturalness of dialogue systems, they remain incapable of adapting to specific situations. For such systems to adapt to various situations, we first need to identify the situations so that they can recognize them for adaptation. Unfortunately, current investigation into the full nature of such communication is insufficient due to the difficulty of data collection.

Recent years have seen the development of a large-scale Corpus of Everyday Japanese Conversation (CEJC) [26]. Figure 1 shows examples of the type of data in it. This corpus, which contains audio-visual recordings of spontaneous conversations held in a variety of everyday dialogue situations (e.g., chatting with friends at a restaurant or meeting colleagues in an office), is expected to accelerate research related to daily-life conversations. Several studies have already investigated linguistic phenomena using CEJC, including distal demonstratives [11], self-addressed questions [16], benefactive constructions [15], and prosody [53]. Due to these studies, particular conversational phenomena in everyday conversations have gradually been clarified. Many of these investigations are based on case studies, and in a few of them, data-driven analyses have been conducted. For example, Iseki et al. [23] demonstrated that the distribution of dialogue acts depends on a dialogue style. Murai [34] focused on the speech styles of participants and conducted a macroscopic analysis based on the ending particles of utterances. However, it remains unclear how conversations differ from situation to situation.

Fig. 1.

Fig. 1. Examples of dialogue data in CEJC with dialogue situation labels (images were anonymized for publication).

In this study, we identify the common factors that represent the linguistic variations of everyday conversations to explain the differences among those held in various situations. We conducted factor analysis of CEJC using the semantic labels of functional expressions extracted from dialogue transcriptions. These labels represent a wide range of a speaker’s subjective information, such as modality, thoughts, and communicative intentions in addition to grammatical tenses and facts. Since conversations are formed by the exchange of such information, these features provide useful cues for capturing the true nature of daily conversations. We also conducted cross-corpus analyses to confirm the generalizability of the factors obtained from the analysis. For such analyses, we used the BTSJ–Japanese natural conversation corpus with transcripts and recordings (BTSJ) and the Nagoya University conversational corpus (NUCC).

This article is an extended version of our earlier report [9]. Here, we elaborated the interpretation of the factors to more accurately describe the salient feature set and conducted cross-corpus analyses to confirm the generalizability of the factors obtained from our previous analysis.

Skip 2RELATED STUDIES Section

2 RELATED STUDIES

2.1 Macroscopic Analysis of Dialogue Corpus

Various approaches have been taken to analyze the global characteristics of a text corpus, such as factor analysis [4], principal component analysis [6], and cluster analysis [32]. In particular, factor analysis is used to discover the potential factors that influence observed variables. Biber [4] applied it to various texts and identified six underlying dimensions to discriminate speech and written texts: “Involved versus Informational Production,” “Narrative versus Non-Narrative Concerns,” “Explicit versus Situation-Dependent Reference,” “Overt Expression of Persuasion,” “Abstract versus Non-abstract Information,” and “On-Line Informational Elaboration.” Paiva [36] conducted analysis targeted on pharmaceutical leaflets and obtained two factors: “Involvement versus Abstraction” and “Nominal Style, and Explicit Referencing versus Verbal Style, and Pronominalised Referencing.” Macroscopic analysis has investigated many kinds of dialogue, such as online news commentary [14], TV shows [2], and spoken British and American English [20]. Ehret and Taboada [14] compared online news comments and face-to-face conversations based on factor analysis. Al-Surmi [2] employed Biber’s model and clarified the differences between natural conversations and TV sitcoms and soap operas. In addition, Ward [49] investigated the individual interaction styles of speakers using macroscopic analysis with prosodic configurations. Inspired by these studies, in this work, we apply factor analysis to conversations in various situations using the semantic labels of functional expressions as features. The main difference between this study and those of Biber and Paiva is the target text data. Biber used texts of a variety of genres such as press reportage, academic prose, mystery fiction, face-to-face conversation, and interviews to investigate the differences between speech and written texts. The target texts of Paiva’s study were pharmaceutical leaflets. In contrast, our study focuses on transcriptions of everyday conversations of various situations, which we believe will yield the factors that represent variations in dialogues.

2.2 Analysis of a Specific Dialogue Corpus

Dialogue corpora that specialize in particular dialogue styles have been constructed and used for analysis and experimentation. For example, the Switchboard corpus [19] is a large-scale speech corpus that collects conversations between two speakers over the phone. In it, the speakers engage in chats about given topics. The Switchboard corpus provides rich annotations of discourse and prosody for analysis [7]. The AMI meeting corpus [31] and the KyuTech corpus [51, 52] collected meeting dialogues. Another study focused on reference interviews [42], which are conversations that listen to the opinions of others to direct them to a useful source of information. Since the above studies targeted a specific dialogue style, it is unclear how the conversations differ between styles and between situations.

2.3 Analysis of Everyday Conversations

Several studies have also specifically analyzed everyday conversations, which is the target of our study. Eggings and Slade [13] classified spoken language samples into casual and pragmatic conversations and identified several interactional patterns of the former type. Tsutsui [46] categorized the chain structures from which casual conversations are composed and clarified the language form required for each chain from the viewpoint of language education. Biber et al. [5] investigated and categorized discourse units using the cluster analysis of informal conversations. Although these studies shed light on aspects of everyday conversations, the variation of the target situations is limited to casual and informal conversations.

For everyday conversations, the Spoken British National Corpus 2014 (BNC2014) [29] was constructed. Its conversations were recorded in informal settings among friends and family members. This corpus is the successor to BNC1994 and can be used to investigate how British English has changed since the early 2000s. BNC2014’s design that resembles that of CEJC does not contain any information regarding dialogue styles. Various analyses of everyday conversation have been conducted using this corpus. For example, Fuchs [17] investigated the influence of age, gender, social class, and dialect on the use of intensifiers in private conversations and whether their use has changed over the past two decades. Aijmer [1] focused on the intensifiers in present-day English and analyzed the process of undergoing delexicalization and grammaticalization. In addition, much analysis has focused on specific elements of the current usage of British English, such as its expressions of formulaic politeness [10], its use of stative progressives [39], and hedging strategies [25]. In particular, Love et al. [28] compared the frequency of certain grammatical categories (e.g., verbs or personal pronouns) between such situations as recoding locations and activities.

In contrast to these studies using BNC, we conducted a macroscopic analysis for the everyday conversation. Our analysis identifies the factors that discriminate among a wide range of conversations, such as dialogue styles, recording locations, activities, and relationships between participants based on a bottom-up approach using a large-scale corpus of everyday conversations. The interpretation of such extracted linguistic factors will provide perspectives about the characteristics of everyday conversations in various situations.

Skip 3CORPUS OF EVERYDAY JAPANESE CONVERSATIONS Section

3 CORPUS OF EVERYDAY JAPANESE CONVERSATIONS

The CEJC [26] was constructed using video cameras and integrated circuit (IC) recorders to capture conversations embedded in naturally occurring daily activities. We used its pre-released edition, which contains 152 hours of conversations. This edition includes data of everyday conversations recorded by 33 informants selected based on balanced gender and age groups.

The audio data were recorded with IC recorders (Sony ICD-SX734) for individual participants as well as a central IC recorder (Sony ICD-SX1000) placed in the center of the dialogue scenes. A Panasonic HX-A500 was used for video recording outdoors and moving situations, and a spherical camera (Kodak PIXPRO SP360 4K) and two portable video cameras (GoPro Hero3+) were used for the other situations. A single GoPro Hero3+ was employed for many of the recordings. The corpus contains transcriptions with detailed annotations, including speaker labels and the starting and ending times of utterances. It has 427 dialogues. The CEJC contains 1,462 speakers, 400k utterances, and 1,840k tokens. In this study, we also used the dialogue transcriptions for our analyses.

Many kinds of metadata were attached to each record. As labels of dialogue situations, we used a conversational style (Style), the place where the conversation occurred (Place), the kind of activity conducted while talking (Activity), and the relationship between the participants (Relation). Since the number of samples of some labels was small, we combined them. For labels under “Place,” we combined “School” and “Workplace” into “School/Workplace.” The labels related to “Indoor” were integrated into a single category. The labels related to “Facility” (except “Restaurant”) were combined into “Other_facilities.” For the labels under “Activity,” we combined the following: “Leisure activities” and “Leisure activities with transportation,” “Housework” and “Housework with meals,” and “Social participation” and “Social participation with meals.” The labels under “Relation,” were merged into “Social relationships” except for “Family” and “Friends.” If more than one label is given as “Relation,” then we selected the label with the largest interpersonal distance; that is, “Friends” was selected when the labels contain both “Family” and “Friends.” The reason for this selection is that we believed that the conversations are influenced by the participants who have the most distant social relationships. The definitive dialogue situation labels are summarized in Table 1. We also show the correspondence of the original and combined labels in Tables A.1 and A.2 in the appendix.

Table 1.
SituationLabels
StyleMeeting, Discussion, Chat
PlaceCar, School/Workplace, Indoors, Outdoors, Other_facilities, Restaurant, Home
ActivityLeisure activities, Work, Social life, Meals, Social life with meals, Working over meals, Rest, Studying, Housework, Social participation, Transportation, Professional services, Extra-curricular activities, Others
RelationSocial relationships, Friends, Family

Table 1. Dialogue Situation Labels for CEJC Analysis

Skip 4METHODOLOGY Section

4 METHODOLOGY

In this section, we describe the procedure of factor analysis. First, we overview this method and the semantic labels of functional expressions, which are features of the analysis. Then we explain how to interpret the obtained factors.

4.1 Factor Analysis

In factor analysis, common factors, which are aspects shared by multiple observables, are extracted. We extracted the semantic labels of functional expressions and personal pronouns (described in Section 5.1) and used their frequency in dialogues as observed variables. The observed variables are represented by a linear combination of factor scores and factor loadings: (1) \(\begin{equation} \boldsymbol {x}_i = A \boldsymbol {f}_i+D \boldsymbol {\epsilon }_i. \end{equation}\)

Here \(\boldsymbol {x}_i \in \mathbb {R}^{j}\) is the observed variables and \(\boldsymbol {f}_i \in \mathbb {R}^{N}\) represents the factor scores of the \(i\)th sample. \(\boldsymbol {\epsilon }_i \in \mathbb {R}^{j}\) is the unique factor of the \(i\)th sample. \(A \in \mathbb {R}^{j \times N}\) is the factor-loading matrix. \(D \in \mathbb {R}^{j \times N}\) is the unique factor-loading matrix. \(N\) is the number of factors and \(j\) is the feature dimension. The axes of the extracted factors are often rotated to improve the dimension’s interpretability.

This article’s analysis is based on exploratory factor analysis, which means that the number of factors must be set empirically, since no factor structure is assumed in advance. Generally, researchers determine the number of factors by referring to the eigenvalues and the factor loadings obtained from the analysis.

4.2 Features

Feature selection is important for factor analysis. Japanese has many compound particles that are composed of multiple morphemes and work as a single function word. Function words and such compound particles are treated as functional expressions. A functional expression, which is defined as “an expression that has a certain function in a sentence,” can be grouped with similar meanings as a set of synonymous expressions. This grouping is called a semantic label. To assign labels, we used a semantic tagger of functional expressions in Japanese predicative phrases. An inventory of the 89 tagger’s labels used in this article was determined based on the dictionary of “Japanese Expression Sentence Types” [33]. Table 2 shows an example of the labels.

Table 2.
Semantic labels:
topic, reason, possibility, purpose, state, nominalization, meaninglessness, parallel, in addition to, unnecessity, prohibition, inevitability, impossibility, comparison, negated intention, negation, repetition, decision, continuation (from), do a favor for, simultaneity, coordinate, obligation, hearsay, addition, degree, politeness, continuation (toward), target, experience, receive a favor, respect, unexpectedness, conjecture, situation, resultative (teoku-form), restricted coordination, conjunction, subordinate conjunction, endpoint, recipient, spontaneous, advance, after, causative, improbability, restrictive, continuation, continuation (from), trial, excessive, emphasis, permission, contrastive conjunction, contrastive subordination, interrogation, starting point, wish, interjection, completion, completion, invitation, reminiscence, probability, quotation, intention, request
Personal pronouns:
first-person pronoun, second-person pronoun, third-person pronoun, infinitive
  • Semantic labels were taken from a dictionary proposed in Matsuyoshi’s study [30]. Personal pronoun features were adopted from Biber’s study [4].

Table 2. Semantic Labels of Functional Expressions and Personal Pronouns Employed as Features

  • Semantic labels were taken from a dictionary proposed in Matsuyoshi’s study [30]. Personal pronoun features were adopted from Biber’s study [4].

Various kinds of subjective information (e.g., modality, thoughts, and communicative intentions), grammatical tenses, and facts are represented by functional expressions that follow predicates. For example, the sentence

Nani wo yaitan desu ka?

Could you tell me what you have baked?

has functional expressions “ta,” “desu,” and “ka” that respectively denote completion, politeness, and interrogation. Therefore, the tagger assigns multiple labels of completion, politeness, and interrogation to this sentence. In addition to the semantic labels, we also use tags about personal pronouns for feature sets like in a previous study [4].

In our analysis, we first counted the frequency of these labels for each dialogue. Then the frequency vector of the labels was used as observed variables.

4.3 Factor Interpretation

Once the factor loadings are obtained, the factors are interpreted based on factor scores and salient features. Here the factor loading of a feature reflects the extent to which the variation in its frequency correlates with the factor’s overall variation. Therefore, the characteristics of the factor can be explained by extracting the features with a large absolute factor loading. For example, Biber [4] excluded the features of a loading with an absolute value less than 0.30 and defined the remaining features as salient ones. Following his study, we defined salient features as those whose factor loadings have absolute values of 0.30 or higher.

Factor scores are the weights on the latent factors of each dialogue. The higher the factor score is, the higher the degree to which the dialogue is influenced by that factor. We can examine the impact of each factor on the categories by averaging the factor scores belonging to the same category. In this article, we adopt the average factor score for each situation label to discuss the dialogue characteristics in each situation.

Skip 5EXPERIMENT TO INDUCE DIMENSIONS AND RELATIONS OF EVERYDAY CONVERSATIONS Section

5 EXPERIMENT TO INDUCE DIMENSIONS AND RELATIONS OF EVERYDAY CONVERSATIONS

We conducted factor analysis using the features described in the previous section. This section first describes how we conducted it with the features of semantic labels using CEJC as our target corpus. Then we summarize the factors we found and describe the grounds of the interpretation.

5.1 Preparation for Factor Analysis

For the analysis, we extracted the semantic labels of the functional expressions and the personal pronouns from each transcript. Here, the characteristics of situations seem to appear in the overall interactional patterns of the dialogue rather than in a short time span such as several utterance interchanges. Therefore, we conducted a macroscopic analysis dealing with a single dialogue as one sample. The number of samples was 427. We used a semantic tagger of functional expressions in Japanese predicative phrases1 [22, 24] to extract the semantic labels from the dialogues. The personal pronouns were extracted on the basis of the part-of-speech tags determined by a morphological analysis function of the same tagger. Pronouns determined by the morphological analysis were converted to each type of personal pronoun (e.g., first- or second-person pronoun) according to predetermined rules. In our analysis, the frequencies of these labels were used as the features. Labels that appeared fewer than 20 times in the corpus were excluded to avoid noise. Table 2 summarizes the features used for the analysis. The frequency of each label was normalized to a text length of 1,000 words based on an earlier study [4] and standardized to a z-score.

For factor analysis, we used the factor_analyzer2 package of Python. The factor loading matrix was estimated based on the unweighted least squares method, which estimates the factor loading matrix to minimize the errors between actual correlation matrix \(\mathbf {R}\) and correlation matrix \(\mathbf {\Sigma }\) expressed by the model. From Equation (1), \(\mathbf {\Sigma }\) is expressed as \(\Sigma =AA^{\prime }-D^2\) considering the factors to be uncorrelated with each other from the assumption of factor analysis. Then the objective function of the unweighted least squares method is denoted: (2) \(\begin{equation} g_{uls}=\text{tr}[(R-\Sigma)(R-\Sigma)^{\prime }] = \text{tr}[(R-(AA^{\prime }+D^2))^2]. \end{equation}\)

The package used in this article solved the optimized problem with the L-BFGS-B algorithm. The number of iterations was set to 1,000. The L-BFGS-B algorithm [27] is one of the quasi-Newton methods using the BFGS formula [35]. A Newton method iteratively finds solutions to the equations by updating the parameters using the inverse of the Hessian matrix. This is because the positive-definiteness or regularity of the Hessian matrix is not always guaranteed, and there is a case where the Newton method does not converge. However, the quasi-Newton methods use a matrix approximating the Hessian matrix to update the parameters. This approximation matrix must satisfy the secant equation and be positive-definite symmetric; the BFGS formula is the most commonly used matrix that satisfies these conditions.

Then the Varimax rotation was adopted for the extracted factors. This rotation matrix assumes an orthogonal rotation, where all the factors remain uncorrelated with one another. According to a previous study [4], we used the features with factor loadings of absolute values of 0.30 or higher as salient features. The factor scores of each dialogue were calculated by a regression method, which estimates the factor scores to minimize the errors between the actual factor scores and the estimated ones. Estimated factor scores \(\hat{F}\) are calcurated as (3) \(\begin{equation} \hat{F}=XR^{-1}A. \end{equation}\)

Here \(X\) is the matrix of the observed variables whose \(i\)th element is \(\mathbf {x}_i\). To discuss the dialogue variation situation by situation, the obtained factor scores were averaged for each situation label (see Table 1 for the situation labels used for the analysis).

5.2 Determination of Factors

There are various methods for determining the number of factors. For example, one approach is a scree test, which is a scheme for identifying the factor number based on a scree plot: a plot that arranges in descending order the eigenvalues obtained from principal component analysis. In this method, the number of eigenvalues up to the point just before a sudden decrease in value is selected as the factor number. Determining the number of factors from such objective scores as minimum average partial (MAP) and Bayesian information criterion (BIC) is another major approach. In this study, we first selected the candidate factor numbers using these methods and compared the result while changing the number of the factors around the candidate numbers. A definitive number of factors is determined from the number of salient features on each factor and its interpretation.

Figure 2 is a scree plot obtained from the CEJC. The first and second eigenvalues account for the majority of the shared variances. In addition, there are large decreases in the values between the second and third and the seventh and eighth eigenvalues. The scree plot’s trend resembles a previous study [4], where the first and second factors accounted for the greatest proportion of variance, and the clearest breaks in the plot occur between the fourth and fifth factors and the seventh and eighth factors. Biber [4] argued that “Extracting too few factors will result in loss of information, because the constructs underlying the excluded factors will be over looked,” and determined the number of the factors as six based on a comparison with different numbers of factors. We followed Biber’s study and selected seven as one of the candidates for the number of factors.

Fig. 2.

Fig. 2. Scree plot of eigenvalues.

Table 3 shows the changes in various objective scores when the number of factors is changed for the target data. We used the Very Simple Structure criterion (VSS) function in the psych package of R,3 which is a function for determining the appropriate number of factors, to calculate these quantitative values. Descriptions of the objective scores are as follows:

Table 3.
Num. factorsVSS1VSS2MAPBICSABIC
10.140.000.0065\(-\)9643\(-\)1761
20.200.270.0054\(-\)9979\(-\)2322
30.230.310.0052\(-\)9967\(-\)2532
40.220.300.0053\(-\)9813\(-\)2596
50.240.320.0054\(-\)9655\(-\)2654
60.240.340.0054\(-\)9554\(-\)2766
70.250.350.0054\(-\)9391\(-\)2813
80.260.350.0055\(-\)9215\(-\)2843
90.270.370.0057\(-\)8976\(-\)2807
100.260.360.0058\(-\)8726\(-\)2757
  • VSS1 and VSS2 show very simple structure [40] with the complexity one and two. SABIC represents sample size adjusted BIC. Bold represents most appropriate number of factors in respective criterion.

Table 3. Various Criteria for Determining Number of Factors

  • VSS1 and VSS2 show very simple structure [40] with the complexity one and two. SABIC represents sample size adjusted BIC. Bold represents most appropriate number of factors in respective criterion.

  • Very Simple Structure criterion [40]:

    VSS is calculated by comparing the original correlation matrix (\(R\)) with that reproduced by replacing the \(k-v\) smallest elements in each row of the factor matrix (\(S\)). Here, \(v\) is the complexity. The VSS criterion compares the fit of the simplified model to the original correlations, (4) \(\begin{equation} VSS = 1-SS_R^{*}/SS_R. \end{equation}\)

    \(SS_R\) and \(SS_R^{*}\) are the summation of the squared elements of \(R\) and \(R^{*}\). Here, \(R^{*}\) is the residual matrix \(R^{*} = R - SS^{\prime }\). VSS tends to peak at the optimal number of factors.

  • Minimum Average Partial [48]:

    MAP is the minimum average partial when modeling the data by a specified factor number. If the factor number is small or excessively large, then MAP becomes large.

  • Bayesian Information Criterion [43]:

    BIC is one of the information criteria and is calculated as (5) \(\begin{equation} BIC = - 2 \log L + p \ln (N). \end{equation}\)

    Here, \(L\) is the likelihood of the estimated model with p free parameters and \(\ln (N)\) is the logarithmic function of the total sample size \(N\).

  • The sample-size adjusted BIC (SABIC) [44]:

    SABIC is the criterion adjusting the penalty term of BIC, and it is calculated as (6) \(\begin{equation} SABIC = - 2 \log L + p \ln \frac{N+2}{24}. \end{equation}\)

In Table 3, VSS1 and VSS2 show a very simple structure with complexity one and two. SABIC represents the sample size adjusted BIC. The appropriate number of factors varied according to the criterion. VSS1, VSS2, and SABIC suggest eight or nine as the number of the factors, and MAP and BIC suggest two or three. Since we were concerned that we might lose information when the number of factors was two or three, because that number is too few, we determined eight and nine as factor number candidates. In future studies, we plan to conduct evaluations using other factor analysis methods, such as those based on maximum-likelihood, as a way to better examine the goodness of fitting.

We compared the salient features of each factor while changing the number of factors. We compared the factor numbers around the candidate factor number (between five and nine). The kinds of salient features that appeared in the factors themselves were not significantly different in either condition, although their constituents were slightly different. When the number of the factors was five and six, the factors, which correspond to Factor 6 in our definitive solution, tended to disappear. This factor represents a highly interactive, thus important dialogic situation, such as when the participants refer directly to addressees, making five and six factor solutions insufficient. In contrast, the factors that have only two salient features appeared with eight and nine factors. Biber excluded the solutions if the salient features were too small for a meaningful interpretation of the construct underlying a factor [4]. Following Biber’s study again, we determined that seven was the definitive number of factors.

We summarize the salient features of the seven extracted factors in Table 4 and list the average factor scores for all the situation labels in Table 5. In Table 4, bold indicates the factor loadings of the salient features. Bold in Table 5 indicates the top-five scores of the labels for each factor. From the factor loadings and scores, we identified seven factors: “Explanation,” “Request,” “Narrative,” “Politeness,” “Affective,” “Involvement,” and “Suggestion.” The rationale for this interpretation is explained in the following section. Four of the extracted factors (“Explanation,” “Request,” “Narrative,” and “Suggestion”) are assumed to correspond to major dialogue/speech acts [3, 45]. Therefore, we interpreted these four as being related to dialogue purposes. In contrast, the remaining three (“Politeness,” “Affective,” and “Involvement”) are assumed to be factors related to dialogue manners, that is, how to interact with a dialogue partner.

Table 4.
Factor 1Factor 2Factor 3Factor 4Factor 5Factor 6Factor 7
FeatureExplanationRequestNarrativePolitenessAffectiveInvolvementSuggestion
quotation\({\bf 0.68}\)0.030.130.190.18\(-0.18\)\(-0.05\)
subordinate conjunction\({\bf 0.48}\)0.080.030.100.01\(-0.17\)0.13
contrastive conjunction\({\bf 0.47}\)0.000.020.010.29\(-0.17\)\(-0.25\)
reason\({\bf 0.45}\)0.020.28\(-0.02\)0.030.040.02
topic\({\bf 0.40}\)0.06\(-0.00\)0.010.10\(-0.03\)\(-0.11\)
nominalization\({\bf 0.33}\)\(-0.06\)0.090.07\(-0.04\)0.090.06
addition\({\bf 0.32}\)0.02\(-0.03\)\(-0.03\)0.160.01\(-0.04\)
obligation\({\bf 0.32}\)\(-0.06\)0.02\(-0.06\)0.01\(-0.12\)0.11
emphasis\(-0.14\)\({\bf 0.61}\)\(-0.07\)0.08\(-0.03\)0.12\(-0.01\)
request\(-0.07\)\({\bf 0.57}\)\(-0.06\)0.15\(-0.12\)0.060.21
conjunction0.14\({\bf 0.53}\)\(-0.04\)\(-0.06\)0.09\(-0.12\)0.10
degree0.00\({\bf 0.47}\)\(-0.03\)0.040.04\(-0.17\)\(-0.17\)
in addition to0.20\({\bf 0.40}\)\(-0.03\)0.010.08\(-0.11\)\(-0.08\)
continuation (from)0.26\({\bf 0.38}\)\(-0.04\)\(-0.12\)\(-0.03\)\(-0.11\)0.14
excessive\(-0.01\)\({\bf 0.31}\)\(-0.05\)0.000.000.15\(-0.16\)
completion\(-0.25\)\(-0.06\)\({\bf 0.54}\)\(-0.34\)0.050.08\(-0.09\)
continuation0.040.06\({\bf 0.43}\)\(-0.17\)0.270.05\(-0.28\)
after\(-0.06\)\(-0.04\)\({\bf 0.38}\)0.05\(-0.08\)\(-0.12\)\(-0.00\)
negation0.05\(-0.02\)\({\bf 0.37}\)\(-0.14\)0.180.26\(-0.07\)
do a favor for (kureru-form)0.240.02\({\bf 0.36}\)0.060.12\(-0.03\)\(-0.02\)
indefinite\(-0.13\)\(-0.12\)\({\bf 0.34}\)0.01\(-0.05\)0.12\(-0.10\)
probability0.10\(-0.03\)\({\bf 0.33}\)\(-0.02\)0.07\(-0.01\)0.12
politeness\(-0.06\)0.17\(-0.24\)\({\bf 0.67}\)\(-0.19\)\(-0.06\)0.09
interrogation0.00\(-0.10\)\(-0.02\)\({\bf 0.63}\)\(-0.03\)\(-0.23\)\(-0.15\)
do a favor for (ageru-form)0.100.290.01\({\bf 0.46}\)0.00\(-0.01\)0.24
causative0.02\(-0.00\)\(-0.05\)\({\bf 0.42}\)0.030.130.07
possibility0.020.000.00\({\bf 0.30}\)0.03\(-0.12\)0.17
wish0.08\(-0.00\)\(-0.19\)0.04\({\bf 0.62}\)\(-0.11\)0.09
comparison\(-0.06\)0.34\(-0.07\)0.02\({\bf 0.49}\)0.040.14
conjecture0.15\(-0.03\)0.110.07\({\bf 0.48}\)0.02\(-0.25\)
restrictive0.070.010.08\(-0.07\)\({\bf 0.34}\)\(-0.05\)0.01
2nd-person pronouns\(-0.02\)0.01\(-0.11\)\(-0.12\)\(-0.04\)\({\bf 0.55}\)\(-0.08\)
interjection\(-0.23\)\(-0.25\)0.28\(-0.06\)0.36\({\bf 0.48}\)\(-0.14\)
1st-person pronouns0.09\(-0.11\)\(-0.19\)\(-0.06\)\(-0.00\)\({\bf 0.40}\)\(-0.14\)
state\(-0.13\)\(-0.04\)0.08\(-0.11\)\(-0.12\)\({\bf 0.33}\)0.00
invitation\(-0.02\)\(-0.11\)\(-0.07\)0.060.110.12\({\bf 0.43}\)
trial\(-0.13\)0.14\(-0.09\)0.020.24\(-0.11\)\({\bf 0.36}\)
permission0.110.04\(-0.06\)0.100.010.01\({\bf 0.31}\)
decision0.20\(-0.23\)0.050.360.14\(-0.04\)\({\bf -0.47}\)
  • Bold indicates salient features of each factor.

Table 4. Factor Loading of Features

  • Bold indicates salient features of each factor.

Table 5.
Factor 1Factor 2Factor 3Factor 4Factor 5Factor 6Factor 7
SituationLabelExplanationRequestNarrativePolitenessAffectiveInvolvementSuggestion
StyleMeeting\({\bf 0.904}\)0.027\(-0.291\)0.293\(-0.280\)\(-0.373\)0.359
Discussion0.239\({\bf 0.286}\)\(-0.077\)0.257\(-0.117\)\(-0.236\)\({\bf 0.487}\)
Chat\(-0.128\)\(-0.080\)0.041\(-0.091\)0.0510.090\(-0.158\)
PlaceCar\(-0.072\)\(-0.168\)\({\bf 0.678}\)\(-0.102\)0.0570.071\(-0.282\)
School/Workspace0.417\(-0.015\)\(-0.152\)\({\bf 0.428}\)\(-0.049\)\(-0.341\)0.167
Indoors\(-0.196\)\(-0.043\)0.001\(-0.265\)\(-0.194\)\(-0.082\)0.012
Outdoors\(-0.560\)\(-0.155\)\({\bf 0.254}\)\(-0.217\)\({\bf 0.223}\)\(-0.012\)0.179
Other_facilities\(-0.063\)\({\bf 0.429}\)0.061\({\bf 0.742}\)\(-0.062\)\(-0.145\)\(-0.218\)
Restaurant0.1170.000\(-0.203\)\(-0.047\)0.0190.096\(-0.143\)
Home\(-0.166\)\(-0.075\)0.160\(-0.333\)0.094\({\bf 0.239}\)0.090
ActivityLeisure activities\(-0.899\)\({\bf 1.450}\)\(-0.151\)\(-0.007\)0.056\({\bf 0.197}\)\({\bf 0.549}\)
Work\({\bf 0.566}\)\(-0.091\)\(-0.207\)\({\bf 0.467}\)\(-0.119\)\(-0.391\)0.136
Social life0.040\(-0.104\)0.081\(-0.040\)\(-0.009\)\(-0.103\)\(-0.341\)
Meals\(-0.229\)\(-0.089\)\({\bf 0.167}\)\(-0.349\)\(-0.300\)\({\bf 0.317}\)\(-0.106\)
Working over meals\(-0.105\)0.177\(-0.041\)\(-0.317\)\(-0.221\)\({\bf 0.299}\)\(-0.154\)
Social life with meals\(-0.130\)\(-0.084\)\(-0.244\)\(-0.047\)\(-0.008\)0.121\(-0.141\)
Rest\(-0.043\)\(-0.040\)0.084\(-0.341\)\({\bf 0.263}\)\({\bf 0.142}\)\(-0.005\)
Studying\(-0.186\)\({\bf 0.612}\)0.1620.300\({\bf 0.606}\)\(-0.171\)\({\bf 1.099}\)
Housework\(-0.475\)\(-0.027\)0.068\(-0.401\)0.013\(-0.034\)\({\bf 0.588}\)
Social participation\({\bf 0.445}\)0.1220.1470.407\(-0.244\)\(-0.048\)0.347
Transportation\(-0.017\)\(-0.123\)\({\bf 0.632}\)\(-0.124\)0.1440.030\(-0.306\)
Professional services0.1890.076\(-0.131\)\({\bf 1.537}\)\({\bf 0.288}\)\(-0.141\)\(-0.597\)
Extra-curricular activities\({\bf 0.474}\)\(-0.166\)0.0140.202\(-0.193\)\(-0.311\)0.333
Other\({\bf 0.922}\)\({\bf 0.203}\)\(-0.503\)\(-0.181\)\(-0.131\)\(-0.434\)\({\bf 0.996}\)
RelationSocial relationships0.2690.101\(-0.297\)\({\bf 0.493}\)\(-0.125\)\(-0.179\)\(-0.051\)
Friends0.0360.004\(-0.109\)\(-0.045\)\({\bf 0.220}\)\(-0.001\)\(-0.032\)
Family\(-0.242\)\(-0.081\)\({\bf 0.336}\)\(-0.334\)\(-0.121\)0.1380.070
  • Bold indicates top-five label scores for each factor.

Table 5. Average Factor Scores of Each Situation Label

  • Bold indicates top-five label scores for each factor.

Below we precisely explain our interpretation of each factor. All the utterance examples were translated from Japanese. The square brackets show the dialogue index, and within the parentheses, we show the semantic labels extracted from the utterances. The labels described in bold indicate the salient features.

5.3 Factor Interpretation

Factor 1: Explanation

As shown in Table 4, the salient features of Factor 1 contain the semantic labels given to the utterances that state the basis for an opinion, such as quotation, subordinate conjunction, contrastive conjunction, and reason. The following is a typical dialogue example:

[T013_018]

IC01: Ore Yoshinoya, ore wa Yoshinoya ga ichiban uma i to omotte nda ke do.

Hmm, I think Yoshinoya is my favorite restaurant.

(first-person pronoun, quotation, contrastive conjunction)

IC02: Yoshinoya ga ichiban mazu kute.

I think Yoshinoya is the worst.

IC02: Demo yasu i kara kutte ta kedo ne.

Although I used to go there because its food is pretty cheap.

(reason, continuation, completion, contrastive conjunction, interjection)

Another example is a conversation regarding a day’s schedule:

[K006_003b]

IC02: Ohiru gurai ni kaette ki te no n de ku ru ka waka ra n.

I’ll get back around noon, but I don’t know if I’m going drinking.

(continuation (to), negation)

IC02: Uchiage.

At a closing party.

IC02: Ma denwa mo ne kou syocchu de re ru wake demo na i shi.

Ah, I don’t always answer my phone.

(recipient, decision, topic, negation, addition)

IC02: Sorede ashita kurutte itte ta kara ne.

That’s why he said he’d be here tomorrow.

(continuation, completion, reason, interjection)

As shown in these examples, the salient features of this factor tend to appear in conversations that explain the speaker’s preferences and private information. From the factor scores in Table 5, “Meeting” has a higher score in terms of “Style.” The other labels with high scores are related to “Activity,” such as “Work,” “Social participation,” “Extra-curricular activities,” and “Other.” The utterances aimed at explanation are more likely to appear in such situations. Considering the salient features and the situations with high scores, we interpreted this factor as “Explanation.”

Factor 2: Request

The salient features of Factor 2 include emphasis, request, conjunction, degree, in addition to, continuation (from), and excessive. They appear when a speaker requests something from his or her listeners. The following is a typical dialogue example containing a request:

[T011_015]

IC01: Zenzen honn daizi ni shi nai no.

You don’t take very good care of your books.

(negation, decision)

IC04: Hee i.

Alright.

IC02: Harry Potter mo yo n de nee jan.

You haven’t even read Harry Potter, have you?

(continuation, negation, interjection)

IC01: Sou da yo hondana shimatte yo.

Put it back on the shelf.

(hearsay, decision, interjection)

IC02: Katazu Katazuke te yo.

Yes, please put it back.

(request, interjection)

Other salient features frequently appear in conversations for giving instructions. A conjunction is used to connect consecutive events in Japanese. This label frequently appears in the instructions of procedures. The following two examples are utterances from instructors about dyeing clothing and exercising. These examples obtained high scores for this factor:

[K005_021b]

N10A: Jaa Shibotte kudasa i.

Then please squeeze.

(subordinate conjunction, degree, continuation (from), request)

...

N10A: IC05-san to N30A-kun, ja futte.

IC05-san and N30A-kun, next shake it off.

(conjunction)

...

N10A: Mo me ba mo mu hodo iroga haitte i ku.

The more you rub, the more completely the cloth will be dyed.

(subordinate conjunction, degree, continuation (from))

[T017_011b]

IC02: Jaa kondo ashi raku ni shi te kakato no ue ni o shiri sutoon to nose te ryoute mae ni noba shi te n datsuryoku desu ne.

Now, relax your feet, place your buttocks on your heels, stretch your legs out in front of you, and relax.

(conjunction, decision, interjection).

...

IC02: Daraan toshite kudasa i.

Just relax.

(emphasis, request)

...

IC02: Ja te wo go zhibun no hou ni hikiyose te yukkuri to karada wo oko shi te i ki masu.

Next, pull your hands toward you and slowly raise yourself.

(conjunction, continuation (from))

From these examples, since the salient features of this factor accompanied request, they play a role in requests.

From the factor scores, “Other_facilities,” “Leisure activities,” “Studying,” and “Other” obtained high scores. In such situations, conversations frequently occur whose style is “Discussion.” Since conversations like discussions tend to occur under such situations as “Other_facilities” and “Studying,” this factor’s scores were high. In these situations, many utterances are requests. “Leisure activities” obtained a high score, because directive utterances frequently occur in a sports scene. The following is an utterance example:

[T003_019]

IC03: Yoku mi te yo kao wo ugoka shi cha dame da yo.

Look. Keep your head still.

(request, interjection)

Based on the above observations, we named this factor “Request.”

Factor 3: Narrative

In Factor 3, such semantic labels as completion, continuation, after, and negation become salient features. These salient features are similar to the “Narrative versus Non-Narrative Concerns” in an earlier study [4]. This factor’s score increases when a participant concisely relates a series of events, facts, and so on. The following is a typical dialogue example:

[T013_021]

IC02: Gassyuku wo ano mukashi wa suiei to are ga atta n da kedo.

We used to have swimming and skiing camps.

(completion, decision, contrastive conjunction)

IC01: Sou da ne su sukii a sou da yo.

That’s right. Yes, skiing.

(hearsay, decision, interjection)

IC02: Choudo gakuen funsou no zidai da kara nanka tochuu de chuushi ni natta n da na.

But that was during the student protests, so they were canceled.

(decision, reason, completion, interjection)

Participants tend to share recent and past events around them under a situation where the score for this factor is high (e.g., talking with family members). Table 5 shows that the labels with high factor scores are “Car,” “Outdoors,” “Meals,” “Transportation,” and “Family.” Family members tend to share events when traveling, driving, and sitting. Therefore, we interpreted this factor as “Narrative.”

In addition, conversations fueled by external observations of objects in the environment often occur when driving. The following is a typical dialogue example:

[T011_012]

IC02: Are michi kawatta.

Huh? The road’s been widened.

(completion)

IC03: Uun.

Umm.

IC02: A kocchi no michi de seikai datta ne.

Oh, I am going the right way.

(decision, completion, interjection)

Such conversations resemble narrative conversations and increase the score of this factor.

Factor 4: Politeness

Factor 4 is characterized by such salient features as politeness, interrogation, do a favor for (ageru-form), and causative. Its most characteristic feature is politeness. The salient features of this factor (except for politeness) are labels that frequently co-occur with honorific expression in Japanese. The following is a typical dialogue example:

[T009_022]

Z201: Hi i te mora e ma su.

Could you grind it for me?

(do a favor for (ageru-form), politeness)

IC01: Hai, hai, nihyaku guramu de yoroshi idesu ka.

Of course, 200 grams, right?

(interrogation)

The following is another utterance sampled from the same dialogue.

[T009_022]

IC02: Taihen o ma ta se ita shi mashita.

Thank you for your patience.

(causative, politeness)

Thus, these features express politeness on the whole. For the labels under “Relation,” this factor directly reflects interpersonal distance; the scores are higher for “Social relationships,” “Friends,” and “Family,” in that order. The “School/Workplace,” “Other_facilities,” “Work,” and “Professional services” scores were also high. In these situations, since the conversational participants tended to have a business association or a hierarchical relationship, this factor directly reflects the interpersonal relationships among dialogue participants. Based on the above observations, we named this factor “Politeness.”

Factor 5: Affective

In Factor 5, the salient features are the semantic labels that represent relatively ambiguous utterances, such as wish, comparison, and conjecture. The following is a typical dialogue that contains these labels:

[T009_005a]

IC01: Atashi ima ANA no intaan da shi te te sa.

I’m currently applying for an internship at ANA.

(first-person pronouns, continuation, conjunction, interjection)

IC02: ANA tte hikouki?

The airline company?

IC01: Hikouki, Hikouki.

Yes.

IC01: Hayaku denwa ko nai ka na mitai omotte.

I’m hoping to get a call soon.

(wish, comparison)

The wish and comparison labels tend to appear in utterances that convey a speaker’s feeling to the interlocutors in Japanese.

In addition, conjecture appeared in a scene like this example:

[T010_007]

IC01: Boshu kakete nittei tsutae te teema wo tsutae te to.

Call and ask for an application, and give them the schedule and the topic.

(conjunction, continuation)

IC02: Dou deshou?

What do you think?

(conjecture)

IC01: Iya omoshiro i to omoi masu.

It sounds interesting.

(quotation, politeness)

As shown in this example, an utterance that contains this label plays a role in requesting empathy in human–human conversations. Therefore, its salient features are related to conversations during which participants exchange emotions and feelings.

From the factor scores, “Friends” has a high score in terms of “Relation.” In conversations between friends, utterances through which emotions are shared play an important role in maintaining relationships. As for other situations, “Outdoors,” “Rest,” “Studying,” and “Professional services” (e.g., visiting a doctor at the hospital or going to a hairdresser) obtained high scores. All of these situations tend to involve conversations among friends and dialogues between people who care for each other. From these observations, we interpreted this factor as “Affective”4 in a sense similar to “Emotional.”

Note that the score of “Family” on this factor is not as high as that of “Friends.” The reason for this may be that family members already share their feelings to some extent, and thus it is not necessary for them to state their feelings explicitly. In future studies, we plan to investigate in more detail the characteristics of conversations between individuals holding various relationships including family ties and friendship.

Factor 6: Involvement

The salient features of Factor 6 are first- and second-person pronouns, interjection, and state. Among the features, first- and second-person pronouns refer directly to an addressor and an addressee and are thus used frequently in highly interactive discourse. These are the same features as the first factor of a previous study [4], called “Informational versus Involved Production.” The following is a typical dialogue example:

[T003_017]

IC03: Kou maitoshi sa kooras bu no pianisuto ga bando yatte kudasa ru koto ni suru.

Get the chorus club’s pianist to play in the band every year.

(do a favor for (kureru-form), intention)

IC02: Un.

Hmm.

IC04: Sokka yatte kudasa re ba i i ne.

I see. I hope she agrees.

(do a favor for (kureru-form), invitation, interjection)

IC01: Watashi mo sou omotta.

Me, too.

(first-person pronoun, completion)

In the above example, an interjection is appended to the ending particle that requests or expresses agreement with the other party. In addition, state is used for utterances that solicit something in everyday conversations from a person with whom a close relationships is shared, such as a family members.

[T016_002]

IC02 Omae ga ya re yo.

You do it.

(second-person pronoun, state)

IC03 Yatte ya ru yo.

Ok, I will.

(do a favor for (ageru-form), interjection)

These features suggest that this factor represents an intricate conversation.

The factor scores on this axis were high for “Home,” “Leisure activities,” “Meals,” “Work over meals,” and “Rest.” In such situations, intricate conversations tend to occur, such as chats with family members. Considering the salient features and situations that have high scores, we described this factor as “Involvement.”

Factor 7: Suggestion

The salient features of the last factor are invitation, trial, permission, and decision. It represents utterances about making recommendations and gaining permission. The salient features contain labels of trial that express a performance of an action or request an action from someone, such as “try to have someone do something, or try to do something by myself.” The following is an example of a dialogue that contains an invitation:

[T014_011a]

IC04: Ano ike na kattara shicho ate no ankeeto demo i i shi.

Ah, if you can’t go, send a questionnaire to the head of the department.

(negation, subordinate conjunction, addition)

IC04: Sokorahen mo kou soudan shi te mi tara dou kashira.

You might want to talk to him, too.

(trial, invitation, interrogation)

IC03: Naruhodone.

I see.

In contrast, the following is an example of permission:

[K002_014]

IC03: Chotto ake te ake te mi te i i?

Can I open it?

(conjunction, trial, permission)

In everyday conversation, utterances that contain permission have a role that makes some sort of suggestion in a collaborative context. For example, the above example corresponds to “Let’s open it up.” Therefore, this axis expresses a request with almost no coercion. The factor scores were high for “Discussion” in terms of “Style.” In addition, “Leisure activities,” “Studying,” “Housework,” and “Other” had high scores. Conversation about suggestions is an important element of discussions. Similarly, the utterances for such purposes among participants frequently occur in conversations about “Leisure activity,” “Studying,” and “Housework.” Based on the above observations, we named this factor “Suggestion.”

5.4 Characteristics of Each Situation Label

Our analysis has shown that everyday conversations are composed of a combination of dialogue purposes and manners and have seven components. We believe that our interpretation of the results has a certain validity regarding the factor scores, the factor loadings, and the dialogue examples. In this section, we discuss the characteristics of each situation label by referring to Table 5.

For the labels under “Style,” we found that conversations held for explanations were conducted in “Meeting.” In addition, “Discussion” is formed by combining requests and suggestions. These interpretations closely reflect the characteristics of actual conversational styles. In contrast, the “Chat” score is not high on any factor, suggesting that such conversations have the characteristics of all the extracted factors rather than being distinct from them. This result is reasonable based on the definition of “Chat” in Japanese, which is a conversation about miscellaneous matters.

For the labels under “Place” and “Activity,” the conversations differ situation by situation, since the factors that had high scores were different from each other. However, some labels reveal a similar trend in their scores. For example, we found a similar trend in the scores among “Outdoors,” “Indoors,” and “Home” for “Place.” Furthermore, “Work,” “Social participation,” and “Extra-curricular activities” resembled each other for “Activity.” Here chats with family members account for a large proportion of the former group, and discussions between social relationships account for a large proportion of the latter group. These results suggest that dialogue styles and relationships with interlocutors significantly impact the dialogues.

For the labels under “Relation,” we identified the intuitive characteristics of each one. “Social relationships,” “Friend,” and “Family” obtained high scores for “Politeness,” “Affective,” and “Narrative.” Although these techniques have already been introduced to some dialogue systems (e.g., an empathetic dialogue system [38]), our analysis shows that the importance of these factors changes depending on the relationships of the dialogue partners.

5.5 Comparing Factors with Previous Studies

Comparing the factors obtained in this study with those in previous studies [4, 36], some of the axes are common; others are different. One characteristic of our results is that the factors related to the dialogue purpose were extracted, represented by “Explanation,” “Request,” and “Suggestion.” This result indicates that the dialogue purpose plays an important role in everyday conversations. In contrast, Biber extracted the factors related to how to provide information, such as “Involved versus Informational Production” and “On-Line Informational Elaboration,” and those related to expressions, references, and style, such as “Explicit versus Situation-Dependent Reference,” “Overt Expression of Persuasion,” and “Abstract versus Non-abstract Information.” In our study, we also obtained factors that related to how to talk with partners. Among them, the “Politeness” factor, which represents distant social relations, is a characteristic factor of everyday conversations in which people participate in various social relations.

In contrast, Paiva [36] identified two factors, “Involvement versus Abstraction” and “Nominal Style, and Explicit Referencing versus Verbal Style, and Pronominalised Reference.” The first factor has some commonality with our model, which was constructed from everyday conversations (i.e., “Involvement”). Similarly to Biber, Paiva did not identify the factors related to the dialogue purpose, although we did.

From the above, we believe that the present study provides a more comprehensive set of factors to explain everyday conversations than previous studies.

Skip 6CROSS-CORPUS ANALYSES Section

6 CROSS-CORPUS ANALYSES

We conducted cross-corpus analyses to confirm the generalizability of our seven obtained factors. If the scores of the factors exhibit the identical trend as the original corpus (i.e., the CEJC), even in different corpora, then that suggests that the seven factors robustly explain dialogue characteristics. For cross-corpus analyses, we used the BTSJ [47] and the NUCC [18]. These corpora contain natural conversations among Japanese participants in certain situations.

6.1 BTSJ–Japanese Natural Conversation Corpus with Transcripts and Recordings

The BTSJ corpus was constructed to fuel the studies of the pragmatic analysis of Japanese conversations and contains natural conversations between two speakers. Its dialogues, which were recorded without prior planning or design, include telephone conversations, role-play dialogues, and natural conversations. In addition to conversations between Japanese participants, it contains many conversations in contact situations between native and non-native speakers. A wide range of situation labels is assigned to the dialogues, especially in terms of “Relation.” For example, the corpus contains dialogues between new acquaintances, friends, opposite genders, same genders, and a teacher and a student. The total number of dialogues is 446, and their total duration is 112.5 hours. In this study, we used face-to-face dialogues between native Japanese speakers for our analysis. The total number of selected samples was 254. This subset of the corpus contains 508 speakers, 120 utterances, and 1,080k tokens.

Among the metadata appended to the corpus, we used the conversation genre and the relationship between speakers as situation labels, which correspond to “Style” and “Relation” in the situation labels of CEJC. Table 6 shows the situation labels of BTSJ. Those about acquainted speakers were divided into “Friends” and “Acquaintances” based on the dialogue descriptions. “Acquaintances” is appended to the dialogues between a teacher and a student. BTSJ has several labels that are not contained in CEJC: “Debate,” “Tutoring,” and “Invitation” in “Style” and “Acquaintances” and “Strangers” in “Relation.” Table A.3 in the appendix shows the statistics of the labels.

Table 6.
SituationLabels
StyleChat, Debate, Tutoring, Invitation
RelationFriends, Acquaintances, Strangers

Table 6. Dialogue Situation Labels for BTSJ Analysis

6.2 Nagoya University Conversational Corpus

NUCC is a corpus of chat-talks between Japanese speakers. The dialogues were recorded using a tape recorder or a portable mini-disc recorder. Their dialogue durations ranged from 30 minutes to an hour. The recording places are unlimited, and the participants generally selected quiet places. In principle, the number of participants in a conversation is 2, although some dialogues involved three or four people. After the recordings, 129 dialogues were transcribed. Speakers from teenagers to seniors in their 90s participated in the recordings. The dialogue style was “Chat” without topic restrictions. The dialogue data were assigned situational labels regarding places and relationships between the participants. The situation labels of “Place” contain “University” or “Curry shop,” and the labels of “Relation” contain “Junior high school classmates” or “Italian class classmates.” The relationships of the participants were between both close friends, and new acquaintances, between laboratory members, and between older and younger students. NUCC contains 198 speakers, 130k utterances, and 1,170k tokens.

Since NUCC’s situation descriptions may be too detailed, we integrated them into fewer categories. The NUCC dialogues were mainly recorded by graduate students on the university campus. Therefore, all the labels about university facilities were integrated into “School/Workspace.” All the conversations in eating places were assigned to “Restaurant,” such as hamburger shop, sushi restaurant, and coffee shop. The labels about the homes of the participants were combined into “Home.” In terms of “Relation,” we integrated the situation labels into five labels. First, such relationships as classmates and friends were labeled as “Friends.” We grouped the connections between colleagues into “Social relationships.” This group contains hierarchical relationships between older and younger students. “Strangers” is used for conversations during first-time meetings. Conversations involving close relatives, such as parent/child or married couples, were integrated into “Family.” In addition, there are two conversations between “Romantic relationships.” The kinds of NUCC situation labels are shown in Table 7. “Romantic relationships” and “Stranger” in “Relation” are labels that were not included in the CEJC. Table A.4 in the appendix shows the statistics of the NUCC labels.

Table 7.
SituationLabels
PlaceCar, School/Workspace, Outdoors, Restaurant, Home
RelationSocial relationships, Friends, Family, Romantic relationships, Strangers

Table 7. Dialogue Situation Labels for NUCC Analysis

6.3 Conditions for Calculating Factor Scores

We extracted the semantic labels of functional expressions and personal pronouns from the BTSJ and NUCC transcriptions with a semantic tagger of the functional expressions in the Japanese predicative phrases mentioned in Section 5.1. Then we counted the label frequencies. The labels that appeared fewer than 20 times in the CEJC corpus were excluded from the features. The frequency of each label was normalized to a text length of 1,000 words and standardized to a z-score. The average and standard deviations obtained from the CEJC features were used for standardization.

The factor scores in the cross-corpus analyses were calculated using these feature vectors and the factor loading obtained in Section 5. We calculated the factor scores with a regression method. Finally, the factor scores were averaged for each situation.

6.4 Analysis Results

Tables 8 and 9 show the factor scores when adapting the seven factors obtained from the CEJC dialogues to the target corpora. The score in the table is the average for each situation. Bold indicates the top-three label scores for each factor.

Table 8.
Factor 1Factor 2Factor 3Factor 4Factor 5Factor 6Factor 7
SituationLabelExplanationRequestNarrativePolitenessAffectiveInvolvementSuggestion
StyleChat0.199\(-0.181\)\(-0.567\)0.421\(-0.032\)\(-0.438\)\(-0.807\)
Debate\({\bf 1.526}\)\(-0.129\)\(-0.769\)\(-0.274\)\({\bf 0.218}\)\({\bf 0.339}\)\(-0.213\)
Tutoring\({\bf 1.144}\)\(-0.226\)\(-0.731\)\({\bf 0.432}\)\(-1.084\)\(-0.615\)\({\bf 0.010}\)
Invitation0.261\(-0.074\)\(-0.090\)\(-0.202\)\({\bf 1.289}\)\(-0.045\)\(-0.469\)
RelationFriends0.525\(-0.114\)\(-0.326\)\(-0.341\)\({\bf 0.360}\)\({\bf 0.128}\)\(-0.319\)
Acquaintances\({\bf 1.144}\)\(-0.226\)\(-0.731\)\({\bf 0.432}\)\(-1.084\)\(-0.615\)\({\bf 0.010}\)
Strangers0.201\(-0.240\)\(-0.891\)\({\bf 1.108}\)\(-0.345\)\(-0.874\)\(-1.207\)
  • Table shows top-three label scores with positive values by bold to focus on characteristic situations on each factor.

Table 8. Average Factor Scores of Each Situation Label (BTSJ)

  • Table shows top-three label scores with positive values by bold to focus on characteristic situations on each factor.

Table 9.
Factor 1Factor 2Factor 3Factor 4Factor 5Factor 6Factor 7
SituationLabelExplanationRequestNarrativePolitenessAffectiveInvolvementSuggestion
PlaceCar0.6390.2730.452\(-0.084\)\({\bf 0.449}\)\({\bf 0.699}\)\(-0.671\)
School/Workspace0.7040.0600.007\({\bf 0.317}\)\({\bf 0.350}\)\(-0.237\)\(-0.931\)
Outdoors\({\bf 1.212}\)\({\bf 0.584}\)\(-0.167\)\(-0.271\)\(-0.445\)\(-1.096\)\(-0.864\)
Restaurant\({\bf 0.840}\)0.2800.2370.0080.162\(-0.140\)\(-0.818\)
Home0.629\({\bf 0.291}\)\({\bf 0.501}\)\(-0.260\)0.064\({\bf 0.325}\)\(-0.302\)
RelationSocial relationships0.8120.2050.154\({\bf 0.439}\)\(-0.127\)\(-0.420\)\(-0.815\)
Friends0.7060.1770.202\(-0.088\)\({\bf 0.416}\)0.094\(-0.755\)
Family0.442\({\bf 0.446}\)\({\bf 0.758}\)\(-0.397\)\(-0.148\)0.118\(-0.134\)
Romantic relationships0.5370.052\({\bf 0.606}\)\(-0.242\)0.341\({\bf 2.277}\)\(-0.401\)
Strangers\({\bf 1.013}\)0.022\(-0.327\)\({\bf 1.703}\)\(-0.119\)\(-1.093\)\(-1.596\)
  • Bold indicates top-three label scores for each factor. Here we focused on positive values.

Table 9. Average Factor Scores of Each Situation Label (NUCC)

  • Bold indicates top-three label scores for each factor. Here we focused on positive values.

As shown in Table 8, the model obtained from the CEJC shows similar trends in factor scores with the original corpus even in other corpora, indicating that it robustly explains everyday conversations. First, the scores of “Explanation” for “Debate” and “Tutoring” are particularly high in terms of “Style.” Since these dialogues contain a number of utterances about the purpose of explanations, this result is adequate. In addition, the score of “Involvement” was higher in “Debate,” because the participants often expressed their own opinions. Here is an example of an interaction in “Debate”:

[215-16-JFB026-JF099]

JFB026: Daigaku wa nyuugaku ga muzukashi i hou ga yoi noka, sotsugyo ga muzukashi i hou ga yoi noka.

Here’s our topic: “Should universities be more difficult to enter or to graduate from?”

(interrogation)

JF099: Atashi wa, sotsugyo ga muzukashi i hou ga i i to omo i masu.

I think that being more difficult to graduate from is better.

(first-person pronoun, quotation, politeness)

Here the square brackets indicate the dialogue index and the parentheses show the semantic labels extracted from the utterances. A label in bold indicates a salient feature. In contrast, the “Politeness” score was higher for “Tutoring,” because all its dialogues in the BTSJ concerned advise from teachers to students about writing their theses. In terms of “Relation,” “Affective” obtained the high score in “Friends” as well as in the CEJC. The score of “Politeness” was higher for “Acquaintances,” which has similar characteristics to “Social relationships.” Therefore, the seven extracted factors effectively captured the differences of the relationships between participants, even in a different corpus. Unlike CEJC, “Friends” obtained the high score for “Involvement,” because BTSJ contains many debates (among friends) that rarely appeared in CEJC. For “Strangers,” which is not a situation found in CEJC, the “Politeness” score exceeds that of “Acquaintances.” This result well represents the difference of the interpersonal distance between “Strangers” and “Acquaintances.”

The situation labels of NUCC are related to “Place” and “Relation.” In terms of “Place,” the similarities were obtained with CEJC, such as the high score of “Involvement” for “Car” and “Home.” This result shows again that a dialogue’s characteristics are greatly influenced by its style and the relationship between the participants. We also obtained intuitive results, similarly to the CEJC for “Relation.” The “Politeness,” “Affective,” and “Narrative” scores were high for “Social relationships,” “Friends,” and “Family.” These factors basically represent the general characteristics of “Relation.” We also found a satisfactory result for “Stranger,” which is a label that is not found in CEJC. “Stranger” had a higher “Politeness” score than “Social relationships.” This result coincides with the BTSJ result. In addition, interesting results were observed for “Romantic relationships,” although the number of labels of this situation was small. For them the score for “Narrative” was close to that of “Family,” and the score for “Affective” was close to that of “Friends,” suggesting that this group’s interaction has intermediate characteristics between “Family” and “Friends.” Furthermore, this group’s speakers tend to talk about each other; their “Involvement” scores are high.

These results suggest that the seven factors obtained from CEJC have a certain generality for explaining conversation variations. The structures for dialogue purposes and manners expressed by the obtained factors can be applied to other natural conversation corpora, and might be useful for investigating conversation characteristics.

Skip 7SUMMARY AND FUTURE STUDIES Section

7 SUMMARY AND FUTURE STUDIES

We conducted factor analysis using CEJC to identify the characteristics of everyday conversations. We employed the semantic labels of functional expressions as features and identified seven factors that distinguish everyday conversations in various situations. Four of the extracted factors were axes related to a dialogue’s purpose: “Explanation,” “Request,” “Narrative,” and “Suggestion.” Three other factors were found to be related to a dialogue’s manners: “Politeness,” “Affective,” and “Involvement.” Our findings suggest that everyday conversations can be explained by a combination of dialogue purposes and manners.

In addition, this article applied the factors obtained from the analysis of daily conversation to two corpora of natural conversations: BTSJ and NUCC. Our analyses showed that the scores of the seven factors of CEJC had the same trend as the original corpus, even in different corpora in terms of dialogue styles and relationships between participants. Therefore, we confirmed that the factors extracted in this study have a certain generality for explaining the characteristics of natural Japanese conversation.

The findings of this article can be applied for constructing dialogue-based applications. For example, we previously described a dialogue situation recognition method for constructing a dialogue system for everyday conversations [8]. A dialogue system’s design must cover the extracted purposes and manners to participate effectively in everyday human–human conversations. In recent years, BlenderBot [41], which was trained to acquire such multiple skills required for chats as empathy and knowledge-based conversation, has achieved natural response generation. Our analysis suggests that the naturalness of the conversations of such dialogue systems can be improved for everyday situations by training them with dialogues that cover the factors corresponding to each situation. Related to the issue of language models, Hovy and Yang argued that such large-scale language models have limitations in their ability to deal with several social aspects of language [21]. For this reason, many studies have examined various methods to introduce social factors (such as intimacy [37], politeness [12], and empathy [50]) to language generation models. As future work, we intend to examine the relationship between such social factors and our study’s factors in more detail.

In future studies, we plan to collect dialogue data that satisfy the acquired factors and examine the response-generation methods of dialogue systems that engage in everyday conversations. It is also important to investigate the elaboration and cultural dependency of the analysis results. There is a possibility that there are useful features other than the semantic labels of the functional expressions; we will consider other features in the future. Furthermore, since our analysis was specific to Japanese, it is unclear whether the seven factors are language independent. We plan to conduct a cross-cultural analysis in the future.

Footnotes

  1. 1 https://www.rd.ntt/e/research/MD0057.html.

    Footnote
  2. 2 https://factor-analyzer.readthedocs.io/en/latest/factor_analyzer.html.

    Footnote
  3. 3 https://www.rdocumentation.org/packages/psych/versions/2.1.9/topics/VSS.

    Footnote
  4. 4 Although in a previous study we described this factor as “Empathy” [9], we chose a more comprehensive label that represents the salient features for expressing feelings and emotions, such as wish and comparison.

    Footnote

REFERENCES

  1. [1] Aijmer Karin. 2018. “That’s well bad”: Some new intensifiers in spoken British English. In Corpus Approaches to Contemporary British Speech. 6095.Google ScholarGoogle Scholar
  2. [2] Al-Surmi Mansoor. 2012. Authenticity and TV shows: A multidimensional analysis perspective. Tesol Quart. 46, 4 (2012), 671694.Google ScholarGoogle Scholar
  3. [3] Austin John. 1975. How to Do Things with Words. Oxford University Press.Google ScholarGoogle ScholarCross RefCross Ref
  4. [4] Biber Douglas. 1991. Variation across Speech and Writing. Cambridge University Press.Google ScholarGoogle Scholar
  5. [5] Biber Douglas, Egbert Jesse, Keller Daniel, and Wizner Stacey. 2021. Towards a taxonomy of conversational discourse types: An empirical corpus-based analysis. J. Pragmat. 171 (2021), 2035.Google ScholarGoogle ScholarCross RefCross Ref
  6. [6] Burrows John F.. 1986. Modal verbs and moral principles: An aspect of Jane Austen’s style. Liter. Ling. Comput. 1, 1 (1986), 923.Google ScholarGoogle ScholarDigital LibraryDigital Library
  7. [7] Calhoun Sasha, Carletta Jean, Brenier Jason M., Mayo Neil, Jurafsky Dan, Steedman Mark, and Beaver David. 2010. The NXT-format switchboard corpus: A rich resource for investigating the syntax, semantics, pragmatics and prosody of dialogue. Lang. Resourc. Eval. 44, 4 (2010), 387419.Google ScholarGoogle ScholarDigital LibraryDigital Library
  8. [8] Chiba Yuya and Higashinaka Ryuichiro. 2021. Dialogue situation recognition for everyday conversation using multimodal information. In Proceedings of the Conference of the International Speech Communication Association (INTERSPEECH’21). 241245.Google ScholarGoogle ScholarCross RefCross Ref
  9. [9] Chiba Yuya and Higashinaka Ryuichiro. 2021. Variation across everyday conversations: Factor analysis of conversations using semantic categories of functional expressions. In Proceedings of the Annual Meeting of Pacific Asia Conference on Language, Information and Computation (PACLIC’21). 10 pages.Google ScholarGoogle Scholar
  10. [10] Culpeper Jonathan and Gillings Mathew. 2018. Politeness variation in England: A north-south divide? In Corpus Approaches to Contemporary British Speech. 3359.Google ScholarGoogle Scholar
  11. [11] Daiju Saori and Ono Tsuyoshi. 2018. Studying Japanese distal demonstrative are using video corpus. In Proceedings of the International Conference on Language Resources and Evaluation (LREC’18). 14.Google ScholarGoogle Scholar
  12. [12] Danescu-Niculescu-Mizil Cristian, Sudhof Moritz, Jurafsky Dan, Leskovec Jure, and Potts Christopher. 2013. A computational approach to politeness with application to social factors. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL’13). 250259.Google ScholarGoogle Scholar
  13. [13] Eggins Suzanne and Slade Diana. 2005. Analysing Casual Conversation. Equinox Publishing Ltd.Google ScholarGoogle Scholar
  14. [14] Ehret Katharina and Taboada Maite. 2020. Are online news comments like face-to-face conversation?: A multi-dimensional analysis of an emerging register. Regist. Stud. 2, 1 (2020), 136.Google ScholarGoogle ScholarCross RefCross Ref
  15. [15] Endo Tomoko. 2021. The Japanese benefactive-te ageru construction in family and adult interactions. J. Pragmat. 172 (2021), 239253.Google ScholarGoogle ScholarCross RefCross Ref
  16. [16] Endo Tomoko and Yokomori Daisuke. 2020. Self-addressed Questions as Fixed Expressions for Epistemic Stance Marking in Japanese Conversation. Vol. 315. John Benjamins Publishing Company. 203–236.Google ScholarGoogle ScholarCross RefCross Ref
  17. [17] Fuchs Robert. 2017. Do women (still) use more intensifiers than men?: Recent change in the sociolinguistics of intensifiers in British English. Int. J. Corpus Linguist. 22, 3 (2017), 345374.Google ScholarGoogle ScholarCross RefCross Ref
  18. [18] Fujimura Itsuko, Chiba Shoju, and Ohso Mieko. 2012. Lexical and grammatical features of spoken and written Japanese in contrast: Exploring a lexical profiling approach to comparing spoken and written corpora. In Proceedings of the VIIth GSCP International Conference. Speech and Corpora. 393398.Google ScholarGoogle Scholar
  19. [19] Godfrey John J., Holliman Edward C., and McDaniel Jane. 1992. SWITCHBOARD: Telephone speech corpus for research and development. In Proceedings of the International Conference on Acoustics, Speech, & Signal Processing (ICASSP’92), Vol. 1. IEEE Computer Society, 517520.Google ScholarGoogle ScholarCross RefCross Ref
  20. [20] Helt Marie E.. 2014. A multi-dimensional comparison of British and American spoken English. In Variation in English: Multi-dimensional Studies. Routledge, 171183.Google ScholarGoogle Scholar
  21. [21] Hovy Dirk and Yang Diyi. 2021. The importance of modeling social factors of language: Theory and practice. In Proceedings of the Annual Conference of the North American Chapter of the Association for Computational Linguistics (NAACL-HLT’21). 588602.Google ScholarGoogle ScholarCross RefCross Ref
  22. [22] Imamura Kenji, Izumi Tomoko, Kikui Genichiro, and Sato Satoshi. 2011. Semantic label tagger for predicate functional expressions (in Japanese). In Proceedings of the International Conference on Natural Language Processing (NLP’11). 308311.Google ScholarGoogle Scholar
  23. [23] Iseki Yuriko, Kadota Keisuke, and Den Yasuharu. 2019. Characteristics of everyday conversation derived from the analysis of dialog act annotation. In Proceedings of the Oriental International Committee for the Co-ordination and Standardisation of Speech Databases (O-COCOSDA’19). 16.Google ScholarGoogle ScholarCross RefCross Ref
  24. [24] Izumi Tomoko. 2014. Normalization and Similarity Recognition of Complex Predicate Phrases Based on Linguistically-motivated Evidence. Ph.D. Dissertation. Kyoto University.Google ScholarGoogle Scholar
  25. [25] Johansen Stine Hulleberg. 2020. A contrastive approach to the types of hedging strategies used in Norwegian and English informal spoken conversations. Contrast. Pragmat. 2, 1 (2020), 81105.Google ScholarGoogle ScholarCross RefCross Ref
  26. [26] Koiso Hanae, Den Yasuharu, Iseki Yuriko, Kashino Wakako, Kawabata Yoshiko, et al. 2018. Construction of the corpus of everyday Japanese conversation: An interim report. In Proceedings of the International Conference on Language Resources and Evaluation (LREC’19). 42594264.Google ScholarGoogle Scholar
  27. [27] Liu Dong C. and Nocedal Jorge. 1989. On the limited memory BFGS method for large scale optimization. Math. Program. 45, 1 (1989), 503528.Google ScholarGoogle ScholarCross RefCross Ref
  28. [28] Love Robbie, Brezina Vaclav, McEnery Tony, Hawtin Abi, Hardie Andrew, and Dembry Claire. 2019. Functional variation in the spoken BNC2014 and the potential for register analysis. Regist. Stud. 1, 2 (2019), 296317.Google ScholarGoogle Scholar
  29. [29] Love Robbie, Dembry Claire, Hardie Andrew, Brezina Vaclav, and McEnery Tony. 2017. The spoken BNC2014: Designing and building a spoken corpus of everyday conversations. Int. J. Corpus Linguist. 22, 3 (2017), 319344.Google ScholarGoogle Scholar
  30. [30] Matsuyoshi Suguru, Sato Satoshi, and Utsuro Takehito. 2007. A dictionary of Japanese functional expressions with hierarchical organization (in Japanese). J. Nat. Lang. Process. 14, 5 (2007), 123146.Google ScholarGoogle ScholarCross RefCross Ref
  31. [31] McCowan Iain, Carletta Jean, Kraaij Wessel, Ashby Simone, Bourban S, Flynn M, Guillemot M, Hain Thomas, Kadlec J, Karaiskos Vasilis, et al. 2005. The AMI meeting corpus. In Proceedings of the International Conference on Methods and Techniques in Behavioral Research, Vol. 88. 137140.Google ScholarGoogle Scholar
  32. [32] Moisl Hermann. 2015. Cluster Analysis for Corpus Linguistics. Vol. 66. Walter de Gruyter GmbH & Co KG.Google ScholarGoogle ScholarCross RefCross Ref
  33. [33] Morita Yoshiyuki and Matsuki Masae. 1989. Japanese Expression Sentence Types. Aruku.Google ScholarGoogle Scholar
  34. [34] Murai Hajime. 2019. Japanese daily utterance styles: A factor analysis based on balanced corpus. In Proceedings of the Annual Meeting of Pacific Asia Conference on Language, Information and Computation (PACLIC’19). 503510.Google ScholarGoogle Scholar
  35. [35] Nocedal Jorge and Wright Stephen J.. 1999. Numerical Optimization. Springer.Google ScholarGoogle ScholarDigital LibraryDigital Library
  36. [36] Paiva Daniel. 2000. Investigating style in a corpus of pharmaceutical leaflets: Results of a factor analysis. In Proceedings of the Student Workshop of the 38th Annual Meeting of the Association for Computational Linguistics (ACL’00), Vol. 38. 5259.Google ScholarGoogle Scholar
  37. [37] Pei Jiaxin and Jurgens David. 2020. Quantifying intimacy in language. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP’20). 53075326.Google ScholarGoogle Scholar
  38. [38] Rashkin Hannah, Smith Eric Michael, Li Margaret, and Boureau Y-Lan. 2019. Towards empathetic open-domain conversation models: A new benchmark and dataset. In Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL’19). 53705381.Google ScholarGoogle ScholarCross RefCross Ref
  39. [39] Rautionaho Paula and Fuchs Robert. 2021. Recent change in stative progressives: A collostructional investigation of British English in 1994 and 2014. Engl. Lang. Ling. 25, 1 (2021), 3560.Google ScholarGoogle ScholarCross RefCross Ref
  40. [40] Revelle William and Rocklin Thomas. 1979. Very simple structure: An alternative procedure for estimating the optimal number of interpretable factors. Multivar. Behav. Res. 14, 4 (1979), 403414.Google ScholarGoogle ScholarCross RefCross Ref
  41. [41] Roller Stephen, Dinan Emily, Goyal Naman, Ju Da, Williamson Mary, et al. 2021. Recipes for building an open-domain chatbot. In Proceedings of the Annual Meeting of the European chapter of the Association for Computational Linguistics (EACL’21). 300325.Google ScholarGoogle ScholarCross RefCross Ref
  42. [42] Ross Catherine Sheldrick, Nilsen Kirsti, and Radford Marie L.. 2019. Conducting the Reference Interview. American Library Association.Google ScholarGoogle Scholar
  43. [43] Schwarz Gideon. 1978. Estimating the dimension of a model. Ann. Stat. (1978), 461464.Google ScholarGoogle ScholarCross RefCross Ref
  44. [44] Sclove Stanley L.. 1987. Application of model-selection criteria to some problems in multivariate analysis. Psychometrika 52, 3 (1987), 333343.Google ScholarGoogle Scholar
  45. [45] Searle John R. and Searle John Rogers. 1969. Speech Acts: An Essay in the Philosophy of Language. Cambridge University Press.Google ScholarGoogle ScholarCross RefCross Ref
  46. [46] Tsutsui Sayo. 2012. Structural Analysis in Japanese Casual Conversation. Kurosio Publishers. [in Japanese]Google ScholarGoogle Scholar
  47. [47] Usami Mayumi. 2018. BTSJ–Japanese natural conversation corpus with transcripts and recordings (2018). In NINJAL Institute-based Projects: Multiple Approaches to Analyzing the Communication of Japanese Language Learners.Google ScholarGoogle Scholar
  48. [48] Velicer Wayne F.. 1976. Determining the number of components from the matrix of partial correlations. Psychometrika 41, 3 (1976), 321327.Google ScholarGoogle ScholarCross RefCross Ref
  49. [49] Ward Nigel. 2021. Individual interaction styles: Evidence from a spoken chat corpus. In Proceedings of the Annual Conference of the Special Interest Group on Discourse and Dialogue (SIGDIAL’21). 2731.Google ScholarGoogle Scholar
  50. [50] Welivita Anuradha and Pu Pearl. 2020. A taxonomy of empathetic response intents in human social conversations. In Proceedings of the International Conference on Computational Linguistics (COLING’20). 48864899.Google ScholarGoogle ScholarCross RefCross Ref
  51. [51] Yamamura Takashi and Shimada Kazutaka. 2018. Annotation and analysis of extractive summaries for the Kyutech corpus. In Proceedings of the International Conference on Language Resources and Evaluation (LREC’18). 32163220.Google ScholarGoogle Scholar
  52. [52] Yamamura Takashi, Shimada Kazutaka, and Kawahara Shintaro. 2016. The Kyutech corpus and topic segmentation using a combined method. In Proceedings of the Asian Language Resources Conference (ALR’16). 95104.Google ScholarGoogle Scholar
  53. [53] Yuichi Ishimoto and Hanae Koiso. 2019. Prosodic diversity according to relationship among participants in everyday Japanese conversation. Proceedings of the International Symposium on Linguistic Patterns in Spontaneous Speech (LPSS’19). 6266.Google ScholarGoogle Scholar

Index Terms

  1. Analyzing Variations of Everyday Japanese Conversations Based on Semantic Labels of Functional Expressions

    Recommendations

    Comments

    Login options

    Check if you have access through your login credentials or your institution to get full access on this article.

    Sign in

    Full Access

    • Published in

      cover image ACM Transactions on Asian and Low-Resource Language Information Processing
      ACM Transactions on Asian and Low-Resource Language Information Processing  Volume 22, Issue 2
      February 2023
      624 pages
      ISSN:2375-4699
      EISSN:2375-4702
      DOI:10.1145/3572719
      Issue’s Table of Contents

      Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for components of this work owned by others than the author(s) must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. Request permissions from [email protected].

      Publisher

      Association for Computing Machinery

      New York, NY, United States

      Publication History

      • Published: 10 March 2023
      • Online AM: 9 August 2022
      • Accepted: 25 July 2022
      • Revised: 24 July 2022
      • Received: 30 December 2021
      Published in tallip Volume 22, Issue 2

      Permissions

      Request permissions about this article.

      Request Permissions

      Check for updates

      Qualifiers

      • research-article

    PDF Format

    View or Download as a PDF file.

    PDF

    eReader

    View online with eReader.

    eReader

    HTML Format

    View this article in HTML Format .

    View HTML Format
    About Cookies On This Site

    We use cookies to ensure that we give you the best experience on our website.

    Learn more

    Got it!