Translation-based connotation visualization for classical poetic Japanese vocabulary of the Kokin Wakashū ca. 905

2.1 Aspects of Connotation

In a basic definition, denotation is a word’s explicit meaning, whereas connotation encompasses sociocultural and individual affiliations (Chandler 2002, 173–174). Nevertheless, the definition and scope of connotation vary among scholars. The scope pertains to sociolinguistic aspects (e.g., Bloomfield 1933 and Hjelmslev 1969), emotional aspects e.g., Eco 1976), and associative aspects (e.g., Rössler 1979). According to Eco (1976), semantic relationships such as hyponyms, hypernyms, and antonyms also pertain to the realm of connotation. In communication, pragmatic aspects of meaning, a speaker’s attitude and a listener’s value judgment of the perceived speech is also connotation (Mounin 1976, 159–160). They are independent between the speaker and the listener. From a speech without any intended connotation by the speaker, the listener can perceive unintended connotation (Rössler 1979, 101). Moreover, connotation is understood as inseparable from denotation (Stede 1999, 91; Stubbs 2002, 198; Voloshinov 1986, 105) and connotative meanings are inexhaustible (Chandler 2002, 139). Therefore, operationalizing connotation is challenging.

On the other hand, most of the aspects of connotation mentioned above are non-literal, except for some associations like collocates or phraseology as connotation. The non-literal aspect is a vital aspect we can use, as such non-literal elements can become literal when explicated for a better understanding of cross-cultural communication. Denotation and connotation exist at the conceptual level; they are inseparable and thus nonoperational without access to the target speech community to conduct psychological experiments or questionnaires. Conversely, literal and non-literal elements are physically written/unwritten; they are separable and thus operational when explanatory material is available. Hence, we may view non-literal elements as a physical world projection of the concept of connotation to aid operationalization (Table 1). Although the two are not equivalent, non-literal elements can intuitively reflect certain aspects of connotation. Dictionaries, introductory books, philological annotations, and translations can serve as explanatory materials.

Table 1: Relationship between denotation/connotation and literal/non-literal: Dashed lines indicate the instability of the division; solid lines indicate the stable division.

The operationalization of lexical connotation has been attempted within computational linguistics and corpus linguistics. The operationalization mainly focuses on certain aspects such as emotional value, sentiment, valence, impact, and collocates (e.g., Allaway and McKeown 2021, Rashkin et al. 2016, and Stubbs 2002), while the intuitive aspect, the non-literal aspect of connotation, is not widely utilized.

2.2 Dictionary and Translation as Explanatory Materials

Many materials can serve to turn non-literal elements into literal ones. Among the materials, dictionaries and translations can provide a systematic approach to classical Japanese poetic vocabulary, whereas the others typically select certain poems or words to offer specific explanations and therefore are selective. Subsection 2.2 discusses a comprehensive dictionary for classical poetic Japanese (Katagiri 1983) and explains why we need to use additional information in translations as an independent supplement for the dictionary.

Katagiri (1983) includes 830 lexical entries. The dictionary does not explicitly distinguish between connotation and denotation, yet covers a broad spectrum of meaning, encompassing the two. Connotations in the dictionary pertain to collocates, phraseological patterns, and figurative/rhetorical usages. For instance, its description of “sakura-bana” (en. cherry blossoms) incorporates collocates such as “chiru” (en. falling) and associations such as the impermanence of life and the passing of spring (Katagiri 1983, 172–173).

However, dictionary compilers inevitably apply their conscious knowledge in the dictionary compilation. For example, compilers decide which words to include, which examples and meanings to emphasize. Whereas, when actually translating Japanese poetry, relying solely on the dictionary proves insufficient for producing comprehensive translations. As Masao Takeoka, one of the translators of the Kokinshū, noted:

The translation is not solely an introduction or explanation of the original text’s “plot”. It is, after comprehending the author’s way of perception and feeling from all aspects, entirely transforming the perceptions and feelings into the same or as similar as possible expressions that exist in contemporary language. (Takeoka 1976a, 11, translation by the authors)

To provide a comprehensive understanding of the “perceptions and feelings” of poets in translations, beyond consciously aligning corresponding equivalent words, translators must simultaneously incorporate those perceptions and feelings about each poetic word into their translations. These incorporations are rarely covered in the dictionary, as they remain unconscious for compilers until they are translated within actual contexts. Let us have a look at the following example of a parallel text.

In this parallel text, the translation not only renders the two parallel images, long rain, and reverie, of the kakekotoba “nagame” by adding corresponding words “monoomoi no nagame” (en. gaze with reverie), but also uses additional words to highlight the aspiration to “cross the river” for the word “namidagawa” (en. river of tears). Although the dictionary covers rhetorical usages like “nagame”, an important aspect of connotation, understanding the connotation of the dilemma of longing to cross the river to reach a lover, yet unable to, is challenging through the dictionary alone. We, therefore, pay attention to the explanatory addition in translations. The objective of providing an independent supplement to the dictionary necessitates the visualization of the unconscious knowledge exposed by translators. Through a dictionary-independent visualization system, novice readers can query all lexical entries in the Kokinshū, not just the consciously crucial lexical entries and usages pre-selected by compilers.

2.3 Non-Literal Elements in Schramm’s Communication Model

Based on the above two sections, the operationalization of the concept connotation will be based on the operationalization of non-literal elements revealed in the translations. This section introduces a specific way of addressing non-literal elements in parallel texts with hints from Schramm’s (1954) communication model.

Schramm’s model is considered to be more suitable for this study compared to other communication models because it uses fewer concepts and adopts a concept, the field of experience of participants, which is useful for explaining the success or failure of communication (Yamamoto 2005, 26). Stuart Hall’s (2018) encoding/decoding model of communication, which is well-established in the cultural studies community, also includes variables that influence communication. However, it serves different analytical purposes less suited to the current study.

Schramm’s model features five components: source, encoder, signal, decoder, and destination (Figure 1). The source encodes a message that the destination decodes. In human communication, senders encode thoughts into the language (the signal), and receivers decode it (Schramm 1954, 6).

Figure 1: Communication process by Schramm (1954, 4): A communicative process consists of source, encoder, signal, decoder, and destination.

Yamamoto (2005) proposes that the translation and reading behaviors of classical Japanese poetry be considered communication processes. In the process of reading, the poet is the encoder who transfers thoughts into a poem; the reader (novice/expert reader) is the decoder who processes the information in the poem into their understanding; the poem is a signal in the communication process (Figure 2). The translation process follows the reading process: After processing the information of the poem into understanding as a decoder, the expert turns into an encoder who transfers the understanding of the poem into contemporary translations.

Figure 2: Reading poetry as communicative processes based on Yamamoto (2005, 27): (a) reading classical Japanese poetry by a novice reader; (b) reading classical Japanese poetry by an expert; circles indicate the field of experience.

The difference in the reading process between a novice reader and an expert reader lies in the “field of experience” (Schramm 1954, 6) which refers to accumulated experience. When a novice reader attempts to understand a poem as shown in Figure 2a, the novice reader may fail to understand the poem as the novice reader in the twentieth-century shares no field of experience with a poet who lived in the tenth century. Conversely, an expert in classical Japanese poetry, having accumulated relevant knowledge from numerous materials, shares a broader field of experience (Figure 2b). Hence, an expert can decode classical Japanese poetry better than a novice reader.

Sharing a similar field of experience means that most of the information can be understood non-literally; on the other hand, sharing a different field of experience means most information must be explained literally to be understood. Between the two communication processes with distinct fields of experience, the process of translation by experts acts as a bridge (Figure 3). In the translation process, to share the information in the original poem smoothly with novice readers, experts translate literal elements and add non-literal elements into translations.

Figure 3: Non-literal elements in the schema of Schramm’s communication process (reused from Yamamoto 2005, 28): We simplify non-literal elements as the residual between two kinds of signals – the poem and the translation; CT indicates contemporary translations; OP indicates original poems.

On this basis, we can formalize the non-literal elements (Equation 1), where an original poem (OP) is a set consisting of literal elements and a contemporary translation (CT) is a set consisting of elements in OP and additional non-literal elements. Such formalization turns the extraction of non-literal elements into a well-defined technical problem of calculating the complement set.

$\text{Non-literal elements}=\text{CT}\setminus \text{OP}$ (1)

2.4 Summary and Potential Issues

In section 2, we assumed that through non-literal elements revealed in translations, we can visualize a portion of lexical connotation. Moreover, non-literal elements revealed in translations can provide a heuristic for the unconscious aspect of connotation that is not covered by dictionaries. We can extract the non-literal elements by computing the set difference between translations and original poems, guided by the hint from Schramm’s communication model.

On the other hand, we must highlight some issues when utilizing non-literal elements for connotation visualization.

The first is the coverage issue, which is unsolvable because “no inventory of the connotative meanings generated by any sign could ever be complete” (Chandler 2002, 139). That is, through non-literal elements, we cannot visualize all aspects of connotation and cannot reproduce all connotations described in the dictionary (Figure 4).

Figure 4: Relationship between connotation and non-literal elements, and relationship between connotation in translations and dictionaries: (a) non-literal elements can reflect some aspects of connotation, but they can also include non-connotative elements and cannot cover the whole connotation; (b) Connotations revealed in translations cover some of the connotations described in dictionaries and visualize the unconscious aspects not explicitly mentioned in the dictionaries; the dashed line represents an open set; the solid line represents a closed set.

The second is the “impurity” issue. The impurity mainly refers to the inclusion of function words used to facilitate the translations in the extracted non-literal elements (Figure 4a), which represent the syntactic differences between target languages and source languages. In the following sections introducing the methodology, we will explain how we handle these “impurities”.

3.1 Materials

3.1.1 The Kokinshū Text Data

The Kokinshū comprises 1,111 poems. We use the initial 1,000 poems¹, which are classified as tanka to maintain uniformity in poem length.

We use the Kokinshū text data from the Hachidaishu Vocabulary Dataset (Hodošček and Yamamoto 2022).² The dataset provides a TEI format version and a space-delimited format version. For compatibility with the translation data format, we choose the space-delimited format (refer to Figure 5 for details). The space-delimited format version is generated by automatically processing using a domain-specific morphological analysis system for classical Japanese poetry (Yamamoto 2007) and a semantic category code annotation system (Yamamoto 2009). The semantic category codes are based on an old version of the Word List by Semantic Principles (WLSP; Nakano et al. 1994), a collection of words classified and organized by semantic categories. The dataset includes both compound tokens and their constituents, i.e., decomposed simplex tokens, and assigns multiple semantic category codes to each polysemous token.

Figure 5: Space-delimited format of the Hachidaishu Vocabulary Dataset (Yamamoto and Hodošček 2021): A line consists of seven columns separated by spaces. The first column 01:000001:0007 consists of three fields separated by colons: 1) anthology, 2) poem number, and 3) serial ID of the token. The anthology ID 01 indicates the Kokinshū. the second column indicates the type of token: A is a single token; B is a compound token; C is a breakdown of B. A00 indicates a single token; A01 indicates a single token with another meaning; B00 indicates a compound token; B01 indicates a compound token that has another meaning; C00 indicates the first element of the B00/B01.. breakdown; C01 indicates the second element of the B00/B01.. breakdown. the third column BG-02-1527-01-0102: Classification ID based on semantic categories; 4–9th column indicates respectively: a Part-of-Speech number, a form that appears in literary works, a lemma in kanji script, a lemma in kana script, conjugated form in kanji writing form, conjugated form in kana writing form.

We use semantic category codes during processing instead of lemmata. For each compound token, we use the token rather than its decompositions. In cases where a token has multiple semantic category codes, we use the first code. We do not exclude any stop words in the preprocessing since function words in classical Japanese poems can function as content words simultaneously.³

3.2 Implementation A: Misalignment-Based Visualization

Implementation A builds on Chen et al. (2022), which visualizes the non-literal elements with a statistical word alignment model, IBM model 2 (Brown et al. 1993). This method applies the set difference in the form of misalignment extraction.⁵ We train two statistical word alignment models: one trained with parallel texts in source-to-target order (Model A); the other trained with parallel texts in target-to-source order (Model B). From the alignments inferred by Model B, we deduct alignments inferred by both Model A and Model B. The remaining pairs form a misalignment of Model B. In each pair in the misalignment, the target word is not a precise translation for the source word, yet statistically and semantically related to the source word. We view the target word as a non-literal element added for the source word to transmit information that requires verbalizing in the poem to novice readers. As observed in subsection 2.4, non-literal elements can be function words, which do not represent connotation. We omit those function words from the extracted non-literal elements with post-processing. Detailed steps are as follows.

Step A: Training of Word Alignment Models. We apply IBM Model 2 as the word alignment algorithm to the parallel corpus.⁶ We train two models: a source-to-target model (Model A) and a target-to-source model (Model B). We use the IBM Model 2 from NLTK 3.7.0, a natural language processing toolkit, with Python 3.8. The intersection of alignments by Model A and B results in a 69.85% precision while Model B results in a significantly lower precision of 38.25% when iteration is 8, indicating we can extract numerous misaligned patterns.⁷ For the intersection part, the algorithm itself is very open to improvement to guarantee that most of the sure links are ruled out; for the Model B part, the accuracy criterion should be loosened in the implementation to add more misaligned links.

Step B: Extraction of Misalignment. For each parallel text containing the queried word, we apply the word alignment models trained in Step A to the parallel text in the source-to-target direction (Model A) and the target-to-source direction (Model B). For a queried poetic word, there are three types of alignments with words in translations inferred: (a) an aligned translation word inferred by Model A; (b) one or several aligned translation words inferred by Model B; (c) an intersection of the above two, which indicates the potentially precise alignment. Aligned translation words in (b) can contain incorrect but statistically related translation words for the queried poetic word. To obtain misalignment, we subtract (c) from (b). Figure 6 illustrates the procedure.

Figure 6: Procedure of the set difference of aligned results of the poetic word “ominaeshi” 女郎花 (en. golden valerian) in the 226^th Kokinshū poem and its translation by Kaneko (1933) (figure based from Chen et al. 2022): The first row is the aligned results of (a); the second row is the aligned results of (b); the third row is (c), the intersection of (a) and (b); the final row is the complementary set of (a) in (b) and hence the non-literal information.

Step C: Post-Processing to Filter Function Words. While Chen et al. (2022) do not explicitly include this step, the study implied that without the step, the non-literal elements in the results inevitably contain function words. Therefore, Step C removes those functional elements from the extracted non-literal elements, retaining only the connotative elements. The filtering is based on part of speech. Function words that are not nouns, verbs, adjectives, or adverbs are excluded before visualization.

Step D: Network Visualization in Three Phases. We execute steps B and C on each parallel texts and build an aggregate visualization by accumulating misalignments, following a bottom-up strategy. This means that we can deconstruct the visualization, phase by phase (Figure 7): Phase 1 visualizes the non-literal elements from a single parallel text; Phase 2 visualizes the non-literal elements from ten parallel texts of the same poem (a single poem containing the queried word with its ten translation texts); Phase 3 visualizes the non-literal elements from all parallel texts that include the queried word. Phase 3 is the overall aggregate visualization of the non-literal elements for the queried word. The aggregate visualization shows which non-literal elements of the queried poetic word are commonly added by various translators. On the other hand, Phases 1 and 2 depict non-literal elements for the word at the poem level, aiding in contextual checks.

Figure 7: Phase-by-phase misalignment visualization for “ominaeshi” (女郎花, en. golden valerian): (a) the detailed supplementary translation words for golden valerian in a specific parallel text; (b) collective visualization of supplementary translation words by different translators for golden valerian in a specific poem; (c) overall visualization of supplementary translation words for golden valerian in all the poems.

In summary, Implementation A visualizes non-literal elements for poetic words as misaligned semantic additions in translations. Implementation A can trace non-literal elements of a poetic word in detail with its three-phase visualization: It shows from which translation version a non-literal element originates, and from which poem the non-literal element comes. On the other hand, Implementation A is designed to visualize non-literal information with a strict criterion – if there is no explanatory addition in the translations directly for the queried poetic word, the visualization outputs nothing. This does not mean the poetic word has no connotation; rather, it means all ten translators commonly consider that the poetic word is understandable to novice readers in a poem or multiple poems without any explanation. In other words, the connotation of the word is largely shared between poets and contemporary readers. In this case, visualizable connotation via non-literal elements is limited, which can be a drawback. Implementation B in the next section can address this drawback.

Implementation A is available on Github.⁸ The repository provides a dashboard built with Dash 2.7.1 in Python 3.8. This dashboard integrates all the three phases discussed in Implementation A. Upon entering the dashboard, we can first access the query interface to search for words in classical Japanese poetry. The query results in a table composed of parallel texts where the searched word appears. We can select a certain version of the translation (the default version is KNK/Kaneko 1933). By clicking on the word alignment visualization tab, we can see the word alignment outcomes of the selected parallel text (Phase 1 visualization). By clicking on the misalignment network visualization tab, we can see the Phases 2 and 3 visualizations.

3.3 Implementation B: Co-Occurrence Pattern-Based Visualization

Implementation B builds on Yamamoto (2005), which is an information-theoretical-based approach. Unlike Implementation A, the analysis unit of Implementation B is the co-occurrence pattern rather than the word. In Implementation B, both the poem set (OP) and translation set (CT) are sets of co-occurrence patterns. The set difference is, therefore, used to visualize co-occurrence patterns that exist only in CT.

In Implementation B, we use co-occurrence patterns as units for subsequent reasons. Firstly, if we merely calculate the set difference using two bag-of-words, the remaining words in the complementary set for CT cannot keep their minimum contexts in OP. This diminishes the interpretability of the visualized non-literal elements. Secondly, without co-occurrence, we cannot tell whether an extracted non-literal element is added for the queried word. Unlike Implementation A, which associates non-literal elements with queried words as misaligned links, the set difference of two bag-of-words can fail to indicate which elements link to the queried word and how. Thirdly, the co-occurrence-based implementation can tackle the weaknesses of Implementation A: When there is no explanatory addition directly for a queried poetic word, Implementation A may visualize nothing for the word; inversely, Implementation B can still visualize some elements. In the instance of the co-occurrence pattern, for a poetic word p_a, if it shares a co-occurrence relation with a word p_b in both OP and CT, the relation (p_a, p_b) will not be visualized directly for p_a; however, when p_b holds a co-occurrence pattern with an additional element c in translations, p_b can remain as the non-literal element in the visualization for poetic word p_a in the form of co-occurrence pattern (p_b, c). That is, Implementation B can visualize not only co-occurred words in poems (e.g., p_b) but also indirectly related textual additions in translations (e.g., c) as non-literal elements⁹ for the queried word p_a. Such non-literal patterns are not remnants of the direct co-occurrence with the queried word in original poems, but minimum context indirectly related to the queried word that occurs in translations.

Implementation B also has the issue that non-literal elements may contain non-connotative impurity. To tackle this issue, unlike the post-processing in Implementation A, Implementation B filters co-occurrence patterns with a flexible keyness threshold of patterns. The detailed processes are as follows.

Step A: Construction of Two Sets of Co-Occurrence Patterns for Queried Words. Suppose we query a poetic word. We keep each original poem and each translation that contains the queried poetic word. From the retained poems, we collect all the co-occurrence patterns, regardless of the linear distance between two co-occurring words, and form the queried Poem set (OP) of co-occurrence patterns. The same is carried out on the retained translations to form the queried Translation set (CT).

Step B: Co-Occurrence Pattern Keyness Calculation. We calculate the keyness of each co-occurrence pattern in OP and CT relative to the whole poems/translations respectively. The keyness is for filtering function words and unimportant words. The relative keyness is modeled by cw (Equation 2).¹⁰

$cw(t_1,t_2,d)=(1+\log ctf(t_1,t_2,d))\cdot \sqrt{idf(t_1)\cdot idf(t_2)}$ 2

$idf(t)=\log \frac{N}{df(t)}$ 3

where ctf(t₁, t₂, d) is the co-occurrence frequency of words t₁ and t₂ in a document d; df(t) is the document frequency of word t. We here opt to use cw because the cw considers not only the keyness of co-occurrence patterns as a whole but also the keyness of each word in each pattern simultaneously. $\sqrt{idf(t_1)\cdot idf(t_2)}$ emphasizes the keyness of t₁ and t₂ should be high simultaneously for document d; otherwise, (t₁, t₂) may not be a key pattern. By doing so, we can avoid unimportant function words from being included in key patterns. Conversely, general keyword extraction methods for single words may ignore the keyness of each word in patterns, leading to high keyness for patterns like high idf content word with low idf function word.¹¹

Step C: Filtering Co-Occurrence Patterns. We filter co-occurrence patterns where cw is lower than a specific threshold to automatically remove “impurity” before the calculation of set difference. According to Yamamoto and Hodošček (2018), cw is normally distributed; after standardization, cw can sample key patterns consisting only of content words by setting the threshold as one standard deviation of the normal distribution. In this study, however, we do not use the threshold of one standard deviation. To simultaneously reduce the overlap of labels in visualization, we adjusted the threshold flexibly. Because of the different sizes (usually a ratio near 1:3 or 1:4) between the OP and CT, the thresholds for the two sets are set differently. The thresholds for CT should be 3 or 4 times higher than that of OP.

Step D: Set Difference. We subtract the intersection of the two filtered sets of co-occurrence patterns from the filtered CT. The residuals are patterns that saliently occur in translations but not in originals. Figure 8 shows the procedure of the set difference. Details of the example complementary set can be seen in Table 4.¹²

Figure 8: Procedure of set difference for the salient co-occurrence network of “ume” (梅, en. plum) poems and their translations: Co-occurrence patterns are connected by edges; nodes represent words; Square nodes indicate words that exist only in CT; elliptical nodes indicate poetic words.

Step E: Network Visualization. The visualization uses the dot language to visualize important co-occurrence patterns that are added in translations. These co-occurrence patterns form networks and present a global view of additional information and the relation between the queried word and the additional information.

In Implementation B, the complementary set contains two types of non-literal elements: (a) explanatory additions that are added directly for the queried words (e.g., “break off” for “plum” in translations); (b) collocates with the queried words in original poems, which hold explanatory additions in translations (e.g., “woven hat”-”hide” for “plum”, in which “hide” is the explanatory addition for the “woven hat” in translations). Different from Implementation A, non-literal elements visualized by Implementation B include not only words but also direct and indirect relations.

We provide a simple web application to visualize key co-occurrence patterns for classical poetic words.¹³ The web application can also be applied to translations, but translations are currently protected by copyright. Therefore, the set difference of co-occurrence patterns is not publicly available.

3.4 Summary

Implementation A and B are two different strategies to identify non-literal elements based on the same perspective, Schramm’s communication model. We summarize the differences in Table 5.

Implementation A uses words as the unit while Implementation B uses co-occurrence patterns. To identify intersections of OP and CT, Implementation A uses potentially proper word alignment and Implementation B uses semantic category code to determine the semantic equivalence between old and contemporary Japanese. Therefore, Implementation A does not require a prior language-specific semantic annotation but Implementation B requires. To filter function words from non-literal elements, Implementation A uses part-of-speech while Implementation B uses a flexible keyness threshold. The two set difference-based implementations are also different in the level of set difference. That is, Implementation A firstly calculates the set difference at the parallel text level and then aggregates the set difference as a whole; Implementation B, on the other hand, calculates the set difference at the corpus level and directly returns the aggregate set difference. Hence, implementation B cannot provide breakdown visualizations as Implementation A.

4.1 Case-Specific Observations Compared with the Dictionary

“Ume” (en. plum; Figure 9). The translation word “fragrance” (ja. 香/移り香/薫る) was near the hub of the “plum” network generated by Implementation A as well as Implementation B. According to Katagiri (1983, 82–83), praising the fragrance of “plum” became common in classical Japanese poetry after the Kokinshū. The results agreed with the core description of the poetic word “plum”. Moreover, Implementation B presented a potentially significant node, “woven hat”, which is not directly mentioned in the “plum” entry of the dictionary. The dictionary contains an entry “ume no hanagasa” (梅の花笠, en. woven hat of plum blossoms), which often suggests a woven hat stitched by a warbler (Katagiri 1983, 83). Given the connection between “plum” and “woven hat”; “plum” can also retain a link with “warbler”, a frequently utilized pair in classical Japanese poetry and Japanese visual arts. Besides, an explanatory addition “distinction” ties “plum” with “snow” indirectly, from which we might deduce that the cluster portrays a situation where it is difficult to distinguish between “snow” and “plum”. Similarly, in Figure 9b, we can deduce from which classical Japanese poem each cluster of the network originates. For instance, the cluster “snow-break off-distinction-plum-which” could stem from No. 337¹⁵ in the Kokinshū, and the cluster “plum-woven hat-warbler-hide-sew” could stem from the No. 36.¹⁶ Conversely, fig. 9a fails to replicate the context of the original poem, whereas it can display the core explanatory addition (“fragrance”) as a hub for various translation versions more clearly.

Figure 9: Network visualizations of “ume” (梅, en. plum): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

“Ominaeshi” (en. golden valerian; Figure 10). The explanatory “woman” was the hub in the “golden valerian” network formed by Implementation A (Figure 10a). The visualization showed that the flower holds a feminine image due to its “name” (ja. 名) of “woman” (ja. 女/をみな). This outcome aligned with Katagiri (1983, 481–482): During the Heian period (794–1185), the majority of poems concerning golden valerian depicted the flower as “woman”. Figure 10b also included the connotative term “woman” although “woman” was not in a central position in the network. Besides, the hypernym “flower” is frequently added in translations (Figure 10a) and Figure 10b visualized the antonym terms “male” and the “Mountain of Man” (Mt. Otoko, ja. 男山), which are also aspects of connotation in some theories (Eco 1976). Moreover, each cluster in Figure 10b can reflect part of the narratives of corresponding poems, poems No. 226, No. 227, and No. 230, respectively.

Figure 10: Network visualizations of “ominaeshi” (女郎花, en. golden valerian): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

“Kiku” (chrysanthemum; Figure 11). Both visualizations visualize the connotative term “color change” (ja. 色変わり/移る/変わる), which was described in Katagiri (1983, 127–129): During the Heian period, people admired the gradual change in color of white chrysanthemums to red due to the cold weather. According to Katagiri (1983, 128), after the revival of the Chongyang Festival¹⁷ in the fifth year of Emperor Saga’s reign (814), “chrysanthemum” came to be used in many classical Japanese poems. Chrysanthemum was an essential part of the Japanese Chongyang Festival because it was believed to have the effect of prolonging one’s life. Therefore, “chrysanthemum” was often used in classical Japanese poems to convey the sense of longevity. As for the connotative words regarding longevity, Figure 11a only presented “peak (of blossom season)” (ja. 盛り), which is a common addition at the hub; Figure 11b presented “one thousand years” (ja. 千年), “never get old” (ja. 不老), and “peak (of blossom season)” (ja. 盛り/花盛り) while these words are not at an important position in the network (Figure 11b). Besides, Figure 11b presented “hear” (parallel image for “chrysanthemum” as kakekotoba) as a common addition. Nevertheless, the two methods failed to visualize the relationship between the poetic word “chrysanthemum” and the Chongyang Festival, which should be essential background knowledge unfamiliar to contemporary Japanese novice readers. This was because the Chongyang Festival not appearing in the translation text.

Figure 11: Network visualizations of “kiku” (菊, en. chrysanthemum): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

“Sakura” (en. cherry; Figure 12). The dictionary (Katagiri 1983, 172–173) includes “sakura” as a compound item “sakura-bana” (桜花, en. cherry blossom). According to Katagiri (1983), most cherry blossom poems in the Kokinshū are about falling/scattering (ja. 散る) cherry blossoms which are often related to the transience of human life. We could observe connotative words regarding the sense of falling, such as “to be or to lay scattered about” (ja. 散り乱れる), in Figure 12a. On the other hand, Figure 12a indicated that four translators incorporated “fall” (four added general “fall”; one added 散り果てる, en. all falling) into their translations of “cherry” poems. Moreover, the terms “season”, “now”, and “later” (ja. 以上 and 後) were common additions in Figure 12a, which might suggest that “cherry” as a flower could have a strong relationship with time, evoking a mood of lamenting life’s impermanence and the passing of spring (Katagiri 1983, 173). Furthermore, Figure 12b illustrates the robust connection between “cherry” and “mountain” (inclusive of “mountain breeze”, “Yoshino”, “mountain cherry”). According to Katagiri (1983, 456), “cherry” and “Mt. Yoshino” have maintained a significant relationship since the Shin Kokinshū, the eighth anthology of the Hachidaishū, while instances of the duo are scarce in the Kokinshū. From the non-literal associations visualized in Figure 12b, we discern that “mountain” is a vital context for cherry blossoms. This could provide a foundation for the forthcoming shifts in the usage of “cherry”.

Figure 12: Network visualizations of “sakura” (桜, en. cherry): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

“Matsu” (en. pine; Figure 13). The most salient node in the two networks is “to wait”. The results were consistent with Katagiri (1983, 383–384): The word “pine” is a kakekotoba that holds another meaning of “to wait” (ja. 待つ, the same kana character with “pine” in Japanese), which implies “long-awaited”, to celebrate the eternal nature of the pine tree. Poets associated it with the word “one thousand years old” (ja. 千歳) or with “crane” (ja. 鶴) and “wisteria” (ja. 藤), which also symbolize the one-thousand-year longevity (Katagiri 1983, 383–384). We identified these terms in Figure 13b. On the other hand, Figure 13a displayed “change” and “color”, as frequently added non-literal elements, from which we can deduce that translators aim to highlight the longevity of pine because its “color” never “changes”. Nevertheless, both visualizations failed to visualize many descriptions in the dictionary. For example, poets use the poetic word pine as “evergreen pine” (ja. 常盤の松) or “pine green” (ja. 松の緑). “Pine” is also known for “kadomatsu (gate pine)” (ja. 門松),¹⁸ where the gods resided, and “pine wind” (ja. 松風), the wind that blows through the treetops of pine trees (Katagiri 1983, 384). They are rarely explained in translations.

Figure 13: Network visualizations of “matsu” (松, en. pine): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

“Yamabuki” (en. kerria; Figure 14). Figure 14b showed that “kerria” was “reflected” (ja. 映る) by the “river” or “underwater” (ja. 水底/河底). The results agreed with the comment in Katagiri (1983, 439–440): Since “kerria” often blooms near water, poets usually use “kerria” with the verb “to be reflected/to fade” (ja. 移ろう/映ろう). Conversely, Katagiri (1983, 439–440) notes that since the Heian period, “kerria” had become an important symbol of the end of “spring”, and “kerria” often occurs with the nouns “well (noun)” and “frog”. From these perspectives, neither Implementation A nor Implementation B could visualize such connotative information. Such connotative information is typically represented by phraseological patterns that appear in the original texts and are translated as is, without any additional explanation and context in translations.

Figure 14: Network visualizations of “yamabuki” (山吹, en. kerria): Square nodes indicate additions in translations; elliptical nodes indicate poetic words.

4.2 Summary

Based on the case study of poetic flora words, we can summarize that for each flora word, non-literal elements visualized by the two set difference-based Implementations could serve as hints to the connotation of classical poetic Japanese vocabulary.

Implementation A could divide the additions into common ones (hubs) and individual ones (peripheries in networks), and hence visualize essential explanatory additions to reflect the translators’ common understanding of the poets’ field of experience; Implementation B could visualize the main contexts in original poems linked by explanatory additions in translations for the flora words. From the visualization by Implementation B, we could infer part of the narratives stemming from the original poems. The two methods, especially Implementation B, can cover most of the descriptions regarding connotation in the classical poetic Japanese dictionary (Katagiri 1983), while the non-literal visualization strategy could not visualize two types of connotation: (a) explicit phraseological patterns in the original poem, for which minimal explanation was not added in translations; (b) encyclopedic knowledge that was beyond the texts in translations.

Implementation A occasionally visualized domain-general knowledge as a hub, which was often the hypernym of queried words, such as “flower” for “cherry”. Since such words were not function words, we could not exclude them. Although the hypernym and hyponym for the queried words are viewed as connotations in some definitions, for our objectives, hypernyms and hyponyms are not the ideal connotation candidates. This is because these semantic relations are domain-general rather than domain-specific knowledge only for poetic words.

5.1 Characteristics of the Visualizations

Both Implementations A and B are based on Schramm’s communication model and visualize non-literal elements in complementary sets of original poems and translations. However, they take different strategies to perform set difference calculations and hence have different presentations of connotations.

Implementation A visualizes non-literal elements reflecting rhetorical techniques, hypernyms, and core associations; Implementation B visualizes non-literal elements reflecting rhetorical techniques, hyponyms, antonyms, and a wide array of associations (e.g., phraseological pattern, contextual information). The pragmatic aspects (e.g., translators’ value judgment) of connotation were included in the visualizations; whereas, both Implementation A and B cannot visualize sociolinguistic aspects (connotative terms reflecting gender, style, social class, and region of the poets) and emotional aspects (positive/negative/neutral) of connotation. These aspects are also less mentioned in Katagiri’s dictionary, which might be left for further studies using approaches from corpus-based variationist linguistics and stylistics to explore.

For the objective of supplementing the descriptions for poetic language dictionaries, Implementation B could help incorporate minimal contextual information from original poems through clusters in the co-occurrence networks. On the other hand, Implementation A could demonstrate the consistent use of explanatory additions among experts to reflect core connotations. In contrast, a dictionary with selective consciousness may occasionally overlook some dynamic connotations visualized by the methods.

5.2 Contributions

The two visualizations demonstrated how to utilize explanatory additions in the translations of literary texts to visualize the inaccessible part of the connotation of historical, literary languages. We showed practically that not only dictionaries but also translations are feasible resources as a medium to access connotation in a historical, literary language. Employing traditional statistical natural language processing algorithms and information theory-based methods enhanced our interpretation when the historical literary text data is a low resource.

Although our operationalization of connotation is only an imperfect simplification, it demonstrates that removing all the semantically equivalent elements between the parallel texts and visualizing the leftover additional information is a transparent way of approaching the connotation.

The misalignment-based implementation, which does not use language-specific semantic category annotation, can be applied not only to contemporary Japanese translations of classical Japanese poetry but also to translations in other languages. On the other hand, the co-occurrence-based implementation can cover a large part of the connotation described in the dictionary and also provide essential contexts stemming from original poems.

The two implementations may also provide some theoretical considerations as follows.

Firstly, connotation could be considered a relative concept in practice. The misalignment-based Implementation A yields only a minimal number of non-literal elements related to connotation, suggesting that many connotations may still be accessible to contemporary Japanese people, such as domain-general knowledge. Consequently, translators may choose not to add words for this connotation to translations. In other words, the connotation found at the intersection of contemporary and ancient Japanese remains unseen in the non-literal set defined by Implementation A. However, we could regard this imperfect visualization as a dynamic feature of Implementation A, where connotation is seen as a relative concept. The connotative information in one culture is relative to other cultures. Absolute connotation is considered an open set so far; while we do not have to visualize all the elements of the connotation even when the elements are general among different cultures. What kind of information is the essential connotation should be answered relatively through a specific comparison between the source culture and the target culture.

Secondly, most of the lexical connotations lie in the interrelation of words. When there is no direct explanatory addition for a queried poetic word, Implementation A might not provide a visualization for the word (which does not mean the word has no connotation). In contrast, the co-occurrence patterns in Implementation B can help visualize the unseen elements in Implementation A. The explanatory addition may be not directly for the queried word itself, but for the co-occurrence relation that the queried word holds. This reflects that connotation exists not only in the poetic word itself but is also implied in the interrelation between poetic words.

5.3 Limitations

The operationalization and the two implementations leave the following problems:

Firstly, encyclopedic knowledge and sociocultural context could not be extracted. Our methods could never visualize complete narratives and sociocultural contexts beyond texts. For example, both methods could not visualize the connection between the poetic word “chrysanthemum” and the Chongyang Festival, which could be common tacit knowledge among poets during the Heian period. If such knowledge and sociocultural context do not appear in the translation and original texts, we are unable to visualize them.

Secondly, the presumption that novice readers share no field of experience with poets is extreme. In reality, domain-general connotations are shared between ancient Japanese poets and contemporary novice Japanese readers. Therefore, based on our operationalization, we could not visualize non-literal elements regarding these parts of connotation.

Thirdly, the study did not provide a quantitative evaluation of the implementations. We assumed a stance where we had no knowledge about the subject of study as everything was observed only through visualizations. We did not presume any specific hypotheses to be answered from the methods. This stance makes the methods difficult to evaluate. In future works, we should apply the methods to tackle particular issues. For instance, we can scrutinize philological annotations for specific hypotheses of poetic words and subsequently assess them. Although we do not intend to artificially impose arbitrary limits on the elements that do or do not fit into connotation, we may explore various ways of extracting connotations from other sources, namely, automatic extraction from dictionaries, and compare them with the results of the current methods.

Fourthly, the algorithms used in the two implementations are open to improvement. We used IBM model 2 in the first implementation and cw value in the second implementation. However, the two algorithms were not systematically evaluated and were not the only options. We need to provide additional algorithm options in a suite of the workflow and compare among different algorithms in the future.

Despite the limitations, the methods is an attempt to reflect on the fundamental challenges encountered in the studies of meaning in classical Japanese literary languages and how to deal with them in a principled way. The methods themselves are not the ultimate solution to the challenges, but they can be a starting point for further studies.