2.1 Methodical Basis

[6]Nowadays, the concept of distributional semantics is used widely in the realm of computational linguistics. The basic assumption is that the meaning of a word is established according to how much it is used and how often it co-occurs with other words within a specific context. Words and documents are represented in a high-dimensional space; semantic relationships are inferred from the similarities within that space. For the representation of documents, the frequencies (absolute or relative) of the words in each document are stored as matrices of vectors where each word corresponds to a column of the matrix and every document to a row. The cells of the matrix contain co-occurrence frequencies; pure frequencies are often replaced through degrees of statistical association, such as pointwise mutual information or tf-idf (term frequency–inverse document frequency), in order to counteract the Zipf distribution of words.[16] To represent the meanings of words, the same kind of matrix is created, with the target terms forming rows and contextual words forming columns. Such matrices can serve to compute the distances between single words or texts, to compare them to each other, to cluster them into groups, and to visualize them. As a rule, these very large matrices contain thousands of columns and are sparse, i. e. most of their elements are zero. This calls for reduction to a much smaller number of dimensions in order to be appropriate for the computation of distance or similarity matrices. The resulting low dimensional vectors are often referred to as word or document embeddings and are probably the most common practice for semantic representation in natural language processing (NLP). They are related to, but not identical to topic models. The reduction of dimensions is a purely technical requirement and hardly alters the underlying intention.[17]

[7]The choice of a distributional approach for the task at hand is based on our starting assumption, namely the hypothesis that comedies and tragedies – in accordance with the treatment of each of the different themes – can be differentiated by observing word choice and word usage. Simply put, it can be expected that in Calderónian tragedies, terms such as ›honor‹, ›power‹ and ›death‹ strongly co-occur, while the comedies tend to combine words like ›love‹, ›disguise‹ and ›jealousy‹. This is quite obviously an approach that represents an oversimplification – narrative patterns or plot structures, however, cannot be characterized in this manner. At the same time, the wide success of approaches based on frequency and co-occurrence of words and common methods for author recognition demonstrates that such analyses allow for surprisingly deep understandings even of literary texts.

2.2 Data Basis

[8]Beginning with the fourteen tragedies identified by Sullivan, yet another was added to the examined texts, which had apparently remained unknown to him: Saber del bien y del mal.[18] 15 further dramas, which were identified by qualitative research as comedies and which are often called comedias cómicas (or urbanas or palatinas),[19] make up the counterpart to the tragedies in this body of work. The other 82 Calderónian comedias are available as full digital texts in modernized and normalized Spanish.[20] The spoken texts of the dramatis personae were extracted from all 112 plays and collected for analysis; stage instructions or similar additional texts were not included. The 15 tragedies were each marked with a T and a consecutive number, the comedies with a C, and the remaining 82 plays were marked »Test« and also numbered.[21]

2.3 Research Goal

[9]In the absence of suitably large bodies of dramatic works beyond the Spanish-language world, the classification of genre with word or document embeddings is still relatively new.[22] Thus, the goal of our study is to explore various methods and combinations thereof, and to compare the results. We will compare four approaches, which all follow the same general unobserved schemes: 1) pre-filtering of the vocabulary; 2) calculation of document embeddings, and, if applicable, dimension reduction; 3) clustering of embeddings; 4) visualization und evaluation. Our corpus provides us with an excellent basis, as the categories are known in about a quarter of the plays, but not in the remaining dramas. In this manner, we can simultaneously review the quality of the process (on the basis of the known categories) and obtain findings on the yet unclassified dramas. We find this type of methodical comparison to be important, because it is known that the findings from unobserved distributional methods depend heavily on the parametrization of the process.[23]

2.4 Practical Application

[10]All analyses were implemented with the statistics software R. The pre-processing of the texts was mostly carried out using the R package quanteda, as it also enables the exclusion of Spanish stop words, punctuation and numbers, and the conversion of the prepared corpus of texts to be processed in other packages. As was revealed in the course of exploration, only a small number (viz., 308) of Spanish stop words were retained in the quanteda package. One exploration showed that the exclusion of function words from the matrices did not lead to significantly different results, thus the stop word list was considerably expanded manually.[24] Furthermore, the analysis of the different methods employed, in particular the tf-idf statistics, showed that the grouping results were quite negatively affected by names of characters, places, and countries within the texts, also in their adjectivized form, as these elements of speech tend to reflect idiosyncrasies of single pieces rather than stereotypical genre characteristics. These proper names were likewise – primarily through the list of dramatis personae – compiled and removed from the texts; the number of terms to be excluded from the corpus thus rose above 800, additionally to the 308 stop words contained in the quanteda package. As a rule, the frequency of the words in each drama was calculated, subsequently the frequencies were normalized per document. This took place wherever the distance and similarity matrices for grouping were generated. When calculating the similarity between documents using cosine similarity this could be omitted, because they remain constant in relation to the vector lengths. Consistently throughout the analyses, work was done with inflected or conjugated forms of words; a lemmatization or a stemming of these words was not carried out. In this way linguistic information that might help in the classification of literary genres (and with respect to style, authorial signals or diachronic positionality) was preserved.

3.1 Experiment 0

[11]In a first exploration, we applied a well-established method, Skip-gram,[25] to the body of text in order to assess whether word embeddings could tell us something interesting about the text and which word pairs within the entire body of 112 dramas exhibited the highest number of similarities. We reduced the matrix to the 1,000 terms with the highest log-likelihoods and calculated the cosine similarity between all pairs of vectors. Cosine similarity, or more precisely, the cosine of the angle between two vectors, is a widely used measure of similarity which determines to what extent two vectors ›point‹ in the same direction in the high dimensional space. Cosine ranges between 0 and 1, and a high cosine indicates that two terms are found in similar contexts.

[12]Word pairings with a very high cosine similarity value of more than 0.75 are, for instance, »cielo« and »muerte« (heaven, death), »esperanza« and »desdichas« (hope, despair), »poder« and »temor« (power, fear), »poder« and »gusto« (power, taste), »honor« and »alma« (honor, soul) or »alma« and »muerte« (soul, death). One of the highest cosine similarity values, at 0.96, showed that the word pairing »honor« and »muerte« – honor and death – can be determined as a major theme throughout the entire body of work. Indeed, these first results proved to be surprisingly clear, in that, by using the Skip-gram algorithm, central themes in the Calderónian comedias could be identified, even when they deal with the intersection of social conventions (honor) and individuality (taste, soul, fear, social or actual death).

[13]Conversely, word pairings like »honor« and »poder« (honor, power; 0.58), »amores« and »agravios« (love, infidelity, each in plural form; 0.69), »gracia« and »corte« (grace, court; 0.63) or »gracia« and »culpa« (grace, guilt; 0.60) showed lesser cosine similarity values. Cosine similarity values under 0.5 exhibit only weakly developed commonalities in the contexts; this could be observed for the word pairings »amar« and »honra« (loving, reputation), »muere« and »sepulcro« (he / she / it dies, grave), »muerte« and »engaño« (death, deceit), »mueran« and »suerte« (they may die, fate), »amores« and »honra« (love, reputation) and also »mentira« and »gracia« (lie, grace). First and foremost, it is apparent that the central themes in Calderón’s works (»Amor, honor y poder«[26] – love, honor, and power) do not necessarily have to be interconnected with one another. This can be attributed to the fact that comedies and tragedies can be distinguished from each other through differing combinations of these terms. It is to be expected that the combination »honor« and »poder« is more characteristic of tragedies, and the combination »amar« and »honra« is more characteristic for comedies, but not for the entire body of work. We will come back to this point later.

3.2 Experiment 1

[14]With the first experiment, our goal was to be able to explore the validity of the document embeddings. We take advantage of the known (or: labeled) tragedies and comedies to evaluate our document clusterings as follows, in the spirit of cluster purity[27] analysis: we assign each cluster to the class that the majority of documents with known affiliation belongs to. We then consider the other known classes of documents in this cluster, and compute purity, that is, the degree of agreement between these classes and the majority class, as a measure of success of our clustering. Our setup has the additional aspect that our data set includes documents for which the ›true‹ class is unknown. Since purity only considers documents with known classes, this makes the measure hard to interpret for clusters that consist predominantly or entirely of such documents. For such clusters — which we call underdetermined — we refrain from discussing purity in detail. After carrying out the preprocessing steps described above, we explored the following four methods: 1) Reduction of the matrix through the deletion of words according to their frequency and appearance within the texts; calculating the distance matrix according to relative frequencies, clustering with the Ward.D2 algorithm[28] based on the Euclidian distance. 2) Reduction of the matrix through the deletion of sparse terms which only appear in a few documents, calculation of the distance matrix based on relative frequencies, clustering based on the Euclidian distance with the Ward.D2 distance algorithm. 3) Part-of-speech tagging in each of the dramas, extraction of verbs, nouns and adjectives, calculation of the cosine similarity values between the documents, calculation of the distance matrix, clustering with the Ward.D2 distance algorithm. 4) Calculation of the tf-idf statistics, calculation of the cosine similarity values between the documents, calculation of the distance matrix and clustering with the Ward.D2 distance algorithm. We discuss the results of each method.

[15]The first method represented a conservative approach: only the 1,094 words with a frequency > 120 and appearing in at least half of the documents were included. The document word matrix was filled with mere frequencies; no dimension reduction was carried out. The grouping was carried out through a clustering with the Ward.D2 distance algorithm. Figure 1 shows the resulting dendrogram. Recall that among the documents that form the leaf nodes of the dendrogram, some are known as comedies (CXX), some as tragedies (TXX), but most are unknown regarding their status (»Test«).

Fig. 1: Ward.D2 clustering of 112 Calderónian Comedias. [Lehmann 2022]

[16]Read from left to right, the first cluster represents a pure tragedy cluster which includes 29 dramas; 10 of these had already been characterized as tragedies. The third cluster from the left side depicts a pure comedy cluster; here 22 dramas are included, of which 10 had already been classified as tragedies. The two additional clusters must be described either as undefined or mixed clusters, as they either contain only 1 comedy (second cluster from the left, comprising 39 dramas) and therefore cannot be described as pure, or 4 comedies and 5 tragedies (the cluster to the right, comprising 22 dramas). Together, these two clusters contain more than half of the plays, namely 61 works. We conclude that with regard to the main research question, this approach does not appear to be especially effective, as only 20 of the 30 previously marked dramas (or 67%) were assigned in a clear fashion, while the remaining 10 comedies and tragedies mutually appeared in the clusters. However, the still relatively high dimensionality of the document embeddings makes a failure analysis challenging.

[17]The goal of the second process is to create a low dimensional representation that is easier to interpret, in order to gain more insight into the distribution of the two genres. First, only terms which appear in at least 80% of all of the documents (i. e. in at least 90 plays) are retained; in other words, the sparsity is limited to 20%. This reduces the number of terms to a more compact total of 496. Again, a frequency-based word-document matrix is established and normalized, whereby the frequency of each of the remaining terms in each drama is divided by the sum of frequencies of all the words in the text. Finally, a distance matrix is established, based upon the Euclidian distance, and again, clustering is conducted using the Ward.D2 distance algorithm.

Fig. 2: Ward.D2 clustering of 112 Calderónian Comedias. Euclidian distance on the basis of a sparsity of 20%. [Lehmann 2022]

[18]The dendrogram illustrates three clusters: In the first cluster to the left, all 15 comedies and 16 further dramas appear. The cluster on the right contains 14 tragedies and, likewise, 33 dramas of unknown classification. The cluster in the middle is mixed; it contains 1 tragedy (T4: El mayor monstruo del mundo) and 33 additional dramas of unknown classification. Through this process, which only deals with 496 words, 29 of 30 classified dramas, or 97%, were correctly assigned.[29]

[19]Both of these automatic procedures, in which the fundamental matrices are reduced on the basis of word frequencies, establish a transitional zone between tragedy and comedy. This observation presents us with the question of whether it would be more appropriate, in light of distributional semantics, to consider classifications like ›tragedy‹ and ›comedy‹ as poles between which gradual differences appear, showing the resulting overlap in regard to the applied word selection. In the matter of Calderónian dramas, this seems quite sensible, as themes such as ›honor‹ and ›power‹ can just as well be included in comedic plots as in those of the famous honor tragedies.

[20]Comedies may also present serious subjects in a lighthearted, entertaining manner. For example, power struggles between royal families can be indirectly alluded to within the framework of a mythological play; the allegory would have been quite understandable for the court audience at the time.[30]

[21]One possible fundamental critique on simple document embedding methods, like those we have observed thus far, is the total absence of linguistic structure. For this reason, we made the decision to subject all of the dramas to part-of-speech tagging, including only verbs, nouns and adjectives from each play in the corpus for clustering.[31] For testing the third procedure, therefore, a second corpus is established, in which each of the drama texts include only verbs, nouns and adjectives in their basic forms. All proper names are once more filtered out of the matrix created for this purpose – they had been falsely recognized as adjectives – and subsequently a calculation is made, based on the non-normalized frequencies of the cosine similarities. This similarity matrix is converted to a distance matrix and, once again, clustered with the Ward.D2 algorithm. The results are depicted in a dendrogram.

Fig. 3: Ward.D2 clustering of 112 Calderónian Comedias. Cosine similarity based on verbs, nouns and adjectives. [Lehmann 2022]

[22]The first cluster to the left, which might be identified as a comedy cluster, contains 14 comedies, 5 tragedies (T1: A secreto agravio, secreta venganza; T2: El alcalde de Zalamea; T5: El médico de su honra;[32] T6: El pintor de su deshonra; T13: Las tres justicias en una) and 18 additional plays of unknown classification. The cluster to the right is mostly a tragedy cluster, because it contains 10 tragedies and 49 additional plays, but also 1 comedy (C3: El encanto sin encanto). In the middle between these two categories is an undefined cluster, containing 15 plays marked with »Test«. With regard to the plays identified thus far as tragedies and comedies, 80% of these dramas were correctly clustered; however, this result applies only if clusters are identified by the majority of previously identified dramas.[33]

[23]Taking into consideration the previously tested methods, it seems advisable to focus on every term that carries meaning and thus leads to a differentiation between the categories. The fourth method we tried was based on the tf-idf statistics, thus underlying a measure of association commonly used in text mining, whereby terms can be evaluated for their significance within a document or body of work. With the tf-idf statistics, the weight of each term per document is calculated; the term frequency (tf) is multiplied by the inverse document frequency (idf). The latter depends not on individual documents, but rather on the total number of all documents in the corpus. In this way, the tf-idf statistics considers the relative significance of words which appear frequently in the corpus to determine how relevant the term is for a document within the corpus under study. Once more, the proper names are removed, the cosine similarity for the vectors is calculated, the similarity matrix is converted into a distance matrix and clustering is carried out with a Ward.D2 algorithm. The results are depicted in a dendrogram.

Fig. 4: Ward.D2 clustering of 112 Calderónian Comedias. Cosine similarity on the basis tf-idf values. [Lehmann 2022]

[24]This image shows three clusters: The first one to the left can best be described as a comedy cluster. In addition to all of the 15 comedies, however, it also contains 5 tragedies, exactly the same five ones as in the part-of-speech based analysis conducted previously (T1, T2, T5, T6, T13), as well as 20 other dramas. The cluster on the right, with 8 tragedies and 51 further plays, can be considered a tragedy cluster. The smallest one in the middle is hard to define, since it is only weakly determined and thus cannot be understood as pure; it contains only 2 dramas clearly identified as tragedies and eleven others of unknown classification. In comparison with the dramas already identified as tragedies or comedies, this result shows that 8 of 15 tragedies and, respectively, all comedies have been clustered correctly; this correlates to a recognition rate of 76%.[34] Compared against the models considered above, this recognition rate seems to be satisfactory.

[25]The four methods explored here differ by the choice of data as well as by the choice of distance or similarity metrics. Three of the four generated robust to very good results. The process of employing the strongest matrix reduction produced the best findings. However, only one approach yielded a clustering result that would arguably approximate the classification of researchers applying qualitative analyses.

3.3 Experiment 2

[26]In a second experiment, we assess to what extent the document clusterings we found in the first experiment were based on word choice or word use being consistent with the two genres. To do so, we analyze the word lists upon which the clusters found by the four methods were based. In addition, we calculate the log-likelihood distribution over the vocabulary for the sets of (predicted) comedies and tragedies of each method. This approach determines the 200 words with the highest log-likelihood values for each genre, and these lists can be compared across methods (contrastive vocabulary analysis with word embeddings).

[27]Recall that the first procedure in experiment 1 (Ward.D2 clustering based on the Euclidian distance between normalized word frequencies) created a clustering in which only the first and the fourth clusters could be clearly assessed as comedy or, relatively, tragedy clusters. For both of these clusters, the probability margin for each word is evaluated based on the previously established matrix, and the 15 terms with the highest probability margin for each were investigated. These 15 selected terms for both comedy and tragedy clusters with the highest probability margins give an impression of the cluster formation. For the comedy cluster, the terms »don«, »casa«, »calle«, »papel«, »caballero«, »puerta«, »dama«, »padre«, »hermano«, »saber«, »cuarto«, »amigo«, »hombre«, »sé«, and »señora« (esquire, house, street, paper, knight, door, lady, father, brother, knowledge, room, friend, man, I know, and madam) appeared. Very interesting is the word »papel«, since it points to the paper or card fanning the intrigue; however, beyond this term, the word list does not seem to be significantly distinctive of comedies. By contrast, for the tragedy cluster, the words »rey«, »muerte«, »dios«, »cielo«, »hoy«, »vida«, »sol«, »valor«, »mar«, »tierra«, »gran«, »rigor«, »mundo«, »quiero«, and »poder« (king, death, God, heaven, today, life, sun, value / valor, sea, earth, grand, severity, world, I want, and power) were especially frequent. At any rate, people of high social standing, death, God, valor and power stand out as being characteristic terms relating to these storylines.

[28]The 496 words selected for their sparsity of 20% enable a preview of terms which carry a strong distinction with regard to the separation of comedies and tragedies. For the comedy cluster, meaningful terms like »don«, »casa«, »dama«, »calle«, »puerta«, »sé«, »señor«, »caballero«, »bien«, »cuarto«, »papel«, »señora«, »saber«, »amigo«, and »celos« (esquire, house, lady, street, door, I know, lord, knight, good, room, paper, madam, knowledge, friend, and zeal) are present. For the tragedy cluster, words like »rey«, »señor«, »dios«, »hoy«, »muerte«, »cielo«, »sol«, »quiero«, »rigor«, »mundo«, »gran«, »valor«, »alma«, »viento«, and »sangre« (king, lord, God, today, death, heaven, sun, I want, severity, world, grand, value / valor, soul, wind, and blood) appear. At first glance, the high degree of consistency of both lists of words from the first and second procedures may come as a surprise. Then again, it appears that the high degree of purity in the clustering of the second procedure quite obviously depends upon the condensed and precise selection of distinct terms.

[29]With regard to the third procedure – based upon a part-of-speech tagged corpus – the most frequent words found in the clusters in the underlying matrix illustrate why it does not lead to compelling results: Not surprisingly, the most frequent words here are the verbs »ser« (to be) and »haber« (to have), followed by a list of much less frequent additional verbs, like »ver«, »decir«, »estar«, »dar«, »poder«, »saber«, »hacer«, »tener«, »ir«, »querer«, »venir« (seeing, saying, being, giving, being able, knowing, doing, having, going, wanting, and coming). This is then followed by a list of nouns, like »señor«, »vida«, »cielo« or »don« (lord, life, heaven or esquire). In light of the fact that these frequently used words seem to have little ability to distinguish between comedies and tragedies, the results of the clustering can be described as rather poor.

[30]In the fourth procedure – based on the tf-idf matrix – an approach analogous to methods 1 and 2 is applied. The 15 terms that show the highest probability margin within the comedy cluster are: »don«, »doña«, »tapada«, »hermana«, »calle«, »hermano«, »coche«, »amiga«, »anoche«, »papel«, »cuarto«, »aposento«, »reja«, and »casa« (esquire, lady, veil, sister, street, brother, carriage, friend, last night, paper, room, chamber, grid, and house). In the tragedy cluster, terms such as »arma«, »dioses«, »cristianos«, »templo«, »montes«, »cueva«, »ciencias«, »cruz«, »muro«, »reino«, »pastor«, »rey«, »cristiano«, »cajas«, and »guerra« (arms, gods, Christians, temple, mountains, cave, sciences, cross, wall, kingdom, pastor, king, Christian, crates, and war) are characteristic. While the frequent terms selected for the comedy cluster seem, for the most part, to be less discriminating, save for the typical allusions to veiling and masking or intrigue through forgery, the terms relating to tragedy reflect, at least, military and Christian themes as well as the aristocratic descent of the protagonists.

[31]An open question at this point is how robust these methods are. Thus, in the next step, we test the word lists created in the steps above and base them on a larger body of works. We extend our data basis to clusters, expand the body of plays identified as comedies or tragedies and create two somewhat larger subgroups. From the dramas hitherto marked as »Test«, we choose 16 which were unanimously clustered as being ›tragedy‹ by all four procedures, as well as ten which were unanimously clustered as ›comedy‹. For the comedies, we corroborated this classification on the basis of secondary literature;[35] moreover, all of these dramas were included in the collection of comedies by the editors of the Aguilar edition. In this manner, we generate two new subgroups, one for tragedies, containing 31 plays, and one for comedies, containing 25 plays.[36] Both of these subgroups are converted into matrices using the prevalent preprocessing techniques, whereby all of the terms found in less than four of the plays are filtered out. For the remaining words, the 200 most informative for each subgroup are identified for inclusion, using the log-likelihood function, with which discriminative terms can be found. The comparison of the results for each subgroup shows that only 70 terms appear in both lists, while 130 terms for each (almost exactly two-thirds) are discriminative for either the tragedy or the comedy subgroup.

[32]The analysis of these 130 discriminative terms for each subgroup proves to be very revealing. In the case of the comedies, we discover references to certain themes (»ama«, »amiga«, »carta«, »celoso«, »desdichas«, »desengaño«, »escondido«, »favor«, »joyas«, »juego«, »máscara«, »papeles«, »secreto«, »tapada«, »vestido« – mistress, girlfriend, letter, jealous, misfortune, disappointment, veiled, favor, jewelry, game, mask, papers, secret, hidden / stashed, disguise), typical indications relating to the mythological background of the comedies (»astrólogo«, »duende«, »forastero«, »jardines«, »ninfas« – astrologer, elf / gnome, foreigner, gardens, nymphs) and also the appearance of some rather surprising terms (like »enemigo«, »pendencia«, »razón« or »saber« – enemy, brawl, reason or knowledge).

[33]By contrast, among the tragedies we find references to the (mostly high) standing of the characters (»convento«, »corona«, »emperador«, »esclavo«, »infanta«, »infante«, »majestad«, »reina«, »reinar«, »reino«, »rey«, »tirano«, »villano« – cloister, crown, emperor, slave, infant, infanta, highness, queen, ruling, kingdom, king, tyrant or villain), the contents of the plot (»cristo«, »cruz«, »desdichado«, »divina«, esperanza, »gloria«, »laurel«, »lealtad«, »libertad«, »morir«, »poder«, »salud«, »sangre«, »traíción«, »triste«, »triunfo«, »venganza«, »victoria« – Christ, cross, misery, divine, hope, fame, laurel, devotion, freedom, dying, power, health, blood, treason, sad, triumph or revenge, victory) and a few surprises as well (»ciencias«, »enamorado«, »sueño« – sciences, enamored or dream). Altogether, the word lists determined log-likelihoods in the two subgroups outline the contents of the comedies and tragedies much more precisely than the word lists based on each cluster.

3.4 Experiment 3

[34]In our final experiment, we move beyond the analysis of documents in terms of words, as in experiment 2, to an analysis of the usage of individual words across the two genres. For this purpose, we used the embedding method fastText[37] and the R package of the same name. In each subgroup, the ten nearest neighbor terms of interest are established, so that each word which was identified as pertaining to both genres is visible, along with the terms found closest to it within the text. In contrast to Skip-gram, fastText is more appropriate for smaller bodies of text, as it does not compute an embedding for each word. Instead, embeddings for parts of words are calculated (for instance, for »honor«: »hon«, »ono«, »nor«, etc.) and accumulated to create an embedding for the whole word. In this way, more robust representations emerge for rarely used or unknown words.[38]

[35]In order to contrast the terms in each subgroup, we will illustrate in the following the ten nearest neighbor terms per subgroup together with the similarities for each, whereby the maximum possible similarity is represented by the number 1.

[36]The keyword »honor«, which is found not only in comedies, but also in tragedies, when assessed within the comedy subgroup, shows no common neighboring terms in the tragedy subgroup, nor were they found for the word »hado« (fate). In other words, both terms are used in comedies and tragedies, but within completely different contexts according to each. It becomes apparent that the terms ›honor‹ and ›fate‹ appearing in tragedies are more clearly outlined within the context and the meaning of the terms more precisely defined. For example, ›honor‹, within the context of the tragedy, refers to the loss thereof, or defamation, for which the remedy is obviously associated with possible death.

[37]The many similar word endings in this table may be baffling at first glance, but hardly surprising: All of Calderón’s plays are written in verses. Through this metric alone, the selection of possible neighboring words is drastically limited.[39] To make things worse, the similar inflections and conjugations of the Spanish language also left Calderón with a very narrow selection of possible neighboring words when composing his dramatic works.

[38]Other terms which were used in both subgroups also produce a similar pattern. The words »fineza«, »justicia«, and »amistad« (nicety, justice, friendship) yielded only one or two common neighboring words within both subgroups (represented in bold type); these terms are found in both comedies and tragedies alike, but within very different contexts. While these three terms within the comedic context tend to reflect the profane, their appearance in the tragic context reflects the formal authority of the court and its jurisdiction as well as seriousness and the realm of divine providence and justice.

[39]However, other terms clearly show overlaps with regards to the nearest neighbor terms; for instance, »celos«, »gusto« or »muera« (zeal / jealousy, taste, he / she / it dies) each share three or four nearest neighbor terms within the ten words in the selection.

[40]This analysis illustrates that the differences between tragedies and comedies do not merely consist of different vocabularies, but rather, that even shared vocabularies are substantially used in a different way. The more central for the genre, the more distinguishable the usage – at least, this is the tendency our results have shown so far.