In the field of New Testament Studies, classification of the heterogenous group of non-canonical texts, i.e. apocrypha, is disputed. The main problem is that the taxonomy of modern scholars largely reproduces the ancient classifications which were shaped during the 4th and 5th-century debate of canon which tends to lead to a binary classification between either canonical or noncanonical, and, moreover, to utilize conceptualized labels such as "Gnosticism" and "Encratite" which does not do justice to the complexity of the topical variety in these texts [1,2,3]. To contribute to this debate, we want to create an ontology, i.e., a structured framework, out of (apocryphal) textual data with the overarching goal of establishing a computationally driven classification system for New Testament Apocrypha within the context of the semantic web. We will investigate the topical distribution of four non-canonical acts. Non-canonical acts are stories from roughly 2nd-3rd century CE about early Christian apostles' legendary deeds and speeches [4,2,5]. In this paper, we investigate the stories called Acta Joannis, Acta Thomae, Acta Barnabae and Acta Philippi [6]. These texts have been chosen as tests to create the basis of including more texts. We want to contribute to define categories, textual characteristics, and relationships between texts by building a structured ontology. While the creation of this ontology extends beyond the scope of this paper, our present study of integrating topic modelling in research on New Testament Apocrypha serves a crucial step towards this endeavor. By utilizing ontological relationships and semantic annotations, in this context as topical distributions, this study explores the potential for integrating the classical theological concepts present in New Testament Apocrypha with modern computational methods, thereby bridging the gap between traditional textual analysis and advanced semantic technologies within the context of the semantic web. Our working question is how and in what way can this model contribute to the discussion of classifying non-canonical acts in a corpus of ancient Greek texts?
Our starting point is that digital methods are not quick and magic tools to solve complex questions [7], rather we find that constant and critical exchange between traditional and new methods in qualitative collaboration is the way forward [8,9,10]. Due to the circumstance that this study originates from a research discipline where computational methods are not embedded, we find it important to provide a detailed description -tending to tedious -of the methodological process, since we believe that this can help to bridge the disciplinary divides with a low-practical how-to-style
The first step is to provide our text corpus. The corpus serves in our experiments as the literary context for the target texts, for which reason a historical relation between corpus and target texts are needed. Our target texts, the four apocryphal acts, are written in the Ancient Greek language, for which reason we have gathered a database consisting of 2153 Ancient Greek Texts. These texts are retrieved from the Perseus Corpus, First1KGreek, Pseudepigrapha.org and Deutsche Bibelgesellschaft [11,12]. The selection of this corpus is chosen based on its relevance for our target texts. The corpus largely sets the parameters for our topic modelling at the later stages of the experiments. The Ancient Greek texts are in a solid machine-readable state. For the corpus text to be prepared for calculations we perform a set of preprocessing steps, so that the text is cleaned and lemmatized. We clean out for any Latin characters, digits, extra whitespaces and stop words. The textual cleaning follows the logic of standard natural language processing utilities. The models for cleaning and parsing the text were built with a transformer-based pipeline called OdyCy [13]. This pipeline is a single-transformer pipeline that uses the following workflow: The transformer is built on Ancient Greek Bert [14]. The parser, morphologizer and lemmatizer follows the infrastructure of SpaCy [15]. Before running topic modelling, we had already preprocessed the input text, so we set max_df and min_df to 1.0, since we did not want to ignore any terms in this experiment [16]. When the textual preprocessing is done, we move on to vectorization. Vectorisation is a highly qualitative choice, almost a language philosophical step, in which we must decide how to represent our textual corpus as numerical vectors [17]. In this case, we employ the method of term frequency-inverted document frequency (TF-IDF) which represents the logarithmic scale between a term's frequency and the inverted frequency of total documents in the corpus [18,17,16].
Topic modelling is a much-utilized tool when engaging in natural language processing classification tasks, mostly for the purpose of assigning a category based on the most probable topic to a text in a corpus [19,20,21]. This is also partly our aim, although we do see possibilities of detecting more complex layers in the topical distribution besides a text being only a part of one single category, rather, topic modelling gives due credit to the complex topical distribution and its significance for the position of our target texts in the corpus. Topic modelling is a way of structuring our textual data to be included in a future knowledge graph and similar semantic web technologies. Topic modelling assumes that a corpus of texts consists of topics and that these topics are comprised by the words of the corpus. We utilize the kind of topic modelling called Non-negative Matrix Factorization (NMF) [22,23]. NMF is an approach that decomposes a high-dimensional term-document matrix, i.e., a matrix consisting of the corpus documents in columns, the words in rows, and their occurrence-values in the cell entries. The occurrence-values are, as mentioned above, chosen to be TF-IDF. Based on optimization, the method of NMF calculates topical patterns by associating words with topics and topics with documents. The NMF model, so to say, produces latent topical patterns by grouping similar and co-occurring words in the corpus [24]. These topics are then backtracked to fit to the documents. The number of topics to choose is important for the analytical task. If we choose too many topics, there were no interpretable coherence, so based on our knowledge of the corpus and trial-and-error process the most robust output in topics was 10 topics [25].
The 10 topics and a selection of their top words are: Testament Gospels. These genre categories did not influence the topic model's calculations but can be used by us as a navigating tool to interpret the topics. For example, we can see that Topic 7 and Topic 6 have overlapping words about justice and city, but we can see that the texts that are dominated by Topic 6 are rhetorical texts like Demosthenes, whereas Topic 7 dominate many of Philo of Alexandria's texts as well as Libanius' Declamationes which are more philosophical. These extra steps enable us to distill the words of the topic into a qualified label or notion about how to describe the topic presented.
In our first analysis, we interpret the topical distribution of the Acta Joannis, Acta Thomae, Acta Barnabae and Acta Philippi [27]. The results can be seen in Figure 1, which shows four pie charts over the topical distribution of the four non-canonical acts. Overall, all the four target texts are dominated by topic 2 and generally they share similar groups of topics, but the distribution is not equal. If we wanted to group the target texts together in the corpus on a coarse level, they would be set in the group of topic 2. However, this would not be a discovery. Where topic modelling can lead us further is in the presence and distribution of minor topics compared to each other. The topics of Acta Joannis are distributed over mainly topics 2, 1, 9, 3, 0, 4 and 6. Topic 1 is, like topic 2, also a theological topic. The presence of topic 9, 3 and 4 is revealing of the content of the story, since these topics represent philosophically oriented words which tell about how the story engages in Hellenistic religion and philosophy. Topic 0 is the historical-political topic which is understandable since the text narrates sequences of events and speeches. We also see a small portion of Topic 6, which concerns justice and law, which resonates with a few scenes in the narrative. From the topical distribution, we can generally characterize Acta Joannis as being situated in a Jewish-Christian theological context where Hellenistic anthropocentric philosophy and religion is also present. Concerning Acta Thomae , the topical distribution is similar to that of Acta Joannis, with a large topic 2 and 1, but Acta Thomae has a more equal distribution between its subtopics, 3, 9, 4 and 0. The Acta Thomae-story is set in legendary India, where Thomas is sent off as a missionary. From manual reading, Acta Thomae and Acta Joannis are comparable in the sense of the mix between narrative and preaching, where the preaching might drag in the more philosophically weighted topics, and this relation is also visible in the topical distribution. When we inspect the topics of Acta Philippi, then the distribution is markedly different from the two previous narratives with Topic 0, the historical-political topic being the third most prominent. Although Acta Philippi in its content has a much more adventurous and mythical tone with, for example, the protagonists arriving to a city of snakes, the structural dynamic of this narrative is dominated by sequence narration where we follow event after event [5,2]. This makes the Acta Philippi into a drier, journal-like text. The large presence of Topic 0 is also visible in the topical distribution of Acta Barnabae. This circumstance is probably linked to the fact that Acta Barnabae is situated in a church-political context of legitimizing the so-called autocephalous, i.e., independent ecclesiastical unity in Creta in the 4th century [28].
In the second part of the analysis, we want to address how the topical distribution analyzed in the previous part affects texts' position in the corpus. The position of each of the four target texts are visualized in Figure 2 The clusters are formed based on texts with a dominant topic. Those texts which have an almost unequivocal dominance of one topic, e.g., Aristotle's Problemata and The Book of Jeremiah are placed in the outskirts of each colored cluster dragging away from the center of complexity. Conversely, those texts that are close to the center and also other topical groups display diversity and complexity in their topical distribution. Some clusters appear like relatively demarcated cone-shapes like e.g. Topic 4 of Hellenistic religion and philosophy and Topic 2 of god, people, human, whereas topic 3 of Greek element-philosophy is more flat, which indicates that this topic is dispersed more evenly in the corpus. The topical distribution on the corpus level gives a navigational tool with which texts can be grouped with relative clarity. For example, it is noteworthy that classic, and almost foundational texts, of Ancient Greek language, the Iliad and Odyssey, are clearly situated in topic 4 of Greek religion and philosophy, and that later texts that are trying to imitate these, like Tryphiodorus' Sack of Troy (4th century CE) and Nonnus' Dionysiaca from the 5th century CE, almost a thousand years after Homer, have an almost identical topical distribution. All of the Acts stories are placed in the dark green cluster of Topic 2, the topic of god, people, human. But it is noticeable that they all show diversity in the distribution. Acta Joannis and Acta Thomae are placed more securely in the Topic 2 cluster, whereas Acta Philippi and Acta Barnabae are drawn toward the center which might be explained by their higher percentage of the historical-political topic Topic 0. When texts are placed in this corpus, it becomes important to iterate the basic assumption of topic modelling: that all topics are produced based on words in the corpus. This means that the corpus words are constituent of the created topics which calls for a qualitative choice of corpus texts. Our corpus consists, as mentioned, of texts that historically have shaped our four target texts, either directly or indirectly due to the circumstance that the (mostly anonymous) authors were educated people in the Greco-Roman world about whom it can be assumed that they had basic knowledge of the texts in their historical and literary context. The method of topic modelling, then, almost backtracks the literary world of the authors of our texts, of course, with the important acknowledgement that we do not have all texts which made up the author's literary context.
In this analysis assisted by topic modelling, we were able to characterize and place four noncanonical acts based on their topical distribution. The topical distribution of the analyzed texts will be used to map ontological relationships and enhance semantic annotations in order to classify New Testament Apocrypha, among which the topical distribution is a major component. The results of this topic modelling analysis will thus be able to be included in a future New Testament Apocryphal Ontology. The navigational advantages of topic modelling allowed us to inspect the target texts in a qualitatively selected corpus consisting of texts from a similar historical and literary horizon. This ensured meaningful topics. Instead of characterizing and classifying the texts based on abstractions and taxonomy from the 4th and 5th century church-political discussions on canonization, we could characterize and classify, or at least contribute to these tasks on the basis of raw content, both on a small, close scale in the topical distribution from text to text, but also on a larger scale based on the entire corpus.
The research in this article is funded by the Carlsberg Foundation in the Semper Ardens: Accelerate-project "Computing Antiquity: Computational Research in Ancient Text Corpora. " We would also like to thank Deutsche Bibelgesellschaft, Pseudepigrapha.org, The Perseus-Project and First1KGreek for providing texts.