Using Computational Methods to Explore Law in Sermons

Abstract

An empirical study on the use of law in Christian sermons has so far been a blank space in research, especially when large corpora of sermons are examined. In this article, we present the first findings of the ongoing RUNIP project, in which computer-assisted methods are used and validated in sermon analysis. The process integrates manual coding via MaxQDA with machine learning techniques, notably contextual embeddings derived from Transformer architectures such as SBERT, enabling us to detect patterns across large corpora. We argue that embeddings in text analysis can help to complement a manual, human-based text analysis. Clustering based on sentence embeddings helps identify semantically related sermon passages, although the complexity and length of the original texts, as well as the nuanced theological language, pose challenges to computer-aided analysis. By bridging historical and contemporary sermon analysis with data science methodologies, we demonstrate how an interdisciplinary approach can expand our understanding of how preachers address law, norms, and moral questions in Christian sermons. This is demonstrated by qualitative results from the analysis of the large historical sermon corpus of Friedrich D. E. Schleiermacher.

1. Law in Sermons–An Explorative Approach in Computational Text Analysis

Interest for sermons in society and media is particularly high when a feeling spreads: Someone wants to moralize from the pulpit. Now it’s getting normative! But of course, preachers also want to take ethical and moral responsibility–especially when, for instance, values such as pluralism and democracy are called into question in society (Karle 2017). But on what normative foundations are sermons then based? And doesn’t this always have to do directly or indirectly with the biblical legal tradition? In this paper we address these questions and we explore the extent to which computational analysis can help us find answers.

As far as the research on these questions for Christian sermons is concerned, one has to admit that there is no extensive empirical research done on this topic–at least speaking from a specific German background1. Robust methodological approaches to sermon analysis—particularly computer-assisted procedures able to process large collections of sermon texts—are still lacking. Moreover, it seems that the question of law and norms in sermons remains unanswered in empirical terms despite an obvious public interest. And there might even be a deeper theological reason for that: Theologians often argue that the law and normative language have no place in the pulpit. In this respect, theologians emphasize that the communication of the gospel as freedom is the main focus of a sermon, not rules and norms. There has been a strong theological discourse on the “legalism of the sermon” (“Gesetzlichkeit der Predigt”) in the 20th century in Germany (Hirschmann 2011; Josuttis1 1966; Josuttis2 1995), which in many respects has led to a theological narrowing and has also found expression in the academic area of practical theology. In German homiletics the law often forms an empty space2 or appears as a negative term in the sense of “legalism”. All this leads to at least three problems.

The first one is historic: influential traditions of interpreting the law in sermons, which have developed in Christianity over the centuries, particularly in the Reformed tradition3, are set aside. But historical research in law and religion in recent decades has pointed to the modernizing and democracy-promoting tendencies of these traditions.4

The second problem is a practical one for preachers in the present day: if preachers intend to address current topics in the fields of law (e.g., in the field of human rights, natural law or the right to asylum), they lack in many cases a general hermeneutical approach to biblical law and extra-biblical sources of law. Elaborative concepts of how to preach on law are simply missing in practical theology.

The third problem concerns the credibility of churches in society from a modern or–if you want–postmodern perspective. When the role of Christian churches is called into question in secular societies, especially their often ascribed role as “moral agencies” (Joas 2016), it should be a matter of course to deal with the foundations and sources of norms, that the church relies on and which are perhaps not merely moral but have to with the biblical tradition as well.

These problems were the starting point for the RUNIP5 project together with a new orientation in computer-assisted methods of text analysis. In this article, we provide insights into the project and focus on methodological questions: How can we analyze legal references or legal language in sermons with the help of computers? What are the possible advantages over a manual approach without machine assistance?

2. Starting with a Historical Corpus of Sermons (Friedrich Schleiermacher’s Sermons)

The RUNIP project is about a time comparison and at the same time about training an algorithm on a historical basis. In the first phase of the project we concentrate on the computer-aided analysis of all legal references in a historical corpus of sermons. In our case, we chose the historical corpus of sermons by Friedrich D. E. Schleiermacher (1768–1834). But why exactly have we chosen Friedrich D. E. Schleiermacher and his historical corpus of sermons?

Schleiermacher is certainly one of the most important Protestant theologians. Schleiermacher is often called “Church Father of the 19th century” (Lülmann 1907) in Germany and even considered the “founder of modern Protestant theology” (Scharlemann 2024). He was instrumental in establishing practical theology as a subject in the canon of theology (Schleiermacher 2002). His influence was not limited to theology, but also extended to the church, politics, and society of his time and far beyond (Arndt et al. 2019; Wolfes 2004). He was one of the founders of Humboldt University in Berlin and provided influential ideas for hermeneutics, the German education system (Polke 2017), and church politics (Dinkel 1996; Geck 1997). While Schleiermacher was an important polymath in the truest sense of the word, he was primarily a theologian and brilliant preacher. He integrated many areas of his thinking in his sermons. The fact that several thousand people listened to his sermons back then underlines his importance for theology and society.

It is remarkable that no comprehensive study of his sermons has been done,6 and most especially that the topic of our research: his constant reference of law in his sermons, has been unexplored.7

Fully edited since 2017/18, the sermons are published in Series III of the Kritische Gesamtausgabe (KGA = Critical Complete Edition). Each KGA volume runs 800–1200 pages and, taken together, contains 1072 sermons (not including drafts); a single sermon may stretch to 20 pages. Schleiermacher’s own calendar, however, records more than 3500 preaching occasions, so the existing texts represent only a fraction of his output. Even this reduced corpus is immense–too large for one scholar to handle alone. Our interest is to rediscover the breadth of references to the law within these sermons. Theologically, we hope to learn from these sources for today’s practical theology; from a data-science perspective, we aim to see how far computer-assisted text analysis can achieve deeper or otherwise unattainable insights.

3. How Do We Examine Legal References in Sermons?

How do we examine legal references in sermons? At what point can computer-aided methods help us in the analysis and in what way? Without going into more detail about the forms of transmission of Schleiermacher’s sermons, it should be mentioned that one should distinguish between manuscripts, printed texts, and transcripts of Schleiermacher’s sermons. The fact that the Critical Complete Edition of Schleiermacher’s sermons (KGA III) has recorded all of this in full proved to be an advantage. One disadvantage was that the sermons in the edition were only available in the form of PDFs, which meant that the data preparation of the PDFs in an OCR process and the separation of all metadata was a time-consuming process.

Meanwhile, the qualitative analysis was already underway. How do we examine legal references in sermons? The first task was to find and define the relevant passages in sermons. Here, Schleiermacher’s sprawling language proved to be a difficulty. Sentences can be as long as half a page. In addition, the sentence structure and syntax were more complex than today’s language. At the same time, defining meaningful passages under the unit of one sentence has not proven to be useful. In the first phase of our project we labeled relevant passages by hand with the help of a software tool for qualitive and mixed methods research (MaxQDA) (Rädiker and Kuckartz 2019). A first orientation was the search for key words such as ’law’ (Gesetz, Recht) or ’commandments’ (Gebot) with related word stems.

Methodologically this first step was orientated towards the Grounded Theory, developed by Bernhard Glaser and Anselm Strauss primarily in the field of empirical social studies and interview analysis (Corbin and Strauss 1990; Glaser and Strauss 2017). In recent years it has been used in various studies in the field of corpus linguistics, which can serve as orientation (Scharloth 2018). With Grounded Theory, Glaser and Strauss described a method of theory generation from (as much as possible) qualitative data, whereby the additional use of quantitative data is explicitly not excluded. The collection of data and the analysis is described as a circular process that takes place “from the data” and methodically undergoes a permanent comparison. In this inductive process (Gibbs 2012, pp. 337–43), the data or texts are labeled or coded for the investigation and provided with memos. The ongoing labeling aims to develop concepts as basic units of the analysis: Where can legal references be identified? How does religious communication about the law work in these labeled passages of the sermon? Based on grounded theory, we can speak of categories that once again bundle similarities of concepts on an upper level. In rhetorical terms, we can also speak of topoi, general ‘places’ that recur in the treatment of the topic of law. Overall, a correspondence between form and content is assumed. The connection between semantics and syntax is taken into account just as much as pragmatics of language, even if the overall focus is clearly on semantics. It would then be easiest to speak of “topic bundles” with rhetorical functions.

As the iterative analysis of sermons continues, our goal is to identify categories, relationships, patterns, and variations within the underlying concepts. Labeling 700 sermon passages has made a six-category scheme–or topoi–increasingly evident. These six topoi are outlined below.

Schleiermacher’s sermon passages, which have a primarily (1.) explicative function and can also be described as a form of “teachings about law”, can be classified under the first major category.8 From a rhetorical point of view, they correspond above all to a docere of the sermon. In these passages, Schleiermacher proceeded in his sermons in such a way that he explained and differentiated the law to his listeners. This resulted in a variety of differentiated doctrines. What people find as laws can, for example, have a transient or eternal form, insofar as it corresponds perfectly to God’s will, or it is changeable, mutable, contingent. The law that derives from God (“divine law”) can address man inwardly or outwardly, it can therefore be of a more spiritual nature or is transient and changeable in a more “literal form”. It can correspond to the gospel in its inner core or, in its transient form, only represent the divine law in a “shell” or externally as a law. In relation to the biblical tradition, Schleiermacher was thus also able to address the distinction between the old covenant of Israel and the new covenant in Christ in his sermons. Corresponding sermon passages may be explanatory either in a historical sense or in a systematizing one.

A second large category of all sermon passages with legal references is referred to below as (2.) Christ and the law.9 A further distinction can be made here: In the corpus of sermons, one will find passages and sermon sections in which Schleiermacher concentrated on the relationship that the (historical) Jesus had to Jewish law (Torah). These include questions of legal validity, which concerned, for example, the possibility of retribution (talion) or the marriage commandments. What position did Jesus Christ, whom Schleiermacher primarily called the ‘Redeemer’, take in relation to individual commandments and their interpretation and the Torah as a whole? However, Schleiermacher also viewed the relationship between Christ and the law from a different perspective. If Christ could be interpreted as the fulfillment of the divine law, then the divine law could also be understood as being ordered towards him. And if this was the case, then the depiction of the birth, life, death on the cross, and resurrection of Jesus Christ could also be interpreted in this way from a legal perspective for Schleiermacher. If this also involved questions of human salvation, overlaps with the next category should be noted.

The category (3.) Law, redemption, and faith10 summarizes the legal references in Schleiermacher’s sermons that were dedicated to questions of salvation (soteriology). This refers to sermons and sermon passages that relate the question of human salvation before God to the meaning of the law. In Schleiermacher’s sermons–and also in his other theological works–the term ‘redemption’ is the most comprehensive expression for a context of salvation that refers to both this world and the hereafter or God’s eternity. In Protestantism, starting with the Reformation, ‘justification’ or the doctrine of justification was certainly the more common term and doctrinal context. The Reformation doctrine of justification was always accompanied by juridical ideas: an individual is not granted righteousness before God’s judgment seat through himself, but through faith or trust in the salvation that Jesus Christ has accomplished. In this way, a believer is justified. In accordance with the teachings of the Reformation, faith in salvation through Jesus Christ remained decisive for Schleiermacher, but for him ‘justification’, ‘rebirth’, and ‘sanctification’ were already aspects of redemption in his dogmatics and thus also in his sermons. ‘Redemption’ is therefore the more comprehensive concept of the salvation of people with God. In his sermons, Schleiermacher repeatedly played out the idea that human redemption could always also mean redemption from the law and at the same time with the law. This can be broken down to the question: If God has given mankind not only a transient, but also a good, wise, and eternal law, what role does it play in man’s salvation or redemption? To what extent is an orientation towards commandments and instructions, which God suggests to mankind, decisive for man’s salvation?

If these questions are primarily linked to God’s actions beyond finitude in Schleiermacher’s sermons, the sermons and sermon sections in question are assigned below to the category (4.) Concepts of judgment and election.11 The concept of judgment could easily be misunderstood as the idea of a divine judgment in which sinners and the unrighteous are rejected and the faithful and righteous are chosen for eternal life. Schleiermacher, on the other hand, uses the concept of judgment in a process-oriented way for God’s actions. God’s judgments can already take place in this world if they bring about change with regard to God’s eternal redemption. He thus ties in with end-time ideas of judgment and election but also transforms them at the same time. This category (4.) could also be called eschatological concepts on the law or conceptions of an eternal order of God for human beings, which Schleiermacher spells out in a legal perspective. A key concept in this respect is the “eternal decree”, which can be attributed to God even beyond the temporal experience of human beings. This category includes questions such as: Does the world created by God in which we live correspond to a “plan” of God that is in some way still revealed to us humans as an order or in regularities? Are we humans still free in our will, our decisions, and our actions or are we ourselves only part of an order chosen by God for this world?

In category (5.) Church and congregational ethics12 the organization and shaping of life of Christians becomes more specific: Sermons and sermon passages in which Schleiermacher addressed the legal form of the church, its organization, and offices or aspects of ethics in the church could be assigned to this category. Overlaps with category (2.) arise when sermon passages deal more specifically with aspects of church ethics. This is because Schleiermacher understood the church as a community of faith that follows the community founded by Jesus Christ. Christ’s consciousness of God (a feeling of absolute dependence) in spirit and religious practice is lived on in the church. When conclusions were drawn from the actions and teachings of Jesus Christ to the actions of Christians in the church, Schleiermacher again used legal language and spoke, for example, of a ‘new commandment’ or ‘new law’ that Jesus had given to his church.

The last category (6.) Society and State13 summarizes the legal references in Schleiermacher’s sermons that addressed the worldly order in the state, politics, and society. This would therefore correspond most closely to what we today call the legal system in the modern state. For Schleiermacher, it is not congruent with the legal form and structure of the church, although the two are nevertheless interlinked. In his sermons, for example, Schleiermacher also addressed the criminal law and protective function of the law, the relationship of citizens to the state, and the limits of civil law. Category (6.) also includes political implications in the narrower sense as well as references to the broader contexts of social life, insofar as it concerns a socio-political and not primarily ecclesiastical order.

As described, this is an interim status of our results and methodological approach. An essential step in the second project phase was the addition of algorithm-based labeling to this human-based labeling.

4. Advancing Text Analysis Through Contextual Embeddings and Transformer Architectures

When working with text corpora, uncovering their internal structure poses numerous challenges. While most texts come with some metadata such as the date or author, their content is largly unstructured, and labels for topics or other categories are usually not available. Questions like how often a theme recurs across the corpus or how it is framed are therefore hard to answer. Most often researchers will rely on searching for keywords and rely on them to be guided through the texts. But depending on your research question, this strategy is fragile: it misses synonyms, overlooks spelling or transcription variants, and falters when historical authors use divergent vocabulary. A first remedy in Natural Language Processing (NLP) is stemming, which reduces words to a common root and catches grammatical variants, yet this solves only part of the problem. In recent years, embeddings have gained traction because they identify concepts and map textual relationships far more robustly.

4.1. Word Embeddings

If you argue that the meaning of a word is how it is used in the language (Wittgenstein 1953, §43), then knowing in what context a word is used, is how you understand what the word means. In this sense, focusing on encoding meaning into words14 is the primary objective for embedding algorithms. They provide a numerical representation that encodes the information of their context and help in revealing similarities between words (Bengio 2003, p. 1139). By transforming the vocabulary of a corpus into vectors, their location in vector space encodes semantic and syntactic relationships among those words. This concept is based on the works of (Bengio 2003) and (Collobert and Weston 2008), among others, where it was shown that such word vectors can be used to improve and simplify many NLP applications (Mikolov et al. 2013, p. 2).

By incorporating context, the model encodes synonym similarities while still distinguishing homographs, thereby improving language modeling. If we base the search for a concept or topic not on keywords but on this encoding, then we can identify all areas where these topics are present, regardless of the language in which they are described. The efficiency of embeddings for encoding language make them an important component for Large Language Models and are entwined with how Transformer Models like GPT and BERT work. Having a numerical representation of language enables machine-learning methods that let us assess semantic similarity, cluster, and organize texts with their underlying concepts, and extract information from the corpus in more comprehensive ways.

For our project RUNIP we have gained a lot of insight into specific sermons. Through qualitative analysis several sermons have been annotated, key concepts were identified. Sentences and passages in these sermons have been labeled to mark them as identifiers for specific topics in question. Through the use of embeddings we can now evaluate how much these concepts are present over the whole corpus, what sermons might be of particular interest for us to study and how we can identify them.

4.2. Contextual Embeddings

Combining neural network-based language understanding models with advancements in word embeddings and contextual embeddings can lead to enhanced performance in language understanding (Peters et al. 2018, p. 1). One particularly influential architecture here is BERT (Bidirectional Encoder Representations from Transformers). Introduced by Devlin et al. in 2019, BERT builds on earlier pretrained language modeling methods but differs through its reliance on the Transformer architecture’s self-attention mechanism (Vaswani et al. 2017). Self-attention adjusts previously generated embedding vectors based on the other words in the sequence, making embeddings more specific to each word’s meaning in that context. It does this for all parts of the sequence, thereby producing a rich representation of the entire sequence (Vaswani et al. 2017, p. 2).

During training, the model begins by learning general embeddings that capture word meanings. It then adds positional encodings to each word to retain ordering information (Vaswani et al. 2017, p. 6). The training process refines these embeddings within each sequence, resulting in a more contextually accurate representation. This refinement is achieved by taking the dot product of a given token’s embedding (and its positional encoding) with every other token’s embedding (and positional encoding) in the sequence (Vaswani et al. 2017, p. 4). The resulting similarity scores indicate how closely related each token is to the others, and the Transformer’s final embedding for a token is the sum of these dot products (Vaswani et al. 2017, p. 4).15 Because the token itself typically remains the most similar to its own position, its core meaning is preserved. Meanwhile, the embeddings of other words shift to better represent their contextual roles, with more similar tokens exerting stronger influence.

As training proceeds, the model gains a general understanding of tokens through initial embeddings and then refines these representations for all occurrences of the tokens in their respective contexts. In this way, self-attention encodes more accurate information, enabling improved learning outcomes. Because Transformer embeddings better capture contextual information, they propagate through the model’s layers more directly and effectively. With more accurately captured information, the average signal length through the network is shortened, improving efficiency (Vaswani et al. 2017, p. 6). During pretraining, BERT randomly masks tokens and uses both left and right context to predict them, allowing it to capture deep linguistic relationships and nuances—such as polysemy, syntax, and semantic dependencies—in a way that surpasses many earlier models. By factoring in context from all directions, BERT can produce highly accurate, context-aware word representations, achieving strong performance across a wide range of NLP tasks (Devlin et al. 2019, p. 4176). This bidirectional approach, combined with the Transformer’s self-attention, marks a significant milestone in LLM development and paves the way for further advances in language understanding.

4.3. Sentence Embeddings

As single words do not capture complex topics accurately enough, we focused on sentence embeddings: encoding meaning at the sentence level. Algorithms like GloVe (Pennington et al. 2014) and Word2Vec (Mikolov et al. 2013) often lack accuracy in capturing the meaning of entire sentences (Reimers and Gurevych 2019, p. 5). Similarly, utilizing a Transformer model like BERT does not adequately solve the issues of creating good single-sentence encodings (Reimers and Gurevych 2019), because BERT’s architecture does not produce dedicated sentence embeddings during training. Averaging BERT outputs or simply using the first token of a sentence as its representation can yield poorer results than relying on embeddings created with GloVe (Reimers and Gurevych 2019, p. 4).

To address these shortcomings, Reimers and Gurevych introduced Sentence-BERT (SBERT), which fine-tunes a BERT model to produce fixed-length vectors for more effective sentence representation. It is trained on sentence-pair datasets, where the pooled output vectors are optimized to have high cosine similarity for semantically similar sentences (Reimers and Gurevych 2019, p. 3). Through this fine-tuning, SBERT outperforms other BERT models, GloVe, and additional encoders (Reimers and Gurevych 2019, p. 7).

In our work, we used SBERT to create a representation of the sentence meanings within our sermon corpus and to identify recurring themes based on qualitative annotations of key concepts. Sentence-level embeddings offer a more accurate way to capture these nuanced ideas.

5. Clustering with Sentence Embeddings

After extracting and cleaning all available sermons from the original PDF files, we compiled a dataset of 1072 sermons, treating each sermon as a distinct unit. To facilitate deeper analysis, we appended key metadata to each sermon entry, including the date it was delivered, the medium through which it was recorded or published, and the location of delivery. We focused on sermons preserved in a fully written form, as drafts of Friedrich Schleiermacher’s sermons cannot be reliably processed by language models trained on prose.

We retained the labels assigned during earlier qualitative analyses, in which individual sentences were tagged according to specific research topics. We then used SBERT to generate sentence embeddings for the corpus, allowing us to identify unlabeled passages potentially relevant to the same topics. For this task, we relied on Python (Version 3.12.7) and the sentence-transformers (Version 2.7.0) library. Specifically, the pretrained all-mpnet-base-v2 model was employed to map each sentence to a 768-dimensional vector space. Although this model’s training data largely comprises modern texts, its capacity to capture general syntactic and semantic relationships makes it effective even when applied to historical materials. While there do exist models trained on historical corpora, they typically rely on smaller datasets; in our tests, models with broader, modern training data performed better for our purposes.16

Using the resulting sentence embeddings, we then performed a clustering analysis using K-means. Based on standard criteria such as the elbow method and silhouette scores, four clusters emerged as the most effective balance of between-cluster separation and within-cluster cohesion. These four clusters reflect distinct syntactic and structural patterns in the data. As illustrated in Figure 1 and

Table 1. Summary Statistics of Sentence Length (in tokens) per Cluster.

Cluster	Count	Mean	Std	Min	25%	50%	75%	Max
0	29,040	39.38	19.92	4.0	24.0	38.0	52.0	151.0
1	41,725	127.25	68.67	9.0	81.0	111.0	155.0	936.0
2	6261	5.88	2.10	3.0	4.0	5.0	8.0	30.0
3	14,510	124.88	81.87	7.0	67.0	106.0	161.0	901.0

, one clear distinction among these clusters is the variation in average sentence length.

Cluster 2 consists of very short sentences, often related to text-structuring elements such as numbering within the sermons or the conventional closing phrase “Amen”. An examination of the sentences in Cluster 3 reveals that they are primarily references and quotations from the Bible. Although this distinction does not apply to all sentences it still accounts for a significant portion. Clusters 0 and 1 comprise the majority of the corpus. They do not exhibit any inherent substantive meaning; instead, they constitute the main portion of the sermons, accounting for 77% of the total number of tokens. Even a simple clustering approach thus provides structural insights into how the sermon corpus is composed.

Through cosine similarity, we can compare sentence embedding vectors to measure the similarity between sentences and identify those that are more closely related. In the qualitative analysis, broad topics of interest were identified. To illustrate, we focus on the topics of Society and State and Christ and the law for deeper analysis using sentence embeddings. Recognizing the various distinctions within these topics, we conducted separate clustering analyses on all sentence embeddings for these labeled sentences, thus categorizing them by similarity.

This approach offers several advantages. First, it helps identify outliers in our sentences, which can reveal errors in data processing (e.g., incorrectly split sentences leading to inaccurate embeddings). It also highlights sentences that are very dissimilar to the rest of the labeled set. These sentences could represent special cases of interest not observed elsewhere in the corpus, or they might prompt a reconsideration of their labels to reflect new insights more accurately. Sentence embeddings can also be used to locate other similar sentences in the corpus, potentially indicating a more frequent presence of these topics than previously recognized.

Furthermore, the clustering analysis helps identify sentences that are most representative of a topic. To find more sentences in our corpus that exhibit high similarity to these representative clusters, we calculated the embedding vector for each cluster centroid and then filtered out sentences with the highest cosine similarity to these centroids. Rather than focusing on individual sentences, this method allows us to surface sentences that are more characteristic of an entire group.

Figure 2 illustrates the 10 newly identified sentences with the highest cosine similarity scores for a given centroid within our topic. We have also highlighted all of our labeled sentences for that cluster, which serve as the basis for the centroid vector used to locate similar sentences in the corpus. Note that the two-dimensional visualization can capture only 9.65% of the total variance in the 768-dimensional vectors, and thus cannot fully represent every aspect of similarity.

Using the same process, we created a subset of 100 sentences in total, with 10 sentences per cluster centroid, resulting in 50 sentences for each topic. We selected 10 sentences per centroid by examining the distribution of cosine similarity scores between the sentence embeddings and the centroids (see Figure 3).

Table 2. Top 10 sentences similar to cluster 2 centroid for Christ and the law.

Sentence	Similarity	Full Text
1	0.8082	Sehen wir nun auf das, was im Innern des Erlösers vorging, so müssen wir wol glauben, daß er sehr bewegt war.
2	0.7873	Es ist dasselbe wodurch sich der Erlöser in der einzelnen Seele des Menschen verklärt, und wodurch er die ganze Welt umgestaltet und neu geschaffen hat.
3	0.7807	Und wer unter uns vermöchte da fröhlichen Gemüthes zu sein wenn er das leztere befürchtet.
4	0.7773	Der aber führte sie zu dem Erlöser, dem einzig wahren und würdigen Ziele dieser dem Menschen eingepflanzten Sehnsucht.
5	0.7759	Dieses Gewaltige bestand nun eben darin, daß sie das Innere ergriff, daß sie den Menschen ein ganz anderes Ziel vorsteckte, welches sie vorher nicht kannten.
6	0.7734	Wenn wir dem Erlöser im Uebrigen nachfolgen und in seine Fußstapfen treten, so wird das auch hier sein.
7	0.7718	F., laßt uns nun noch zweitens sehen, wie der Erlöser sich in diesem Wechsel nachtheiliger Gerüchte befand.
8	0.7716	Und diese geziemt auch in dieser Hinsicht uns allen eben so wie der Erlöser sie trug.
9	0.7714	Diese haben ganz die Handlungsweise des Erlösers gegen sich, sondern er fragt weder nach dem Einen noch dem Andern.
10	0.7700	Hierbei ermahnt er nun nach dem Geist zu wandeln und so immer mehr frei zu werden von dem Gesetz der Sünde.

presents the sentences identified in this manner for the example under discussion.

6. Results

Identifying 50 sentences for each topic, we evaluated whether they addressed the relevant topics. To clarify references within the sentences, we also included the sentence immediately preceding and following each candidate sentence, thereby enabling a more informed evaluation.

Finally, we present an illustration from our analysis of Schleiermacher’s sermons, demonstrating how human-based labeling and algorithm-based labeling can complement each other. Figure 4 shows a passage from Schleiermacher’s sermon on Mark 1:15–22, delivered on 11 September 1831.

The biblical text interpreted by Schleiermacher describes the beginning of Jesus’ travels in Galilee and the gathering of the first disciples. It also recounts how Jesus went to the synagogue in Capernaum on the Sabbath to teach. His teaching was characterized by “mightiness” (exousia). Notably, Schleiermacher focuses on this final aspect. From the context, it can be inferred that “mightiness” pertains specifically to Jesus’ teaching. However, a manual text analysis relying solely on keyword searches would not have identified this nuance. Only the broader context of the sermon indicates the legal reference. In comparison, the algorithm-based search marked the passage shown in Figure 5.

Using only the sentences identified through our embedding-based search and their immediately surrounding sentences as context, we assigned a rating in

Table 3. Assessment of the identified sentences’ relevance to the topics of interest.

Topic	Relevant	Not Relevant	Uncertain
Society and State	24	10	16
Christ and the law	23	22	5

to indicate how well they matched our topics of interest.

Table 3 shows that fewer than half of the sentences can be identified as pertaining to the topics of interest, with some clear false positives in our search results. Some of these false positives are expected, given that previous cluster analyses revealed sentences of special interest. Because this cluster of special cases represents sentences that are “similar in being different” from the rest, they do not exhibit the same internal cohesion. Consequently, similar sentences would be better identified by selecting certain individual sentences rather than relying on the entire cluster. This factor explains the higher number of irrelevant sentences in the Christ and the law topic, where two clusters consisted of such outliers. To assess the methodological implications, we chose to retain these clusters rather than adjust our search. As for the Uncertain sentences, it was not possible to determine their relevance solely by examining them and their directly neighboring sentences.

7. Discussion

This article has presented the initial findings of the ongoing RUNIP project, concentrating on methodology. A key question is how algorithm-based labeling can augment human labeling in textual analysis. We believe a machine-assisted approach can greatly ease work with large sermon corpora—whether historical or contemporary.

The method goes beyond the usual manual strategies—keyword searches, close- and distant-reading, or sampling. As our example showed, keyword searches alone can be unreliable: how do we surface similar meanings when the decisive words never appear? Embeddings offer a powerful extension.

Sentence embeddings let researchers organize a corpus and locate relevant passages. We have demonstrated that they not only retrieve highly similar sentences but can also reveal the broader themes first spotted through qualitative reading. Still, when we push towards those richer topics and their complexities, the shortcomings of sentence embeddings emerge: however capable the language model, only limited nuance fits into a single dense vector for each sentence.

Given the length of sentences in our corpus and the complexity of the writing style, we will continue refining our approach to processing the corpus. We are currently evaluating results from different pretrained datasets to explore their impact on encoding quality. Moreover, additional refinements to the model architecture may further enhance outcomes—these could involve expanding beyond the original SBERT model to newer variants such as ANCE (Xiong et al. 2020) and TAS-B (Hofstätter et al. 2021). In addition, approaches like ColBERT (Khattab and Zaharia 2020) allow for token-level embeddings, thereby preserving more granular details from the original text for evaluation.

In the next phase of the project, a transformer model trained on the historical sermon corpus will serve as a comparative framework for analyzing legal references in contemporary Protestant sermons. The dataset consists of present-day Protestant sermons retrieved from widely used online sermon repositories; to date, we have collected 1,000 German-language sermons.