Mining Inter-Document Argument Structures in Scientific Papers for an Argument Web

Argument Mining for scientific papers has its roots in Argumentative Zoning [49]: classifying the relevant sentences from scientific articles into one of four categories (Claim, Method, Result, Conclusion). To the best of our knowledge, Sci-Arg [22] is the only dataset for Argument Mining in the scientific domain composed of fully annotated papers in English. This is unsurprising, given that annotating gold standard data is a resource-demanding task. The challenge is even exacerbated in the case of scholarly texts because they demand considerable domain expertise of the annotators [2].

For one of the few datasets [1] available, the authors annotated $60$ abstracts based on a subset of the SciDTB corpus [57]. They add an argumentation layer containing $352$ argumentative components, which are connected by $292$ argumentative relations, resulting in partly annotated papers from computational linguistics. Since abstracts generally do not contain references to other documents, they cannot be used for inter-document relation prediction.

Another dataset consisting of $24$ papers is presented in [20]. The articles from the domain of educational research have their introduction and discussion sections argument-annotated, notably including relations. However, this corpus has the reservation that it is comprised of publications in German, for which Argument Mining approaches are very limited. Instead, our work focuses on English data as it allows us to aim for a broader audience.

While not intended to be purely argumentative, the Citation Typing Ontology (CiTO) [46] allows for the classification of the nature of in-text citations as factual or rhetorical relationships. The latter has more fine-grained types such as supports for positive and refutes for negative, which directly correspond to their argumentative equivalents. Since CiTO assigns types to citations, these relationships inherently involve more than one document, making them inter-document. However, [46] presents no method for any (automated) annotations but solely an ontology. Furthermore, it also lacks a way to annotate intra-document relations since the entirety of CiTO targets citations.

The authors of [11] follow the relation-based Argument Mining paradigm [10]. Its goal is to predict the type of relations between text spans, such as attack, support, or neither. Using LSTMs [17] and GloVe embeddings [36], their approach is explicitly capable of classifying links between “any two texts,” which also implies inter-document settings. The results indicate that their neural network methods improve over traditional classifiers.

The notion of “intertextual correspondence” [53] introduces another idea to connect annotated corpora (i.e., their documents) by exploiting relations, for example, of topical or temporal nature. This can also lead to multi-modal datasets for Argument Mining purposes. However, such links need to be identified manually first. The effectiveness of the approach is demonstrated by fusing a corpus of the debates in the US election of 2016 with commentary and reactions on the online platform Reddit. Such techniques could also be applied to the Sci-Arg dataset with other media or documents, since this inherently would result in inter-document relations.

In [35], the authors investigate the impact of content and context on analyzing argumentative relations. They find that systems that focus too much on the text of the argument components for relation prediction may easily be deceived and make wrong predictions. For example, they may rely on discourse indicators contained in the two text fragments to be analyzed, even though they are not adjacent, and, thus, the discourse indicator does not apply. Therefore, the importance of the context of argumentative units is asserted [35]: the position in the text and the surroundings (pre- and succeeding tokens). They show that systems only relying on the context instead of the content (i.e., component text span) may improve accuracy. Finally, the authors argue that systems dissecting content and context should also be more adequate for handling inter-document relation prediction.

This work builds upon the vision of the Argument Web of Science [45]. While the notion of a scholarly knowledge graph is not novel, building one out of arguments extracted from scientific documents is. Given the prevalent role of argumentation in scientific communication [55], arguments are a suitable vehicle to represent scholarly information. Hence, we follow the Argument Web [39], which is based on Semantic Web knowledge representation technologies [8, 13].

The most prominent approach to a Knowledge Graph for Science has been described in [4]. Moving from document-centrism to a knowledge-based perspective would allow for a systematic organization of scholarly information. These efforts culminate in the Open Research Knowledge Graph (ORKG) [48, 5]: a representation that describes scientific papers in a structured way, facilitating question answering [33], paper comparisons, and visualizations. ORKG distinguishes itself from other large-scale efforts to represent scientific paper content, such as Microsoft Academic Graph (MAG) [56] (which is deprecated and succeeded by OpenAlex [38]), primarily in its granularity and explicit focus on structured, semantic content within papers. While MAG provided a vast bibliographic graph of papers, authors, and citations, ORKG aims to extract and structure the specific findings, methods, and results of scientific contributions, making the scientific content itself machine-readable and comparable. This deep semantic representation supports fine-grained analyses and direct comparisons of research outputs, unlike the primarily metadata-level focus of general bibliographic databases. However, ORKG uses a complex science ontology to represent the knowledge. While we share the goal of representing scientific information as a knowledge graph, our focus is different: we model arguments by means of their components (claims, premises) and two relationships connecting them (attacks, supports).

The idea behind Research Objects [6] contributes to Semantic Web efforts by defining a framework to preserve scientific workflows. It extends traditional workflow ontologies with metadata, annotations, provenance traces, and execution environments to improve reproducibility, reusability, and long-term preservation. A suite of ontologies provides structured descriptions of workflows and their evolution and aligns with other Semantic Web initiatives to structure scientific knowledge for better discovery and reuse. However, the focus is on Research Objects and their workflows, while this paper uses arguments as the core element of scientific knowledge representation.

Nanopublications [15, 21] aim to represent atomic units of knowledge with structured metadata and provenance information. This is to ensure trust, reproducibility, and interoperability. Hence, Nanopublications contribute to the Semantic Web’s support for the scientific process by providing a provenance-centric format for publishing scholarly assertions, enhancing machine-readable scientific communication. However, Nanopublications focus on self-contained statements with metadata such as provenance and publication information. Our approach, in contrast, models arguments explicitly by capturing claims along with their attacking or supporting premises. This makes the logical relationship between statements a central feature.

In the context of knowledge representation and the Semantic Web, the Provenance Ontology (PROV-O) [26] primarily focuses on describing the origin, history, and derivation of entities and activities. While argument relations certainly possess provenance aspects (e.g., who asserted a claim, when, and based on what evidence), PROV-O’s scope is distinct from the explicit modeling of types of argumentative connections (e.g., “supports”, “attacks”) between content units across documents. This is where the necessity of an Argument Web arises. Unlike general knowledge graphs that might link papers by co-authorship or topics, the Argument Web specifically aims to map the persuasive and confrontational dynamics across research. It addresses the limitations of single-document or citation-based views, particularly by capturing the complex inter-document support and attack relations that define scientific discourse. Our framework could be extended to integrate PROV-O for richer provenance tracking of the argumentative links themselves, but its core purpose lies in explicitly mapping the argumentative structure across documents, a functionality that goes beyond the scope of a pure provenance ontology. Our contribution lies in establishing the argumentative structure itself, which is a specialized form of inter-resource relationship not directly captured by general provenance models.

Lastly, SciHyp [52] introduces a fine-grained dataset representing hypotheses explicitly mentioned in scientific publications. It captures and structures scientific hypotheses from $479$ computer science papers, categorized into relation-finding and comparative hypotheses. The dataset was created using a hybrid human-AI pipeline, combining experts, LLMs, and crowd refinements. Furthermore, SciHyp also extends existing hypothesis ontologies to better model their components. Through extensive evaluation, the authors demonstrate that LLMs can assist in hypothesis detection and component extraction, while also showing that human intervention remains crucial for precision. While SciHyp focuses on hypothesis-driven research, this paper aims to enhance the scientific process with a data-centered approach based on scientific arguments.

Predicting relations between argument components is the most complex and challenging part of the Argument Mining pipeline [27, 3]. The goal is to identify related pairs of argumentative components and to classify the nature of this link.

We follow Opitz and Frank [35] and define argumentative relation prediction to be an irreflexive function as follows:

$C$ is the set of argumentative components. Each component $c\in C$ is associated with a specific document $doc(c)$ from which it originates. Components can be either a text span from within $doc(c)$ or a reference component to an external document. For a reference component $c_{ref}$, $doc(c_{ref})$ refers to the document containing the citation. The range $R$ of the function refers to the directed argumentative relation between the components. In this work, $R$ is the set $\{\text{attack, support, no-relation}\}$, where “attacks” indicates a clash, “supports” indicates backing, and “no relation” signifies the absence of an argumentative link between the given components. Applying this function creates (typed) directed edges between argument component pairs, thus allowing us to conceive an argument graph. Table 1 provides examples of component pairs with the relation that connects them.

There are multiple ways to represent an argument, which influences the sets of $C$ and $R$. In the context of this paper, we consider the following representation of an argument: the claim/premise model [54] with attack and support relations. In the context of this paper, the argumentative roles of components within $C$ are further categorized as claim or premise (also called data, evidence, or reason in the literature [27]). A claim is a proposition about a point the arguer tries to make, and a premise is a statement about the validity of a claim. In the context of $f(\cdot )$, $c_1$ is typically the premise and $c_2$ the claim, especially for the support relation. However, there is no formal restriction on the type of components, as a claim may also attack another claim. This follows the conventions set by BAM, our Benchmark for Argument Mining [44], and is grounded in the fact that many other (more detailed) argument representations can be simplified to this model and, thus, allows for a unified evaluation (even if this incurs some information loss).

Building upon this general definition of argumentative relation prediction, we further categorize relations based on the origin and nature of their components, distinguishing between intra-document and inter-document argumentative relations. We use the following definitions:

With the concept described in Definition 2, we can build an Argument Web as shown in Figure 1 since arguments can now be linked across multiple documents. Furthermore, using these definitions, we can form two (disjoint) subsets of relations for any dataset. It is important to clarify that this disjointness applies to the classification of argumentative relations based on the defined criteria of component origin and type (text span vs. reference to another document). This does not prevent the possibility of shared or overlapping content existing between documents that are involved in inter-document argumentative relations. Indeed, such shared content often provides the very basis for one paper to support or attack another. Our definitions categorize the type of argumentative relation established between components, rather than asserting that the underlying documents themselves are semantically or topically distinct. This ensures a clear distinction in how different types of argumentative interactions, within a document versus across documents, are formally represented and evaluated. Annotation examples from the Sci-Arg dataset [22] can be seen in Table 1 for the relation types to predict, as well as the two settings we distinguish.

We based the extended dataset on the freely available Sci-Arg corpus [22], which consists of $40$ computer graphics papers fully argument annotated, including components and relations. Sci-Arg was annotated by one expert and three non-experts in five iterations. The reported Inter Annotator Agreement (IAA) in terms of the $F_1$-measure was reported with two criteria. For the strict version, span and type have to match exactly for components, and for relations, the direction also has to be correct. In the relaxed version, components only have to correspond in type and match half the span. This results in a higher IAA for the relaxed criteria: $72\%$ for the components and $49\%$ for relations. Meanwhile, the strictly measured IAA is $60\%$ for the components and $35\%$ for relations. The new annotations and the code for the automated processes are available in the online repository.²²2https://gitlab.ifi.uzh.ch/DDIS-Public/midas

Figure 1 visualizes the concept behind the extension of the dataset, which is represented by the circle, adding new documents outside of it. From the Sci-Arg annotations, we identified relations that involved components, which are references to other documents. That is, we determined triples consisting of an argumentative unit annotated in the dataset, a relation (type), and an external paper acting as the respective endpoint. Crucially, for these inter-document relations, our annotation captures a document-level relationship, indicating that the referencing document supports or attacks the cited document as a whole, rather than specific claims or spans within it. Hence, we did not identify or annotate specific argumentative spans within the target (cited) papers. Instead, the argumentative relation type (supports or attacks) for these inter-document links was directly inherited from the original intra-document annotation in Sci-Arg. Specifically, if an argumentative component within one of the original $40$ Sci-Arg papers was annotated with a certain relation (e.g., “supports”) to another component, and that second component was identified as a citation to an external document, then that existing “supports” label was applied as the relation between the citing component and the cited document as a whole. Thus, the labels for inter-document relations originate solely from the high-quality, human-derived annotations within the original Sci-Arg corpus, ensuring their trustworthiness without requiring fine-grained annotation of the newly added documents. This process yielded $796$ unique external document references from the initial $40$ Sci-Arg papers that were part of annotated argumentative components. These references formed the basis for expanding our network of inter-document relations.

We assumed that if an in-text citation is annotated as an argumentative component, the original paper the citation is in reference to is meant to be used as the component. Therefore, we extracted all these components that contain a citation based on pattern matching (e.g., “Smith et al., 2000” or “[1]”).³³3The full list of patterns is available in Appendix A. We then manually verified that all the identified components were citations in their context by looking through the $40$ documents in the Sci-Arg corpus. This involved confirming the presence of standard citation formats (e.g., author-year mentions or numerical indices) that directly corresponded to entries in the papers’ bibliographies, ensuring the component’s accurate role as a citation to an external source.

Next, we matched the citations to their reference in the bibliographies in the papers. This was automated wherever possible. For the cases of ambiguity, occurring in approximately 10% of the instances, we determined the references by hand. These typically arose from multiple bibliography entries matching a single citation pattern, from unclear author-year combinations, or special characters not handled correctly during the content extraction. The issues were resolved by careful cross-referencing with the bibliography, the citing document’s content, and, when necessary, external academic search engines, following a predefined set of internal guidelines to ensure consistency.⁴⁴4https://gitlab.ifi.uzh.ch/ruosch/wp3/-/blob/main/data/guidelines.md This way, we could assign the new documents to the annotated argument components.

Then, we resolved the references to their Digital Object Identifiers (DOI) using Crossref’s API.⁵⁵5https://api.crossref.org These API calls were successful for approximately 90% of the identified references. For the remaining cases without a direct DOI, we followed the guidelines detailed in the online repository to derive a DOI or, when none was available, another unique, persistent identifier (e.g., arXiv ID, official publisher URL) that ensured the consistent and unambiguous identification of the target paper. With these, we could now obtain the content of papers as PDFs, if accessible, from which we extracted information about their content and structure using Papermage [29]. We successfully obtained PDF content for $649$ of the $796$ new papers (approximately $81.5\%$ of the newly added documents). From these PDFs, Papermage extracted key structural elements such as paragraphs, sentences, and headings, making their content programmatically accessible for potential future use, even though fine-grained annotation of these new documents was not within the scope of this initial extension.

It is crucial to clarify that for the current scope of this dataset extension, the detailed full-text content extracted by Papermage from these additional $649$ papers was not directly used to determine the argumentative relation types (supports/attacks) between documents. These relation types were derived by performing lookups in the original, human-annotated Sci-Arg data, where the citation acted as an argumentative component. Therefore, the content acquisition and extraction primarily served to identify and confirm the existence and structure of the cited documents and to build a resource for future, more granular analyses. While this approach might be considered “shallow retrieval” in terms of not using the full text for the initial relation labeling, it was a deliberate and necessary choice for the initial, broad-scale expansion of the Argument Web. It allowed us to efficiently establish a foundational network of document-level argumentative links. The availability of this extracted full text for a large portion of the new documents will be crucial for future work aimed at developing more sophisticated, content-based methods for fine-grained inter-document argument mining and automatic relation prediction, moving beyond simple citation-based links.

To finalize the new annotations for the intra- and inter-document argumentative relations, we needed to complete the triples with their predicate (relation type) and assign them to either of the two disjoint subsets. For this, we performed lookups in the original Sci-Arg data using the two component identifiers to get the relationship label connecting the two. Then, we added these triples to their corresponding subset based on Definitions 1 and 2: if both components were from the same document, they were classified as intra, and inter, otherwise.

As a result, we had a new dataset with two additional aspects. First, we added the distinctive notion of intra- and inter-document argumentative relations, producing annotations of these two disjoint subsets. This allows us to evaluate relation prediction approaches in the two settings independently, potentially leading to more in-depth analyses of how different methods work for them. Second, and more importantly, we augmented the original $40$ papers by including an additional $796$ new papers, extending the dataset to 836 total papers.

It is crucial to clarify that this expansion primarily involved identifying external papers cited by the original 40 and determining document-level relationships to them. Thus, while this represents a twentyfold increase in the number of documents, the detailed, claim-level annotations present in the original 40 papers were not replicated for the additional $796$ documents. This targeted approach required substantially less effort per additional document, allowing us to broaden the scope of inter-document relations efficiently. Even though we do not currently make use of the full text of the additional papers for detailed internal annotations, their inclusion is vital. These $796$ documents serve as essential nodes within the expanded inter-document argumentative graph, providing a richer, more representative context for studying cross-document argumentation. They enable the analysis and prediction of how arguments originating from the original $40$ papers relate to a much wider body of scientific literature, making them valuable for understanding the broader Argument Web.

Additional statistics about the original and the extended dataset can be found in Table 2 (the class distributions are shown in Appendix F). The number of total relations remains unchanged in the extension because these $1996$ relations represent all argumentative links identified within and originating from the original $40$ Sci-Arg papers. Our extension primarily involved: 1) identifying which of these existing relations were inter-document (i.e., involved a reference to an external paper), and 2) adding the external documents themselves as nodes to the Argument Web to serve as endpoints for these newly categorized inter-document relations. We did not perform new fine-grained argumentative annotations within the 796 newly added papers. Instead, the total number of relations reflects the comprehensive set of argumentative links present in the original Sci-Arg corpus, now re-categorized and contextualized within a larger network of documents.

We determined $568$ relations that go across document boundaries and $1428$ within documents. If we only allow one relation per direction and type for each pair by consolidating relations where one paper is used to support or attack at multiple locations in another, we reduce the annotations to the document level. This way, we maintain a graph with $836$ nodes and $1109$ edges (i.e., the number of doc-level relations).

We publish these document-level annotations as Resource Description Framework (RDF) data since Semantic Web technologies provide the means to represent machine-readable data that is easy to integrate with other data despite heterogeneity.1 This is particularly useful in this context, as there exists a variety of argument representation models with different levels of granularity [27]. To model the data, we used the CrowdAlytics ontology, which extends other ontologies and integrates ontological components to describe scientific hypotheses and arguments present in scientific publications [51]. Even though we produced an Argument Web, the inter-document relations are indeed only on the document level.

This document-level granularity for inter-document relations was a deliberate design choice, allowing us to scalably extend the dataset’s scope to hundreds of external documents. Unlike the fine-grained, claim-level annotations within documents, identifying specific span-level argumentative connections across disparate documents proved prohibitively complex and resource-intensive for manual annotation, so we only resolved the references to the documents and did not determine the more fine-grained details of which exact proposition the reference is to. This is visualized in Figure 1 by the arrows representing the relations only pointing to the documents and lacking detail outside the circle representing the original corpus. The implication of this design is that our dataset provides a foundational macro-level view of argumentative connections between papers, capturing their overall support or attack roles. While this work focuses on these document-level interactions, future research can leverage this foundation to develop automated methods for identifying more granular, span-level inter-document argumentative links.

For consistency, we thus make this difference in the relations evident by altering their labels slightly. Supports becomes is used as support in, and attacks turns into is used as attack in. This emphasizes the difference between argument component pairs where both are well-specified propositions as opposed to those consisting of one text span and one reference to a different document. Semantic Web technologies facilitate the alignment of the two different granularities of annotations in the same dataset.

This subsection describes the three approaches used to create the baseline for the new dataset. Furthermore, we also explain the evaluation methods that we apply. An overview is shown in Figure 2: the inputs on the left-hand side, the three approaches in the middle, and the evaluation on the right. The first two approaches only receive the two components to predict the relation, while the bottom one also has access to the full text. They are all evaluated by the same methodology to produce comparable results in the F1-score.

Table 3 shows the results for the approaches on the overall dataset. A class-based evaluation can be found in Appendix F. Furthermore, we show more detailed statistics on the distribution of the misclassified samples per relation type of each system in Appendix G.

With a relation prediction score of $0.437$, the naive approach, which leverages only the presence of discourse indicators, clearly – and statistically significantly – outperforms the other more sophisticated neural methods. This may be surprising, but it can be put into perspective with the following three points. First, discourse indicators have been shown to work well for predicting argument relations [25]. Second, it was the only technique that had access to not only the content but also to the context of the argument components. This has been noted to contribute to accuracy positively [35]. Still, the only influence the access to the context had on the approach was that the discourse indicator could be contained in either the sentence holding the components or in the components themselves. Finally, even though the rule-based approach achieved the highest score, the result is nowhere near where it could be, as the $0.437$ score clearly indicates significant room for improvement.

Scoring $0.251$, the transformer-based Argument Mining system [31] came in second place and considerably ahead of the LLM approaches. This result is consistent with the outcome for TRABAM in the original showcase of BAM [44]. Even with the training on the dataset, it appears the system still fails to predict most of the relations for argument component pairs correctly.

The results of the LLM-based approaches are nuanced. Both the zero-shot and few-shot methods show a statistically significant improvement in their mean F1-scores when provided with context (p=0.008 and p=0.012, respectively). However, the effect size of this improvement is markedly different. For the zero-shot LLM, the improvement is negligible ($d$=-0.173), while for the few-shot LLM, the effect is medium ($d$=-0.618). This finding corrects our initial hypothesis, as it indicates that the information from semantically similar examples is effective at improving accuracy, but only when combined with the contextual information from the surrounding sentences.

When we examine the performance with context more closely, the few-shot LLM achieves the best F1-score among the LLM variants ($0.132$). However, this improvement comes with a trade-off: while its recall increases substantially ($0.095\to 0.217$), its precision decreases ($0.141\to 0.099$). This behavior indicates that providing context and few-shot examples enables the model to identify more potential relations but at the cost of generating more false positives. This finding confirms that while LLMs are sensitive to contextual information, they still struggle to accurately and precisely discern the correct relations, particularly without extensive fine-tuning.

A key observation from our class-based results (cf. Appendix F) is the significant disparity in performance between “supports” and “attacks” relations, where models, particularly the LLM-based approach, consistently exhibit much higher F1-scores for “supports”. This phenomenon can be attributed to a combination of factors. Firstly, the inherent class imbalance in our dataset, as detailed in Subsection 3.2 (Table 2), means “supports” relations are substantially more frequent than “attacks”. This naturally biases models towards the majority class. Secondly, identifying “attacks” relations is often an intrinsically harder task in argument mining. They frequently involve more nuanced linguistic cues, require deeper semantic understanding of contradiction or refutation, and may manifest in more diverse textual patterns compared to expressions of support. The confusion matrices (Appendix G) further illuminate this, revealing instances where models tend to conflate “attacks” with “no relation” or even incorrectly classify them as “supports”, highlighting the challenge in accurately discerning these critical dissenting links.

The results for the intra- and inter-document settings are shown in Table 4, revealing that all systems achieve higher scores for the latter. We make pairwise comparisons to see whether these numbers are statistically significantly different for F1.

This is neither the case for the rule-based nor for the few-shot LLM approach. The two represented the two different ends of the scale of the results. The former is at the top ($0.430$ and $0.432$), and the latter is at the bottom ($0.091$ and $0.136$). However, they do not exhibit statistically significantly different scores for the argumentative relation prediction in the intra- and the inter-document settings.

Still, the results of the rule-based approach give insights. They indicate that discourse indicators in scientific papers are a simple but reliable signal for predicting argumentative relations in the two situations. This is shown in both settings by the higher precision than recall of the method, confirming the findings in [24].

Meanwhile, the transformer-based argument miner [31] achieves a distinctly higher score in the inter-document ($0.345$) than in the intra-document ($0.238$) setting. The difference has statistical significance, suggesting that it is better at predicting this type of argumentative relationship. The precision and recall values, with the former being higher than the latter, also indicate that the AM system is able to pick up on the cues in the content of the components. However, it fails to capture most of the argumentative relations.

For the zero-shot LLM approach, the statistically significant difference in the relation prediction scores is even larger: $0.083$ and $0.249$, respectively. The very high recall values stand out for both settings, indicating that it catches many of the argumentative relations to predict, but also produces a lot of garbage annotations, considering the very low precision. The analysis for the few-shot LLM approach is the same as that for the overall dataset. It is inadequate for predicting argumentative relations either way, with some of the lowest F1, precision, and recall scores across the board. Since these systems are based on black-boxes that are the neural network architectures, we can only speculate about the reasons for the disparities. For some LLMs, we have knowledge about the data involved in their pre-training, while others may be completely closed off. However, given that their training involved coherent and, most likely, academic text, we can assume that their capabilities involve recognizing argumentative components such as claims and premises but struggle to pick up on cues for the more complex task of argumentative relation prediction. More investigations in this direction are clearly necessary, also when taking the difficulties of effective prompting into account.

For both the argument miner and the LLM, the results may seem counterintuitive when comparing the relation prediction scores to those of the naive rule-based approach. The components in the inter-document setting tend to contain less semantic and syntactic information, as at least one of them is a reference to another document: a citation. In most cases, that does not give any details about the referenced document apart from possibly author names and the publication year. Neither of which carries significant information without more semantic context. Therefore, the only conclusion is that for the inter-document setting, the complementary component of the pair (i.e., not the citation) contains particularly useful information for predicting the relation. Surprisingly, this would also imply that the cues for the inter-document argumentative relation prediction from only one component can be leveraged more effectively than those from two components in the intra-document setting, resulting in higher scores for the former.

The LLMs show a varied response to the addition of context. The few-shot LLM with context shows no statistically significant difference between intra- and inter-document performance. In contrast, the zero-shot LLM with context exhibits a large and statistically significant difference between intra- and inter-document F1 scores. This indicates that while context helps both approaches, the few-shot learning method, when combined with context, is better able to adapt and apply its learned knowledge more consistently to both intra- and inter-document relationships, reducing the performance gap between the two, as was also the case without context.