Accessing research and innovation information increasingly requires efective retrieval across languages, document types, and levels of textual granularity. In many research ecosystems, content is inherently multilingual and queries are short, concept-driven, and underspecified, posing challenges for traditional lexical retrieval methods, while performance of general-purpose dense retrieval is limited. In this work, we present an empirical evaluation of multilingual dense retrieval for scholarly documents in Catalan, Spanish, and English. We analyse the behaviour of general-purpose and domain-adapted embedding models across monolingual and cross-lingual settings, query types, and query lengths, and compare dense retrieval against strong sparse baselines. Using weakly supervised query-passage and triplet datasets derived from open research information, we show that domain-specific multilingual fine-tuning substantially improves retrieval efectiveness, semantic alignment, and embedding coherence. Our results highlight the importance of domain and multilingual adaptation for robust scholarly information access. These capabilities are particularly important for research mapping and scientometric analysis tools, where retrieval quality can directly influence downstream analytical modules such as topic and collaboration analysis, or research portfolio mapping.
Scientific and technical information is increasingly available through open databases of research projects,
scholarly publications, and patents [
These documents form the basis of research mapping platforms [
The aim of research mapping platforms often goes beyond traditional document ranking and relevance, because search results are also commonly used as input to analytical modules aimed at understanding scientific specialisation, organisational performance, or thematic trends across research portfolios. In this sense, retrieval can be interpred as a form of classification. Users frequently issue brief queries such as ’cancer’, ’artificial intelligence’ or ’blue economy’, expecting to retrieve relevant documents that may be written in diferent languages and described using diverse and more specific scientific terminology. These fine-grained scientific concepts are the basis of scientific queries [ 12]. However, the major dificulty in scholarly information retrieval could be the knowledge behind the words which is expected to be known or understood [13], especially relevant for those queries that are not self-descriptive or only known by domain experts.
While traditional keyword-based information retrieval systems can handle lexical representations [15], they fail to capture the semantic relationships and contextual meaning that characterise modern scientific language—for example, they cannot recognise that ’oncology’ and ’cancer research’ refer to related concepts, or that a query in Spanish should match an English abstract on the same topic, or to process more complex queries like ’AI for energy transition’. Recent advances in multilingual large language models have enabled more natural and concept-oriented interaction with textual databases through semantic search [16]. Embedding-based retrieval [17, 18] represent search queries and documents in a shared semantic space, supporting retrieval beyond exact word matches, allowing proper semantic search. Sentence encoder models are increasingly used from scholarly RAG systems [19], to modern scientometrics topic modelling approaches [20, 21]. While language models are trained to represent context, concept-based queries (e.g. ’cancer’) lack of suficient context to produce informative dense representations.
However, semantic search systems often rely on similarity thresholds or ranking signals derived from embedding similarity, and general-purpose embedding models may introduce biases that afect scholarly retrieval efectiveness. Similarity scores in this analysis are computed using cosine similarity between query and document embeddings produced by SentenceTransformers [17]. As shown in Figure 1, despite being multilingual, the model displays cross-lingual diferences in similarity distributions. In this case, documents written in the query language (English, in this case) tend to receive higher similarity scores, while texts in less-represented languages, such as Catalan, obtain lower scores. While retrieval metrics such as Recall@k and MRR depend on the relative ranking of documents rather than absolute similarity values, systematic shifts in similarity distributions may still influence ranking outcomes when relevant documents consistently receive lower similarity scores than competing candidates. In addition, similarity scores correlate with passage length, with shorter texts (e.g., titles) often ranked ahead of longer passages regardless of their semantic completeness given a concept-based query. This behaviour is something we have observed in the practical development of dense retrieval search systems with industry-standard models, as well as narrow similarity window discrimination for scientific documents with general-purpose models. Although these behaviours are partly expected, this issue is particularly relevant in open scholarly knowledge graphs, where abstract availability has recently decreased due to publisher restrictions [22].
Our goal is to evaluate and improve dense retrieval models for multilingual documents by adapting models with: • scientific domain knowledge, • multilingual alignment across Catalan, Spanish, and English, • ranking-oriented behaviour for search, • fine-grained cosine separation to support classification and analytical tasks.
Our contribution is primarily empirical rather than architectural. Instead of proposing new training
objectives or model architectures, we evaluate multilingual semantic search for scholarly publications
and projects, analysing how general-purpose embedding models behave across languages, passage
lengths, and similarity ranges, and how they can be adapted to the scientific multilingual domain.
This study builds on our previous work on multilingual semantic retrieval and query segmentation
for scientific information access, expanding both evaluation and experiments. The development of
these models is in the context of building new text search capabilities for open research information
platforms in Catalonia, with tools like RIS3MCAT [
These capabilities are directly relevant to scientometric and research policy analysis. Research mapping platforms such as RIS3MCAT integrate search with analytical modules that support the exploration of research portfolios, collaboration networks, thematic specialisation, and funding impact. In such systems, retrieval acts as a filtering and classification step that determines which documents are included in downstream analytical workflows. Consequently, the quality of semantic retrieval directly afects the reliability of scientometric analyses derived from these platforms.
Scholarly information access research has addressed the challenge raised by the rapid growth of
scientific literature, by exploring how to address information needs of researchers [ 23], developing
recommendation systems [24] and discovery tools [25], and using publication textual content and
metadata [26]. Semantic similarity search is dominated by dense retrieval methods, which encode
queries and documents into a shared embedding space and rank candidates by vector similarity [27, 28].
These approaches enable concept-level search beyond lexical overlap, but also face known limitations
[29] in capturing rare terminology and fine-grained distinctions, issues especially relevant for scientific
and multilingual contexts. However, several studies show that dense retrieval does not consistently
outperform strong lexical baselines as BM25 [
A key challenge for dense retrieval of scholarly documents is the lack of annotated data for training
and test. Creating supervised query-passage pairs for scientific literature is costly and generally requires
domain experts [12]. To address this gap, prior studies have explored the challenge of generating
training labels with unsupervised and weakly supervised approaches [
The sentence-transformers framework [17] provides a widely adopted pipeline for training dense
retrievers, typically using Multiple Negatives Ranking Loss (MNRL) [
In the multilingual domain, several model families are particularly relevant. The multilingual E5
models [
Dense retrieval also plays a central role in retrieval-augmented generation (RAG) frameworks,
improving factual accuracy for LLMs [
This section describes the methodology used to evaluate and adapt dense retrieval models for
multilingual scholarly search, with a focus on cross-lingual alignment, ranking behaviour, and sensitivity to
passage length. Following prior studies [
Base Models. We evaluate a set of multilingual or scientific-domain sentence encoder models:
• Multilingual E52 [
We build several weakly supervised datasets from openly available scholarly collections of publications and projects to support training, evaluation and analysis.
Trilingual Research Project Corpus.
This is a dataset of 1.5K publicly funded research projects and is used to analyse retrieval behaviour
across languages and document granularities. It consists of 500 English projects from the European
Commission’s CORDIS platform6, 500 Catalan projects from RIS3CAT[
Query-Passage Dataset.10 Our primary training dataset comprises 76k query-text pairs, equally distributed across English, Catalan, and Spanish. We collect 30K scientific publications in English from several bibliographic databases 11, extracting their titles, abstracts and author keywords. Individual author keywords are treated as queries, while titles and abstracts serve as passages. To obtain multilingual supervision samples, all textual ifelds (author keywords, titles and abstracts) are automatically translated into Catalan and Spanish using machine translation system, using the Google Translate API. The original and translated texts are then aligned to construct both monolingual and cross-lingual query-passage pairs across the three languages. There are no repeated articles between languages. The dataset contains both monolingual and cross-lingual pairs. Approximately 90% of the examples correspond to keyword→text pairs (where the text may consist of the title, abstract, or title+abstract), reflecting the typical use of short concept queries in scholarly search, and the missing abstracts for some records. The remaining 10% consist of title→abstract pairs to preserve document-level semantic similarity during training, and for textual equivalent searches. While automatic translation may introduce some noise or semantic drift, keywords are typically short technical terms, which reduces the likelihood of substantial translation errors. Contrastive training has been shown to be robust to moderate noise in supervision signals. The overall data construction process is illustrated in Figure 2. This dataset can therefore be considered a weakly supervised resource, where author keywords act as implicit relevance signals. In a low-resource multilingual setting, this approach enable the generation of cross-lingual training pairs with a low investment of resources. The dataset is split into 80/10/10 partitions. Splitting is performed at pair level, ensuring that each query-passage pair appears in only one partition. Because queries correspond to author keywords representing scientific concepts, the same query term (e.g. “cancer”) may occur across splits paired with diferent passages, while longer and more specific queries appear less frequently. This setup reflects realistic retrieval scenarios where common concepts may correspond to many diferent documents.
Classification Dataset. 12
We additionally use the classification dataset scidocs-mag [
In order to assess diferent fine-tuning strategies on our query–passage dataset, we experimented with diferent dataset configurations and loss functions. This allows us to analyse how training objectives influence multilingual alignment and ranking behaviour.
Loss Functions. dense retrieval.
We evaluate four complementary loss functions [17] commonly used in
• Multiple Negatives Ranking Loss (MNRL) [
Training Setup. Models are fine-tuned using SentenceTransformer framework [17]. Training is performed for three epochs with identical optimisation setting across models, and evaluated on held-out test data. Detailed hyperparameter settings are reported in Appendix B.
We explore model retrieval capacity, and analyse performance in monolingual and cross-lingual settings, explore the impact of query length, as well as model behaviour for lexical and semantic queries. We evaluate on 5 example queries performance of sparse and dense retrieval and explore how to choose the best similarity thresholds for dense retrieval.
Models are evaluated on the held-out multilingual test split using the following metrics:
• Top- Recall ( ∈ {1, 5, 10}): proportion of queries for which the paired passage appears among
the top- retrieved results.
• Cosine@1: average cosine similarity between positive query–passage pairs, measuring
embedding alignment quality.
• Mean Reciprocal Rank (MRR): evaluates the ranking quality by measuring the inverse rank of
the first correct passage within the top-10 retrieved candidates.
• Neighbourhood Polarity: following [
Table 1 reports retrieval performance of the base and fine-tuned embedding models on our multilingual Query-Passage Dataset test split. We report R@k, MRR, cosine accuracy, and neighbouring polarity. To ensure robust evaluation given the weakly supervised construction of the dataset, retrieval is performed over batches of 64 candidate documents. Because supervision signals are derived from author keywords, relevance annotations are incomplete, given that a document may be relevant to a query even if it is not paired in the dataset. This setting reduces false positives arising from semantically related but non-paired samples. Evaluating over the full corpus would introduce several false negatives and artificially penalise correct semantic matches. Restricting the candidate pool allows us to measure the capacity of the model to distinguish relevant passages from semantically related distractors while mitigating noise from incomplete supervision.
In addition, we also treat as valid positives any passages associated with additional author keywords from the same source document, reflecting the many-to-many nature of author keywords, where a single document may correspond to multiple conceptually related queries. The neighbouring polarity score derived from the classification dataset, measures whether the top- neighbours (here, = 16) share the same scientific field. This provides an external estimate of whether fine-tuned models preserve meaningful semantic structure across scientific domains.
Comparing loss functions, MNR loss consistently provides the largest gains, reflecting its explicit optimisation of relative similarity among in-batch candidates for retrieval. The results indicate that domain-specific and multilingual enrichment significantly improves both ranking and semantic organisation of the embedding space. While E5 achieves the strongest overall performance, all models benefit from fine-tuning with MNR loss, including encoders originally trained only on English data. In contrast, neighbourhood polarity improves to a similar extent across all loss functions, suggesting that most objectives encourage comparable levels of inter-document semantic cohesion, even when the improvements in retrieval are smaller.
In the following sections, we focus on the best-performing fine-tuned models, those with MNR loss, and analyse their behaviour in more detail. 4.2.1. Performance by Query Type To better analyse retrieval behaviour, and to answer the question of how well dense retrieval preserve lexical retrieval capacities, we further distinguish between lexical and semantic query-passage matches. A pair is considered lexical when the query string appears verbatim in the passage (e.g., query “cancer” and passage containing “breast cancer”). Otherwise, it is labelled as semantic when retrieval success requires conceptual inference or paraphrasing (e.g., query “cancer” and passage mentioning “basal cell carcinoma”). This classification is language-agnostic: cross-lingual pairs are still considered lexical if translated forms match directly. Table 2 reports Recall@1 and Recall@10 across lexical and semantic Model – Base Models E5 mRoBERTA DistilRoBERTa Specter – Fine-tuned with ContrastiveLoss E5 mRoBERTA DistilRoBERTa Specter – Fine-tuned with CosineSimilarityLoss E5 mRoBERTA DistilRoBERTa Specter – Fine-tuned with MNRLoss E5 mRoBERTA DistilRoBERTa Specter – Fine-tuned with TripletLoss E5 mRoBERTA DistilRoBERTa
Specter matches, comparing base models and MNRLoss finetuned. Base models show a pronounced gap between lexical and semantic performance, indicating a strong reliance on surface-level term overlap. Fine-tuning with MNRLoss substantially improves retrieval performance for both match types, with lexical and semantic recall doubling in most cases. These results suggest that training on a mixture of lexical and semantic query–passage pairs strengthens both exact-match sensitivity and deeper semantic generalisation.
MRR .70 .59 .58 .47 Model R@1 Base Models Lex.| Sem.
E5 .48 | .34 mRoBERTA .29 | .27 DistilRoBERTa .37 | .29 Specter .21 | .19 Fine-tuned Models E5 .79 | .64 mRoBERTA .69 | .54 DistilRoBERTa .66 | .50 Specter .63 | .51
R@10 Lex.| Sem.
.83 | .73 .68 | .68 .68 | .62 .54 | .55
Model Base Models E5 mRoBERTA DistilRoBERTa Specter Fine-tuned Models E5 mRoBERTA DistilRoBERTa Specter
R@1 Mono.| Cross.
.54 | .27 .31 | .25 .40 | .25 .25 | .15 4.2.2. Performance by Language Configuration To analyse the impact of multilingualism on retrieval quality, we evaluate the models separately under monolingual and cross-lingual pairs. In the monolingual scenario, queries and passages are written in the same language, while the cross-lingual scenario contains pairs where the query and the target text are in diferent languages. This distinction allows us to measure both in-language semantic retrieval and the ability of models to align concepts across languages. Table 3 reports Recall@1 and Recall@10 under both conditions for all base and fine-tuned models. We observe how in fine-tuned models the gap between monolingual and cross-lingual performances are reduced considerably. 4.2.3. Monolingual Performance by Language and Query Length We further analyse monolingual retrieval performance by grouping test pairs according to the language of the target passage. This analysis examines how models handle scientific text in each language independently, isolating retrieval accuracy from cross-lingual alignment efects. Table 4 reports Recall@1 and Recall@10 for all models across the three languages. We observe the English as the dominant language, likely due to a major representation in training datam. However, this gap is reduced after ifne-tuning, by most models Catalan is the worst, possibly due to representation and resource availability. Finally, we analyse in Table 5 retrieval performance in function of query length. Queries are grouped into three categories: short (single token), medium (2-3 tokens), and long ( 4 tokens). This captures the efect of some of the challenges of concept-driven keyword searches, comparing from no context to more descriptive queries. Across all models, fine-tuning yields substantial improvements for all lengths of queries, indicating enhanced robustness to limited contextual information.
Model R@1 Base Models CA| EN| ES E5 .45 | .62 | .54 mRoBERTA .26 | .35 | .32 DistilRoBERTa .31 | .59 | .29 Specter .18 | .34 | .21 Fine-tuned Models E5 .70 | .79 | .75 mRoBERTA .60 | .67 | .67 DistilRoBERTa .57 | .73 | .61 Specter .55 | .71 | .59
R@10 CA| EN| ES .86 | .90 | .89 .66 | .75 | .69 .66 | .84 | .68 .54 | .70 | .55 .78 .69 .68 .68
Long .70 .43 .55 .39
To compare dense retrieval with well-established sparse methods, we suggest and conduct a small-scale analysis using exact keyword matching and BM25 as baselines. We evaluate five representative, conceptdriven queries we have annotated over the trilingual research project corpus, spanning well-established scientific topics, emerging policy-oriented concepts, and semantically complex queries that are not always lexically explicit in project descriptions. Table 6 reports Precision@10 across retrieval methods. While BM25 provides a strong and stable baseline, particularly for topics with consistent terminology, base embedding models do not consistently outperform it. In contrast, fine-tuned embedding models, especially E5, achieve higher precision across all queries, with the largest gains observed for conceptdriven queries of diferent natures. This is a small-scale analysis, but it would be interesting to analyse more deeply going forward, also as query routing strategies. Method Exact Keyword Match BM25 E5 (base) mRoBERTa (base) DistilRoBERTa (base) Specter (base) E5 (ft) mRoBERTa (ft) DistilRoBERTa (ft) Specter (ft)
r ’NuEcnleeargy’ ’SustaFinoaobdl’e ’EBclouneomy’
A key challenge in dense retrieval is determining an appropriate relevance threshold on cosine similarity, particularly when retrieval outputs are used for analytical tasks. Unlike ranking-based evaluation, these applications require a binary decision on document relevance, making threshold selection both critical and model-dependent. To address this question, we leverage the classification dataset to estimate cosine similarity thresholds that maximise the F1 score on the test set. We report, in Table 7, average optimal threshold and corresponding F1 score across 19 subject categories, providing practical orientation and empirical guidelines for selecting similarity thresholds under diferent embedding models.
Model E5 mRoBERTA DistilRoBERTa Specter
Base Threshold .79 .70 .21 .71
To present the efect of threshold selection, we present an example in Figure 3, which visualises cosine similarity distributions with True/False samples for a representative query (Environmental science) under the base and fine-tuned E5 models. The histogram highlights how fine-tuning increases the separation between relevant and non-relevant documents, leading to higher F1 score.
Across all experiments, fine-tuning consistently improves retrieval quality for every model and evaluation setting. Gains are especially pronounced in cross-lingual retrieval, where base encoders struggle to align Catalan, Spanish, and English scientific content. Models show 20–30 point improvements in Recall@1 after contrastive fine-tuning, confirming that domain-specific multilingual adaptation is important for multilingual scientific search.
Improvements extend across match types, while lexical queries are naturally easier, fine-tuning also boosts semantic retrieval capacity, demonstrating that the models learn to generalise beyond surface forms. Importantly, dense models do not lose lexical retrieval capacity, fine-tuning strengthens both lexical and semantic abilities. Even weaker base encoders become competitive multilingual retrievers after adaptation. Monolingual performance by language shows clear asymmetries that reflect underlying resource availability. English remains the easiest setting, with the highest scores even before adaptation. Catalan and Spanish lag behind in base models, particularly Catalan, which sufers from limited representation in openly available corpora. After fine-tuning, however, these gaps narrow substantially: Catalan gains the largest relative improvements, and Spanish reaches parity with English in some models.
These results show that the combination of multilingual contrastive learning and modest domainspecific supervision yields robust multilingual and cross-lingual semantic search capabilities—crucial for accessing R&I information in ecosystems where English, Spanish, and Catalan coexist. These models are able to improve upon strong sparse retrieval baselines. Finally, our analysis highlights that efective scholarly retrieval requires not only strong ranking performance but also interpretable similarity scores. The use of classification dataset threshold identification provides guides for bridging retrieval and analytical applications. Because from a scientometric perspective, improving retrieval quality is critical for research intelligence platforms rely on search results as input for analytical modules and facets that compute indicators such as thematic specialization or collaboration networks. Reliable multilingual retrieval therefore supports more accurate mapping and monitoring of research ecosystems.
In this work, we examine the performance of multilingual embedding models for accessing scientific and innovation data in a trilingual setting characteristic of many R&I information systems. Our results demonstrate that lightweight and domain-adapted models, including Catalan-centric variants, can efectively adapt to domain-specific data. Beyond findings, we contribute new multilingual datasets, model checkpoints, and evaluation resources designed to support future research on cross-lingual scientific information access. Taken together, our work underscores the importance of domain-specific adaptation and robust multilingual alignment for enabling reliable and scalable access to open research information. Beyond improving document retrieval, these capabilities are also relevant for scientometrics and research analysis, where semantic search systems are used to identify and analyse resarch portfolios and collaboration patterns.
During the preparation of this work, the authors used the following generative AI tools and services: ChatGPT, Claude, DeepL, and LanguageTool. These tools were used exclusively to support writingrelated tasks, including grammar and spelling checking, paraphrasing and sentence rephrasing, and general proofreading of the manuscript. In addition, generative AI tools were used for assistance in code development, documentation, and testing during the preparation of experimental scripts. After using these tools/services, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content.
Supported by the Industrial Doctorates Plan of the Department of Research and Universities of the Generalitat de Catalunya, by Departament de Recerca i Universitats de la Generalitat de Catalunya (grant reference 2022/DI /00017).
We thank the anonymous reviewers for their constructive feedback and suggestions, which improved the clarity and quality of this work. [12] Y. Zhang, R. Yang, S. Jiao, S. Kang, J. Han, Scientific paper retrieval with llm-guided semantic-based ranking, arXiv preprint arXiv:2505.21815 (2025). doi:10.48550/arXiv.2505.21815. [13] C. Friedman, P. Kra, A. Rzhetsky, Two biomedical sublanguages: a description based on the theories of zellig harris, Journal of biomedical informatics 35 (2002) 222–235. doi:10.1016/ S1532-0464(03)00012-1. [14] L. Wang, N. Yang, X. Huang, L. Yang, R. Majumder, F. Wei, Multilingual e5 text embeddings: A technical report, arXiv preprint arXiv:2402.05672 (2024). doi:10.48550/arXiv.2402.05672. [15] S. Robertson, H. Zaragoza, The probabilistic relevance framework: BM25 and beyond, volume 4,
Now Publishers Inc, 2009. doi:10.1561/1500000019. [16] A. Biswal, L. Patel, S. Jha, A. Kamsetty, S. Liu, J. E. Gonzalez, C. Guestrin, M. Zaharia, Text2sql is not enough: Unifying ai and databases with tag, arXiv preprint arXiv:2408.14717 (2024). doi:10.48550/arXiv.2408.14717. [17] N. Reimers, I. Gurevych, Sentence-bert: Sentence embeddings using siamese bert-networks, in: Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing, Association for Computational Linguistics, 2019. URL: https://arxiv.org/abs/1908.10084. [18] A. Singh, M. D’Arcy, A. Cohan, D. Downey, S. Feldman, Scirepeval: A multi-format benchmark for scientific document representations, in: Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 2023, pp. 5548–5566. doi:10.18653/v1/2023.emnlp-main. 338. [19] M. D. Skarlinski, S. Cox, J. M. Laurent, J. D. Braza, M. Hinks, M. J. Hammerling, M. Ponnapati, S. G.
Rodriques, A. D. White, Language agents achieve superhuman synthesis of scientific knowledge, arXiv preprint arXiv:2409.13740 (2024). doi:10.48550/arXiv.2409.13740. [20] A. Glazkova, Identifying topics of scientific articles with bert-based approaches and topic modeling, in: Pacific-Asia Conference on Knowledge Discovery and Data Mining, Springer, 2021, pp. 98–105. doi:10.1007/978-3-030-75015-2_10. [21] N. Bovenzi, N. Duran-Silva, F. A. Massucci, F. Multari, J. Pujol-Llatse, Mapping sti ecosystems via open data: overcoming the limitations of conflicting taxonomies. a case study for climate change research in denmark, in: International Conference on Theory and Practice of Digital Libraries, Springer, 2022, pp. 495–499. doi:10.1007/978-3-031-16802-4_52. [22] B. Kramer, More open abstracts? comparing abstract coverage in crossref and openalex, 2024. URL: https://doi.org/10.5281/zenodo.11580550. doi:10.5281/zenodo.11580550. [23] I. Frommholz, P. Mayr, G. Cabanac, S. Verberne, Bibliometric-enhanced information retrieval: 14th international bir workshop (bir 2024), in: European Conference on Information Retrieval, Springer, 2024, pp. 442–446. doi:10.1007/978-3-031-56069-9_61. [24] S.-Y. Yang, C.-L. Hsu, S.-H. Lu, Developing an ontology-supported information recommending system for scholars, in: 2009 Joint Conferences on Pervasive Computing (JCPC), 2009, pp. 223–228. doi:10.1109/JCPC.2009.5420185. [25] S. Volkova, P. Bautista, A. Hiriyanna, G. Ganberg, I. Erickson, Z. Klinefelter, N. Abele, H.-T. Kao, G. Engberson, Cross-disciplinary knowledge retrieval and synthesis: A compound ai architecture for scientific discovery, arXiv preprint arXiv:2511.18298 (2025). doi: 10.48550/arXiv.2511. 18298. [26] T. Strohman, W. B. Croft, D. Jensen, Recommending citations for academic papers, in: Proceedings of the 30th annual international ACM SIGIR conference on Research and development in information retrieval, 2007, pp. 705–706. doi:10.1145/1277741.1277868. [27] V. Karpukhin, B. Oguz, S. Min, P. S. Lewis, L. Wu, S. Edunov, D. Chen, W.-t. Yih, Dense passage retrieval for open-domain question answering, in: EMNLP (1), 2020, pp. 6769–6781. doi:10.18653/ v1/2020.emnlp-main.550. [28] G. Izacard, M. Caron, L. Hosseini, S. Riedel, P. Bojanowski, A. Joulin, E. Grave, Unsupervised dense information retrieval with contrastive learning, arXiv preprint arXiv:2112.09118 (2021). doi:10.48550/arXiv.2112.09118. [29] O. Weller, M. Boratko, I. Naim, J. Lee, On the theoretical limitations of embedding-based retrieval, 2025. URL: https://arxiv.org/abs/2508.21038.
We provide experimental details of our baseline fine-tuning approaches of sentence encoder models for content retrieval. Training was run (using 1x 24 GB GPU) for all models with hyperparameter defined in Table 8 . Parameter Loss function Epochs Batch size Learning rate Selection criterion