Makbule Gulcin Ozsoy

Makbule Gulcin Ozsoy

Large Language Models enable users to access database using natural language interfaces using tools like Text2SQL, Text2SPARQL, and Text2Cypher, which translate user questions into structured database queries. While these systems improve database accessibility, most research focuses on English with limited multilingual support. This work investigates a scalable multilingual Text2Cypher, aiming to support new languages without re-running full fine-tuning, avoiding manual hyper-parameter tuning, and maintaining performance close to joint multilingual fine-tuning. We train language-specific LoRA adapters for English, Spanish, and Turkish and combined them via uniform linear merging or learned fusion MLP with dynamic gating. Experimental results show that the fusion MLP recovers around 75\% of the accuracy gains from joint multilingual fine-tuning while requiring only a smaller subset of the data, outperforming linear merging across all three languages. This approach enables incremental language expansion to new languages by requiring only one LoRA adapter and a lightweight MLP retraining. Learned adapter fusion offers a practical alternative to expensive joint fine-tuning, balancing performance, data efficiency, and scalability for multilingual Text2Cypher task.

Makbule Gulcin Ozsoy

Large language models (LLMs) allow users to query databases using natural language by translating questions into executable queries. Despite strong progress on tasks such as Text2SQL, Text2SPARQL, and Text2Cypher, most existing methods focus on better prompting, fine-tuning, or iterative refinement. However, they often do not explicitly enforce structural constraints, such as syntactic validity and schema consistency. This can reduce reliability, since generated queries must satisfy both syntax rules and database schema constraints to be executable. In this work, we study how structured constraints can be used in test-time inference for Text2Cypher. We focus on post-generation validation to improve query correctness. We extend a confidence-based inference framework with a sequential filtering process that combines confidence scoring, grammar validation, and schema constraints before final aggregation. This lets us analyze how different constraint types affect generated queries. Our experiments with two instruction-tuned models show that grammar-based filtering improves syntactic validity. Schema-aware filtering further improves execution quality by enforcing consistency with the database structure. However, stronger filtering also increases the number of empty predictions and reduces execution coverage. Overall, we show that adding simple structural checks at test time improves the reliability of Text2Cypher generation, and we provide a clearer view of how syntax and schema constraints contribute differently.

Makbule Gulcin Ozsoy

Large language models have significantly improved natural language interfaces to databases by translating user questions into executable queries. In particular, Text2Cypher focuses on generating Cypher queries for graph databases, enabling users to access graph data without query language expertise. Most existing Text2Cypher systems assume a single preselected graph database, where queries are generated over a known schema. However, real-world systems are often distributed across multiple independent graph databases organized by domain or system boundaries, where relevant information may span multiple sources. To address this limitation, we propose a shift from single-database query generation to multi-database query reasoning. Instead of assuming a fixed execution context, the system must reason about (i) relevant databases, (ii) how to decompose a question across them, and (iii) how to integrate partial results. We formalize this setting through a three-phase roadmap: database routing, multi-database decomposition, and heterogeneous query reasoning across database types and query languages. This work provides a structured formulation of multi-database reasoning for Text2Cypher and identifies challenges in source selection, query decomposition, and result integration, aiming to support more realistic and scalable natural language interfaces to graph databases.

Makbule Gulcin Ozsoy, Leila Messallem, Jon Besga et al.

Knowledge graphs use nodes, relationships, and properties to represent arbitrarily complex data. When stored in a graph database, the Cypher query language enables efficient modeling and querying of knowledge graphs. However, using Cypher requires specialized knowledge, which can present a challenge for non-expert users. Our work Text2Cypher aims to bridge this gap by translating natural language queries into Cypher query language and extending the utility of knowledge graphs to non-technical expert users. While large language models (LLMs) can be used for this purpose, they often struggle to capture complex nuances, resulting in incomplete or incorrect outputs. Fine-tuning LLMs on domain-specific datasets has proven to be a more promising approach, but the limited availability of high-quality, publicly available Text2Cypher datasets makes this challenging. In this work, we show how we combined, cleaned and organized several publicly available datasets into a total of 44,387 instances, enabling effective fine-tuning and evaluation. Models fine-tuned on this dataset showed significant performance gains, with improvements in Google-BLEU and Exact Match scores over baseline models, highlighting the importance of high-quality datasets and fine-tuning in improving Text2Cypher performance.

Makbule Gulcin Ozsoy

Database query languages such as SQL for relational databases and Cypher for graph databases have been widely adopted. Recent advancements in large language models (LLMs) enable natural language interactions with databases through models like Text2SQL and Text2Cypher. Fine-tuning these models typically requires large, diverse datasets containing non-trivial examples. However, as dataset size increases, the cost of fine-tuning also rises. This makes smaller, high-quality datasets essential for reducing costs for the same or better performance. In this paper, we propose five hard-example selection techniques for pruning the Text2Cypher dataset, aiming to preserve or improve performance while reducing resource usage. Our results show that these hard-example selection approaches can halve training time and costs with minimal impact on performance, and demonstrates that hard-example selection provides a cost-effective solution.

Makbule Gulcin Ozsoy

Knowledge graphs represent complex data using nodes, relationships, and properties. Cypher, a powerful query language for graph databases, enables efficient modeling and querying. Recent advancements in large language models allow translation of natural language questions into Cypher queries - Text2Cypher. A common approach is incorporating database schema into prompts. However, complex schemas can introduce noise, increase hallucinations, and raise computational costs. Schema filtering addresses these challenges by including only relevant schema elements, improving query generation while reducing token costs. This work explores various schema filtering methods for Text2Cypher task and analyzes their impact on token length, performance, and cost. Results show that schema filtering effectively optimizes Text2Cypher, especially for smaller models. Consistent with prior research, we find that larger models benefit less from schema filtering due to their longer context capabilities. However, schema filtering remains valuable for both larger and smaller models in cost reduction.

Makbule Gulcin Ozsoy, William Tai

Recent advances in large language models (LLMs) have enabled natural language interfaces that translate user questions into database queries, such as Text2SQL, Text2SPARQL, and Text2Cypher. While these interfaces enhance database accessibility, most research today focuses on English, with limited evaluation in other languages. This paper investigates the performance of both foundational and finetuned LLMs on the Text2Cypher task across multiple languages. We create and release a multilingual dataset by translating English questions into Spanish and Turkish while preserving the original Cypher queries, enabling fair cross-lingual comparison. Using standardized prompts and metrics, we evaluate several foundational models and observe a consistent performance pattern: highest on English, followed by Spanish, and lowest on Turkish. We attribute this to differences in training data availability and linguistic features. We also examine the impact of translating task prompts into Spanish and Turkish. Results show little to no change in evaluation metrics, suggesting prompt translation has minor impact. Furthermore, we finetune a foundational model on two datasets: one in English only, and one multilingual. Finetuning on English improves overall accuracy but widens the performance gap between languages. In contrast, multilingual finetuning narrows the gap, resulting in more balanced performance. Our findings highlight the importance for multilingual evaluation and training to build more inclusive and robust query generation systems.

Makbule Gulcin Ozsoy

Recommender systems guide users through vast amounts of information by suggesting items based on their predicted preferences. Collaborative filtering-based deep learning techniques have regained popularity due to their straightforward nature, relying only on user-item interactions. Typically, these systems consist of three main components: an interaction module, a loss function, and a negative sampling strategy. Initially, researchers focused on enhancing performance by developing complex interaction modules. However, there has been a recent shift toward refining loss functions and negative sampling strategies. This shift has led to an increased interest in contrastive learning, which pulls similar pairs closer while pushing dissimilar ones apart. Contrastive learning may bring challenges like high memory demands and under-utilization of some negative samples. The proposed Multi-Margin Cosine Loss (MMCL) addresses these challenges by introducing multiple margins and varying weights for negative samples. It efficiently utilizes not only the hardest negatives but also other non-trivial negatives, offers a simpler yet effective loss function that outperforms more complex methods, especially when resources are limited. Experiments on two well-known datasets demonstrated that MMCL achieved up to a 20\% performance improvement compared to a baseline loss function when fewer number of negative samples are used.

Makbule Gulcin Ozsoy

Recommender systems (RS) suggest items-based on the estimated preferences of users. Recent RS methods utilise vector space embeddings and deep learning methods to make efficient recommendations. However, most of these methods overlook the sequentiality feature and consider each interaction, e.g., check-in, independent from each other. The proposed method considers the sequentiality of the interactions of users with items and uses them to make recommendations of a list of multi-item sequences. The proposed method uses FastText \cite{bojanowski2016enriching}, a well-known technique in natural language processing (NLP), to model the relationship among the subunits of sequences, e.g., tracks, playlists, and utilises the trained representation as an input to a traditional recommendation method. The recommended lists of multi-item sequences are evaluated by the ROUGE \cite{lin2003automatic,lin2004rouge} metric, which is also commonly used in the NLP literature. The current experimental results reveal that it is possible to recommend a list of multi-item sequences, in addition to the traditional next item recommendation. Also, the usage of FastText, which utilise sub-units of the input sequences, helps to overcome cold-start user problem.

Makbule Gulcin Ozsoy

Venue recommendation systems model the past interactions (i.e., check-ins) of the users and recommend venues. Traditional recommendation systems employ collaborative filtering, content-based filtering or matrix factorization. Recently, vector space embedding and deep learning algorithms are also used for recommendation. In this work, I propose a method for recommending top-k venues by utilizing the sequentiality feature of check-ins and a recent vector space embedding method, namely the FastText. Our proposed method; forms groups of check-ins, learns the vector space representations of the venues and utilizes the learned embeddings to make venue recommendations. I measure the performance of the proposed method using a Foursquare check-in dataset.The results show that the proposed method performs better than the state-of-the-art methods.

Makbule Gulcin Ozsoy

Social network platforms can use the data produced by their users to serve them better. One of the services these platforms provide is recommendation service. Recommendation systems can predict the future preferences of users using their past preferences. In the recommendation systems literature there are various techniques, such as neighborhood based methods, machine-learning based methods and matrix-factorization based methods. In this work, a set of well known methods from natural language processing domain, namely Word2Vec, is applied to recommendation systems domain. Unlike previous works that use Word2Vec for recommendation, this work uses non-textual features, the check-ins, and it recommends venues to visit/check-in to the target users. For the experiments, a Foursquare check-in dataset is used. The results show that use of continuous vector space representations of items modeled by techniques of Word2Vec is promising for making recommendations.