SGPT: A Generative Approach for SPARQL Query Generation From Natural Language Questions

Abstract

SPARQL query generation from natural language questions is complex because it requires an understanding of both the question and the underlying knowledge graph (KG) patterns. Most SPARQL query generation approaches are template-based, tailored to a specific knowledge graph and require pipelines with multiple steps, including entity and relation linking. Template-based approaches are also difficult to adapt for new KGs and require manual efforts from domain experts to construct query templates. To overcome this hurdle, we propose a new approach, dubbed SGPT, that combines the benefits of end-to-end and modular systems and leverages recent advances in large-scale language models. Specifically, we devise a novel embedding technique that can encode linguistic features from the question which enables the system to learn complex question patterns. In addition, we propose training techniques that allow the system to implicitly employ the graph-specific information (i.e., entities and relations) into the language model's parameters and generate SPARQL queries accurately. Finally, we introduce a strategy to adapt standard automatic metrics for evaluating SPARQL query generation. A comprehensive evaluation demonstrates the effectiveness of SGPT over state-of-the-art methods across several benchmark datasets.INDEX TERMS Knowledge based systems, knowledge graph, information retrieval, query generation, language models.

Enhancing SPARQL Query Generation for Knowledge Base Question Answering Systems by Learning to Correct Triplets

“…In this paper, we compare our methods with some classical approaches like AQGnet [35], Multi-hop QGG [51], CLC (+BERT/Tencent Word) [46], and some recently used PLM methods such as BART [16], PGN-BERT (-BERT) [11], SGPT Q,K [52], and T5 [15]. AQG-net [35] employs a generative model based on neural networks to generate an abstract representation of query graphs, capturing the logical structures of queries.…”

Leveraging Text-to-Text Pretrained Language Models for Question Answering in Chemistry

“…Capitalizing on the generative capabilities of LMs, KGQA systems that adopt the translation approach take in natural language questions and feed them into fine-tuned LMs to output SPARQL queries. For many systems, , the grounded queries are generated in a one-shot manner. Meanwhile, LMs could be employed to obtain only query skeletons, which are then grounded with entities and/or relations detected by a separate set of components. , Recognizing that one-shot generation of executable SPARQL queries is prone to KG misalignment especially for unseen entities and relations, but decoupling entity and relation linking from logical form generation opens up more room for errors, researchers behind the system GETT-QA propose a middle ground, whereby in the first step, SPARQL queries are generated with entity and relation slots already filled with their surface forms as found in input questions, and in the second steps, these labels are grounded to actual KG entities and relations.…”

“…Capitalizing on the generative capabilities of LMs, KGQA systems that adopt the translation approach take in natural language questions and feed them into fine-tuned LMs to output SPARQL queries. For many systems, 29,33 the grounded queries are generated in a oneshot manner. Meanwhile, LMs could be employed to obtain only query skeletons, which are then grounded with entities and/or relations detected by a separate set of components.…”

“…More is left to be explored with semantic parsing-based approaches and the use of PLMs. Key approaches to KGQA that have benefited from generative PLMs include end-to-end translation, − the retriever-reader model, − and dynamic logical form induction. , Even though many of these systems have clinched state-of-the-art results on public benchmarks, ,, attempts to quantify their speed to demonstrate practical usage are still inadequate; we are only able to find reports of system speed measured on GPU in two papers. , This casts doubt on whether systems with impressive accuracy can be realistically deployed in a real-world setting, where specialized hardware might not be available. More broadly, we notice that efforts to study the accuracy-latency trade-off for KGQA in a hardware-constrained setting are lacking despite the wealth of similar work in other machine learning domains. − …”

GETT-QA: Graph Embedding based T2T Transformer for Knowledge Graph Question Answering

“…We include both these systems in our evaluation Table 4. SGPT [31] and STAG [29] both use generative methods for forming the query using pre-trained language models, which is similar to what we do, however, neither of them generate the entity or relation label, or the embeddings. Instead STAG uses an external entity linking step, while SGPT attempts to generate entity and relation IDs directly.…”