Kāraka-Based Answer Retrieval for Question Answering in Indic Languages

Proceedings of Recent Advances in Natural Language Processing, pages 109–116 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_012 109 A Review in Knowledge Extraction from Knowledge Bases Fabio Y´a˜nez-Romero University Institute for Computer Research University of Alicante fabio.yanez@ua.es Andres Montoyo, Rafael Mu˜noz, Yoan Guti´errez, Armando Su´arez Department of Computing and Information Systems University of Alicante montoyo@dlsi.ua.es, rafael@dlsi.ua.es, ygutierrez@dlsi.ua.es, armando@dlsi.ua.es Abstract Generative language models achieve the state of the art in many tasks within natural language processing (NLP). Although these models cor- rectly capture syntactic information, they fail to interpret knowledge (semantics). Moreover, the lack of interpretability of these models pro- motes the use of other technologies as a re- placement or complement to generative lan- guage models. This is the case with research focused on incorporating knowledge by resort- ing to knowledge bases mainly in the form of graphs. The generation of large knowl- edge graphs is carried out with unsupervised or semi-supervised techniques, which promotes the validation of this knowledge with the same type of techniques due to the size of the gener- ated databases. In this review, we will explain the different techniques used to test and infer knowledge from graph structures with machine learning algorithms. The motivation of validat- ing and inferring knowledge is to use correct knowledge in subsequent tasks with improved embeddings. 1 Introduction Knowledge bases (KB) are widely used for stor- age information used in different machine learning tasks. Knowledge bases are generally represented by knowledge graphs (KG), which store informa- tion that employ nodes (entities) and edges (re- lations) creating a network. This way of storing knowledge has been popularized in recent years due to it being a more expressive, versatile and scalable than traditional databases (Hogan et al., 2021). The efficient use of knowledge stored in a KG with machine learning models is not a trivial task. Traditional machine learning models, including deep neural networks use vectors as input, while the structure of KGs is more complex and can’t be simplified in a vector, due to the need for rep- resenting nodes, edges, connectivity, global rela- tions inside the graph and features of every element (Sanchez-Lengeling et al., 2021). Methodologies used for extracting knowledge from KGs focus on creating latent vectors with the graph information (embeddings) or using neural networks specially designed for dealing with the graph structure. Many KBs are developed using non supervised machine learning techniques, generating massive data in the process. Those methods may cause er- rors when completing the KB due to false relations between nodes. Large KBs also have problems with not useful information introduced for a spe- cific task which can be considered as noise. 2 Knowledge Graph Structure KGs are made up of two sets of elements G = V, E. Where V is the set of nodes (entities) and E is the set of edges (relations). Where: |V | = N, |E| = R being N and R the number of entities and rela- tions, respectively. A knowledge graph can be classified as homo- geneous or heterogeneous …

Content truncated to 500 words