Logo des Repositoriums

Auflistung nach Schlagwort "word embeddings"

1 - 2 von 2
  • Textdokument
    Explore FREDDY: Fast Word Embeddings in Database Systems
    (BTW 2019, 2019) Günther, Michael; Thiele, Maik; Lehner, Wolfgang; Yanakiev, Zdravko
    Word embeddings encode a lot of semantic as well as syntactic features and therefore are useful in many tasks especially in Natural Language Processing and Information Retrieval. FREDDY (Fast woRd EmbedDings Database sYstems), an extended PostgreSQL database system, allowing the user to analyze structured knowledge in the database relations together with unstructured text corpora encoded as word embedding by introducing novel operations for similarity calculation and analogy inference. Approximation techniques support these operations to perform fast similarity computations on high-dimensional vector spaces. This demo allows exploring the powerful query capabilities of FREDDY on different database schemes and a variety of word embeddings generated on different text corpora. From a systems perspective, the user is able to examine the impact of multiple approximation techniques and their parameters for similarity search on query execution time and precision.
  • Textdokument
    Fast Approximated Nearest Neighbor Joins For Relational Database Systems
    (BTW 2019, 2019) Günther, Michael; Thiele, Maik; Lehner, Wolfgang
    K nearest neighbor search (kNN-Search) is a universal data processing technique and a fundamental operation for word embeddings trained by word2vec or related approaches. The benefits of operations on dense vectors like word embeddings for analytical functionalities of RDBMSs motivate an integration of kNN-Joins. However, kNN-Search, as well as kNN-Joins, have barely been integrated into relational database systems so far. In this paper, we develop an index structure for approximated kNN-Joins working well on high-dimensional data and provide an integration into PostgreSQL. The novel index structure is efficient for different cardinalities of the involved join partners. An evaluation of the system based on applications on word embeddings shows the benefits of such an integrated kNN-Join operation and the performance of the proposed approach.