abstract |
Computer-implemented method, method, apparatus, and program for generating dynamic context-based word vectors. Text data is obtained from websites and written to a local storage as individual text data instances with respective metadata. A neural language model is trained in respect of each text data instance. Queries are received for which word vectors are required in formulating a response. Word vector representation of query string is requested, with a query scope of multiple text data instances. The multiple text data instances are combined in a weighted manner with weightings determined by metadata features such as publication date and publication reputation. Resultant word vectors are more significantly influenced by word usage in text data instances with metadata features indicating greater relevance, such as more recent publication or better reputation. |