Deep learning-based structured information extraction method

Abstract

The invention discloses a deep learning-based structured information extraction method. The method comprises the following steps of: 1) constructing large-scale flag data by adoption of remote monitoring: constructing an extractor by using the existing Wikipedia to provide remote supervision, wherein a signature and a Wikipedia information box comprise entity-related structured facts, the facts are mentioned in a free text part of an entity webpage, and sentences which express the facts in the information box are used as training data; 2) integrating prior knowledge into a structured information extraction model, wherein the prior knowledge comprises type and phrase information, and labels of words in natural language sentences are related to the foregoing words and continuous words thereof; 3) utilizing the past and future input features by using a bidirectional hidden LSTM layer, and inputting each sentence into a bidirectional LSTM model in a sequence form; and 4) finally outputting a marker sequence, wherein the marker sequence comprises true or false markers and has a length same as the length of the input word sequence.

Description

technical field [0001] The invention belongs to the field of information processing, and in particular relates to a method and system for extracting structured information based on deep learning. Background technique [0002] In the extraction of structured information, many studies have done a lot of work in collecting the structured knowledge of entities from corpora, such as Kylin and DBpedia. These knowledge bases, also known as knowledge graphs, contain rich facts about entities, such as Barack Obama's birthplace being Honolulu. We usually refer to entities as subjects (s), properties or aspects as predicates (p) and values ​​as objects (o). Extracting structured facts (in the form of SPO triples) from corpora has increasingly attracted much research interest due to the wide application of knowledge graphs. In this technique, we focus on the problem of simultaneously extracting structured facts for a large number of predicates with tens of millions of entities. [00...

AI Technical Summary

Problems solved by technology

Second, solutions based on supervised learning require manually labeled data

[0007] 2. A sentence about a subject can contain multiple objects for a predicate, which further complicates the pattern and confuses the learner

[0008] 3. The objects we want to extract may not necessarily appear in a single sentence

Images