Neural Network-based Open Information Extraction And Its Application

Open Information Extraction(Open IE)aims to extract structured informational tuples from unstructured text,which can find out valuable information automatically and is one of the most promising method in solving the information overload problem.Currently,this technique is widely used in many industrial applications and plays an important role in many downstream NLP tasks.Previous methods of Open IE mainly rely on neural-network-based sequence labeling model.However,these kinds of approaches cannot identify confused entity mentions accurately,and the extraction speed is relative slow leading to problems when the systems are applied on super large-scale datasets and real-time applications.On the other hand,the usage of Open IE technique on Query-Based Text Summarization task is not well studied.Previous summarization methods mainly base on neural-network-based generative model,which cannot ensure the factual correctness and completeness of generated summaries.To solve the above problems,this research proposes two novel Open IE algorithms and one Open IE-based text summarization model.1.To enable the model to identify confused entity mentions,we propose a multi-level features based Open IE model.First,we propose predicate-specific embedding layer to involve sentence-level features in the argument boundary identification process.Then,we introduce coattention mechanism to build span-level features to classify the category of entity mentions.The proposed method significantly outperforms previous Open IE systems on OIE2016 and ReOIE2016 benchmarks.2.To accelerate the information extraction process,we propose an Open IE model with a sliding window-based grouping algorithm.The proposed method first uses a simplified SPO tagging schema to identify all information phrase within the inputs,and then introduce the sliding window-based grouping algorithm to group the information into tuples.The experimental results on SAOKE benchmark show that this approach shorten the overall extraction time significantly with a comparable extraction precision.3.To solve the unfaithful and unmeaningful generation problem of query-based text summarization systems,we proposed a knowledge-augmented query-oriented text summarization model.First,the model constructs knowledge graph via Open IE technique,and propose the Graph Self Attention(GSA)module to select useful part of the knowledge graph.Then,we introduce the Conditional Self Attention(CSA)module to generate summaries according to the topic in which the query is concerned.