Research On Methods Of Event Factuality Identification

Event factuality expresses the factual nature of events in texts,i.e.,it describes whether an event is a fact,a possibility,or an impossible situation.The information of event factuality,the common phenomenon in natural language texts,reflects the different stances hold by participants about the credibility of events,or the factual status of events in various contexts.The identification of event factuality is fundamental for many relative tasks in the field of natural language processing,such as question-answer system,opinion detection,sentiment analysis,and rumor identification.This disseration launches the research of event factuality mainly from sentence level and discourse level,and focuses on the following three sub-tasks: The first is the speculation and negation scope detection,which is in charge of the identification of the linguistic scopes of speculative and negative cues in sentences.This disseration proposes the CNN and LSTM neural network models for scope detection,which can achieve excellent performance on Bio Scope corpus.The second is the identification of sentence-level event factuality,which aims to identify the factuality of events committed by relevant source within sentences.This disseration proposes a generative adversarial network with auxiliary classification for sentence-level event factuality,which is superior to the baselines.The third is identification of discourse-level event factuality,which aims at determining the event factuality for the perspective of discourses.This disseration constructs a sentence-level event factuality corpus for the first time,and proposes a neural network based model with adversarial training for identification of discourse-level event factuality.The experimental results show that our model is superior to other baselines.Specifically,the main contents of this thesis can be summarized as follows:(1)Speculation and negation scope detection based on syntactic paths via neural networks.Related work has proven that syntactic features are important information for speculation and negation scope detection.However,they required much human intervention when developing features.This disseration proposes two neural network models for speculation and negation scope detection,and considers the syntactic paths from cues to tokens as the main syntactic features.Particularly,the Convolutional Neural Network(CNN)model regards the scope detection problem as a token classification task,and extracts the syntactic feature representations from syntactic paths via CNN.The Long Short-Term Memory(LSTM)neural network regards the scope detection as a sequence labeling task,which learns the syntactic feature representations from syntactic paths via one LSTM network,and determines the labels of tokens in the sentence via another LSTM network.The experimental results show that our neural network models can achieve excellent performances on Bio Scope corpus.(2)Sentence-level event factuality identification via a generative adversarial network with auxiliary classification.This disseration firstly proposes a two-step supervised framework to identify event factuality,in which we first extract basic information related with factuality,i.e.,events,predicates,sources,and speculative or negative cues,and then utilize a Generative Adversarial Network with Auxiliary Classification for Event Factuality identification(EF-ACGAN).The dependency syntactic paths from basic information to events are considered as the main syctactic features.The generator in this model can generate syntactic paths that are close to the distribution of real ones,and can offer more useful syntactic information.Moreover,to identify speculative and negative factuality values more effectively,we design the auxiliary classification with two outputs in EF-ACGAN.Experimental results on Fact Bank show that EF-AC-GAN outperforms the baselines significantly.(3)Construction of discourse-level event factuality corpus.Currently,the scarcity of relative corpus seriously limits the advance of the research of discourse-level event factuality.Therefore,we construct English and Chinese discourse-level event factuality corpora,which are the first English and Chinese corpus as far as we know and contain1727 English and 4649 Chinese news documents.The statistics on the corpora and the experimental results show that these corpora can sufficiently reflect linguistic characteristics of news texts,and provide adequate and effective support on corpus resource for the research.(4)Document-level event factuality identification via neural networks with inter-sequence attentions and adversarial training.This disseration proposes a neural network model with adversarial training for discourse-level event factuality identification,which learns feature representations within sentences and dependency paths via attention-based an LSTM network,and extracts feature representations among sequences via an inter-sequence attention-based layer.In addition,to enhance the robustness,this model adds small adversarial perturbation to the word embeddings and utilizes adversarial training.Experimental results on the constructed corpora show considering adversarial training and contexts of sentences containing events can both improve the performace.The research of event factuality is still in the initial developing stage at present.This thesis focuses on event factuality identification,which is innovative in the methodology and resources for sentence-level and discourse-level event factuality identification task.This thesis can not only be beneficial to the relative research in the field of event factuality identification,but also promote the progress of relative tasks in the field of natural language processing.