| The purpose of stock movement prediction is to predict the variation of stock price on a certain day within a lookback window,which is one of the important research of quantitative investment.Accurately predicting stock market movements can help investors make the right investment decisions and maximize gains while minimizing losses.The stock price of a corporate is affected by a variety of factors,including historical trends,major news events,etc.Most researchers predict stock movement by modelling historical stock price and introducing financial news and correlation between stocks.Most stock graphs are composed of simple pair-wise graphs.Actually,the stock market contains numerous complex high-order relations,and simple graph neural network is difficult to capture the complex stock market correlations.It is crucial to build a complete and accurate stock relation graph.Currently,only a few datasets provide market-industry relation graphs,which limits the generalization ability of predictive models;while the key to building stock relation graphs is to mine the relations between financial entities from source texts.To further optimize the stock movement prediction task,we research sentence-level relation extraction,which not only further mine the deep semantic information,but also extract the relation between entities in the sentence to provide technical support for expanding the stock relation network.So far,most relation extraction models capture the syntactic features of sentences through syntactic dependency tree to extract entity relations.However,these model ignore the influence of words outside the syntactic dependency tree,and location of entities on the final relation prediction.Therefore,we have improved the stock movement prediction and the sentence-level relation extraction in a targeted manner,and the contributions can be summarized in two aspects:●A novel model(HGCN)for financial text-oriented stock movement prediction is proposed:a spatiotemporal hypergraph convolution network based on hierarchical attention mechanism,which can better encodes stock price volatility.The price encoder obtains the aggregate representation of historical stock prices within the lookback window,and the fintext encoder obtains the text features occurring in the corresponding lookback window through the universal sentence encoder.Finally,the output features of the two encoders are processed by bilinear pooling,and fed into the graph encoding module.We construct a hypergraph structure based on the Global Industry Classification Standard(GICS)for the U.S.stock market.Our proposed hypergraph convolutional network(HGCN)updates the node representation by aggregating adj acent nodes within the hyperedge through message passing.The classification performance and earnings performance of the U.S.stock market are analyzed,the experimental results show that our method is relatively better than the current state-of-the-art methods.●A novel model(ESC-GCN)for sentence-level relation extraction is proposed:a contextualized graph convolutional network based on entity-aware self-attention mechanism,which can effectively integrate the syntactic structure of input sentences and the semantic context information of sequences,obtained higher classification performance.Specifically,relative position self-attention mechanism acquires semantic representation related to entity position,while contextualized graph convolution network captures abundant intra-sentence dependencies between words through pruning operations.Finally,the entity-aware attention layer dynamically selects more crucial words for the final relation prediction.Experiments on TACRED and SemEval-2010 datasets for general relation extraction show that our model performs well,especially in the long sentence.In addition,we also extracted the quantitative financial industry relations augmented stock relation graph from the financial-RE dataset FinRED,which appropriately improved the prediction accuracy of the stock movement prediction task. |
PDF Full Text Request