LNG unloading system plays an important role in the transport and storage medium of the process. Once the unloading system occurs exceptions, such as pressure fluctuations, uneven pre-cooling and unexpected temperature rises, these phenomena will affect the safe operation of LNG terminals. In the process of LNG unloading, there is a coupling relationship between the changing of pressure, flow rate, temperature and other parameters. The accurate cause and development direction of abnormal working conditions are conducive to the safe operation and maintenance of the unloading system. According to the accident statistics of 38 LNG receiving stations and peak shaving stations by The LNG International Importer's Group(GIIGNL), the accident rate of the unloading system was 27%. Therefore, the realization of the extraction and identification of the causality of abnormal working conditions of the LNG unloading system has great significance for the traceability and early warning of the accident. As a crucial type of relationship, causality plays a key role in many fields such as relational reasoning and so on. Therefore, extracting causality is a basic task in text mining. At present, in China, the analysis of abnormal working conditions of LNG unloading system adopts the methods of HAZOP and FMEA, and a large amount of causality is contained in the text data. To solve the problem of automatically extracting the causality of the LNG unloading system text data, this paper adopts the method of sequence labeling to extract the causal relationship entity and determine its direction. Sequence to sequence (Seq2Seq) architecture and bidirectional long short memory network (Bi-LSTM) are used to train input word vectors and predict the development of causal relationship. In order to solve the problem of unclear objects and inaccurate predictions caused by redundant causal nodes, attention mechanism is applied, and the prediction accuracy reaches 95.23%. In the sensitivity analysis, the recall rate and the granularity of causal relationship information perform well, which indicates that the algorithm has generalization and applicability in the establishment of the causal knowledge base of abnormal conditions in the unloading system.