Causal Relation Extraction Based On Cascading Conditional Random Fields

Causal analysis is the key to solve problem,and is also a common analysis method of the innovation theory of TRIZ.At present,domestic and foreign certain research work for causal relation identification has been carried out,and achieved certain results.However,there are still shortcomings in the causal identification of event sequences,including the lack of refinement of causal event pairs and long-distance dependence.In this paper,these two points were studied in depth:(1)In terms of causal event pair's description: This paper defines the causal relation expression of the innovation problem.The cause event and outcome event are divided into entities,attributes and performance levels according to the expression of the function in the innovation problem,and the performance level is classified,and finally causal event pairs form the normative expression.(2)In terms of long-distance dependence: In this paper,we propose Cascade Skip-chain Conditional Random Fields model.In the low-level model of causality identification,the linear chain conditional random field model is selected,the Skip-chain Conditional Random Fields model is used in the high-level model,and the Skip-chain Conditional Random Fields model breaks the strongest dependent relation among adjacent nodes.Then,Experiments were carried out on a variety of real corpus and the results show that the Cascade Skip-chain Conditional Random Fields model used in this paper can not only identify the causal role,but also effectively solve the problem of long-distance dependence in causal relation identification.(3)Combining the work of(1)and(2),we have designed a causal relation extraction system.The system preprocesses the text of the patent and then uses the low-level Linear Conditional Random Fields model to label the causal boundaries of the pre-processed candidate sets.Secondly,we have filtered noise and extended the labeled results as a new feature to the high-level Skip-chain Conditional Random Fields.The random field model is used to identify causal roles.High-level results are denominated and filtered noise.Finally the causal roles are extracted and the performance level is classified.