Method and program for extracting causal relationships

Abstract

The system accurately extracts causal relationships between three or more items, taking into account background knowledge. [Solution] For multiple items, the first probability of the causal direction between all pairs of items is calculated using the output from a system that outputs the causal direction between input items. The second probability of the causal direction between all pairs of items is calculated based on the data of each of the multiple items. Based on the third probability obtained by integrating the first and second probabilities, the causal order of the multiple items is identified, and the causal relationships of the multiple items are extracted according to the causal order.

Claims

1. For multiple items, the first probability of causal relationship between all pairs of items is calculated using the output from a system that outputs causal relationships between input items. The second degree of certainty regarding the causal direction between all the aforementioned pairs of items is calculated based on the data of each of the multiple items. Based on the third probability obtained by integrating the first probability and the second probability, the causal order of the multiple items is identified. The causal relationships of the multiple items are extracted according to the aforementioned causal order. A method for extracting causal relationships, in which a computer performs a process that includes [specific actions].

2. The system is a large-scale language model, as described in claim 1, for extracting causal relationships.

3. The process for calculating the first probability includes taking the multiple items as input, querying the system multiple times to determine whether or not there is a causal relationship between all the pairs of items, and calculating the first probability as the ratio of the number of times the system outputs that there is a cause-and-effect relationship to the number of queries. A method for extracting causal relationships according to claim 1 or claim 2.

4. The method for extracting causal relationships according to claim 1 or claim 2, wherein the process for calculating the second probability is calculated using a statistical causal exploration method.

5. The causal relationship extraction method according to claim 4, wherein the causal search method is the ICA-LiNGAM method.

6. The method for extracting causal relationships according to claim 1 or claim 2, wherein the third degree of certainty is the sum, product, weighted sum, or weighted product of the first degree of certainty and the second degree of certainty.

7. The method for extracting causal relationships according to claim 1 or 2, wherein the process for identifying the causal order includes generating a directed graph in which nodes are connected by edges in order of the third probability between two items corresponding to the nodes, in descending order of probability, and arranging each node of the directed graph according to the direction of the edges.

8. The method for extracting causal relationships according to claim 1 or claim 2, wherein the process for extracting the presence and strength of causal relationships between the aforementioned plurality of items includes dividing each of the plurality of items into item groups that include items preceding each item in the causal order, extracting causal relationships between items for each item group, and superimposing the causal relationships for each item group to extract the causal relationships between the plurality of items.

9. The method for extracting causal relationships according to claim 8, wherein the causal relationship is defined as having a first item as the dependent variable, one or more second items that are causally earlier in order than the first item as independent variables, and when the dependent variable is explained by one or more selected independent variables, the method extracts whether or not there is a causal relationship in which the second item corresponding to the selected independent variable is the cause and the first item corresponding to the dependent variable is the effect, and the strength of the causal relationship, expressed as the degree of influence of the selected independent variable on the explanation of the dependent variable, and uses the first or third degree of certainty as an index indicating the likelihood of selection as an independent variable.

10. For multiple items, the first probability of causal relationship between all pairs of items is calculated using the output from a system that outputs causal relationships between input items. The second degree of certainty regarding the causal direction between all the aforementioned pairs of items is calculated based on the data of each of the multiple items. Based on the third probability obtained by integrating the first probability and the second probability, the causal order of the multiple items is identified. The causal relationships of the multiple items are extracted according to the aforementioned causal order. A causal relationship extraction program that causes a computer to perform a process that includes [specific actions].

Images