The previous post discussed a paper that formalized the usage of Graphs for describing reasoning chain relationships. It introduced the TRN graph and described two different types of nodes and three different edges.The nodes were either text nodes and section nodes and the edges were structural, similarity and causal. This graph is constructed from the text with the help of a syntactic parser which adds the text nodes Section nodes are attached to the sentence nodes and all the text nodes are added under these. Similarity measures help add edges for similarity to the node pairs. Only those measures that are above a threshold are added. Graphs that are extracted from different sections are combined Next Causal relations are added with the help of a discourse parser. With the structural, similarity and causal edges added we now have a complete TRN graph. Then with this graph, we can choose source destination pairs and find paths. Paths are combined, post-processed and visualized.
The dataset used for this study was the aviation investigation reports from Transportation Safety Board of Canada. Each report in this collection is an incident report and has sections titled summary, factual information, analysis and findings. The documents are available for download from the TSB of Canada and then they were pre-processed by the authors using GATE NLP platform. The preprocessing steps included tokenization, sentence splitting and part of speech tagging. The sections on the summary is a brief description, the section on the factual information is the details of the aircraft or circumstance, the section on the analysis is a discussion on what might have happened and the section on the findings describes the causes. Therefore the reasoning chain was chosen as the shortest path from Summary to Causes. The constraints for the reasoning chain path was defined in the Backus-Naur Form as one comprising of a summary-path with an edge to a factual-path with an edge to an analysis path with an edge to the causes-path add the intermediaries in this representation are optional. Each of the paths can be recursively defined as one or more nodes with edges. While these edges can be one of the three aforementioned edges such as partof-edge, contains-edge, similar-edge and cause-edge, only the causes path explicitly has part-of edges. A post-processing steps shrinks the path by eliminating transitives. Then the compressed graphs are laid out for visualization.
The dataset used for this study was the aviation investigation reports from Transportation Safety Board of Canada. Each report in this collection is an incident report and has sections titled summary, factual information, analysis and findings. The documents are available for download from the TSB of Canada and then they were pre-processed by the authors using GATE NLP platform. The preprocessing steps included tokenization, sentence splitting and part of speech tagging. The sections on the summary is a brief description, the section on the factual information is the details of the aircraft or circumstance, the section on the analysis is a discussion on what might have happened and the section on the findings describes the causes. Therefore the reasoning chain was chosen as the shortest path from Summary to Causes. The constraints for the reasoning chain path was defined in the Backus-Naur Form as one comprising of a summary-path with an edge to a factual-path with an edge to an analysis path with an edge to the causes-path add the intermediaries in this representation are optional. Each of the paths can be recursively defined as one or more nodes with edges. While these edges can be one of the three aforementioned edges such as partof-edge, contains-edge, similar-edge and cause-edge, only the causes path explicitly has part-of edges. A post-processing steps shrinks the path by eliminating transitives. Then the compressed graphs are laid out for visualization.
No comments:
Post a Comment