Extracting information from narratives: an application to aviation safety reports

Abstract

Aviation safety reports 1,2 are the best available source of information explaining why a flight incident happened. However, stream of consciousness permeates the narratives making the automation of the information extraction task difficult. We propose an approach and infrastructure based on a common pattern specification language to capture relevant information via normalized template expression matching in context. Template expression matching handles variants of multi-word expressions. Normalization improves the likelihood of correct hits by standardizing and cleaning the vocabulary used in narratives. Checking for the presence of negative modifiers in the proximity of a potential hit reduces the chance of false hits. We present the above approach in the context of a specific application that is the extraction of human performance factors from NASA ASRS reports. While knowledge infusion from experts plays a critical role during the learning phase, early results show that in a production mode, the automated process provides information that is consistent with analyses by human subjects.

Cause Identification from Aviation Safety Incident Reports via Weakly Supervised Semantic Lexicon Construction

“…we had not eaten in about 7 hours. 3 Posse et al (2005) identify 14 most important cause types, or shaping factors, that can influence the occurrence of the aviation safety incident described in an ASRS report. These shaping factors are the contextual factors that influenced the reporter's behavior in the incident and thus contributed to the occurrence of the incident.…”

“…Previous work on cause identification for the ASRS reports was done primarily by the researchers at NASA (see Posse et al, 2005) and, to our knowledge, has involved manual analysis of the reports. Specifically, NASA brought together experts on aviation safety, human factors, linguistics and English language to participate in a series of brainstorming sessions, and generated a collection of seed keywords, simple expressions and template expressions related to each shaping factor.…”

“…The first approach is a heuristic approach, which, motivated by Posse et al (2005), labels a report with a shaping factor if it contains at least a word or a phrase that is relevant to the shaping factor. Unlike Posse et al's work, where these relevant words and phrases employed by the heuristic procedure are all manually identified, we automatically acquire these words and phrases via the semisupervised semantic lexicon learning procedure described above.…”

“…It is important to point out that we are by no means arguing that these features are sufficient for cause identification. However, the use of these simple features is relevant for the task and is motivated by the work performed by the NASA researchers, who, as mentioned above, have manually identified seed words and phrases for each shaping factor (Posse et al, 2005). Our semantic lexicon learning procedure precisely aims to learn such words and phrases.…”

Identification of Human Factors in Aviation Incidents Using a Data Stream Approach

“…Posse et al pointed out that machine learning techniques could extract information from the aviation safety reports automatically and reduce human involvement [16]. Two machine learning (ML) methods, classification and clustering, were used in identification of human factors in aviation incidents.…”