Case-Based Approach to Automated Natural Language Generation for Obituaries

Abstract

Automated generation of human readable text from structured information is challenging because grammatical rules are complex making good quality outputs difficult to achieve. Textual Case-Based Reasoning provides one approach in which the text from previously solved examples with similar inputs is reused as a template solution to generate text for the current problem. Natural Language Generation also poses a challenge when evaluating the quality of the text generated due to the high cost of human labelling and the variety in potential good quality solutions. In this paper, we propose two case-based approaches for reusing text to automatically generate an obituary from a set of input attribute-value pairs. The case-base is acquired by crawling and then tagging existing solutions published on the web to create cases as problem-solution pairs. We evaluate the quality of the text generation system with a novel unsupervised case alignment metric using normalised discounted cumulative gain which is compared to a supervised approach and human evaluation. Initial results show that our proposed evaluation measure is effective and correlates well with average attribute error evaluation which is a crude surrogate to human feedback. The system is being deployed in a real-world application with a startup company in Aberdeen to produce automated obituaries.

Interpreting Patient Descriptions using Distantly Supervised Similar Case Retrieval

“…Beyond the legal domain, the idea of exploiting similar cases has recently been used for question answering [51], semantic parsing [72], text generation [64] and language modelling [31] among many others. Within the biomedical domain, one relevant line of research aims to capture the similarity between different patients to predict, for example, a diagnosis or treatment [24,26,47], usually by learning a dense vector representation of each patient.…”

A Case-Based Approach to Data-to-Text Generation

“…D2T has also been studied in CBR with systems ranging from weather forecasts to obituary generation [2,1,30], although these systems have been limited to generating smaller texts describing very few entities. As with most CBR systems, similarity is a key component of CBR-D2T systems as effective similarity measures ensure relevant previously solved problems are reused.…”

“…We use the method proposed in [30] for measuring the case-alignment of the case-base. The alignment score is then used as a loss function for a Particle-Swarm Optimiser whose parameters are the features' weights for a case-base.…”