Summary
Asia-Pacific Network Operations and Management Symposium
2022
Session Number:PS1
Session:
Number:PS1-10
Evaluation of the application of Seq2Seq towards the automatic derivation of a set of APIs that satisfy requirements from only ambiguous requirements
Shingo Omata, Satoshi Kondoh,
pp.-
Publication Date:2022/09/28
Online ISSN:2188-5079
PDF download
Summary:
Evaluation of the application of Seq2Seq towards the automatic derivation of a set of APIs that satisfy requirements from only ambiguous requirements Abstract:Full automation of customer order processing requires technology to accurately understand the content of customer orders and to design appropriate operational processes and the order of execution of APIs. Although previous studies have made progress in examining and introducing automation of routine operations, the task of making requirements from ambiguous customer requests has not been fully automated. This study aims to fully automate the derivation of the API to be used from only ambiguous customer requirements. As a means to achieve this, the application of Seq2Seq, which transforms one series of data into another, was evaluated. In this paper, two models were prepared, one with LSTM applied to the Encoder of Seq2Seq and the other with GRU applied to the Encoder of Seq2Seq. By training each Seq2Seq with the eTOM-API mapping data published by the Tele-Management Forum (TMF), we evaluated whether the constructed Seq2Seq models could derive the appropriate API from any natural language. The goal of this study is to eventually derive a set of appropriate APIs from only ambiguous requests with an accuracy of more than 90%. To achieve the goal, we would like to determine how much data should eventually be available for Seq2Seq. Therefore, in the experimental part, we assessed to what extent the amount of training data affects the accuracy. Before the experiment, it was assumed that an accuracy of around 28.7% could be achieved by making maximum use of the TMF eTOM-API mapping data. However, the results of the experiment showed no clear improvement in accuracy from data volumes of several hundred to several thousand orders of magnitude. Therefore, it was not possible to estimate the amount of data required to achieve the goal. In the future, the possibility of setting up new features to relate between eTOM-API mapping data and weighting which descriptions are important for API derivation in the eTOM definition text will be investigated.