Skip to main content

Répondre automatiquement aux incidents IT par le Deep Learning et proposition d’une nouvelle approche question answering basée sur SpanBERT et LSTM

Engineer: Nour Eddine ZEKAOUI
Organisation: GRUPO AVALON
Language: French
Promotion: 2021
Year: 3

Abstract #

Question answering has become one of the main objectives of companies. It automatically

provides immediate answers to problems that a user may encounter when interacting with a

service offered by a company. Traditional methods of incident response have several

limitations that can impose damage and loss, and which can lead to critical shutdowns of

business processes. In this regard, the host organization « Grupo AVALON » which operates

in the IT field has decided to invest in artificial intelligence, intending to implement a question

answering solution to answer automatically to IT incidents reported by its customers and

employees. Therefore, to meet the needs of the host organization, we proposed a solution

based on the latest models that marked the new era of natural language processing to provide

rapid, automatic, and correct responses to incidents, through the use of textual data.

On the other hand, for a precise definition of the objectives, we carried out a study and a

needs analysis, with the intention of best meeting the expectations concerning the project.

Then, we were inspired by a recent literature review, which allowed us to analyze the different

scientific approaches adopted to solve the problem, namely, Deep Learning models and

models based on transformers like BERT and its variants.

Finally, we developed and compared the Deep Learning models, namely, BETO, BERT,

mBERT, SpanBERT, then we proposed new approaches based on the hybridization of these

models. In particular, BETO-LSTM, BETO-BiLSTM, BERT-LSTM, BERT-BiLSTM, SpanBERTLSTM, SpanBERT-BiLSTM, SpanBERTLarge-LSTM, and SpanBERTLarge-LSTM and this, for two

types of datasets. To conclude, that the BETO model is the one that best models the incident

dataset with an F1 score of 88.77%. Then, thanks to our approach, we were able to exceed

and outperform the advanced models, namely, BERT and its variant SpanBERT, by our

SpanBERTLarge-LSTM model which gave the best result with an Exact Match of 84.5094% on

the benchmarking question answering dataset SQuaD.