Troubleshooting radio devices that power modern mobile networks is currently a manual and labour-intensive process, where speed is crucial to minimize network downtime. Therefore, there is a strong interest in building a retrieval system capable of intelligent search and retrieval in radio logs. To facilitate effective retrieval, several retrievers were evaluated using different configurations. First, a RoBERTa language model was further pretrained on a dataset of unannotated radio logs. Then, a proprietary annotated retrieval dataset named the Event LogRetrieval (ELR) dataset was collected and utilized along with the MS MARCOretrieval dataset for training and evaluating the retrieval models. The evaluation compared different retrieval paradigms for log retrieval; evaluated the impact of further pretraining the language model on log data; and investigated which con-figuration yielded the best performance. The results of the investigation show that the late interaction retrieval paradigm used by the ColBERT model performs best for log retrieval. The results also showed that while further pretraining the language model on logs did improve the representations of log data, it did not improve the performance of the implemented retriever. The investigation into the retrieval datasets showed that fine-tuning on the small ELR dataset is insufficient and that fine-tuning on the larger MS MARCO dataset yielded better performance. The best performance was seen though when first fine-tuning on MSMARCO and then on ELR.
| Identifer | oai:union.ndltd.org:UPSALLA1/oai:DiVA.org:liu-204855 |
| Date | January 2024 |
| Creators | Nossborn, Victor |
| Publisher | Linköpings universitet, Kommunikationssystem |
| Source Sets | DiVA Archive at Upsalla University |
| Language | English |
| Detected Language | English |
| Type | Student thesis, info:eu-repo/semantics/bachelorThesis, text |
| Format | application/pdf |
| Rights | info:eu-repo/semantics/openAccess |
Page generated in 0.0018 seconds