Leveraging Pre-trained Models for Failure Analysis Triplets Generation
This work addresses a domain-specific problem in the semiconductor industry by applying existing methods to new data, making it incremental in nature.
The paper tackled the problem of generating Failure Analysis Triplets (FATs) for semiconductor defect analysis by leveraging pre-trained language models, finding that GPT2 outperformed BERT, BART, and GPT3 by a large margin on ROUGE metrics and introduced a new evaluation metric, LESE, that aligns closely with human judgment.
Pre-trained Language Models recently gained traction in the Natural Language Processing (NLP) domain for text summarization, generation and question-answering tasks. This stems from the innovation introduced in Transformer models and their overwhelming performance compared with Recurrent Neural Network Models (Long Short Term Memory (LSTM)). In this paper, we leverage the attention mechanism of pre-trained causal language models such as Transformer model for the downstream task of generating Failure Analysis Triplets (FATs) - a sequence of steps for analyzing defected components in the semiconductor industry. We compare different transformer models for this generative task and observe that Generative Pre-trained Transformer 2 (GPT2) outperformed other transformer model for the failure analysis triplet generation (FATG) task. In particular, we observe that GPT2 (trained on 1.5B parameters) outperforms pre-trained BERT, BART and GPT3 by a large margin on ROUGE. Furthermore, we introduce Levenshstein Sequential Evaluation metric (LESE) for better evaluation of the structured FAT data and show that it compares exactly with human judgment than existing metrics.