A Two-Stage DistillBERT-BART Hybrid Pipeline for Efficient, Reliable, Stable Text Summarization

This paper introduces a two-stage DistillBERT-BART hybrid pipeline designed to address the critical balance between factual accuracy, logical coherence, and readability flow in an automated text summarization. While a single direct summarization path produces either complete fact checked or highly readable summaries, its deployment is frequently challenged by disjointed outputs, high inference latency, and a tendency to generate hallucinations. To alleviate these limitations, the proposed framework is a two-stage process; an extractive selection stage and an abstractive rewriting stage. DistillBERT model was used for extracting and scoring the most salient sentences from the source document to generate a draft summary set in preparation for the second stage. BART-base model then receives this draft to generate a coherent and readable final summary. The two-stage summarization pipeline significantly reduced the load on the abstractive model by pre-selecting sentences and removing noise and irrelevant data before generating the final outputs. The new model was trained on the publicly available cnotesum dataset. Experimental evaluations were conducted on the generated summaries using both quantitative and qualitative text summarization metrics. ROUGE and BLEU metrics were used to determine the n-gram overlap and lexical similarity and model deviation of the summaries generated, while BERTScore, fluency, and hallucination rate were used to assess the semantic quality of the generated summaries. The final results were further benchmarked against two baseline models to establish the challenges of using a single type of summarization pipeline. The results showed that our two-stage pipeline achieved an outstanding performance with a high BLEU score of 68.4%, which indicated that the model successfully captured all important words. Its high BERTScore of 87.3% showed that the newly generated words used during rewriting retained the same meaning as they appeared in the original context. Furthermore, its high fluency score of 75.2% indicated that the generated outputs were highly readable and cohesive. Finally, its low hallucination rate of 6.0% confirmed that there were few or no fabricated words in the final summaries. Furthermore, human evaluations and factual consistency metrics confirmed that restricting the summarization to purely extractive model generated summaries that were factually correct but are disjointed and lacks coherence and fluent reading. Conversely, using purely abstractive model for such summarization will produce summaries that are logically coherent and fluently readable but it hallucinated heavily, it fabricated words that were not present in the original text. Hence, this paper has established that the abstractive text summarization model performed best when stabilized with extractive techniques.