Reference Summary
Jump to navigation
Jump to search
A Reference Summary is a text summary that serves as a gold standard or benchmark for evaluating Supervised Summarization tasks.
- Context:
- It can (typically) be produced by humans to represent the main points or the essence of a source text or document in a concise manner.
- It can (often) be used in the training phase of a supervised learning model, where the model learns to predict summaries that are as close as possible to these reference summaries.
- It can be utilized in various evaluation metrics for summarization, such as ROUGE or BLEU scores, to quantitatively measure the quality of machine-generated summaries by comparing them against the reference summaries.
- It can vary in form, ranging from extractive summaries, which consist of verbatim excerpts from the source text, to abstractive summaries, which may rephrase, infer, or generate new sentences to convey the source text's content.
- It can sometimes highlight the challenges in summarization tasks, such as capturing nuances, maintaining coherence, and ensuring factual accuracy.
- ...
- Example(s):
- A human-written summary of a news article used as a reference in the evaluation of a news summarization algorithm.
- A set of concise abstracts of scientific papers used to train and test a model that generates abstracts for scientific articles.
- ...
- Counter-Example(s):
- An automatically generated summary without human vetting.
- A paraphrased version of a text that significantly alters the original content's meaning or intent.
- See: Natural Language Processing, Automatic Summarization, Text Summarization, Evaluation Metric.