Ani Nenkova: Predicting sentence specificity, with applications to news summarization

Value of the Information & Strength of the Argument

The talk provides valuable insights into a novel task: automatically identifying sentence specificity. The argumentation is solid, with multiple evaluations: cross-validation, human agreement, and application to summarization. The speaker is transparent about limitations, such as the difficulty of generalizing from discourse pairs to isolated sentences. The application to summarization is a concrete demonstration of utility. However, the talk is a research presentation, not a peer-reviewed paper, so some claims are preliminary.

Scientific Rigor, Source Quality, Title Accuracy

The talk is scientifically rigorous, with clear methodology and evaluation. The primary source is the PDTB, and the speaker cites related work on subjectivity lexicons and word specificity. The title accurately reflects the content. No comments were provided, so no analysis of public reception is possible.

The title accurately reflects the content: the talk focuses on predicting sentence specificity and its application to summarization.

8/10

The talk presents a supervised classifier for sentence specificity, with detailed methodology, multiple evaluations, and discussion of limitations. The speaker is a recognized researcher in NLP. However, it is a conference talk, not a peer-reviewed paper, and some claims are based on unpublished work.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 02:00Introduction: motivation for predicting sentence specificity, link to text quality and information retrieval.
• 10:00Description of the Penn Discourse Treebank (PDTB) and discourse relations, focusing on implicit relations.
• 20:00Extraction of training data from PDTB: instantiations and restatement specifications.
• 30:00Feature engineering: length, polarity, language model probabilities, word specificity, syntax.
• 40:00Results: classifier achieves 76% accuracy on instantiations; feature analysis.
• 50:00Generalization to new domains and human agreement experiments.
• 1:00:00Application to news summarization: selecting specific sentences improves summaries.
• 1:10:00Discussion of limitations, future work, and open questions.

• CLSP Seminar Abstract — Official abstract of the talk.

• Penn Discourse Treebank — The primary data source for training the classifier.