ebook img

Final Summarization System LING 573 Deliverable #04 PDF

88 Pages·2015·1.92 MB·English
by  
Save to my drive
Quick download
Download
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Final Summarization System LING 573 Deliverable #04

. Final Summarization System LING 573 Deliverable #04 . Eric Garnick ( Outline . System Overview Content Selection Information Ordering Content Realization Issues and Deadends ROUGE scores In�uences Future Directions . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 System Overview . Python 3, nltk Gigaword corpus 3 high-level, independent components: content selection, ordering, realization Barzilay, et. al. 2002 �nal ROUGE F scores for evaltest data: R1 R2 R3 R4 0.31068 0.09708 0.03683 0.01701 . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Content Selection . log likelihood ratio (Gigaword corpus) tokenization, data cleanup topic signatures (cluster title, article headlines) tf-idf cosine similarity scores . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Information Ordering . added three different methods to re-order summaries, all part of a discrete step: chronological (Barzilay, 2002; publication date and relative position in document) majority (Barzilay, 2002; heuristic topological search, relative sentence positions) traveling sales person (Conroy, 2006; consider all possible pairs, minimize distances) . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Information Ordering Evaluation . Earlier we evaluated reorderings by random spot check, gauging an overall \feel", which ranked majority highest. Now we have two different review methods: manual, semi-blinded evaluation, 3 levels of comprehensibility (Barzilay, 2002) automatic, correlation to gold-standard summaries . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Information Ordering Evaluation: Manual . System Comprehensible Somewhat Comp. Incomp. Original 8 9 4 Majority 11 7 3 Chronological 6 11 4 Similarity 6 10 5 Majority the least bad? . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Information Ordering Evaluation: Automatic . Re-ordering processes applied to the gold-standards, and then calculated Spearman’s rank correlation. Dataset Algorithm rho p-value evaltest chrono 0.0604 0.3190 evaltest majority -0.0384 0.5270 evaltest pass 0.8825 4.4456e-91 evaltest tsp 0.4102 1.5206e-12 devtest chrono 0.03787 0.5265 devtest majority 0.0379 0.5259 devtest pass 0.9816 1.3553e-203 devtest tsp 0.2655 6.1884e-06 TSP the least bad? . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Content Realization . Sentence �ltering: pruning extraneous words (email addresses, phone numbers, paraentheticals, &c.) found by regexes remove attributions remove adverbials remove ALL CAPS items (but with an exception list) exclude sentences without any verbs (as determined by SpaCy POS tagging) . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4 Issues and Deadends . Items that were tried and abandoned, or were deemed unhelpful: co-reference resolution sentence position weighting stemming / lemmatization augmented chronological re-ordering various parameter tunings . . . . . . . . . . . . . . . . . . . . .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. .. egarnick, jtm37, olgaw LING 573 D4

See more

The list of books you might like

Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.