170 likes | 260 Vues
Explore the use of BLEU metric in automatic text summarization for evaluating content compression levels and human agreement on extracts. Learn about evaluation experiments setup, software features, and conclusions for future improvements.
E N D
Colouring Summaries BLEU Katerina Pastra and Horacio Saggion Department of Computer Science, Natural Language Processing Group, University of Sheffield, U.K. Pastra and Saggion, EACL 2003
Machine Translation vs. Summarization MT: accurate and fluent translation of source doc Auto Sum: informative, reduced version of source • We will focus on: • Automatically generated extracts • Single-document Summarization - Sentence level compression • Automatic content-based evaluation • Reuse of evaluation metrics across NLP areas Pastra and Saggion, EACL 2003
The challenge MT: demanding content evaluation Extracts: is their evaluation trivial by definition ??? • Idiosyncrasies of the extract evaluation task: • Compression level and rate • High human disagreement on extract adequacy • Could an MT evaluation metric be ported to Automatic Summarization (extract) evaluation ? • If so, which testing parameters should be considered? Pastra and Saggion, EACL 2003
BLEU • Developed for MT evaluation (Papineni et al. ’01) • => achieves high correlation with human judgement • => is reliable even when run • >> on different documents • >> against different number of model references • i.e. reliability is not affected by the use of • either multiple references or just a single one Pastra and Saggion, EACL 2003
Using BLEU in NLP • NLG (Zajic and Dorr, 2002) • Summarization (Lin and Hovy, 2002) >> 0.66 correlation in single-document summaries at 100 words compression rate against a single- reference summary >>0.82 correlation when multiple-judged document units (sort of multiple references) used Lin-Hovy conclude: The use of a single reference affects reliability Pastra and Saggion, EACL 2003
Evaluation Experiments set up • Variables: • compression rate, text cluster, gold standard • HKNews Corpus (English - Chinese) • 18K documents in English • 40 thematic clusters = 400 documents • each sentence in the cluster assessed by 3 judges • with utility values (0-10) • encoded in XML Pastra and Saggion, EACL 2003
Evaluation Software • Semantic tagging and Statistical Analysis Software • Features: position, similarity with document, similarity with query, term distribution, NE scores, etc. (all normalised) • Features are linearly combined to obtain sentence scores and sentence extracts • Gate & Summarization classes Pastra and Saggion, EACL 2003
Gold standards and summarisers • QB = Query-sentence similarity summary • Simple 1 = Doc-sentence similarity summary • Simple 2 = Lead-based summary • Simple 3 = End-of-document summary • Reference n = utility based extract based on the • utility given by judge n (n = 1,2,3) • Reference all = utility based extract based on the • sum of utilities given by the n judges Pastra and Saggion, EACL 2003
Experiment 1 • 2 references compared against the third in 5 • different compression rates in two text clusters • (all available combinations) Are the results BLEU gives on inter-annotator agreement consistent ? => Inconsistency both across text clusters and within clusters at different compression rates (the latter more consistent than the former) => Reliability of BLEU in Sum seems to depend on values of the variables used. If so, how could one identify the appropriate values? Pastra and Saggion, EACL 2003
Experiment 1 • 2 references compared against the third in 5 • different compression rates in two text clusters • (all available combinations) Pastra and Saggion, EACL 2003
Experiment 2 For reference X within cluster Y across compression rates the ranking of the systems is not consistent Pastra and Saggion, EACL 2003
Experiment 3 For reference X at compression Y across clusters the ranking of the systems is not consistent Pastra and Saggion, EACL 2003
Experiment 4 For reference ALL across clusters at multiple compression rates the ranking of the systems is(more) consistent Pastra and Saggion, EACL 2003
Experiment 4 (cont.) Is there a way to use BLEU with a single reference summary and still get reliable results back? Pastra and Saggion, EACL 2003
Notes on BLEU • Fails to capture semantic equivalences between n-grams in both their various lexical and syntactical manifestations Examples: “Of the 9 ,928drug abusers reported in first half of the year, 1,445 or 14 .6% were aged under 21.” vs. “...number of reported abusers” “This represents a decrease of 17% over the 1 ,740 young drug abusers in the first half of 1998.” Pastra and Saggion, EACL 2003
Conclusions • Use of multiple reference summaries needed • when using BLEU in Summarization • Lack of such resources could probably be • overcome using the average rank aggregation • technique • Future work: • Scaling up of the experiments • Correlation of BLEU with other content-based • metrics used in Summarization Pastra and Saggion, EACL 2003