Automatic Evaluation of Linguistic Quality in Multi- Document Summarization Pitler, Louis, Nenkova 2010 Presented by Dan Feblowitz and Jeremy B. Merrill

Motivation Automatic evaluation of content selection is already done. Automatic evaluation of content selection is already done. ROUGE: automated metric for info content. (Lin and Hovy, 2003; Lin, 2004) ROUGE: automated metric for info content. (Lin and Hovy, 2003; Lin, 2004) No automatic evaluation of linguistic quality available. No automatic evaluation of linguistic quality available. We want fluent and easy-to-read summaries. We want fluent and easy-to-read summaries. How to test? How to test?

Intuitions: Aspects of Ling Quality Grammaticality Grammaticality The Police found no second armed man. LOS ANGELES -- A sniping incident Sunday damaged helicopter. The Police found no second armed man. LOS ANGELES -- A sniping incident Sunday damaged helicopter. Non-redundancy Non-redundancy Bill Clinton ate a banana yesterday. Bill Clinton liked it. Bill Clinton was in Los Angeles. Bill Clinton ate a banana yesterday. Bill Clinton liked it. Bill Clinton was in Los Angeles. Referential Clarity Referential Clarity The beer scavvy participant, a 20-year-old male, was arrested Saturday. “This was really irresponsible,” she said. The beer scavvy participant, a 20-year-old male, was arrested Saturday. “This was really irresponsible,” she said. Focus Focus To show solidarity with dining hall workers, Bill Clinton ate a banana. He was at Frary. Frary contains a mural by some Mexican muralist. To show solidarity with dining hall workers, Bill Clinton ate a banana. He was at Frary. Frary contains a mural by some Mexican muralist. Structure and Coherence Structure and Coherence Harvey Mudd was founded in It is a engineering college. It has eight dorms. Its founder was named Harvey. Harvey Mudd was founded in It is a engineering college. It has eight dorms. Its founder was named Harvey.

Correlation Among Aspects Referential Clarity, Focus and Structure are significantly correlated with each other. (Along with a few more significant correlations.) Referential Clarity, Focus and Structure are significantly correlated with each other. (Along with a few more significant correlations.) Linguistic quality rankings correlate positively with content quality rankings. Linguistic quality rankings correlate positively with content quality rankings. Human rankers. Human rankers.

Goal Find automated measures that correlate with the intuition-based aspects. Find automated measures that correlate with the intuition-based aspects. System-level evaluation System-level evaluation Input-level evaluation Input-level evaluation

Automated Measures Language Modeling: Gigaword corpus /1-,2-,3-gram Language Modeling: Gigaword corpus /1-,2-,3-gram Entity explanation: Named Entities, NP Syntax Entity explanation: Named Entities, NP Syntax Cohesive devices: demonstratives, pronouns, definite descriptions, sentence-initial discourse connectives Cohesive devices: demonstratives, pronouns, definite descriptions, sentence-initial discourse connectives Sentence fluency: length, fragments, etc. Sentence fluency: length, fragments, etc. Coh-Metrix: Psycholinguistic readability measures Coh-Metrix: Psycholinguistic readability measures Word Coherence Word Coherence Treat adjacent sentences as parallel texts Treat adjacent sentences as parallel texts Calculate “translation model’’ in each direction Calculate “translation model’’ in each direction

Automated Measures (cont) Continuity Continuity Summarization specific : Measures likelihood that discourse connectives retain their context. Does previous sentence in summary match previous sentence in input? Summarization specific : Measures likelihood that discourse connectives retain their context. Does previous sentence in summary match previous sentence in input? Cosine similarity of words across adjacent sentences. Cosine similarity of words across adjacent sentences. Coreference: Pronoun resolution system. Probability of antecedent presence in sentence, previous sentence. Coreference: Pronoun resolution system. Probability of antecedent presence in sentence, previous sentence. Entity coherence Entity coherence Matrix of entities’ grammatical roles; measure transition probabilities among entity’s role in adjacent sentence. Matrix of entities’ grammatical roles; measure transition probabilities among entity’s role in adjacent sentence.

Experiment Setup Data from summarization task of 2006/2007 Document Understanding Conference Data from summarization task of 2006/2007 Document Understanding Conference 2006 (training/dev sets) 50 inputs, 35 systems tested 2006 (training/dev sets) 50 inputs, 35 systems tested Jackknifing Jackknifing 2007 (test set) 45 inputs, 32 systems 2007 (test set) 45 inputs, 32 systems One ranker for each feature group, plus meta- ranker. One ranker for each feature group, plus meta- ranker. Rank systems/summaries relative to a gold standard human ranking based on each automated measure. Rank systems/summaries relative to a gold standard human ranking based on each automated measure. Find correlations with human ranking on aspects. Find correlations with human ranking on aspects.

Results (System-Level) Prediction Accuracy Prediction Accuracy Percentage of pairwise comparisons matching gold standard. Percentage of pairwise comparisons matching gold standard. Baseline: 50% (random) Baseline: 50% (random) System-level: (for summarization system) System-level: (for summarization system) Prediction accuracies around 90% for all aspects Prediction accuracies around 90% for all aspects Sentence fluency method single best correlation with Grammaticality. Meta-ranker has best overall correlation. Sentence fluency method single best correlation with Grammaticality. Meta-ranker has best overall correlation. Continuity method best correlates with Non- Redundancy, Referential Clarity, Focus, Structure. Continuity method best correlates with Non- Redundancy, Referential Clarity, Focus, Structure.

Results (Input-Level) Input-level: (for each summary) Input-level: (for each summary) Prediction accuracies around 70% -- harder task. Prediction accuracies around 70% -- harder task. Sentence fluency method single best correlation with grammaticality. Sentence fluency method single best correlation with grammaticality. Coh-Metrix single best correlation with Non- Redundancy Coh-Metrix single best correlation with Non- Redundancy Continuity best correlates with Referential Clarity, Focus, Structure. Continuity best correlates with Referential Clarity, Focus, Structure. Meta-ranker best correlation for all aspects. Meta-ranker best correlation for all aspects.

Results (Human-Written) Input-level analysis on human-written, abstractive summaries. Input-level analysis on human-written, abstractive summaries. Abstractive: Rewritten content Abstractive: Rewritten content Extractive: Extracts subset of content, i.e. picking sentences Extractive: Extracts subset of content, i.e. picking sentences Grammaticality: NP Syntax (64.6%) Non-redundancy: Coherence devices (68.6%) Referential Clarity: Sentence Fluency, Meta-Ranker (80.4%) Focus: Sentence Fluency, LMs (71.9%) Structure: LMs (78.4%) Grammaticality: NP Syntax (64.6%) Non-redundancy: Coherence devices (68.6%) Referential Clarity: Sentence Fluency, Meta-Ranker (80.4%) Focus: Sentence Fluency, LMs (71.9%) Structure: LMs (78.4%)

Components of Continuity Subsets of features in continuity block removed one- at-a-time to measure effect of each. Cosine similarity had greatest effect (-10%) Summary-specific features were second (-7%) Removing coreference features had no effect.

Conclusions Continuity features correlate with linguistic quality of machine-written summaries. Continuity features correlate with linguistic quality of machine-written summaries. Sentence fluency features correlate with grammaticality. Sentence fluency features correlate with grammaticality. LM and entity coherence features also correlate relatively strongly. LM and entity coherence features also correlate relatively strongly. This will make testing systems easier. Hooray! This will make testing systems easier. Hooray!

Questions?