Summarization Summarization Automatic Summarization (and other stuff) Taylor Berg-Kirkpatrick CS 288 UC Berkeley Multidocument Summarization Sentence Extraction Sentence Extraction Lindsay Lohan pleaded not guilty Lindsay Lohan pleaded not guilty Wednesday to felony grand theft Wednesday to felony grand theft of a 2,500 necklace, a case that of a 2,500 necklace, a case that could return the troubled starlet could return the troubled starlet to jail rather than the big screen. to jail rather than the big screen. Saying it appeared that Lohan had violated her probation in a 2007 drunken driving case the judge set bail at $40,000 and warned that if Lohan was accused of breaking the law while free he would have her held without bail.
Sentence Extraction Maximum Marginal Relevance Maximum Marginal Relevance S 9 Lindsay Lohan pleaded not guilty Wednesday to felony grand theft of a 2,500 necklace, a case that S 2 S 5 could return the troubled starlet to jail rather than the big screen. S 4 Saying it appeared that Lohan had S 1 violated her probation in a 2007 S 7 S 8 drunken driving case the judge S 3 set bail at $40,000 and warned that if Lohan was accused of S 6 breaking the law while free he would have her held without bail. The Mean Girls star is due back in court on Feb. 23 an important hearing in which Lohan could opt to end the case early. Maximum Marginal Relevance Maximum Marginal Relevance Maximum Marginal Relevance S 9 S 9 S 9 S 2 S 5 S 2 S 5 S 2 S 5 S 4 S 4 S 4 S 1 S 1 S 1 S 7 S 7 S 7 S 8 S 8 S 8 S 3 S 3 S 3 S 6 S 6 S 6 S 1 : She traveled to France on Friday. S 1 : She traveled to France on Friday.
Maximum Marginal Relevance Maximum Marginal Relevance Maximum Marginal Relevance S 9 S 9 S 9 S 2 S 5 S 2 S 5 S 2 S 5 S 4 S 4 S 4 S 1 S 1 S 1 S 7 S 7 S 7 S 8 S 8 S 8 S 3 S 3 S 3 S 6 S 6 S 6 S 1 : She traveled to France on Friday. S 1 : She traveled to France on Friday. S 1 : She traveled to France on Friday. S 4 : On Friday, She took a trip to France. S 4 : On Friday, She took a trip to France. S 4 : On Friday, She took a trip to France. S 8 : She plans to stay for two weeks. Maximum Marginal Relevance Maximum Marginal Relevance Maximum Marginal Relevance S 9 S 9 S 9 S 2 S 5 S 2 S 5 S 2 S 5 S 4 S 4 S 4 S 1 S 1 S 1 S 7 S 7 S 7 S 8 S 8 S 8 S 3 S 3 S 3 S 6 S 6 S 6 S 1 : She traveled to France on Friday. S 4 : On Friday, She took a trip to France. S 8 : She plans to stay for two weeks.
Maximum Marginal Relevance Maximum Marginal Relevance Maximum Marginal Relevance S 9 S 9 S 9 S 2 S 5 S 2 S 5 S 2 S 5 S 4 S 4 S 4 S 1 S 1 S 1 S 7 S 7 S 7 S 8 S 8 S 8 S 3 S 3 S 3 S 6 S 6 S 6 � argmax λ · sim( S i , D ) − (1 − λ ) · max (sim( S i , S j )) j i ∈ D \ S [Carbonell and Goldstein, 1998] Maximum Marginal Relevance Maximum Marginal Relevance Max Coverage S 9 S 9 S 2 S 5 S 2 S 5 S 4 S 4 S 1 S 1 S 7 S 7 S 8 S 8 S 3 S 3 She stopped in France. In France she remained. S 6 S 6 Centrality Centrality Redundancy � � argmax λ · sim( S i , D ) − (1 − λ ) · max (sim( S i , S j )) argmax λ · sim( S i , D ) − (1 − λ ) · max (sim( S i , S j )) j j i ∈ D \ S i ∈ D \ S [Carbonell and Goldstein, 1998] [Carbonell and Goldstein, 1998]
Max Coverage Max Coverage Max Coverage She stopped in France. In France she remained. She stopped in France. In France she remained. She stopped in France. In France she remained. Max Coverage Max Coverage Max Coverage A summary is a set of cuts: s She stopped in France. In France she remained. She stopped in France. In France she remained. She stopped in France. In France she remained. (she, stopped) (in, france) (she, remained) (stopped, in) (france, she)
Max Coverage Max Coverage Max Coverage She stopped in France. In France she remained. She stopped in France. In France she remained. She stopped in France. In France she remained. (she, stopped) (in, france) (she, remained) (she, stopped) (in, france) (she, remained) (she, stopped) (in, france) (she, remained) (stopped, in) (france, she) (stopped, in) (france, she) (stopped, in) (france, she) Set of bigrams covered by summary: B ( s ) Max Coverage Max Coverage Max Coverage X coverage( s ) coverage( s ) value( b ) max max max s s s b ∈ B ( s ) length( s ) ≤ L max length( s ) ≤ L max s.t. s.t.
Max Coverage Max Coverage X X X value( b ) value( b ) value( b ) max max max s s s b ∈ B ( s ) b ∈ B ( s ) b ∈ B ( s ) length( s ) ≤ L max length( s ) ≤ L max length( s ) ≤ L max s.t. s.t. s.t. value( b ) = freq( b ) value( b ) = freq( b ) [Gillick and Favre 2008] [Gillick and Favre 2008] ILP for Decoding ILP for Decoding ILP for Decoding X X X value( b ) z b · value( b ) z b · value( b ) max max max s s,z s,z b ∈ B ( s ) b b length( s ) ≤ L max length( s ) ≤ L max length( s ) ≤ L max s.t. s.t. s.t. bigrams in are covered s
ILP for Decoding ILP for Decoding Linear Model for Extraction X X X z b · value( b ) z b · value( b ) value( b ) max max max s,z s,z s b b b ∈ B ( s ) X length( s ) ≤ L max l n s n ≤ L max length( s ) ≤ L max s.t. s.t. s.t. n bigrams in are covered ∀ n,b s s n Q nb ≤ z b value( b ) = freq( b ) only bigrams in are covered X ∀ b s n Q nb ≥ z b s n Linear Model for Extraction Sentence Extraction Extraction and Compression Lindsay Lohan pleaded not guilty Lindsay Lohan pleaded not guilty Wednesday to felony grand theft Wednesday to felony grand theft X value( b ) of a 2,500 necklace, a case that of a 2,500 necklace, a case that max could return the troubled starlet could return the troubled starlet s to jail rather than the big screen. to jail rather than the big screen. b ∈ B ( s ) Saying it appeared that Lohan had Saying it appeared that Lohan had violated her probation in a 2007 violated her probation in a 2007 length( s ) ≤ L max s.t. drunken driving case the judge drunken driving case the judge set bail at $40,000 and warned set bail at $40,000 and warned that if Lohan was accused of that if Lohan was accused of breaking the law while free he breaking the law while free he Parameterize using features: would have her held without bail. would have her held without bail. The Mean Girls star is due back The Mean Girls star is due back in court on Feb. 23 an important in court on Feb. 23 an important value( b ) = w > f ( b ) hearing in which Lohan could opt hearing in which Lohan could opt to end the case early. to end the case early.
Extraction and Compression Extraction and Compression Joint Extractive / Compressive Model Lindsay Lohan pleaded not guilty Wednesday to felony grand theft of a 2,500 necklace, a case that could return the troubled starlet to jail rather than the big screen. Saying it appeared that Lohan had violated her probation in a 2007 drunken driving case the judge She stopped in France. In France she remained. She stopped in France. In France she remained. set bail at $40,000 and warned that if Lohan was accused of breaking the law while free he would have her held without bail. The Mean Girls star is due back in court on Feb. 23 an important hearing in which Lohan could opt to end the case early. [Martins and Smith 2009] [Woodsend and Lapata 2010] Joint Extractive / Compressive Model Joint Extractive / Compressive Model Joint Extractive / Compressive Model s She stopped in France. In France she remained. She stopped in France. In France she remained. She stopped in France. In France she remained.
Joint Extractive / Compressive Model Joint Extractive / Compressive Model Joint Extractive / Compressive Model s s s She stopped in France. In France she remained. She stopped in France. In France she remained. She stopped in France. In France she remained. (she, stopped) (in, france) (she, remained) (she, stopped) (in, france) (she, remained) (she, stopped) (in, france) (she, remained) (stopped, in) (france, she) (stopped, in) (france, she) (stopped, in) (france, she) B ( s ) B ( s ) Joint Extractive / Compressive Model Joint Extractive / Compressive Model Joint Extractive / Compressive Model h h i h i X X X X X value( b ) value( b ) value( b ) value( c ) value( c ) + + max max max s s s b ∈ B ( s ) b ∈ B ( s ) c ∈ s b ∈ B ( s ) c ∈ s Parameterize using features: value( b ) = w > f ( b ) value( c ) = w > f ( c )
Learning Learning Features Linear prediction: Linear prediction: score( s ) = w > f ( s ) score( s ) = w > f ( s ) Feature function factors: X X f ( s ) = f ( b ) f ( c ) + b ∈ B ( s ) c ∈ s Features Features Features Bigram Features f ( b ) Bigram Features f ( b ) Cut Features f ( c ) Bigram Features f ( b ) Cut Features f ( c ) COUNT: Bucketed document counts Stop word indicators STOP: POSITION: First document position indicators CONJ: All two- and three-way conjunctions of above BIAS: Always one
Recommend
More recommend
