: ) B 2.7) A B 2.7) 3) 4) 5) substring) subsequence) A LCStr, LCS s = s 1,..., s m, t = t 1,..., t m character) mo

Transcription

1 1,a) [1] BCCWJ) A B C 3 A B B C ) 1 NINJAL, Tachikawa, Tokyo , Japan 2 MEXT, Chiyoda, Tokyo , Japan a) masayu-a@ninjal.ac.jp ) A, B, C Information Structure) Information Status) Topic) Focus) [2] old) new) accessible) c 2015 Information Processing Society of Japan 1

2 : ) B 2.7) A B 2.7) 3) 4) 5) substring) subsequence) A LCStr, LCS s = s 1,..., s m, t = t 1,..., t m character) morpheme) string) character) character-based string) morpheme) morpheme-based) substring) n n-gram s i n-gram s i,...,i n+1 subsequence) p p-mer s p-mer i = i 1,..., i p 1 i 1 < i 2 < < i p s ) s[ i] Longest Common String: LCStr) Longest Common String) abbreviation LCS Longest Common Subsequence) LCS LCStr, LCS s, t : LCStrs, t) = arg max n s i,...,i n+1 j,s i,...,i n+1=t j,...,j n+1 s, t LCStr ) : c 2015 Information Processing Society of Japan 2

3 LCStr s, t) = max n i, j,s i,...,i n+1=t j,...,j n+1 [0,1] : Score LCStr s, t) = 2 LCStr s + t Longest Common Subsequence: LCS) Levenshtein s, t Longest Common Subsequence: LCS) : LCSs, t) = arg max i s[ i] j,s[ i]=t[ j] s, t LCS ) : LCSs, t) = max i i, j:s[ i]=t[ j] [0,1] : Score LCS s, t) = 2 LCS s + t ) Levenshtein ) LCS : d Levenshtein s, t) = s + t 2 LCS LCS Ulam LCS LCStr LCS LCS LCS LCSC, R) R ) LCSC, R) C ) LCSC, R) R = LCSC, R) C = arg max i j j j1) i, j,c[ i]=r[ j] arg max i i i i1) i, j,c[ i]=r[ j] R WLCS C, R) P WLCS C, R) R WLCS C, R) = α LCS RC,R) LCS LCS R P WLCS C, R) = α LCS CC,R) LCS LCS s Score γ) WLCS C, R) = 1 + γ2 )R WLCS C, R)P WLCS C, R) R WLCS C, R) + γ 2 P WLCS C, R) 2.3 / ROUGE-L [3] ROUGE-L [3] LCS) Score γ) ROUGE-L C, R) = 1 + γ2 ) R LCS C, R) P LCS C, R) R LCS C, R) + γ 2 P LCS C, R) R LCS C, R) P LCS C, R) : R LCS C, R) = LCSC,R) R P LCS C, R) = LCSC,R) C c i C r j R LCS ROUGE-W [3] LCS LCSC, R) R + LCSC, R) C fx) : fx + y) > fx) + fy), x > 0, y > 0, x N, y N N ) LCS ROUGE-W fx) = x α Score γ) C, R) WLCS ROUGE-N [3], [4] ROUGE-N [3], [4] n-gram c 2015 Information Processing Society of Japan 3

4 Score R) C, R) = ROUGE-N e n-gram clip C,R) e e n-gramr) e e n-gramc) C n-gram n-gramr) R n-gram n-gram clip C, R) n-gram e n-gramc) e n-gramr) : n-gram clip C, R) { n-gramc) if n-gramc) n-gramr) = n-gramr) otherwise ROUGE-SU) [3], [5] ROUGE-S 2-mer Score γ) ROUGE-S C, R) = 1 + γ2 )P s C, R)R s C, R) R S C, R) + γ 2 P S C, R) P S C, R) R S C, R) : e 2-mer clip C,R) P S C, R) = e e 2-merC) e 2-mer clip C,R) R S C, R) = e e 2-merR) p-merc): p-mer p-merr): p-mer p-mer clip C, R) p-mer e p-merc) p-mer e p-merr) : p-mer clip C, R) { p-merc) if p-merc) p-merr) = p-merr) otherwise ROUGE-SU ROUGE-S p = 2 p ESK [6] ESK [6] p-mer e e e Score p-mer ESK = C, R) u p-merc) v p-merr) λ e p) u u ) + u,u p-merc) λ e p δu, v) u v v,v p-merr) λ e p) v v ) [6] 2-mer BLEU[7] BLEU [7] n n-gram ω n ) P n-gram C, R) = BLEU e n-gram clip C,R) e e n-gramc) N Score BLEU C, R) = BP C, R) exp ω n log P n-gram C, R)) BLEU n=1 N n ω n = 1 Brevity Penalty BP) BPC, R) = { IMPACT [8] e 1 if C > R exp1 r c ) if C R IMPACT[8] LCS LCStr RN α r e β )) 1 r=0 e LCStrC R IP C, R) = r),r r) ) β R β RN α i e β )) 1 r=0 e LCStrC P IP C, R) = r),r r) ) β C β α rr RN) c 2015 Information Processing Society of Japan 4

5 α < 1.0) β LCStr β > 1.0) C 1) = C R 1) = R C r) = C r 1) \ {LCStrC r 1), R r 1) )} R r) = R r 1) \ {LCStrC r 1), R r 1) )} Score IP = 1 + γ2 )R IP P IP R IP + γ 2 P IP LCStr LCStr RN RIBES [9] RIBES [9] Score RIBES = d Kendall 1-gram align C, R)) P RIBES C, R)) α ) BPC, R) ) β d Kendall µ, ν) µ, ν Kendall 1-gram align µ, ν) [9] wonder 2 1-gram ) P RIBES C, R) = 1-gram align C,R) C 1-gram align µ, ν) wonder ) α β BLEU BP P RIBES C, R) : Score Kendall LRscore C, R) = α BP C, R) d Kendall Ĉ, ˆR) α)score BLEU ) [11] [0,1] ) : Score K s, t) = K s, t) K s, s) K t, t) - γ : Score γ) 1 + γ 2 )K s, t) K s, t) = K s, s)) 2 + γ 2 K t, t)) All String Kernel or Exact Matching Kernel) n u F all str P RIBES C, R) = Score P ) C, R) ROUGE-1 e e 1-gram clip C,R) = e LRscore [10] e 1-gramR) LRscore [10] Hamming Kendall Score Hamming LRscore C, R) = α BP C, R) d Hamming Ĉ, ˆR)+ 1 α)score BLEU Φ str : σ F all str R σ Φ str = ϕ us)) u σ Kn-grams, t) = Φ str s), Φ str t) F all str = u σ ϕ us)ϕ t s) ϕ us) = {i s i... = u} : K all seq s, t) = min s, t ) n=1 s n+1 i=1 t n+1 j=1 δs i...i+n 1, t i...i+n 1 ) 2002 c 2015 Information Processing Society of Japan 5

6 n-gram Length Weighted All String Kernel or Length Weighted Exact Matching Kernel) K all seq s, t) = min s, t ) n=1 s n+1 i=1 t n+1 j=1 ω s δs i...i+n 1, t i...i+n 1 ) ω n n IMPACT n- Suffix Tree n- n-gram Spectrum Kernel) n n-gram) n u Fn-gram Ψ seq : σ F all seq R σ Ψ seqs) = ψ vs)) v σ ψ vs) = { i s[ i] = v} K all seq s, t) = Ψ seqs), Ψ seqt) F all seq = v σ ψ vs) ψ vt) ψ vs) = { i s[ i] = v} K all seq s, t) O s t ) ϵ K all seq s, ϵ) = K all seq t, ϵ) = 1 K all seq s, t) K all seq s a, t) = K all seq s, t) + 1 i t,j:t j=a K all seq s, t i...j 1) K all seq s a, t) = K all seq s, t i...j 1 ) K all seq s a, t b) = K all seq s a, t) + δa, b)ks, t) t p p-mer) p v Fp-mer s Φ n str : σ Fn-gram R σ n Φ n str = ϕn us)) u σ n Kn-grams, t) = Φ n str s), Φn str t) Fn-gram = u σ p ϕ n us)ϕ n t s) ϕ n us) = {i s i...i+n 1 = u} : Kn-grams, t) = s n+1 i=1 t n+1 j=1 δs i...i+n 1, t j...j+n 1 ) ROUGE-N Kn-gramC, R) n-gram Kn-grams, t) 1-gram 1-mer BLEU BP v F all seq Ψ p seq : σ Fp-mer R σ p Ψ p seq s) = ψp vs)) v σ ψvs) = { i s[ i] = v} Kp-mers, t) = Ψ p seq s), Ψp seq t) Fp-mer = ψvs) p ψvt) p v σ p ψvs) p = { i s[ i] = v} ROUGE-S K 2-mer C, R) 2-mer ROUGE-SU K 1-mer,2-mer C, R) 1-mer, 2-mer Kp-mers, t) : p-mer λ ESK [6] p v Fp-mer c 2015 Information Processing Society of Japan 6

7 1 ) ) [0, 1] [0, ] [0, ] [ 1, 1] IMPACT [8] Score γ) K ) all str n-gram) ROUGE-N BLEU [7] Score γ) Kn-gram LRscore [10] n Score γ) K ) all seq p-mer) ROUGE-SU) [3], [5] Score γ) Kp-mer ESK [6] Score γ) Kgap p-mer Score rank θ p-mer p-mer ) Score footrule d footruleθ=1) RIBES? [9] Score Spearman d Spearmanθ=2) 2) Spearman s ρ Pearson s LRscore [10] Score Hamming d Hamming RIBES [9] Score Kendall d Kendall Kendall s τ ) LRscore [10] d Caylay d Ulam ) ROUGE-L Score LCS d Levenshtein ) ROUGE-W [3] Score γ) WLCS ) Score LCStr Kendall d Kendall 1, 4, 3, 2), 1, 2, 3, 4)) = ) ) ) d Kendall 2, 3, 1, 4), 1, 2, 3, 4)) = ) ) Caylay d Caylay 1, 4, 3, 2), 1, 2, 3, 4)) = ) Ulam d Ulam 1, 4, 3, 2), 1, 2, 3, 4)) = ) ) d Caylay 2, 3, 1, 4), 1, 2, 3, 4)) = ) ) d Ulam 2, 3, 1, 4), 1, 2, 3, 4)) = ) 1 c 2015 Information Processing Society of Japan 7

8 Kgap p-mers, t) = Ψ gap seq p p s), Ψgap seq t) Fp-mer = v σ p ψ gap p v s) ψ gap p t) ψv gap p s) = i:v=s[ i] λl i) li) = s i1,...,i v i = i 1,..., i v ) [12] m µ, ν S m m θ- : m d Rank θ µ, ν) = µi) νi) θ ) 1/θ i=1 θ = 1 Spearman footrule m d Footrule µ, ν) = µi) νi) ) i=1 θ = 2 Euclid Euclid 2 Spearman m d Spearman µ, ν) = µi) νi) 2 ) i=1 Spearman Euclid 2 [-1, 1] Spearman ρ Spearman s ρ = 1 6 d Spearman µ, ν) m 3 m µ, ν Pearson *1 Hamming d Hamming µ, ν) = v m δµi), νi)) i=1 Hamming 1) * µ ν Levenshtein 1 Kendall : Kendall d Kendall Swap) Kendall mm 1) 2 d Kendall = minarg max δπ q q=1 π 2k q, k q +1)) µ, ν)) q d Kendall = m m i=1 j=i+1 χi, j) χ i, j 0 1 : χ = { 1 if µi) µj))νi) νj)) < 0, 0 if µi) µj))νi) νj)) 0 [0,1] : Score Kendall = 1 2 d Kendall µ, ν) m 2 m [-1,1] Kendall τ Kendall s τ = 1 4 d Kendall µ, ν) m 2 m Cayley : Cayley d Caylay Swap) d Caylay = minarg max δπ q q=1 π 2k q, l q )) µ, ν)) q Ulam : Ulam d Ulam i, i + 1,..., j 1, j c 2015 Information Processing Society of Japan 8

9 µ ν Ulam d Ulam µ, ν) = m LCSµ, ν) [0,1] : Score Ulam µ, ν) = 1 d Ulam µ, ν) m LCSµ, ν) = m = Score LCS µ, ν) [13] swap Kendall Ulam Spearman s ρ Kendall s τ Daniels : 1 3m + 2) 2m + 1) τ m 2 m 2 ρ 1 m 1 3τ 2ρ 1 d Caylay d Kendall Footrule Kendall Cayley Diaconis-Graham inequality): d Kendall + d Caylay d Footrule 2 d Kendall Spearman Kendall Durbin-Stuart inequality): 4 3 d Kendall 1 + d Kendall m ) d Spearman [14] 1 Score {Score } ) ω Score = log Score = ω ΠScore ω 1 ω w log Score ) substring : n-gram ) subsequence : p-mer ) Ulam BCCWJ-SUMM) GROSS) RETELLING) 3 ) ) BCCWJ-SUMM C BCCWJ-SUMM C BCCWJ Yahoo! 15 ) BCCWJ 1 19 BCCWJ PN A) PC c 2015 Information Processing Society of Japan 9

10 2 BCCWJ-SUMM C BCCWJ-SUMM L GROSS C 71,111,113 GROSS L 4 7,6,3 RETELLING I 10 5 RETELLING K 3 3,3,3 3 RETELLING M BCCWJ-SUMM L BCCWJ-SUMM C FileID A A B B C C C D D D D E E E E E E F F BCCWJ-SUMM L BCCWJ-SUMM L BCCWJ BCCWJ- SUMM C ) FileID A A B B C C C Q ) : BCCWJ-SUMM LP)) BCCWJ-SUMM LT)) GROSS C GROSS C Yahoo! 15 ) 6.6) 6.4) 6.0) 3 * :71:111 :113295/300) GROSS L GROSS L ) GROSS C ) 6.4) 6.0) *2 [15] c 2015 Information Processing Society of Japan 10

11 max 227, min 85 ) 1 : GROSS LP)) GROSS LT)) RETELLING I Retelling [16], [17] [18] ) ) ) ) ) RETELLING IT)) RETELLING K [19] 3 * ) ) ) RETELLING K * : RETELLING KP)) RETELLING KT)) RETELLING M [20] ), 10 ) min:150 max:451 ) 107 min:74 max:152 ) ) 1 : RETELLING MP)) RETELLING MT)) n-gram 1,2,3,4) char/mrph) n-gram 2, 3, 4) char/mrph) p-mer 2,3,4) char/mrph) p-mer 2, 3, 4) char/mrph) 1-gram +Footrule char/mrph) =Spearman) 1-gram +Kendall char/mrph) A 1,A 2 Mean) SD) c m MeCab-0.98+IPADIC ) p n-gram1),n-gram2),pmer2),kendall unigramn-gram1)) GROSS LT) BCCWJ-SUMM LT) Bigramn-gram2)), skip-bigramp-mer2)) c 2015 Information Processing Society of Japan 11

12 Bigramn-gram2)) skip-bigramp-mer2)) bigram Kendall bi-gram BCCWJ-SUMM C BCCWJ-SUMM LP), GROSS C GROSS LP)) ) F ) ) * 4 BCCWJ-SUMM LP) GROSS LP) RETELLING KP) RETELLING MP) BCCWJ-SUMM LP) GROSS LP) Kendall char n-gram2,3,4) char, n- gram2,3,4, 2, 3, 4) mrph, Kendall mrph Footrule mrph, BCCWJ-SUMM LP) RETELLING KP) n-gram3,4) mrph BCCWJ-SUMM LP) RETELLING KM) GROSS LP) RETELLING {K,M}P) RETELLING KP) RETELLING MP) n-gram 3, 4) mrph,p-mer3,4, 3, 4) n-gram1) *4 = ) BCCWJ-SUMM LT) GROSS LT) RETELLING IT) RETELLING KT) RETELLING MT) BCCWJ-SUMM LT) GROSS LT) Kendall char n-gram2,3,4) char, n- gram2,3,4, 2, 3, 4) mrph, Kendall mrph Footrule mrph, BCCWJ-SUMM LT) RETELLING {I,K,M}T) GROSS LT) RETELLING {I,K,M}T) RETELLING IT) RETELLING KT) n-gram1,4, 2) char, p- mer2, 2) char RETELLING IT) RETELLING MT) Kendall char RETELLING IT) RETELLING MT) n-gram2, 2, 3, 4) char, p-mer2,3,4, 2, 3, 4) char n-gram1,2, 2, 3, 4) mrph, p- mer2,3,4, 2, 3, 4) mrph RETELLING {I,K}) RETELLING M) RETELLING I) RETELLING K) BCCWJ-SUMM C GROSS C c 2015 Information Processing Society of Japan 12

13 $"!!##!",!##!"+!##!"*!##!")!##!"!##!"'!##!"&!##!"%!##!"$!##!"!!## -../ # -../ # $"!!##!",!##!"+!##!"*!##!")!##!"!##!"'!##!"&!##!"%!##!"$!##!"!!## -../ # -../ #!"#$%&')*&$+,! -../ :9# ;<=225.# ;<= # ;<=22567:9# -../ :9# ;<=225.# ;<= # ;<=22567:9# 3 <>:>66?@;5A7:9# <>:>66?@;54789# <>:>66?@;547:9# <>:>66?@;5?7:9# <>:>66?@;5A789# <>:>66?@;5A7:9# <>:>66?@;54789# <>:>66?@;547:9# ) BCCWJ-SUMM C BCCWJ-SUMM LP) n-gram2) char, n-gram3) char, n-gram4) char BCCWJ-SUMM C) BCCWJ-SUMM LP)) ) GROSS C GROSS LP) n-gram2,3,4) char, $"!!##!",!##!"+!##!"*!##!")!##!"!##!"'!## 4BCD#!"&!##!"%!##!"$!## 2E#!"!!## $"!!##!",!##!"+!##!"*!##!")!##!"!##!"'!## 4BCD#!"&!##!"%!##!"$!## 2E#!"!!## -../ # -../ #!"#$%&')*&$+,! -../ :9# ;<=225.# ;<= # ;<=22567:9# <>:>66?@;5?7:9# <>:>66?@;5A789#!"#$%&&')*+! <>:>66?@;5A7:9# <>:>66?@;54789# <>:>66?@;547:9# -../ # -../ # -../ :9# ;<=225.# ;<= # ;<=22567:9# <>:>66?@;5?7:9# <>:>66?@;5A789# <>:>66?@;5A7:9# <>:>66?@;54789# <>:>66?@;547:9# n-gram1),n-gram2),p-mer2),kendall: ) n-gram2,3,4) mrph, Footrule mrph, Kendall mrph wikipedia BCCWJ-SUMM LP) BCCWJ-SUMM LT), GROSS LP) GROSS LT), RETELLING KP) RETELLING KT), RETELLING MP) RETELLING MT) BCD# 2E# 4BCD# n-gram n-gram p-mer, Footrule, Kendall n-gram, p-mer n, p n-gram, p-mer n or p) n or p) n or p) n-gram, p-mer n-gram1) * Kendall * n- gram1) * Kendall * 2E# c 2015 Information Processing Society of Japan 13

14 2.8 n-gram p-mer 7 ) Information Structure)[21] : Information Status) Topic) Focus) ) Götze[2] BCCWJ-SUMM 3.1 Götze[2] 5 Götze ) discourse referents) 5 Götze ) [2] Layers Tags Description giv Given ) Information Status) acc Accessible) new cat nil New ) cataphor ) ab Aboutness topic non-referential ) Topic) fs Frame setting topic nf New Information Focus Focus) cf Contrastive Focus ) information status) retrievability) given) accessible) new Prince [22] Prince { discourse-old) discoursenew)} { hearer-old) hearer-new)} 6) Prince [22] giv =, ) acc =, ) new =, ) Prince, ) topic) Jacobs[23] aboutness topic) frame setting topic) 2 what the sentence is about ) the frame within which the sentence holds ) : Frame-setting [23], p.656) X,Y) X Y Y X focus) contrast focus) Götze[2] Cook [24] c 2015 Information Processing Society of Japan 14

15 6 Prince Prince giv ) evoked acc) unused - new ) brand-new 588 aboutness topics Fleiss κ) NP NP new new - ) acc-inf new - acc-inf new Götze[2] 7 giv-active: giv-inactive *5 : acc-sit: ) acc-aggr: [2] Peter went shopping with Maria. They bought many flowers. acc-inf set-rel) giv acc-inf acc-inf: [2] part-whole: The garden beautiful. Its entrance is just across this river. set-rel: The flowers in the garden blossom. The flowers near the gate blossom violet. set-rel: The children swam in the lake. The family experienced a beautiful day. entity-attribute: The flowers enchanted Peter. Their scent was wonderful referent given) *5 inactive discourse-new + hearer-new semi-active textually accessible c 2015 Information Processing Society of Japan 15

16 7 Coarse) Fine) Description giv underspecified) given: giv-active giv-inactive active: inactive: acc underspecified) accessible: new nil cat) acc-sit acc-aggr acc-inf acc-gen situationally accessiblly accessible: aggregation: inferable: part-whole, set-relsubset/superset),entityattribute ) general: new: non-referential: cataphor: acc-gen: Type) ) Token) acc-gen acc-gen ) new ) new new: nil cat catphora) *6 3.3 MAMA [25] 1 [2] { ) *6 2 ) } discourse-old) ) giv-active ) giv-inactive giv-active ) acc-aggr discourse-new) acc-sit acc-inf activeness) - ) hearer-old) ) acc-gen ) ) hearer-new) new 3 ) ) ) ) 4 8 Fine)) Fine) c 2015 Information Processing Society of Japan 16

17 9 4 B-C ) B-C a-g a-i g-a g-i new nil acc-gen 3 3 acc-inf giv-active 1 1 giv-inactive new nil Coarse) 2 A, B, C 2 Cohen s)κ 3 Fleiss s)κ 1 B-C κ 9 4 B-C nil ) acc ) 3.4 ) giv ) acc ) BCCWJ NAIST [26] acc-aggr [27] 4. 2 ) 4.1 Kennedy [29] Dundee Eye- Tracking Corpus Dundee Eye-Tracking Corpus Demberg [30] Gibson Dependency Locality TheoryDLT) integration cost[31] Hale suprisal[32] Dundee Corpus Roland [33] Demberg [34] Dundee c 2015 Information Processing Society of Japan 17

18 8 Cohen s κ) A-B) B-C) C-A) Fleiss s κ 1 Fine) Coarse) Fine) = Coarse) Fine) = Coarse) Fine) Coarse) Corpus [35] Linger *7 5 1 Yes/No Question SRResearch EyeLinkCL 5 1 Yes/No Question 5 1/2 interest area grid ) interest area interest area BCCWJ 4.4 * : 5 : {,,,, } 5 : 15 [15] *8 [37] [38] *8 goitokusei/goi-test.html c 2015 Information Processing Society of Japan 18

19 First pass time Total time Regression path time ) ) ) A 3 A 4,A 5, ) accessible 4 [39] b) 1a) [40] 1) a. The lady i [ that GAP i visited the banker ]... b. The lady i [ that the banker visited GAP i ]... )[41][42] ) [31] 6 ) 6 ) 2) a. GAP i i... b. GAP i i... [43] Roland [44] c 2015 Information Processing Society of Japan 19

20 1a) 57% 43% 1b) 2% 98% 4a) 4b) Roland [44] abstract ) 3) a. ) The banker was very friendly. b. ) There was a dinner party on Saturday night. 4) a. The lady i [ that GAP i visited the banker ] enjoyed the dinner very much. b. The lady i [ that the banker visited the GAP i ] enjoyed the dinner [45] 4 ) BCCWJ % 30% 20% 80% 5) [45]p.63) ) 5)1-a. 1-b. 2-a. 2-b. 1 ) 1 ) 2 ) ) 2 ) ) 3. 3 ) [46][47][48] Gordon [46] 6) Gordon [46] *9 ) 6) a. It was the barber/john that saw the lawyer/bill in the parking lot. b. It was the barber/john that the lawyer/bill saw in the parking lot. [49][50][51] [50] 5 ) 7a) 7b) [50] 5 ) ) 7) a. b. [49] %) %) 656 6a) 31% 6b) 8% 6c) 45% [49] ) 8) a. GAP i i b. GAP i i *9 Gordon description: the barber, the lawyer ) name: John, Bill ) c 2015 Information Processing Society of Japan 20

21 c. ) GAP i 9 15 i d. - / + / [49] ) /357) /114) [52] [52] ) 9) a. 2-b. focus) ) ) ) ) ) accessible accessible-new kappa PC BCCWJ BCCWJ [53] c 2015 Information Processing Society of Japan 21

22 5.2 BCCWJ-Trans 11 BCCWJ-Trans A : σ : s = s 1,..., s m, t = t 1,..., t m character), character-based): σ s i σ σ morpheme), morpheme-based): σ s i σ σ string): character) character-based string) morpheme) morpheme-based) substring): n n-gram s i n-gram s i,...,i n+1 subsequence): p p-mer s p-mer i = i 1,..., i p 1 i 1 < i 2 < < i p s ) s[ i] / reference): / R / candidate): / / C distance): X 2 d : X X R dx, y) 0), x = y dx, y) 0), dx, y) = dy, x)), dx, y) + dy, z) dx, z)) absolute value): 0 x θ- norm): x =, x 1,..., x n θ- x θ = n i=1 x i p ) 1/p θ x ) 2- : ) similarity): correlation): [-1,1] kernel function): ) Ks, t) cosine Ks,t) Ks,s) Kt,t) ) score): [0,1] score prefix): suffix): subset): k k-element rank vector): i i m S m µ = µ1),..., µm) µi) i order vector): i i m T m µi) µ 1 = µ 1 1),..., µ 1 m) µ 1 i) i ) concordant): i j c 2015 Information Processing Society of Japan 22

23 11 BCCWJ-Trans OY 1, OC 1, PN 1, PB 1, PM 1, OW 1 ) OY 1, OC 1, PN 1, PB 1, PM 1, OW OY 6, OC 6, PN 1, PB 1, PM 1, OW OY 3, OC 3, PN 1, PB 1, PM 1, OW 1 : OY, OC, PN, PB, PM, OW µi) µj))νi) νj)) 0 discordant): : insertion) deletion) substitution) : symmetric group) : :permutation) µ 1 = µ 1 1),..., µ 1 m) µ 1 k 1 ), µ 1 k 2 ),..., µ 1 k r ) µ 1 k 1 ) µ 1 k 2 ), µ 1 k 2 ) µ 1 k 3 ),... µ 1 k 1 ) µ 1 k 2 )... µ 1 k r ) µ 1 k 2 ) µ 1 k 3 )... µ 1 k 1 ) π r = k 1, k 2,..., k r ) 2 ) transposition) π 2 = i, j) adjacent transposition) π 2 = i, i + 1) δ : δi, j) = A.2 { ) 1 i = j) 0 i j) A 1 A 2 BCCWJ-SUMM LT) RETELLING KT),RETELLING MT)) A.3 A 3 0-origin) ) A 4,A 5 ) ) ) 95% [-2.0, 2.0] * 2.5 **) NTT CS JSPS B) , B) B) [1] Maekawa, K., Yamazaki, M., Ogiso, T., Maruyama, T., Ogura, H., Kashino, W., Koiso, H., Yamaguchi, M., Tanaka, M. and Den, Y.: Balanced corpus of contemporary written Japanese, Language Resources and Evaluation, Vol. 48, pp ). [2] Götze, M., Weskott, T., Endriss, C., Fiedler, I., Hinterwimmer, S., Petrova, S., Schwarz, A., Skopeteas, S. and Stoel, R.: Information structure, Interdisciplinary Studies on Information Structure Dipper, S., Göetze, M., Skopeteas, S. and Stoel, R., eds.), Working Papers of the SFB 632, Vol. 7, Universiätsverlag, Potsdam, 2nd edition updated version 2014) edition, chapter Information Structure, pp ). [3] Lin, C.-Y.: ROUGE: A Package for Automatic Evaluation of Summaries, Proc. of Workshop on Summarization Branches Out, Post Conference Workshop of ACL 2004, pp ). [4] Lin, C.-Y. and Hovy, E.: Automatic Evaluation of Summaries Using N-gram Co-occurrence Statistics, Proc. of the 4th Meeting of the North American Chapter of the Association for Computational Linguistics and Human Language Technology, pp ). [5] Lin, C.-Y. and Och, F. J.: Automatic Evluation of Machine Translation Quality Using Longest Common Subsequence and Skip-Bigram Statistics, Proceedings of c 2015 Information Processing Society of Japan 23

24 A 1 score BCCWJ-SUMM C BCCWJ-SUMM LP) BCCWJ-SUMM LT) GROSS C GROSS LP) GROSS LT) Mean SD Mean SD Mean SD Mean SD Mean SD Mean SD n-gram1) c n-gram2) c n-gram3) c n-gram4) c n-gram 2) c n-gram 3) c n-gram 4) c p-mer2) c p-mer3) c p-mer4) c p-mer 2) c p-mer 3) c p-mer 4) c Footrule c Kendall c n-gram1) m n-gram2) m n-gram3) m n-gram4) m n-gram 2) m n-gram 3) m n-gram 4) m p-mer2) m p-mer3) m p-mer4) m p-mer 2) m p-mer 3) m p-mer 4) m Footrule m Kendall m the 42nd Annual Meeting on Association for Computational Linguistics, pp ). [6] Vol. 47, No. 6, pp ). [7] Papineni, K., Roukos, S., Ward, T. and Zhu, W.-J.: Bleu: a Method for Automatic Evaluation of Machine Translation, Technical report, IBM Research Report RC22176 W ) 2001). [8] Echizen-ya, H. and Araki, K.: Automatic Evaluation of Machine Translation based on Recursive Acquisition of an Intuitive Common Parts Continuum, Proceedings of the MT Summit XII Workshop on Patent Translation, pp ). [9] Kevin, D. Vol. 21, No. 3, pp ). [10] Birch, A. and Osborne, M.: LRscore for Evaluation Lexical and Reordering Quality in MT, Proceedings of the Joint 5th Workshop on Statistical Machine Translation and MetricsMATR, pp ). [11] Shawe-Taylor, J. Cristianini, N. ) Kernel Methods for Pattern Analysis) chapter 11 : 2010). [12] SIG-DMSM-A ). [13] Nivre, J.: Non-Projective Dependency Parsing in Expected Linear Time, Proceedings of the Joint Conference of the 47th Annual Meeting of the ACL and the 4th International Joint Conference on Natural Language Processing of the AFNLP, pp ). [14] Vol. 22, No. 2, pp ). [15] 1 CD-ROM 1999). [16] 31 pp ). [17] ). [18] ). [19] ). [20] ). [21] Lambrecht, K.: Information Structure and Sentence Form, Cambredge University Press 1994). [22] Prince, E. F.: Discourse description: Diverse linguistic c 2015 Information Processing Society of Japan 24

25 A 2 score RETELLING IT) RETELLING KP) RETELLING KT) RETELLING MP) RETELLING MT) Mean SD Mean SD Mean SD Mean SD Mean SD n-gram1) c n-gram2) c n-gram3) c n-gram4) c n-gram 2) c n-gram 3) c n-gram 4) c p-mer2) c p-mer3) c p-mer4) c p-mer 2) c p-mer 3) c p-mer 4) c Footrule c Kendall c n-gram1) m n-gram2) m n-gram3) m n-gram4) m n-gram 2) m n-gram 3) m n-gram 4) m p-mer2) m p-mer3) m p-mer4) m p-mer 2) m p-mer 3) m p-mer 4) m Footrule m Kendall m A 3 ) ID c 2015 Information Processing Society of Japan 25

26 A 4 :) ID ID First Pass Total Reg. Path analyses of a fund-raising text, chapter The ZPG letter: Subjects, definiteness, and information status, pp , Benjamins 1992). [23] Jacobs, J.: The dimensions of topic-comment, Linguistics, Vol. 39, pp ). [24] Cook, P. and Bildhauer, F.: Identifying aboutness topics : two annotation experiments, Dialogue and Discourse, Vol. 4, No. 2, pp ). [25] Pustejovsky, J. and Stubbs, A.: Natural Language Annotatino for Machine Learning A Guide to Corpus- Buildling for Applications, O Reilly 2012). [26] NAIST Vol. 17, No. 2, pp ). [27] 2009). [28] ). [29] Kennedy, A. and Pynte, J.: Parafoveal-on-foveal effects in normal reading, Vision Research, Vol. 45, pp ). [30] Demberg, V. and Keller, F.: Data from eye-tracking corpora as evidence for theories of syntactic processing complexity, Cognition, Vol. 109, No. 2, pp ). [31] Gibson, E.: Linguistic complexity: Locality of syntactic dependencies, Cognition, Vol. 68, pp ). [32] Hale, J.: A probabilistic earley parser as a psycholinguistic model, Proc. of the second conference of the North American chapter of the association for computational linguistics, Vol. 2, pp ). [33] Roland, D., Mauner, G., O Maera, C. and Yun, H.: Discourse expectations and relative clause processing, Journal of Memory and Language, Vol. 66, No. 3, pp ). [34] Demberg, V. and Keller, F.: Eye-tracking evidence for integration cost effects in corpus data, Proc. of the 29th meeting of the cognitive science society CogSci-07) 2007). [35] Just, M. A., Carpenter, P. A. and Woolley, J. D.: Paradigms and Processes in Reading Comprehension, Journal of Experimental Psychology: General, Vol. 3, pp ). [36] ChaKi.NET 2) 5 pp ). [37] Amano, S. and Kondo, T.: Estimation of mental lexicon size with word familiarity database, Proceedings of International Conference on Spoken Language Processing, Vol. 5, pp ). [38] 2002). [39] 3 Part ). [40] King, J. and Just, M. A.: Individual Differences in Syntactic Processing: The role of Working Memory, Journal of Memory and Language, Vol. 30, pp ). [41] Hawkins, J. A.: Processing Complexity and Filler-Gap Dependencies Across Grammars, Language, Vol. 75, pp ). [42] O Grady, W.: Syntactic Development, The University of Chicago Press 1997). [43] Miyamoto, E. T. and Nakamura, M.: Subject/object asymmetries in the processing of relative clauses in Japanese, Proceedings of the 22nd West Coast Conference on Formal Linguistics Garding, G. and Tsujimura, M., eds.), pp ). [44] Roland, D., O Meara, C., Yun, H. and Mauner, G.: Processing object relative clauses: Discourse or frequency, Poster presented at the CUNY sentence processing conference 2007). [45] 2011). [46] Gordon, P. C., Hendrick, R. and Johnson, M.: Memory Interference During Language Processing, Jounal of Experimental Psychology: Learning, Memory, and Cognition, Vol. 27, No. 6, pp ). [47] Gordon, P. C., Hendrick, R. and Levine, W. H.: Memory-load -interference in Syntacitic Processing, Psychological Science, Vol. 13, No. 5, pp ). c 2015 Information Processing Society of Japan 26

27 A 5 :) ID ID First Pass Total Reg. Path ** ** * * [48] Waters, G., Caplan, D. and Yampolsky, S.: On-line syntactic processing under concurrent memory load, Psychonomic Bulletin and Review, Vol. 10, pp ). [49] Kahraman, B., Sato, A., Ono, H. and Sakai, H.: Why object clefts are easier than subject clefts in Japanese: Frequency or expectation?, IEICE Technical Report, Vol. 111, No. 170, pp ). [50] Kahraman, B.: Processing gap-filler dependencies in Japanese and Turkish: Regarding the incrementality of sentence processsing, PhD Thesis, 2011). [51] Kahraman, B., Sato, A., Ono, H. and Sakai, H.: Incremental processing of gap-filler dependencies: Evidence from the processing of subject and object clefts in Japanese, The Proceedings of the 12th Tokyo Conference on Psycholinguistics, pp ). [52] gap-filler ERP 146 pp ). [53] Vol. 3, No. 2, pp ). [54] Vol. 23, No. 1, pp ). c 2015 Information Processing Society of Japan 27