1. Trang chủ
  2. » Giáo Dục - Đào Tạo

Nghiên cứu, phát triển một số phương pháp tóm tắt văn bản sử dụng kĩ thuật học sâu

181 3 0

Đang tải... (xem toàn văn)

Tài liệu hạn chế xem trước, để xem đầy đủ mời bạn chọn Tải xuống

THÔNG TIN TÀI LIỆU

L IăCAMă OAN Tôi tên L u Minh Tu n, xin cam đoan đơy lƠ cơng trình nghiên c u c a b n thơn nghiên c u sinh th i gian h c t p vƠ nghiên c u d i s h ng d n c a ng i h ng d n khoa h c Các k t qu nghiên c u trình bƠy lu n án lƠ trung th c, xác vƠ ch a t ng đ c công b b t k cơng trình nƠo khác Các k t qu nghiên c u vi t chung v i tác gi khác đ u đ c s đ ng ý c a đ ng tác gi tr c đ a vƠo lu n án Các k t qu s d ng đ tham kh o đ u đ c trích d n đ y đ vƠ theo quy đ nh Hà N i, ngày tháng n m 2022 Nghiên c u sinh L u Minh Tu n NG IH NG D N KHOA H C L IăC Mă Nă Trong trình h c t p nghiên c u, nghiên c u sinh đư nh n đ c nhi u s giúp đ vƠ ý ki n đóng góp q báu c a th y L i đ u tiên, nghiên c u sinh xin đ c bƠy t lịng kính tr ng vƠ bi t n sơu s c t i PGS TS Lê Thanh H ng ng i h ng d n khoa h c đư t n tình ch b o, h ng d n đ nghiên c u sinh có th hoàn thành đ c lu n án nƠy Nghiên c u sinh xin g i l i c m n chân thành t i PGS TS Nguy n Bình Minh v nh ng ý ki n đóng góp quý báu nh ng bu i xê mi na khoa h c nhi u gian nan su t th i gian nghiên c u hoàn thành lu n án Nghiên c u sinh xin g i l i c m n đ n th y cô c a B môn H th ng thông tin, Vi n Công ngh thơng tin vƠ truy n thơng, Phịng đƠo t o, Tr ng i h c Bách khoa HƠ N i, n i nghiên c u sinh h c t p vƠ nghiên c u đư t o u ki n t t nh t có th cho nghiên c u sinh su t th i gian h c t p vƠ nghiên c u đ nghiên c u sinh có th hoƠn thƠnh lu n án ti n s m t cách t t nh t Xin c m n Ban giám hi u Tr ng i h c Kinh t Qu c dơn, Ban Lưnh đ o Vi n Công ngh thông tin vƠ Kinh t s , th y cô B môn Công ngh thông tin - n i nghiên c u sinh công tác vƠ đ ng nghi p đư quan tơm giúp đ , t o u ki n t t nh t đ nghiên c u sinh có th hoàn thành t t k ho ch h c t p vƠ nghiên c u c a L i cu i, nghiên c u sinh xin chơn thƠnh c m n thƠnh viên gia đình, ng i thơn, b n bè đư dành cho nghiên c u sinh nh ng tình c m t t đ p, đ ng viên, giúp đ nghiên c u sinh v t qua nh ng khó kh n trình h c t p vƠ nghiên c u đ đ t đ c k t qu nh ngày hơm ơy c ng lƠ quƠ tinh th n mƠ nghiên c u sinh trơn tr ng g i t ng đ n thƠnh viên gia đình vƠ ng i thơn M t l n n a nghiên c u sinh xin chơn thƠnh c m n! M CăL C DANH M C CÁC KÝ HI U VÀ CH VI T T T iv DANH M C CÁC B NG vi DANH M C CÁC HÌNH V , TH ix DANH M C KÝ HI U TOÁN H C xi M U Ch ng T NG QUAN V TÓM T T V N B N 11 1.1 Gi i thi u v tóm t t v n b n 11 1.1.1 Gi i thi u bƠi tốn tóm t t v n b n 11 1.1.2 Phơn lo i bƠi tốn tóm t t v n b n 12 1.1.3 Các b c th c hi n tóm t t v n b n 13 1.1.4 M t s đ c tr ng c a v n b n 13 1.2 M t s ph ng pháp đánh giá v n b n tóm t t t đ ng 14 1.2.1 Ph 1.2.2 Ph 1.2.3 Ph ng pháp d a đ t ng t v n i dung 14 ng pháp d a đ t ng quan phù h p 14 ng pháp ROUGE 15 1.3 Các ph ng pháp k t h p v n b n tóm t t đa v n b n 18 1.4 Các ph ng pháp tóm t t v n b n h ng trích rút c s 20 1.4.1 PageRank 20 1.4.2 TextRank 20 1.4.3 LexRank 20 1.4.4 Lead-Based 21 1.5 Các b d li u th nghi m 21 1.5.1 Các b d li u v n b n ti ng Anh 21 1.5.2 Các b d li u v n b n ti ng Vi t 24 1.6 K t lu n ch Ch ng 25 ng CÁC KI N TH C N N T NG 27 2.1 Các k thu t h c sâu c s 27 2.1.1 M ng Perceptron nhi u l p 27 2.1.2 M ng n ron tích ch p 28 2.1.3 M ng n ron h i quy 30 2.1.4 Các bi n th c a RNN 31 2.1.5 Mơ hình chu i sang chu i c b n 35 2.1.6 C ch ý 36 2.1.7 C ch t ý vƠ mơ hình Transformer 39 2.2 Các mơ hình ngơn ng d a h c sâu đ c hu n luy n tr c 42 2.2.1 Mư hóa t 42 i 2.2.2 Ph ng pháp Word2Vec 42 2.2.3 Mơ hình BERT 44 2.2.4 Các phiên b n ch y u c a mơ hình BERT 47 2.3 K thu t h c t ng c ng Q-Learning 49 2.3.1 H c t ng c ng Q-Learning 49 2.3.2 Thu t toán h c t ng c ng Deep Q-Learning 50 2.4 Tìm ki m Beam 51 2.5 Ph ng pháp đ liên quan c n biên t i đa 51 2.6 K t lu n ch Ch H ng 53 ng PHÁT TRI N CÁC PH NG PHÁP TĨM T T NV NB N NG TRÍCH RÚT 54 3.1 Gi i thi u toán h ng ti p c n 54 3.2 Mơ hình tóm t t đ n v n b n h ng trích rút RoPhoBERT_MLP_ESDS 56 3.2.1 Gi i thi u mô hình 56 3.2.2 Mơ hình tóm t t v n b n đ xu t 56 3.2.3 Th nghi m mơ hình 59 3.2.4 ánh giá vƠ so sánh k t qu 63 3.3 Mơ hình tóm t t đ n v n b n h ng trích rút mBERT_CNN_ESDS 64 3.3.1 Gi i thi u mơ hình 64 3.3.2 Mơ hình tóm t t v n b n đ xu t 64 3.3.3 Th nghi m mơ hình 67 3.3.4 ánh giá vƠ so sánh k t qu 71 3.4 Mơ hình tóm t t đ n v n b n h ng trích rút mBERT-Tiny_ seq2seq_DeepQL_ESDS 71 3.4.1 Gi i thi u mô hình 71 3.4.2 Mơ hình tóm t t v n b n đ xu t 72 3.4.3 Hu n luy n mơ hình v i k thu t h c t ng c ng 74 3.4.4 Th nghi m mơ hình 75 3.4.5 ánh giá vƠ so sánh k t qu 78 3.5 So sánh đánh giá ba mơ hình tóm t t đ n v n b n h ng trích rút đ xu t 79 3.6 K t lu n ch Ch H ng 80 ng PHÁT TRI N PH NG PHÁP TÓM T T N V N B N NG TÓM L C 82 4.1 Gi i thi u toán h ng ti p c n 82 4.2 Mơ hình tóm t t c s 84 4.2.1 Mơ hình seq2seq c a mơ hình 84 ii 4.2.2 C ch ý áp d ng mơ hình 85 4.2.3 M ng chép t - sinh t 86 4.2.4 C ch bao ph 87 4.3 Mơ hình tóm t t đ n v n b n h ng tóm l c PG_Feature_ASDS 87 4.3.1 Các đ c tr ng đ xu t thêm m i cho mơ hình 88 4.3.2 Mơ hình tóm t t đ n v n b n h ng tóm l c đ xu t 89 4.4 Th nghi m mơ hình 89 4.4.1 Các b d li u th nghi m 89 4.4.2 Ti n x lý d li u 90 4.4.3 Thi t k th nghi m 90 4.5 ánh giá so sánh k t qu 91 4.6 K t lu n ch Ch ng 93 ng PHÁT TRI N CÁC PH NG PHÁP TÓM T T A V N B N 94 5.1 Gi i thi u tốn tóm t t đa v n b n h ng ti p c n 94 5.2 Mơ hình tóm t t đa v n b n h ng trích rút Kmeans_Centroid_EMDS 95 5.2.1 Gi i thi u mơ hình 95 5.2.2 Các thƠnh ph n c a mơ hình 96 5.2.3 Mơ hình tóm t t đa v n b n đ xu t 100 5.2.4 Th nghi m mơ hình vƠ k t qu 104 5.2.5 So sánh vƠ đánh giá k t qu 107 5.3 Các mơ hình tóm t t đa v n b n h ng tóm l c d a mơ hình tóm t t đ n v n b n đ c hu n luy n tr c 110 5.3.1 t v n đ 110 5.3.2 Mơ hình tóm t t đa v n b n h ng tóm l c d a mơ hình tóm t t đ n v n b n h ng tóm l c đ c hu n luy n tr c PG_Feature_AMDS 110 5.3.3 Mơ hình tóm t t đa v n b n h ng tóm l c d a mơ hình tóm t t đ n v n b n h n h p đ c hu n luy n tr c Ext_Abs_AMDS-mds-mmr 121 5.4 K t lu n ch ng 132 K T LU N 133 DANH M C CÁC CƠNG TRÌNH Ã CÔNG B 136 TÀI LI U THAM KH O 137 PH L C Ph l c A: V n b n ngu n c a v n b n tóm t t ví d Ph l c B: Bi u đ phân b c a b d li u th nghi m Ph l c C: V n b n ngu n c a m u tóm t t th nghi m iii DANHăM CăCÁCăKụăHI UăVÀăCH ăVI TăT T T vi t t t AMDS ASDS Beam BERT BERT-Tiny biGRU biLSTM BPTT CNN DE Decoder DeepQL EMDS Encoder ESDS GLUE GRU ILP IR LCS LDA LSA LSTM mBERT mds MLP MMR NLP Ti ng Anh Ý ngh a ti ng Vi t Tóm t t đa v n b n h ng tóm l c Tóm t t đ n v n b n h ng tóm l c Thu t tốn tìm ki m Beam Mơ hình bi u di n mư hóa hai chi u t Transformer Mơ hình BERT thu nh Bidirectional Gated Recurrent Unit M ng GRU chi u Bidirectional Long Short Term M ng LSTM chi u Memory Backpropagation Through Time Thu t toán lan truy n ng c liên h i Covolutional Neural Network M ng n ron tích ch p Document Embedding Mư hóa v n b n Decoder B gi i mư Deep Q-Learning Thu t toán h c t ng c ng d a m ng n ron sơu Extractive Multi-Document Tóm t t đa v n b n h ng Summarization trích rút Encoder B mư hóa Extractive Single-Document Tóm t t đ n v n b n h ng Summarization trích rút General Language Understanding ánh giá hi u ngôn ng Evaluation chung Gated Recurrent Unit Mơ hình đ n v h i quy có ki m sốt chi u Interger Linear Programming Quy ho ch n tính nguyên Information Retrieval Tìm ki m thơng tin Longest Common Subsequence Dưy chung l n nh t Latent Dirichlet Allocation Mơ hình ch đ n Latent Semantic Analysis Phơn tích ng ngh a ti m n Long Short Term Memory Mơ hình m ng b nh ng n h n dƠi h n chi u BERT multilingual Mơ hình BERT đa ngơn ng multi-document summaryzation Tóm t t đa v n b n Multi Layer Perceptron M ng Perceptron nhi u l p Maximal Marginal Relevance liên quan c n biên t i đa Natural Language Processing X lý ngôn ng t nhiên Abstractive Multi-Document Summarization Abstractive Single-Document Summarization Beam Search Bidirectional Encoder Representation from Transformers iv OOV PE PG Khơng có b t v ng Mư hóa v trí Mơ hình m ng Pointer – Generator Mơ hình t i u c a BERT h tr cho ti ng Vi t Hu n luy n tr c H c t ng c ng M ng n ron h i quy Mơ hình t i u c a BERT h tr cho ti ng Anh for đo Rouge Out Of Vocabulary Position Embedding Pointer – Generator PhoBERT PT RL RNN RoBERTa Pre-trained Reinforcement Learning Recurent Neural Network ROUGE Recall-Oriented Understudy Gisting Evaluation Sentence Embedding Mã hóa câu Sequence to sequence Mơ hình chu i sang chu i Stanford Question Answering Dataset B d li u h i đáp c a Stanford Support Vector Machine Mơ hình máy véc t h tr Term Frequency – Inverse Document Tr ng s c a t (m c đ Frequency quan tr ng c a t ) v n b nc am tt pv nb n Tóm t t v n b n warmup Quá trình hu n luy n ban đ u (s b c) v i t l h c nh Word Embedding Mư hóa t SE seq2seq SQuAD SVM TF-IDF TTVB warmup WE v DANHăM CăCÁCăB NG B ng 1.1 Ví d minh h a m t v n b n tóm t t c a v n b n ti ng Anh 12 B ng 1.2 Ví d minh h a m t v n b n tóm t t c a v n b n ti ng Vi t 12 B ng 1.3 Ph ng pháp phân chia b d li u CNN/Daily Mail 21 B ng 1.4 Th ng kê thông tin c a hai b d li u CNN Daily Mail 22 B ng 1.5 Th ng kê thông tin tóm t t c a b d li u DUC 2001 DUC 2002 s d ng cho tóm t t đ n v n b n 22 B ng 1.6 Th ng kê thơng tin tóm t t c a b d li u DUC 2004 23 B ng 1.7 Th ng kê thơng tin tóm t t c a t p d li u Main task c a b d li u DUC 2007 23 B ng 1.8 Th ng kê thơng tin tóm t t c a b d li u Baomoi 24 B ng 1.9 Th ng kê thông tin tóm t t c a b d li u Corpus_TMV 25 B ng 1.10 Th ng kê thơng tin tóm t t c a b d li u ViMs 25 B ng 3.1 K t qu th nghi m m t s ph ng pháp tóm t t v n b n c s Ký hi u ‘*’ th hi n ph ng pháp đ c tri n khai th nghi m b d li u t ng ng 60 B ng 3.2 Giá tr siêu tham s th i gian hu n luy n mơ hình xây d ng 61 B ng 3.3 K t qu th nghi m c a mơ hình xây d ng Ký hi u ‘-’ bi u di n mơ hình mà lu n án không th nghi m b d li u t ng ng 61 B ng 3.4 M t m u tóm t t b d li u CNN 62 B ng 3.5 M t m u tóm t t b d li u Baomoi 63 B ng 3.6 So sánh đánh giá hi u qu ph ng pháp Ký hi u ‘*’,‘-’ bi u di n ph ng pháp đ c th nghi m, không đ c th nghi m b d li u t ng ng 63 B ng 3.7 Các k t qu th nghi m c a mơ hình xây d ng 69 B ng 3.8 K t qu th nghi m ph ng pháp b d li u DUC 2001 DUC 2002 Ký hi u ‘*’, ‘-’ bi u di n ph ng pháp đ c th nghi m, không đ c th nghi m b d li u t ng ng 69 B ng 3.9 M t m u tóm t t b d li u CNN 70 B ng 3.10 M t m u tóm t t b d li u Baomoi 70 B ng 3.11 So sáng đánh giá hi u qu c a ph ng pháp Ký hi u ‘*’, ‘-’ bi u di n ph ng pháp đ c th nghi m, không đ c th nghi m b d li u t ng ng 71 B ng 3.12 B ng giá tr siêu tham s cài đ t cho mơ hình hu n luy n v i k thu t h c t ng c ng Deep Q-Learning 74 vi B ng 3.13 Giá tr siêu tham s th i gian hu n luy n mơ hình xây d ng 76 B ng 3.14 K t qu th nghi m c a mơ hình xây d ng 77 B ng 3.15 M t m u tóm t t b d li u CNN 77 B ng 3.16 M t m u tóm t t b d li u Baomoi 78 B ng 3.17 So sánh đánh giá k t qu ph ng pháp 78 B ng 3.18 So sánh đánh giá đ xác c a mơ hình đ xu t 79 B ng 3.19 So sánh đánh giá th i gian th c hi n c a mơ hình đ xu t 80 B ng 4.1 K t qu th nghi m c a mơ hình b d li u CNN/Daily Mail Ký hi u ‘(*)’ mơ hình c a See c ng s [43] 91 B ng 4.2 K t qu th nghi m c a mơ hình b d li u Baomoi Ký hi u ‘(*)’ mơ hình c a See c ng s [43] 91 B ng 4.3 M u tóm t t g m b n tóm t t tham chi u, b n tóm t t c a mơ hình [43] b n tóm t t c a mơ hình đ xu t b d li u CNN/Daily Mail 92 B ng 4.4 M u tóm t t g m b n tóm t t tham chi u, b n tóm t t c a mơ hình [43] b n tóm t t c a mơ hình đ xu t b d li u Baomoi 93 B ng 5.1 Giá tr c a tham s s d ng th nghi m mơ hình Ký hi u ‘-‘ bi u di n mơ hình khơng s d ng tham s t ng ng 105 B ng 5.2 K t qu th nghi m mơ hình xây d ng hai b d li u 106 B ng 5.3 K t qu th nghi m ph ng pháp tóm t t b d li u DUC 2007 106 B ng 5.4 So sánh đánh giá k t qu c a ph ng pháp Ký hi u ‘-’ bi u di n ph ng pháp không đ c th nghi m b d li u t ng ng 107 B ng 5.5 Các m u tóm t t c a c m D0716D b d li u DUC 2007 c a mơ hình đ xu t ng i 109 B ng 5.6 Các m u tóm t t c a c m Cluster_2 b d li u Corpus_TMV c a mô hình đ xu t ng i 109 B ng 5.6 Giá tr siêu tham s th i gian hu n luy n mơ hình Warmup trình hu n luy n ban đ u v i t l h c nh đ hi u ch nh c ch ý 118 B ng 5.7 K t qu th nghi m c a mơ hình tóm t t đ n v n b n h ng tóm l c PG_Feature_ASDS b d li u CNN Baomoi 118 B ng 5.8 K t qu th nghi m mơ hình PG_Feature_AMDS b DUC 2004 Corpus_TMV s d ng mơ hình PG_Feature_ASDS ch a đ c hu n luy n ti p b DUC 2007 b ViMs t ng ng 118 B ng 5.9 K t qu th nghi m mơ hình PG_Feature_AMDS b DUC 2004 Corpus_TMV s d ng mơ hình PG_Feature_ASDS đ c hu n luy n ti p b DUC 2007 b ViMs t ng ng 119 B ng 5.10 M t m u th nghi m b d li u DUC 2004 119 vii B ng 5.11 M t m u th nghi m b d li u Corpus_TMV 120 B ng 5.12 So sáng đánh giá k t qu c a ph ng pháp Ký hi u ‘-’ bi u di n ph ng pháp không đ c th nghi m b d li u t ng ng 120 B ng 5.13 Giá tr siêu tham s th i gian hu n luy n mơ hình 128 B ng 5.14 K t qu th nghi m c a mơ hình tóm t t đ n v n b n b d li u CNN/Daily Mail Baomoi 129 B ng 5.15 K t qu th nghi m mơ hình xây d ng b d li u DUC 2004 Corpus_TMV 129 B ng 5.16 M t m u th nghi m b d li u DUC 2004 130 B ng 5.17 M t m u th nghi m b d li u Corpus_TMV 131 B ng 5.18 So sáng đánh giá k t qu c a ph ng pháp Ký hi u ‘-’ bi u di n ph ng pháp không đ c th nghi m b d li u t ng ng 131 viii B.3 B d li u DUC 2001 Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.4 B d li u DUC 2002 Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.5 B d li u DUC 2004 Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.6 B d li u DUC 2007 Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.7 B d li u Baomoi Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.8 B d li u Corpus_TMV Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t B.9 B d li u ViMs Bi u đ phân b đ dài trung bình v n b n ngu n Bi u đ phân b đ dài trung bình v n b n tóm t t Ph ăl căC: V năb năngu năc aăcácăm u tómăt tăth ănghi m C.1 V n b n ngu n c a m u th RoPhoBERT_MLP_ESDS nghi m c a mơ hình  N i dung v n b n ngu n b d li u CNN Lois Lilienstein , co - star of " Sharon , Lois & Bram 's Elephant Show " - the Canadian preschool show that ran on Nickelodeon during the early 1990s - has died , aged 78 Her son , David Lilienstein , told CBC News that his mother died in Toronto on Wednesday night from a rare form of cancer first diagnosed last October " She knew it was happening , she was at peace with it , and she died very peacefully and not in pain , " he told the Canadian “ broadcaster Liz Smith at 92 Lilienstein was born in Chicago in 1936 and moved to Toronto in 1996 with her son and her husband , Ernest She joined Sharon Hampson and Bram Morrison to launch their singing trio in 1978 The TV show " Sharon , Lois & Bram 's Elephant Show " ran on CBC in Canada from 1984 to 1989 , before continuing in reruns on Nickelodeon stateside until 1995 The 30 most popular film producers in Hollywood Sally Catto , general manager , programming at CBC Television , paid tribute to Lilienstein for her contribution to Canadian kids TV " Sharon , Lois and Bram got their start at CBC back in the ' 80s with the Elephant Show , and they went on to become icons to children all across North America through various programs and specials Lois will be fondly remembered by her many friends at CBC , " Catto told The Hollywood Reporter ' Ghostbusters ' budget cut The Canadian variety show was best known for songs like " Skinnamarink " and " Peanut Butter " Lilienstein left the Sharon , Lois & Bram performance trio in 2000 The Hollywood Reporter All rights reserved ”  N i dung v n b n ngu n b d li u Baomoi Nhóm tra viên c a C quan N ng l ng Nguyên t Qu c t ( IAEA ) ngày 18/1 t i th đô Tehran c a Iran đ giám sát vi c th c thi tho thu n t m th i gi a n c nhóm P 5+1 Hãng thông t n Fars c a Iran cho bi t , nhóm tra k s h t nhân Massimo Aparo d n đ u s b t đ u n p báo cáo cho IAEA vào ngày 20/1 t i – th i m mà tho thu n gi a Iran P 5+1 th c có hi u l c i s Iran t i IAEA Reza Najafi tr l i h p báo t i tr s IAEA Vienna , Áo ( nh AFP ) Các tra s th sát hai nhà máy h t nhân Natanz Fordow c a Iran đ đ m b o r ng n c s d ng ho t đ ng s n xu t uranium c p đ 20% , đ ng th i vơ hi u hố kho uranium đ c làm giàu b ng cách pha loãng Cùng ngày , IAEA cho bi t , c quan hi n c n thêm ti n tài tr kho ng tri u Euro vòng tháng t i đ có th trang tr i chi phí cho ho t đ ng giám sát Ngày 24/11/2013 , Iran Nhóm P 5+1 ( g m Anh , Pháp , M , Nga , Trung Qu c c ) đ t đ c m t tho thu n s b đ gi i quy t b t c đàm phán v ch ng trình h t nhân gây tranh cãi c a Iran Theo tho thu n , Iran s t m d ng m t ph n ch ng trình h t nhân c a đ đ i l y vi c ph ng Tây n i l ng bi n pháp tr ng ph t kinh t nh m vào n c / H ng Anh / VOV-Trung tâm Tin Theo Reuters “ ” C.2 V n b n ngu n c a m u th mBERT_CNN_ESDS 10 nghi m c a mơ hình  N i dung v n b n ngu n b d li u CNN Panama City, Panama (CNN) Ending a decade - long standstill in U.S - Cuba relations, President Barack Obama met for an hour Saturday with his Cuban counterpart Raul Castro, the first time the two nations' top leaders have sat down for substantive talks in more than 50 years The meeting in a small conference room on the sidelines of the Summit of the Americas came as the two countries work to end the Cold War enmity that had led to a total freeze of diplomatic ties And while both leaders proclaimed progress had been made, a key stumbling block Cuba's place on the U.S list of countries that sponsor terror remained unresolved "This is obviously an historic meeting, "Obama said at the beginning of his session with Castro, claiming that decades of strain had done little to benefit either Cubans or citizens of the nited States "It was time for us to try something new," he said "We are now in a position to move on a path toward the future "Castro, who earlier in the day said he trusted Obama, acknowledged there would be difficult stumbling blocks as his nation works to repair ties with the United States But he said those differences could be surmounted "We are willing to discuss everything, but we need to be patient, very patient," Castro said "We might disagree on something today on which we could agree tomorrow "Speaking to reporters after his session with Castro, Obama said the meeting was "candid and fruitful" and could prove to be a "turning point" in his push to defrost ties with Cuba But he said he hadn't yet decided whether to remove Cuba's designation as a state sponsor of terror, an outcome that had previously been expected during the summit The State Department provided Obama with a review of the terror status this week "I want to make sure I have a chance to read it, study it before we announce publicly what the policy outcome is going to be," Obama said "But in terms of the overall direction of Cuba policy, I think there is a strong majority both in the United States and in Cuba that says our ability to engage, to open up commerce and travel and people to people exchanges is ultimately going to be good for Cuban people "On Friday night, Obama and Castro greeted each other courteously amid an explosion of camera flashes, shaking hands before dining at the inaugural session of the conference The two sat at the same table but not directly next to one another Before Obama arrived in Panama on Wednesday, he spoke with Castro by phone, laying the groundwork for what will become a new era of relations between the neighboring countries "The Cold War has been over for a long time, "Obama said during opening remarks at the summit Saturday "I'm not interested in having battles, frankly, that began before I was born "That exhortation, however, seemed to be lost on Castro himself, who expanded “ 11 what was meant to be a six - minute speech into a 50 minute address lecturing leaders on Cuba's revolution and giving a litany of perceived grievances to Cuba over the past 50 years But he distinguished Obama from past American presidents, saying he respected Obama's move toward reconciliation "In my opinion, President Obama in an honest man, "Castro said through an interpreter "I admire him, and I think his behavior has a lot to with his humble background " S administration official said Castro’s long list of grievances was expected, despite the move toward diplomatic ties "(What’s) unique and new is what he said about the president, "the official said of Castro’s praise for Obama Obama announced in December that he was seeking to renew diplomatic relations with Cuba after half a century of strife, including eventually opening embassies in Washington and Havana Obama set to test engagement doctrine with Cuba in Panama His meeting with Castro on Saturday isn’t being billed as a formal bilateral session, but Obama’s aides are still characterizing the event as the highest - level engagement with the Cuban government since then - Vice President Richard Nixon met with Fidel Castro in 1959 "We’re in new territory here, "Ben Rhodes, Obama’s deputy national security adviser, said Friday "The reason we’re here is because the President strongly believes that an approach that was focused entirely on isolation, focused entirely on seeking to cut off the Cuban people from the United States of America had failed "The overtures to Cuba have not been universally popular in the United States; some lawmakers were irate that Obama was seeking to engage what they regard as a corrupt government "A recommendation to remove Cuba from the list of State Sponsors of Terrorism would represent another significant misstep in a misguided policy, "Sen Bob Menendez, a Democrat who used to the chair the Foreign Relations Committee, wrote in a statement last week In Latin America, however, Obama was receiving a warm welcome after announcing he was seeking to engage Havana in talks over reopening embassies and removing barriers to commerce and travel things you wanted to ask about the Cuban embargo He noted to applause during a session Friday that this was the first summit with Cuba in attendance And he’s cast the decision to reopen the S relationship with Cuba as beneficial to the entire hemisphere, which has also embraced his immigration executive action But even as Obama landed in Panama, the longstanding gulfs between the two countries ' governments were on display Dissidents opposed to Castro’s regime were violently accosted this week by supporters of the Cuban government, a scuffle the White House said was unacceptable "As we move toward the process of normalization, we’ll have our differences, government to government, with Cuba on many issues just as we differ at times with other nations within the Americas, just as we 12 differ with our closest allies," Obama said at a meeting of civil society leaders Friday "There’s nothing wrong with that." "But I 'm here to say that when we speak out, we’re going to so because the nited States of America does believe, and will always stand for, a certain set of universal values, "he said The long history between the U.S and Cuba Obama closed out his time in Panama with a news conference where he covered topics ranging from Hillary Clinton’s expected presidential announcement to his framework deal with Iran on its nuclear program The President had pointed criticism for Sen John McCain, R Arizona Earlier this week, McCain accused Secretary of State John Kerry of intentionally mischaracterizing what the sides had agreed to in the Iran nuclear deal "John Kerry is delusional, "McCain said on the Hugh Hewitt show, a conservative talk radio program, adding that the view from the Supreme Leader of Iran of the provisions agreed to" is probably right, " rather than what the United States maintains are the agreed provisions While discussing the Iran agreement Saturday, Obama brought up those remarks without being asked "When I hear someone like Sen McCain recently suggest that our secretary of state, John Kerry, who served in the United States Senate, (is) a Vietnam veteran, who’s provided exemplary service to this nation, is somehow less trustworthy of the interpretation of what’s in a political agreement than the Supreme Leader of Iran, that’s an indication of the degree to which partisanship has crossed all boundaries," he said at the news conference After the President’s remarks, McCain tweeted "So Pres Obama goes to # Panama, meets with Castro and attacks me - I 'm sure Raul is pleased." As for his 2008 Democratic rival, Obama said, "If she decides to run, if she makes an announcement, she’s going to have some strong messages to deliver," he said ”  N i dung v n b n ngu n b d li u Baomoi V tai n n nói x y vào kho ng 23h15 ngày 19/1, t i c u B n Thu , Qu c l 1A Theo đó, vào kho ng th i gian trên, chi c xe mang BKS: 37 A-048 45 u n xe l u thông theo h ng B c - Nam b t ng lao lên d i phân cách đ ng ô tô xe máy Hi n tr ng v tai n n Sau cú đâm m nh, đ u chi c xe b h h ng, toàn b chi c xe b n m gác d i phân cách đ ng Nh ng ng i ng i chi c xe khơng có b th ng Do bu n ng , thay sang đ ng cho ô tô, chi c xe ch n đ ng riêng b ng cách leo lên dãi phân cách Ngay sau v tai n n x y ra, ch xe ti n hành g i c u h giao thông đ n đ gi i c u chi c xe Tuy nhiên, ph i m t g n gi đ ng h sau, chi c xe m i tách kh i d i phân cách Sau "l " leo lên d i phân cách, đ u c a chi c xe b h h ng n ng V tai n n x y đ u c u nên “ 13 n giao thông qua b h n lo n Các nhân viên c a tr m soát vé c u B n Thu ph i u ti t giao thông, tránh ùn t c nghiêm tr ng n đ ng Sau b tai n n, ch xe g i n nh xe c u h đ n đ gi i quy t v vi c ” C.3 V n b n ngu n c a m u th Tiny_sep2seq_DeepQL_ASDS nghi m c a mơ hình mBERT-  N i dung v n b n ngu n b d li u CNN Duhok , Iraqi Kurdistan ( CNN ) The children laugh and shriek , as some of them seem to always have the capacity to no matter how depressing the circumstances Their bright clothes provide splashes of color against the otherwise drab monotone white of the endless rows of tents A small group plays with rocks , replacements for the toys they left behind when they fled , while others clamber through a jagged tear in the wire fence surrounding the refugee camp The Shariya refugee camp opened around six months ago , made up of some 4,000 tents and counting Thousands of Yazidis now call this corner of Iraqi Kurdistan home , about 18 miles ( 30 kilometers ) from one of the front lines with ISIS , where one can hear the occasional reverberation in the distance of what we are told are airstrikes The vast majority of the camp 's occupants are from the town of Sinjar , which is near the border with Syrian Kurdistan , and fled the ISIS assault there back in August But not everyone escaped ISIS took thousands of Yazidis captive The fighters separated the young women and girls , some as young as years old , to be sold as slaves , for their " masters " to use as concubines Men faced a choice : Convert to Islam or be shot Mahmoud was out running errands when ISIS fighters arrived , taking his wife , Ahlam , their three children the youngest of which was just a month old and his elderly parents " They took our phones , jewelry , money , " Ahlam recalls " They had guns They forced us at gunpoint into big trailer trucks " They were taken to a school turned prison in Tal Afar From there , the family was moved from village to village and at one stage taken to Mosul " They wrote everyone 's name down and they asked where we want to work , in the fields , as cleaners or as herders , " she says Ahlam and her family chose to herd goats They were then taken to a Shia village whose residents had fled , where they were part of a group of around 40 living in one house In the home , Ahlam found a cell phone left behind by its former occupants and called her husband " I said we are alive but we are prisoners " Ahlam 's husband , who up until that moment had lost the will to live , thinking his family was dead , says he cried out of happiness despite his pain Ahlam “ 14 would call when she could , briefly , after midnight , hiding under her bedcovers If she was caught with a phone , she would be killed The village itself was a massive prison , its entrances guarded by ISIS fighters She recalls that two men , in their late 40s or 50s , tried to escape When they were caught , their bones were broken , their bodies tied to the back of a truck and then driven through the streets The Yazidi captives were forced to watch the gruesome spectacle The men 's corpses were then tossed into a ditch and an order given not to bury them One night , some of the Yazidi men risked their lives to toss dirt onto the bodies , to give those slain what dignity they could Ahlam tells us that about a week before we met , ISIS fighters came by and took away her in - laws and the other elderly people living in the house " We did n't know where they were taking them , we thought we would be next , " she remembers So she and the rest of the group realized that they had to try to flee " We decided that either we survive or we n't " They left at midnight Ahlam cradled the baby , as her two other children , ages and years old , clutched at her clothes She prayed the baby would n't cry , that the children could keep walking They knew the general direction to take , but not the exact route , and they could only hope it was toward freedom " When the sun started to come up , I thought that 's it , we are going to get caught , " Ahlam says " And what am I going to with the kids ? I ca n't carry all three of them and run " Luckily , Ahlam never had to answer that impossible question The group made it into Iraqi Kurdistan The couple ca n't put into words their emotions when they were reunited Mahmoud , gently caressing his daughter 's palm , says he could hardly believe that the woman whose stunning eyes and gentle words he had fallen in love with , and their three children , were by his side again It had been eight agonizing months But their joy was tainted by fear for Mahmoud 's parents A few days after Mahmoud and Ahlam were reunited , ISIS released 217 captives No one is disclosing exactly why Among them were 60 children , a handful of men and women , and the rest were elderly including Mahmoud 's parents " We did n't know if they were going to slaughter us or what they were going to with us , " Mahmoud 's father says " They moved us around a lot , and at one stage we stayed in in one place for three days " But Ahlam 's parents are still with ISIS Vian Dakhil , a Yazidi member of Iraq 's parliament , speculates that the Yazidis ISIS released are individuals they were struggling to care for Dakhil is part of Iraqi Prime Minister Haider al - Abadi 's delegation to the United States this week She will be addressing the United Nations pleading for more international support for Iraq and more airstrikes to help defeat ISIS and she will address the plight of the 15 Yazidis " We sometimes say that we wish we had been massacred This would be better than being kidnapped and raped We prefer death now over the fate of what is happening to these girls and women , " Dakhil says " It is now on the government and on the international community to focus on this How to get these captives back It 's inconceivable that in the 21st century , something like this is happening as if we were living in the Stone Age " Ahlam says she was spared because she was breastfeeding and she had young children , which , we 're told , makes her impure and therefore unable to be used as a sex slave For Ahlam , what she went through was not the hardest part of her ordeal It 's the moment when the ISIS fighters began taking away the girls and young women She 's haunted by their screams , the image of them being dragged away sobbing and screaming ”  N i dung v n b n ngu n b d li u Baomoi Ch c ch n r ng SV San Siro s l i r c sáng mà tifosi c a h s ch ng ki n m t đ u tiên c a c u c u th Clarence Seedorf c ng v HLV tr ng c a AC Milan Tuy nhiên , nhi m v c a c u c u th ngu i Hà Lan lúc r t n ng n AC Milan hi n khơng cịn có s c m nh nh th i anh cịn khoác áo n a Hi n t i đ i bóng áo đ đen thành Milan ch đ ng khiêm t n v trí th 11 sau 19 vòng đ u T c h k t thúc lu t n a cu i c a BXH Ph i ti p Verona - CLB thi đ u r t n đ nh mùa m t ki m tra th c s v i Clarence Seedorf c ng v m i mà anh ch a h có kinh nghi m Hi n t i đ i khách Verona có 32 m đ ng th , s m c a AC Milan ch 22 Th nên ch có chi n th ng nh ng tr n đ u l i , Milan m i hy v ng chen chân vào nhóm d cúp châu Âu Hy v ng r ng tr n đ u , đ i bóng ch sân San Siro s làm cho khán đài c a h l i r c sáng tr l i n a đ u BXH , Juventus g n nh s ch c ch n s có đ c tr n v n m đ v ng vàng đ u , mà h đ c thi đ u sân nhà ch ph i ti p m t Sampdoria c i mùa Vì v y , thày trị ơng Antonio Conte hồn toan t tin s có tr n th ng th 18 mùa gi i tr n đ u vào đêm th B y t i Trong , v i AS Roma , nhi m v c a h c ng d dàng nh Juventus h đ c thi đ u sân nhà Olympico đ i th c a đ i bóng áo bã tr u thành Rome ch Livorno - đ i bóng m i ch có 13 m hi n đ ng th 19 Th nên , kho ng cách m so v i Juventus ch c ch n s đ c gi nguyên sau vòng đ u V i Napoli l i khác , thày trị ơng Benitez s nhi u g p khó kh n h ph i hành quân đ n Bologna cho dù th i m , Bologna có phong đ khơng t t ph i v t l n v i nhóm c m đèn đ Ph i th c s t p trung bi t tôn tr ng đ i th may Napoli m i ni hy v ng gi kho ng cách v i đ i x p Juventus “ 16 Roma vòng đ u khép l i Trong , Fiorentina l i có chuy n làm khách đ c cho “ d th ” h n Thày trị ơng Vincenzo Montella ch ph i ch i sân đ i bóng cu i b ng Catania C p đ u cân s c cân tài nh t vịng đ u có l s thu c v Genoa Inter Milan Hi n Inter đ ng th đ c 32 m ( b ng m v i Verona nh ng h n v hi u s bàn th ng-thua ) , Genoa dù Inter b c nh ng h l i đ c thi đ u sân nhà Th c s , phong đ thi đ u sân khách c a c u th áo xanh đen thành Milan mùa không t t Th nên chuy n làm khách h a h n chuy n “ lành d nhi u ” N u th c s có đ c m tr n đ u thành công v i Inter Milan Các c p đ u cịn l i , kh n ng s khơng có nhi u b t ng ” C.4 V n b n ngu n c a m u th PG_Feature_ASDS nghi m c a mơ hình  N i dung v n b n ngu n b d li u CNN A Texas English high school teacher has been arrested after being accused of having an improper relationship with a male student older than 17 years old Mary Todd Lowrance , 49 , a teacher at Moises E Molina High School , turned herself into Dallas Independent School District police on Thursday morning , according to Dallas ISD Police Chief Craig Miller Miller said the teacher had been in a relationship with the student for a couple of months Officers learned of the alleged relationship after Lowrance disclosed details to a coworker who then alerted a Dallas ISD officer Mary Todd Lowrance ( above ) , 49 , a Texas English high school teacher was arrested after being accused of having an improper relationship with a male student over the age of 17 , according to authorities If the student had been a juvenile , the case would have been handled by the Dallas Police Department , Miller told The Dallas Morning News When you get 160,000 students and 10,000 teachers , you hear a lot of rumors , and you have to investigate all of them , ' said Miller And in this case , there was something to it The student has been offered counseling We 're doing everything we can for him as well Miller said that after Lowrance 's coworker alerted authorities , it prompted an investigation where interviews were conducted with Lowrance , the student and two witnesses The Dallas ISD police then got an arrest warrant Miller said Lowrance has been on leave for several weeks while investigators worked on the case , according to NBCDFW After she turned herself into authorities , she was booked into the Dallas County jail around 11 am on Thursday on a $ 5,000 bond and was released at 4.40 pm , according to county records Calls made to the “ 17 Dallas Independent School District police on Friday were not immediately returned to Daily Mail Online Lowrance , a teacher at Moises E Molina High School ( file photo above ) , turned herself in on Thursday morning and was released at 4.40 pm , according to county records ”  N i dung v n b n ngu n b d li u Baomoi S vi c b phanh phui vào ngày 30/8 v a qua C _quan an tồn th c ph m , phịng cơng_th ng , cơng_an quy n th xã ph i h p u tra s vi c t i hi n tr ng , 1.000 kg d u b n v a đ c chuy n t i b i m t xe_t i ch hàng Theo ngu n_tin thân c y , s d u có xu t x t m t x ng ch bi n t i mã m , giao châu Ngày 31/8 , m t t cơng_tác g m phịng an tồn th c ph m , cơng an quy n mã m , giao châu nhanh chóng u tra làm rõ chân t ng v vi c a ch đ c cung c p t ngu n tin khơng cịn t n t i b i x ng s n xu t l n m t t m tr c vài ngày Nh ng t i mã m v n m t x ng khác n m đ ng giao bình c bi t , v n m t x ng ch bi n th c ph m h p pháp c a x ng b c đô , giao châu Tr c , x ng kinh doanh s n ph m nông nghi p Vài ngày tr c , m t x ng s n xu t d u b n tóp m b ng th t l n b y nh y da gà chuy n t i , núp bóng doanh nghi p h p pháp đ ho t đ ng Chình ình tr c c a x ng t m bi n l n v i dịng ch : “ x ng gia cơng th c ph m b c đô , giao châu ” Trong gian x ng có di n tích h n 100 m2 , hai chi c n i l n đ y p nh ng mi ng th t b y nh y đ c tr n đ u b ng m t chi c x ng l n Nh ng b ch nilong đ ng m ch t đ ng ng n ngang , ru i nh ng bâu đen Mé phía đơng có ba chi c n i l n , bên c nh m t gian nhà ngói ng p ng a khói Khơng ch d u n siêu b n , tóp m đen c ng s n ph m đ c lò t i x ng X ng s n xu t d u b n tóp m đen t th t b y nh y da gà Ch x ng th a nh n khơng có gi y phép s n xu t , ch bi n d u n nh ng v n ho t_đ ng suôn_s Lo i d u siêu b n sau đ c chuy n t i t c m c , đ o nh ng ch không kh ng đ nh đ a ch tiêu th m t x ng s n xu t bánh trung thu Tuy nhiên , theo u tra , nh ng thùng d u t i hoàn toàn trùng kh p v i 1.000 kg d u b n đ c m t xe t i ch hàng đ m i t i x ng bánh trung thu g n đ ng di khang , t c m c “ ch a t ch thu đ c s nh t ký giao hàng c a x ng d u , nên t m th i ch a k t lu n x ng bánh trung thu t i t c m c có s d ng lo i d u không S vi c v n trình u tra ” , m t nhân viên thu c t u tra giao châu cho bi t “ n u không l p t c lo i b lo i d u n siêu b n , ch c ch n x ng s n xu t s t n_d ng đ ch bi n th c ph m gây nguy h i t i s c kho ng i tiêu_dùng ” , m t ng i dân t i t c m c b c xúc chia s v i báo chí Theo cát miên báo đ t vi t “ ” 18 C.5 V n b n ngu n c a m u th Kmeans_Centroid_EMDS nghi m c a mơ hình  N i dung v n b n ngu n b d li u DUC 2007 Chi ti t xem c m d li u D0716D c a b d li u DUC 2007 [75]  N i dung v n b n ngu n b d li u Corpus_TMV Chi ti t xem c m d li u Cluster_2 c a b d li u Corpus_TMV [76] C.6 V n b n ngu n c a m u th PG_Feature_AMDS nghi m c a mơ hình  N i dung v n b n ngu n b d li u DUC 2004 Chi ti t xem c m d li u D30003T c a b d li u DUC 2004 [74]  N i dung v n b n ngu n b d li u Corpus_TMV Chi ti t xem c m d li u Cluster_59 c a b d li u Corpus_TMV [76] C.7 V n b n ngu n c a m u th Ext_Abs_AMDS nghi m c a mơ hình  N i dung v n b n ngu n b d li u DUC 2004 Chi ti t xem c m d li u D30001T c a b d li u DUC 2004 [74]  N i dung v n b n ngu n b d li u Corpus_TMV Chi ti t xem c m d li u Cluster_75 c a b d li u Corpus_TMV [76] 19 ... ng b n tóm t t h th ng tóm t t sinh H u h t nghiên c u đư c g ng phát tri n m i vƠ c i ti n ph ng pháp tóm t t v n b n hi n có nh m t o h th ng tóm t t v n b n hi u qu Các ph ng pháp tóm t t... trích rút, c n nghiên c u đ xu t c i thi n h th ng tóm t t d a ph ng pháp tóm t t h ng tóm l c ph ng pháp tóm t t h n h p  Thách th c liên quan đ n đ c tr ng th ng kê ngôn ng : C n phát hi n thêm... l n nghiên c u v tóm t t v n b n hi n đ c th c hi n cho tóm t t v n b n ti ng Anh, nghiên c u v bƠi tốn tóm t t v n b n ti ng Vi t cịn khiêm t n Do đó, vi c nghiên c u phát tri n mô hình tóm

Ngày đăng: 01/03/2022, 16:34

Xem thêm:

TỪ KHÓA LIÊN QUAN

TRÍCH ĐOẠN

TÀI LIỆU CÙNG NGƯỜI DÙNG

  • Đang cập nhật ...

TÀI LIỆU LIÊN QUAN

w