Parsing Discourse Relations Giuseppe Riccardi Signals and Interactive Systems Lab University of Trento, Italy
Behavioral Analytics
Parser Run on Genia Corpus Among 25 cases, 2 homozygous deletions and 1 hemizygous deletion were found in HCC samples. No point mutation was identified in the remaining 22 tumor samples without p16 gene deletions. Hypermethylation was detected in 24% (6/25) of tumor samples. However, the corresponding non-tumor liver tissue specimens were always unmethylated at the p16 locus. Loss of p16 protein expression occurred in 16 of 35 (45.7%) tumor samples, and all the non-tumor liver tissue specimens showed positive p16 staining. For the 25 cases examined for p16 gene alterations, the loss of p16 protein expression was observed in all tumors with p16 gene alterations and also in 3 tumors without p16 gene alterations. (Source: Genia corpus)
Parser Run on Genia Corpus Among 25 cases, 2 homozygous deletions and 1 hemizygous deletion were found in HCC samples. No point mutation was identified in the remaining 22 tumor samples without p16 gene deletions. Hypermethylation was detected in 24% (6/25) of tumor samples. However, the corresponding non-tumor liver tissue specimens were always unmethylated at the p16 locus. Loss of p16 protein expression occurred in 16 of 35 (45.7%) tumor samples, and all the non-tumor liver tissue specimens showed positive p16 staining. For the 25 cases examined for p16 gene alterations, the loss of p16 protein expression was observed in all tumors with p16 gene alterations and also in 3 tumors without p16 gene alterations. (Source: Genia corpus) Parser Output : Hypermethylation was detected in 24 % 6\/25 ) of tumor samples However(Comparison) the corresponding non-tumor liver tissue specimens were always unmethylated at the p16 locus Loss of p16 protein expression occurred in 16 of 35 45.7 % ) tumor samples and(expansion ) all the non-tumor liver tissue specimens showed positive p16 staining
Social Media User Opinions: Negative The acting is below average, even from the likes of Curtis. You're more likely to get a kick out of her work in Halloween H20. Sutherland is wasted and Baldwin, well, he's acting like a Baldwin, of course. The real star here are Stan Winston's robot design, some schnazzy CGI, and the occasional good gore shot, like picking into someone's brain. So, if robots and body parts really turn you on, here's your movie. Otherwise, it's pretty much a sunken ship of a movie. 5/1/12 5
Social Media User Opinions: Positive From here on, the plot takes a back seat, and we are treated to some of the best camera work and action staged. Most all the action is plausible and will hold you at the edge of your seat. There are a few melodramatic parts here, but, they tend to work out well. There is no general antagonist in this film, but the action and suspense makes you forget all about that. Daylight is a great film, I saw a non-matinee showing of it, and I thought it was worth every penny. The characterizations are mostly flat, one dimesional, but they have enough in them to get you to care for some of the characters. Rob Cohen (Dragonheart) does a great job with this film. 5/1/12 6
Discourse Relation Parsing Joint work with Sucheta Ghosh, U. Trento Richard Johansson, U. Trento/U. Gothenburg Sara Tonelli, FBK-Irst Ghosh S., Tonelli S., Riccardi G. and Johansson R., End-to-End Discourse Parser Evaluation, IEEE International Conference on Semantic Computing, Menlo Park, USA, 2011 Ghosh S., Johansson R., Riccardi G. and Tonelli S., Shallow Discourse Parsing with Conditional Random Fields, International Joint Conference on Natural Language Processing, Chiang Mai, Thailand, 2011 Ghosh S., Johansson R., Riccardi G. and Tonelli S., Improving Recall Through Global Constraint Selection, To appear on LREC, 2012 Giuseppe Riccardi
Discourse Parser From raw text extract: Discourse relations: Discourse Predicate (Connective) Connective sense Arg1 Arg2 Explicit Connective Giuseppe Riccardi
Parsing Architecture
Parser end2end Architecture Doc Parser Stanford (K&M) Parse_Tree Chunklink AddDiscourse RootExtract +Morpha By Sabaine Buchholz CoNLL 00 task Pitler & Nenkova 09 Conn. SenseDet. Morph & All Feat Johansson+ Minnen et al Windowing (-2,+2) Arg2 Arg1
Features: Example
Selected Features: Arg1 Features used for Arg1 and Arg2 segmentation and labeling. F1. Token (T) F2. Sense of Connective (CONN) F3. IOB chain (IOB) F4. PoS tag F5. Lemma (L) F6. Inflection (INFL) F7. Main verb of main clause (MV) F8. Boolean feature for MV (BMV) Additional feature used only for Arg1 F9. Previous Sentence (PREV) F10. Arg2 Labels
Inter vs Intra Sentence Arguments Illustration: PREV Feature This &ilm should be brilliant. Howeverr, it can t hold up. 13
Inter vs Intra Sentence Arguments Illustration: PREV Feature However However However However However 0 0 0 0 0 This &ilm should be brilliant. Howeverr, it can t hold up. 14
Inter vs Intra Sentence Arguments 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0.77 0.68 0.61 0.52 0.36 0.27 P R F1 Intra+Prev Inter+Prev - PREV +PREV Illustration: PREV Feature However However However However However 0 0 0 0 0 This &ilm should be brilliant. Howeverr, it can t hold up. 15
Selected Features: Arg2 Features used for Arg1 and Arg2 segmentation and labeling. F1. Token (T) F2. Sense of Connective (CONN) F3. IOB chain (IOB) Ghosh S., Johansson R., Riccardi G. and Tonelli S., Shallow Discourse Parsing with Conditional Random Fields, International Joint Conference on Natural Language Processing, Chiang Mai, Thailand, 2011
Parser Evaluation Giuseppe Riccardi 17
Parser Evaluation Giuseppe Riccardi 18
Lightweight Features -Reduce dimensionality of IOB chain features -Control robustness of parser (wrt to syntactic parse) -Binary features selected from IOB chain Giuseppe Riccardi, UNITN 19
Lightweight Features IOB Chain feature replaced by two pairs of Boolean features (1) The second top parent node whether starting (B) or not (2) The third top parent node whether starting (B) or not (3) The second top parent node whether ending (E) or not (4) The third top parent node whether ending (E) or not Example: Tree diagram showed IOB feature for token flashed is I-S/E-VP/E-SBAR/E-S/C-VP Replacing Boolean feature for flashed respectively: (1) 0 ( ß E-VP ) (2) 0 ( ß E-SBAR ) (3) 1 ( ß E-VP ) (4) 1 ( ß E-SBAR ) Giuseppe Riccardi, UNITN 20
Parser Evaluation: Arg2 Exact Match P R F1 Baseline 0.53 0.46 0.49 Gold - Standard 0.84 0.74 0.79 Gold-Lightweight 0.80 0.74 0.77 AutoConn+GoldSPT 0.82 0.70 0.76 GoldConn+AutoSPT 0.76 0.61 0.68 Lightweight(Auto) 0.72 0.56 0.63 Giuseppe Riccardi 21
N-Best Parse Re-ranking Ø Online Passive-Aggressive Perceptron Ø Structured Voted Perceptron Ø Linear Preference Learning Support Vector Machine Ø Linear Best vs. Rest Support Vector Machine End2End Disc Parse 22
N-Best ReRanking with Global Constraints Ø GF0. Log Posteriors Ø GF1. Overgeneration. Ø GF2. Undergeneration. Ø GF3. Intersentential Arg2. Ø GF4. Arg1 after the connective sentence Ø GF5. Argument overlapping with the connective. Ø GF6. Argument begins with I- tag Ø GF7. Argument begins with E- tag End2End Disc Parse 23
N-Best ReRanking with Global Constraints Exact Arg1 Arg2 P R F1 P R F1 Baseline 69.88 48.51 57.26 83.44 75.14 79.07 Online PA 66.10 53.92 59.39(16) 82.59 76.39 79.37(4) Struct Per 67.18 52.64 59.03(4) 82.96 76.28 79.48(8) Bestvs Rest 66.19 52.83 58.94(8) 81.69 77.14 79.35(4) Pref-Linear 66.54 53.31 59.20(4) 82.82 76.28 79.42(4) Exact Match Scores. Used n- best list numbers in parenthesis End2End Disc Parse 24
Research Challenges Speech, Dialog and Discourse Speech Signal vs Linguistic correlates Eat your porridge! You re not going to football practice Parser Trade-off btw coverage and agreement Robustness of features Semantic Annotation Domain/Genre Adaptation
Research Challenges Speech, Dialog and Discourse Acoustics vs lexical correlates Eat your porridge! You re not going to football practice Parser Trade-off amongst sense-depth, coverage, agreement Robustness of features Semantic Annotation Domain/Genre Adaptation
Publications Speech (LUNA Corpus) Tonelli S., Riccardi G., Prasad R. and Joshi A. "Annotation of Discourse Relations for Conversational Spoken Dialogs", LREC Valletta, 2010. Text (PDTB corpus) Ghosh S., Johansson R., Riccardi G. and Tonelli S., Shallow Discourse Parsing with Conditional Random Fields, International Joint Conference on Natural Language Processing, Chiang Mai, Thailand, 2011 Ghosh S., Tonelli S., Riccardi G. and Johansson R., End-to-End Discourse Parser Evaluation, IEEE International Conference on Semantic Computing, Menlo Park, USA, 2011 Giuseppe Riccardi 27