EMPIRICAL INVESTIGATIONS OF ANAPHORA AND SALIENCE Massimo Poesio Università di Trento and University of Essex Vilem Mathesius Lectures Praha, 2007 Plan of the series Wednesday: Annotating context dependence, and particularly anaphora Yesterday: Using anaphorically annotated corpora to investigate local & global salience Today: Using anaphorically annotated corpora to investigate anaphora resolution Today’s lecture The Vieira / Poesio work on robust definite description resolution Bridging references Discourse-new (If time allows):Task-oriented evaluation Massimo Poesio: Add better examples (e.g., from The book of evidence) Preliminary corpus study (Poesio and Vieira, 1998) Annotators asked to classify about 1,000 definite descriptions from the ACL/DCI corpus (Wall Street Journal texts) into three classes: DIRECT ANAPHORA: a house … the house DISCOURSE-NEW: the belief that ginseng tastes like spinach is more widespread than one would expect BRIDGING DESCRIPTIONS: the flat … the living room; the car … the vehicle Poesio and Vieira, 1998 Results: More than half of the def descriptions are first-mention Subjects didn’t always agree on the classification of an antecedent (bridging descriptions: ~8%) The Vieira / Poesio system for robust definite description resolution Follows a SHALLOW PROCESSING approach (Carter, 1987; Mitkov, 1998): it only uses Structural information (extracted from Penn Treebank) Existing lexical sources (WordNet) (Very little) hand-coded information (Vieira & Poesio, 1996 / Vieira, 1998 / Vieira & Poesio, 2001) Methods for resolving direct anaphors DIRECT ANAPHORA: the red car, the car, the blue car: premodification heuristics segmentation: approximated with ‘loose’ windows Methods for resolving discoursenew definite descriptions DISCOURSE-NEW DEFINITES the first man on the Moon, the fact that Ginseng tastes of spinach: a list of the most common functional predicates (fact, result, belief) and modifiers (first, last, only… ) heuristics based on structural information (e.g., establishing relative clauses) A `knowledge-based’ classification of bridging descriptions (Vieira, 1998) Based on LEXICAL RELATIONS such as synonymy, hyponymy, and meronimy, available from a lexical resource such as WordNet the flat … the living room The antecedent is introduced by a PROPER NAME Bach … the composer The anchor is a NOMINAL MODIFIER introduced as part of the description of a discourse entity: selling discount packages … the discounts … continued (cases NOT attempted by our system) The anchor is introduced by a VP: Kadane oil is currently drilling two oil wells. The activity… The anchor is not explicitly mentioned in the text, but is a `discourse topic’ the industry (in a text about oil companies) The resolution depends on more general commonsense knowledge last week’s earthquake … the suffering people Distribution of bridging descriptions Class Total Percentage Syn/Hyp/Mer 12/14/12 19% Names 49 24% Compound Nouns 25 12% Events 40 20% Discourse Topic 15 7% Inference 37 18% Total 204 100% The (hand-coded) decision tree 1. 2. 3. 4. Apply ‘safe’ discourse-new recognition heuristics Attempt to resolve as same-head anaphora Attempt to classify as discourse new Attempt to resolve as bridging description. Search backward 1 sentence at a time and apply heuristics in the following order: 1. 2. 3. Named entity recognition heuristics – R=.66, P=.95 Heuristics for identifying compound nouns acting as anchors – R=.36 Access WordNet – R, P about .28 Overall Results Evaluated on a ‘test corpus’ of 464 definite descriptions Overall results: R P F Version 1 53% 76% 62% Version 2 57% 70% 62% D-N def 77% 77% 77% Overall Results Results for each type of definite description: R P F Direct 62% anaphora 83% 71% Disc new Bridging 72% 38% 70% 32.9% 69% 29% Questions raised by the Vieira / Poesio work Do these results hold for larger datasets? Do discourse-new detectors help? Bridging: – – – How to define the phenomenon? Where to get the information? How to combine salience with lexical & commonsense knowledge? Can such a system be helpful for applications? Mereological bridging references Cartonnier (Filing Cabinet) with Clock This piece of mid-eighteenth-century furniture was meant to be used like a modern filing cabinet; papers were placed in leatherfronted cardboard boxes (now missing) that were fitted into the openshelves. A large table decorated in the same manner would have been placed in front for working with those papers. Access to the cartonnier's lower half can only be gained by the doors at the sides, because the table would have blocked the front. PREVIOUS RESULTS A series of experiments using the Poesio / Vieira dataset, containing 204 bridging references, including 39 `WordNet’ bridges (Vieira and Poesio, 2000, but also Carter 1985, Hobbs - a number of papers-, etc) need lexical knowledge But: even large lexical resources such as WordNet not enough, particularly for mereological references (Poesio et al, 1997; Vieira and Poesio, 2000; Poesio, 2003; Garcia-Almanza, 2003) Partial solution: use lexical acquisition (HAL, Hearst-style construction method). Best results (for mereology): construction-style FINDING MERONYMICAL RELATIONS USING SYNTACTIC INFORMATION Some syntactic constructions suggest semantic relations – Ishikawa 1998, Poesio et al 2002: use syntactic constructions to extract mereological information from corpora – – – (Cfr. Hearst 1992, 1998 for hyponyms) The WINDOW of the CAR The CAR’s WINDOW The CAR WINDOW See also Berland & Charniak 1999, Girju et al 2002 LEXICAL RESOURCES FOR BRIDGING: A SUMMARY Class Syn Hyp Mer Total WN Total 12 14 12 38 WordNet 4 (33.3%) 8 (57.1%) 3 (33.3%) 15 (39%) HAL 4 (33.3%) 2 (14.3%) 2 (16.7%) 8 (22.2%) Constructions 1 (8.3%) 0 8 (66.7%) 9 (23.7%) (All using the Vieira / Poesio dataset.) FOCUSING AND MEREOLOGICAL BRIDGES Cartonnier (Filing Cabinet) with Clock This piece of mid-eighteenth-century furniture was meant to be used like a modern filing cabinet; papers were placed in leatherfronted cardboard boxes (now missing) that were fitted into the openshelves. A large table decorated in the same manner would have been placed in front for working with those papers. Access to the cartonnier's lower half can only be gained by the doors at the sides, because the table would have blocked the front. (See Sidner, 1979; Markert et al, 1995.) FOCUS (CB) TRACKING + GOOGLE SEARCH (POESIO, 2003) Analyzed 169 associative BDs in GNOME corpus (58 mereology) Correlation between distance and focusing (Poesio et al, 2004) and choice of anchor – – – 77.5% anchor same or previous sentence; 95.8% in last five sentences CB(U-1) anchor for only 33.6% of BDs, but 89% of anchors had been CB or CP Using `Google distance’ to choose among salient anchor candidates FINDING MEREOLOGICAL RELATIONS USING GOOGLE Lexical vicinity measure (for MERONYMS) between NBD and NPA – Search in Google for “the NBD of the NPA” (cfr. Ishikawa, 1998; Poesio et al, 2002) – E.g., “the drawer of the cabinet” Choose as anchor the PA whose NPA results in the greater number of hits Preliminary results for associative BDs: around 70% P/R (by hand) See also: Markert et al, 2003, 2005; Modjeska et al, 2003 NEW EXPERIMENTS (Poesio et al, 2004) Using the GNOME corpus – – – Completely automatic feature extraction – – 58 mereological bridging refs realized by the-nps 153 mereological bridging references in total Reliably annotated Google & WordNet for lexical distance Using (an approximation of) salience Using machine learning to combine the features More (and reliably annotated) data: the GNOME corpus Texts from 3 genres (museum descriptions, pharmaceutical leaflets, tutorial dialogues) Reliably annotated syntactic, semantic and discourse information – – – grammatical function, agreement features anaphoric relations uniqueness, ontological information, animacy, genericity, … Reliable annotation of bridging references http://cswww.essex.ac.uk/Research/NLE/corpora/GNOME METHODS Salience features: – – – Lexical distance: – – – Utterance distance First mention ‘Global first mention’ (approximate CB) WordNet (using a pure hypernym-based search strategy) Google Tried both separately and together Statistical classifiers: MLP, Naïve Bayes – (MatLab / Weka ML Library) Lexical Distance 1 (WordNet) Computing WordNet Distance: Get the head noun of the anaphor and find all the (noun) senses for the head noun. Get all the noun senses for the head noun of the potential antecedent under consideration. Retrieve the hypernym trees from WordNet for each sense of anaphor and the antecedent. Traverse each unique path in these trees and find a common parent for the anaphor and the antecedent; count the no. of nodes they are apart. Select the least distance path across all combinations. If no common parent is found, assign an hypothetical distance (30). Lexical Distance, 1: WordNet Lexical Distance 2 (Google) As in (Poesio, 2003) But use Google API to access the Google search engine Computing Google hits: – Get the head noun for BR and potential candidate. – Check whether the potential candidate is a mass or count noun. – If count, build the query as “the body of the person” and search for the pattern. – Retrieve the no. of Google hits WN vs GOOGLE Description No path in WordNet No path in WordNet between BD and correct anchor Anchor with Min WN Distance correct Zero Google Hits Zero Google Hits for correct anchor Max Google Hits identify correct candidate Results 503/1720 10/58 8/58 1089/1720 24/58 8/58 BASELINES BASELINE ACCURACY Random choice (previous 5) 4% Random choice (previous) 19% Random choice among FM 21.3% Min Google Distance 13.8% Min WN Distance 13.8% FM entity in previous sentence 31% Min Google in previous sentence 17.2% Min WN in previous sentence 25.9% Min Google among FM 12% Min WN among FM 24.1% RESULTS (58 THE-NPs, 50:50) WN DISTANCE GOOGLE DISTANCE MatLab NN, self-tuned 92 (79.3%) 89 (76.7%) Weka NN Algorithm 91 (78.4%) 86 (74.1%) Weka Naïve Bayes 88 (75.9%) 85 (73.3%) Prec Recall F WN distance 75.4% 84.5% 79.6% Google distance 70.6% 86.2% 77.6% MORE RESULTS 1:3 dataset: WN distance Google distance Accuracy F 80.6% 55.7% 82% 56.7% all 153 mereological BRs: Accuracy F WN distance 224 (74.2%) 76.3% Google distance 230 (75.2%) 75.8% MEREOLOGICAL BDS REALIZED WITH BARE-NPS The combination of rare and expensive materials used on this cabinet indicates that it was a particularly expensive commission. The four Japanese lacquer panels date from the mid- to late 1600s and were created with a technique known as kijimaki-e. For this type of lacquer, artisans sanded plain wood to heighten its strong grain and used it as the background of each panel. They then added the scenic elements of landscape, plants, and animals in raised lacquer. Although this technique was common in Japan, such large panels were rarely incorporated into French eighteenth-century furniture. Heavy Ionic pilasters, whose copper-filled flutes give an added rich color and contrast to the gilt-bronze mounts, flank the panels. Yellow jasper, a semiprecious stone, rather than the usual marble, forms the top. HARDER TEST Using classifiers trained on balanced /slightly unbalanced data (the-nps) on unbalanced ones (10-fold cross validation) Distance Balance Accuracy on balanced F on bal Accuracy on unbal F on unbal WN 1:1 1:3 70.2% 75.9% .7 .4 80.2% 91.7% .2 0 Google 1:1 1:3 64.4% 79.8% .7 .5 63.6% 88.4% .1 .3 WN + Google 1:1 1:3 66.3% 77.9% .6 .4 65.3% 92.5% .2 .5 DISCUSSION Previous results: – – This work: – Construction-based techniques provide adequate lexical resources, particularly when using Web as corpus But need to combine lexical knowledge and salience modeling Combining (simple) salience with lexical resources results in significant improvements Future work: – – Larger dataset Better approximation of focusing Back to discourse-new detection The GUITAR system Recent results GUITAR (Kabadjov, to appear) A robust, usable anaphora resolution system designed to work as part of an XML pipeline Incorporates: – – – Several versions – – – Pronouns: the Mitkov algorithm Definite descriptions: the Vieira / Poesio algorithm Proper nouns: the Bontcheva alg. Version 1: (Poesio & Kabadjov, 2004): direct anaphora Version 2: DN detection Version 3: proper name resolution Freely available from http://privatewww.essex.ac.uk/~malexa/GuiTAR/ DISCOURSE-NEW DEFINITE DESCRIPTIONS (1) Toni Johnson pulls a tape measure across the front of what was once a stately Victorian home. (2) The Federal Communications Commission allowed American Telephone & Telegraph Co. to continue offering discount phone services for large-business customers and said it would soon re-examine its regulation of the long-distance market. Poesio and Vieira (1998): about 66% of definite descriptions in their texts (WSJ) are discourse-new WOULD DNEW RECOGNITION HELP? First version of GUITAR without DN detection on subset of DDs in GNOME corpus - 574 DDs, of which - 184 anaphoric (32%) - 390 discourse-new (67.9%) Total Sys Ana Corr NM WM SM R P F 574 (184) 198 457 (119) 38 27 52 79.6 (60.1) 79.6 (64.7) 79.6 (62.3) 26.3% SPURIOUS MATCHES If your doctor has told you in detail HOW MUCH to use and HOW OFTEN then keep to this advice. ….. If you are not sure then follow the advice on the back of this leaflet. GOALS OF THE WORK Vieira and Poesio’s (2000) system incorporated DISCOURSE-NEW DD DETECTORS (P=69, R=72, F=70.5) Two subsequent strands of work: – – Bean and Riloff (1999), Uryupina (2003) developed improved detectors (e.g., Uryupina: F=86.9) Ng and Cardie (2002) questioned whether such detectors improve results Our project: systematic investigation of whether DN detectors actually help – – ACL 04 ref res: features, preliminary results THIS WORK: results of further experiments DN CLASSIFIER: THE UPPER BOUND Current number of SMs: 52/198 (26.3%) If SM = 0, P=R=F overall = 509/574 = 88.7 – (P=R=F on anaphora only: 119/146= 81.5) VIEIRA AND POESIO’S DN DETECTORS Recognize SEMANTICALLY FUNCTIONAL descriptions: SPECIAL PREDICATES / PREDICATE MODIFIERS (HAND-CODED) the front of what was once a stately Victorian home the best chance of saving the youngest children PROPER NAMES. the Federal Communications Commission … LARGER SITUATION descriptions (HAND-CODED): the City, the sun, …. VIEIRA AND POESIO’S DN DETECTORS, II PREDICATIVE descriptions: COPULAR CLAUSES: he is the hardworking son of a Church of Scotland minister …. APPOSITIONS. Peter Kenyon, the Chelsea chief executive … Descriptions ESTABLISHED by modification: The warlords and private militias who were once regarded as the West’s staunchest allies are now a greater threat to the country’s security than the Taliban …. (Guardian, July 13th 2004, p.10) VIEIRA AND POESIO’S DECISION TREES Tried both hand-coded and ML Hand-coded decision tree: 1. Try the DN detectors with highest accuracy (attempt to classify as functional using special predicates, and as predicative by looking for apposition) 2. Attempt to resolve the DD as direct anaphora 3. Try other DN detectors in order: proper name, establishing clauses, proper name modification …. ML DT: swap 1. and 2. VIEIRA AND POESIO’S RESULTS P R F 50.8 100 67.4 DN detection 69 72 70 Hand-coded DT 62 85 71.7 77 77 77 75 75 75 Baseline (partial) Hand-coded DT (total) ID3 BEAN AND RILOFF (1999) Developed a system for identifying DN definites Adopted syntactic heuristics from Vieira and Poesio, and developed several new techniques: SENTENCE-ONE (S1) EXTRACTION identify as discourse-new every description found in first sentence of a text. DEFINITE PROBABILITY create a list of nominal groups encountered at least 5 times with definite article, but never with indefinite VACCINES: block heuristics when prob. too low. BEAN AND RILOFF’S ALGORITHM 1. If the head noun appeared earlier, classify as anaphoric 2. If DD occurs in S1 list, classify as DN unless vaccine 3. Classify DD as DN if one of the following applies: (a) high definite probability; (b) matches a EHP pattern; (c) matches one of the syntactic heuristics 4. Classify as anaphoric BEAN AND RILOFF’S RESULTS P R Baseline 100 72.2 Syn heuristics 43 93.1 66.3 60.7 69.2 84.3 87.3 83.9 Syn Heuristics + S1 + EHP + DO 81.7 82.2 Syn Heuristics + S1+ EHP + DO + V 79.1 84.5 Syn Heuristics + S1 EHP DO NG AND CARDIE (2002) Directly investigate question of whether discourse-new detectors improves performance of anaphora resolution system Dealing with ALL types of anaphoric expressions NG AND CARDIE’S METHODS DN detectors: – – – statistical classifiers trained using C4.5 and RIPPER Features: predicate & superlative detection / head match / position in text of NP Tested over MUC-6 (F=86) and MUC-7 (F=84) 2 architectures for integration of detectors and AR: 1. 2. Run DN detector first, apply AR on NPs classified as anaphoric Run AR if str_match or alias=Y; otherwise, as in 1. NG AND CARDIE’S RESULTS MUC-6 MUC-7 P R F P R F Baseline (no DN detection) DN detection runs first 70.3 58.3 63.8 65.5 58.2 61.6 57.4 71.6 63.7 47.0 77.1 58.4 Same head runs first 63.4 68.3 65.8 59.7 69.3 64.2 NG AND CARDIE’S RESULTS MUC-6 MUC-7 P R F P R F Baseline (no DN detection) DN detection runs first 70.3 58.3 63.8 65.5 58.2 61.6 57.4 71.6 63.7 47.0 77.1 58.4 Same head runs first 63.4 68.3 65.8 59.7 69.3 64.2 URYUPINA’S METHODS A DN statistical classifier trained using RIPPER Trained / tested over Ng and Cardie’s MUC-7 data URYUPINA’S FEATURES: WEB-BASED DEFINITE PROBABILITY " the Y" " the Y" " a Y" " Y" " the H" " the H" " a H" " H" URYUPINA’S RESULTS (DNEW CLASSIFIER) All NPs Def NPs P R F No Def Prob 87.9 86.0 86.9 Def Prob 88.5 84.3 86.3 No Def Prob 82.5 79.3 80.8 Def Prob 84.8 82.3 83.5 (On MUC-7) URYUPINA’S RESULTS (DNEW CLASSIFIER) All NPs Def NPs P R F No Def Prob 87.9 86.0 86.9 Def Prob 88.5 84.3 86.3 No Def Prob 82.5 79.3 80.8 Def Prob 84.8 82.3 83.5 (On MUC-7) PRELIMINARY CONCLUSIONS Quite a lot of agreement on features for DN recognition: – – – – Recognizing predicative NPs Recognizing establishing relatives Recognizing DNEW proper names Identifying functional DDs Automatic detection of these better Using the Web best All these systems integrate DN detection with some form of AR resolution – See Ng’s results concerning how `globally optimized’ classifiers are better than `locally optimized’ ones (ACL 2004) PRELIMINARY CONCLUSIONS, II Ng and Cardie’s results not the last word: – – Performance of their DN detector not as high as Uryupina’s (F=84 vs. F=87 on same dataset, MUC-7) Overall performance of their resolution system not that high best performance: F=65.8 on ALL NPs But on full NPs (i.e., excluding PNs and pronouns): F=31.7 (GUITAR on DDs, unparsed text: F=56.4) Room for improvement A NEW SET OF EXPERIMENTS Incorporate the improvements in DN detection technology to – – the Vieira / Poesio algorithm, as reimplemented in a stateof-the-art `specialized’ AR system, GUITAR a statistical `general purpose’ AR resolver (Uryupina, in progress) Test over a large variety of data – – – New: GNOME corpus (623 DDs) Original Vieira and Poesio dataset (1400 DDs) MUC-7 (for comparison with Ng and Cardie, Uryupina) (3000 DDs) ARCHITECTURE A two-level system: – – – Run GUITAR’s direct anaphora resolution Results used as one of the features of a statistical discourse-new classifier A `globally optimized’ system (Ng, ACL 2004) Trained / tested over – – GNOME corpus Vieira / Poesio dataset, converted to MMAX, converted to MAS-XML (still correcting the annotation) A NEW SET OF FEATURES DIRECT ANAPHORA Run the Vieira / Poesio algorithm; -1 if no result else distance PREDICATIVE NP DETECTOR DD occurs in apposition DD occurs in copular construction PROPER NAMES c-head c-premod Bean and Riloff’s S1 A REVISED SET OF FEATURES (II) FUNCTIONALITY Uryupina’s four definite probabilities (computed off the Web) superlative ESTABLISHING RELATIVE (a single feature) POSITION IN TEXT OF NP (Ng and Cardie) header / first sentence / first para LEARNING A DN CLASSIFIER Use of the data: – – Classifiers: from the Weka package – 8% for parameter tuning 10-fold cross-validation over the rest Decision Tree (C4.5), NN (MLP), SVM 3 evaluations (overall, DN, DA) Performance comparison: t-test (cfr. Dietterich 98 3 EVALUATIONS OVERALL DN sys DAsys DN corr DAcorr P ,R DN sys DAsys DN DA DN DN sys DN corr P ,R DN sys DN DA DAsys DAcorr P ,R DAsys DA RESULTS: OVERALL GuiTAR T Res C P=R=F 574 574 457 79.6 p .1 GuiTAR +MLP GuiTAR +C4.5 574 574 574 574 473 466 82.4 81.18 not sig RESULTS: DNEW CLASSIFICATION P R F A DNC4.5 86.9 92.3 89.3 85.04 DNMLP 86.4 94.6 90.2 85.89 DNSVM 90.0 86.4 88.1 84.15 BASELINE 67.5 100 80.6 67.5 (all DDs are DN) RESULTS: DIRECT ANAPHORA RESOLUTION T Res C NM WM SM P R F GuiTAR 184 198 119 38 27 52 60.1 64.7 62.3 GuiTAR +MLP 184 142 104 60 20 18 74.1 56.5 63.4 GuiTAR +C4.5 184 158 106 56 22 30 68.9 57.7 62.1 GuiTAR +SVM 184 198 119 38 27 52 60.1 64.7 62.3 ERROR ANALYSIS A 65% reduction in spurious matches: – – – “the answer to any of these questions“ “the title of cabinet maker and sculptor to Louis XIV, King of France” “the other half of the plastic“ But: a 58% increase in no matches – “the palm of the hand” THE DECISION TREE DirectAna <= -1? Y N DNEW (339/36) DirectAna <= 20? Y N TheY/A Y <= 201.2? DNEW (11/1) Y N Relative = 0? Y DirectAna <= 12? 1stPar = 0? N DNEW Y ANAPH N DNEW (12/1) RESULTS: THE VIEIRA/POESIO CORPUS Tested on 400 DDs (the ‘test’ corpus) Initial results at DN detection very poor Problem: the two conversions resulted in the loss of much information about modification, particularly relatives Currently correcting the annotation by hand RESULTS: AUTOMATIC PARSING GUITAR without DN detection over the same texts, but using a chunker: 10% less accuracy Main problem: many DDs not detected (particularly possessives) Currently experimenting with full parsers (tried several, settled on Charniak’s) CONCLUSIONS AND DISCUSSION All results so far support the idea that DN detectors improve the performance of AR with DD (if perhaps by only a few percent) Some agreement on what features are useful – One clear lesson: interleave AR and DN detection! But: will need to test on larger corpora (also to improve performance of classifier) Current work: – – Test on unparsed text Test on MUC-7 data Task-based evaluation RANLP / EMNLP slides Conclusions URLs Massimo Poesio: http://cswww.essex.ac.uk/staff/poesio GUITAR: http://privatewww.essex.ac.uk/~malexa/GuiTAR/ WEKA: http://www.cs.waikato.ac.nz/~ml