Combining Lexical and Syntactic Features for Supervised Word

Combining Lexical and Syntactic Features for Supervised Word Sense Disambiguation Masters Thesis : Saif Mohammad Advisor : Dr. Ted Pedersen University of Minnesota, Duluth 1

Path Map n Introduction n Background n Data n Experiments n Conclusions 2

Word Sense Disambiguation Harry cast a bewitching spell n Humans immediately understand spell to mean a charm or incantation. n reading out letter by letter or a period of time ? n n n Words with multiple senses – polysemy, ambiguity! Utilize background knowledge and context. Machines lack background knowledge. n n Automatically identifying the intended sense of a word in written text, based on its context, remains a hard problem. Best accuracies in recent international event, around 65%. 3

Why do we need WSD ! n Information Retrieval n Query: cricket bat n n Machine Translation n Consider English to Hindi translation. n n Documents pertaining to the insect and the mammal, irrelevant. head to sar (upper part of the body) or adhyaksh (leader)? Machine-human interaction n Instructions to machines. n n Interactive home system: turn on the lights Domestic Android: get the door Applications are widespread and will affect our way of life. 4

Terminology n Harry cast a bewitching spell Target word – the word whose intended sense is to be identified. n n Context – the sentence housing the target word and possibly, 1 or 2 sentences around it. n n spell Harry cast a bewitching spell Instance – target word along with its context. WSD is a classification problem wherein the occurrence of the target word is assigned to one of its many possible senses. 5

Corpus-Based Supervised Machine Learning A computer program is said to learn from experience … if its performance at tasks … improves with experience. - Mitchell n n n Task : Word Sense Disambiguation of given test instances. Performance : Ratio of instances correctly disambiguated to the total test instances – accuracy. Experience : Manually created instances such that target words are marked with intended sense – training instances. Harry cast a bewitching spell / incantation 6

Path Map n Introduction n Background n Data n Experiments n Conclusions 7

Decision Trees n A kind of classifier. n n Assigns a class by asking a series of questions. Questions correspond to features of the instance. Question asked depends on answer to previous question. Inverted tree structure. n Interconnected nodes. n Top most node is called the root. n Each node corresponds to a question / feature. Each possible value of feature has corresponding branch. n Leaves terminate every path from root. n n Each leaf is associated with a class. 8

Automating Toy Selection for Max ROOT Moving Parts ? NODES Yes No Car ? No Yes Size ? HATE Small Big SO SO LOVE Color ? Blue Car ? No LOVE HATE Yes Size ? SO SO Small LEAVES Other Red HATE Big LOVE 9

WSD Tree Feature 1 ? 1 0 Feature 2 ? 0 1 Feature 4 ? 0 SENSE 1 Feature 4? 0 Feature 2 ? SENSE 3 1 SENSE 4 1 0 SENSE 1 1 Feature 3 ? SENSE 3 0 SENSE 2 1 SENSE 3 10

Choice of Learning Algorithm n Why use decision trees for WSD ? n n It has drawbacks – training data fragmentation What about other learning algorithms such as neural networks? n Context is a rich source of discrete features. n The learned model likely meaningful. n May provide insight into the interaction of features. Pedersen[2001]*: Choosing the right features is of greater significance than the learning algorithm itself * “A Decision Tree of Bigrams is an Accurate Predictor of Word Sense”, T. Pedersen, In the Proceedings of the Second Meeting of the North American Chapter of the Association for Computational Linguistics (NAACL-01), June 2 -7, 2001, Pittsburgh, PA. 11

Lexical Features n Surface form n n A word we observe in text. Case(n) n n 1. Object of investigation 2. frame or covering 3. A weird person Surface forms : case, cases, casing An occurrence of casing suggests sense 2. Unigrams and Bigrams n One word and two word sequences in text. n The interest rate is low Unigrams: the, interest, rate, is, low Bigrams: the interest, interest rate, rate is, is low n 12

Part of Speech Tagging n Pre-requisite for many natural language tasks. n n Parsing, WSD, Anaphora resolution Brill Tagger* – most widely used tool. n Accuracy around 95%. n Source code available. n Easily understood rules. Harry/NNP cast/VBD a/DT bewitching/JJ spell/NN NNP proper noun, VBD verb past, DT determiner, NN noun * http: //www. cs. jhu. edu/~brill/RBT 1_14. tar. Z 13

Pre-Tagging n Pre-tagging is the act of manually assigning tags to selected words in a text prior to tagging. Mona will sit in the pretty chair//NN this time n n n chair is the pre-tagged word, NN is its pre-tag. Reliable anchors or seeds around which tagging is done. Brill Tagger facilitates pre-tagging. n Pre-tag not always respected! Mona/NNP will/MD sit/VB in/IN the/DT pretty/RB chair//VB this/DT time/NN 14

Contextual Rules n n Initial state tagger – assigns most frequent tag for a type based on entries in a Lexicon (pre-tag respected). Final state tagger – may modify tag of word based on context (pre-tag not given special treatment). Relevant Lexicon Entries Type Most frequent tag chair NN(noun) pretty RB(adverb) Other possible tags VB(verb) JJ(adjective) Relevant Contextual Rules Current Tag NN RB When NEXTTAG DT NEXTTAG NN New Tag VB JJ 15

Guaranteed Pre-Tagging n A patch to the tagger provided – Brill. Patch. n n Application of contextual rules to the pre-tagged words bypassed. Application of contextual rules to non pre-tagged words unchanged. Mona/NNP will/MD sit/VB in/IN the/DT pretty/JJ chair//NN this/DT time/NN n Tag of chair retained as NN. n n Contextual rule to change tag of chair from NN to VB not applied. Tag of pretty transformed. n Contextual rule to change tag of pretty from RB to JJ applied. * ”Guaranteed Pre-Tagging for the Brill Tagger”, Mohammad, S. and Pedersen, T. , In Proceedings of Fourth International Conference of Intelligent Systems and Text Processing, February 2003, Mexico. 16

Part of Speech Features n A word used in different senses is likely to have different sets of pos tags around it. Why did jack turn/VB against/IN his/PRP$ team/NN Why did jack turn/VB left/NN at/IN the/DT crossing n Features used n Individual word POS: P-2, P-1, P 0, P 1, P 2 n n P 1 = JJ implies that the word to the right of the target word is an adjective. A combination of the above. 17

Parse Features n Collins Parser* used to parse the data. n n n Head word of a phrase. n n n Source code available. Uses part of speech tagged data as input. the hard work, the hard surface Phrase itself : noun phrase, verb phrase and so on. Parent : Head word of the parent phrase. n n fasten the line, cross the line Parent phrase. * http: //www. ai. mit. edu/people/mcollins 18

Sample Parse Tree SENTENCE NOUN PHRASE VERB PHRASE Harry cast NNP VBD NOUN PHRASE a bewitching spell DT JJ NN 19

Path Map n Introduction n Background n Data n Experiments n Conclusions 20

Sense-Tagged Data n Senseval-2 data n n Senseval-1 data n n 4328 instances of test data and 8611 instances of training data ranging over 73 different noun, verb and adjectives. 8512 test instances and 13, 276 training instances, ranging over 35 nouns, verbs and adjectives. line, hard, interest, serve data n 4, 149, 4, 337, 4378 and 2476 sense-tagged instances with line, hard, serve and interest as the head words. Around 50, 000 sense-tagged instances in all! 21

Data Processing n n Packages to convert line hard, serve and interest data to Senseval-1 and Senseval-2 data formats. refine preprocesses data in Senseval-2 data format to make it suitable for tagging. n n pos. Senseval part of speech tags any data in Senseval-2 data format. n n Restore one sentence per line and one line per sentence, pre-tag the target words, split long sentences. Brill tagger along with Guaranteed Pre-tagging utilized. parse. Senseval parses data in a format as output by the Brill Tagger. n n Restores xml tags, creating a parsed file in Senseval-2 data format. Uses the Collins Parser. 22

Sample line data instance Original instance: art} aphb 01301041: " There's none there. " He hurried outside to see if there were any dry ones on the line. Senseval-2 data format: <instance id="line-n. art} aphb 01301041: "> <answer instance="line-n. art} aphb 01301041: " senseid="cord"/> <context> <s> " There's none there. " </s> <s> He hurried outside to see if there were any dry ones on the <head>line</head>. </s> </context> </instance> 23

Sample Output from parse. Senseval <instance id=“harry"> <answer instance=“harry" senseid=“incantation"/> <context> Harry cast a bewitching <head>spell</head> </context> </instance> <instance id=“harry"> <answer instance=“harry" senseid=“incantation"/> <context> <P=“TOP~cast~1~1”> <P=“S~cast~2~2”> <P=“NPB~Potter~2~2”> Harry <p=“NNP”/> <P=“VP~cast~2~1”> cast <p=“VB”/> <P=“NPB~spell~3~3”> a <p=“DT”/> bewitching <p=“JJ”/> spell <p=“NN”/> </P> </context> </instance> 24

Issues… n n How is the target word identified in line, hard and serve data? How the data is tokenized for better quality pos tagging and parsing? n How is the data pre-tagged? n How is parse output of Collins Parser interpreted? n n How is the parsed output XML’ized and brought back to Senseval-2 data format? Idiosyncrasies of line, hard, serve, interest, Senseval-1 and Senseval-2 data and how they are handled? 25

Path Map n Introduction n Background n Data n Experiments n Conclusions 26

Lexical: Senseval-1 & Senseval-2 Sval-1 line hard serve interest Majority 47. 7% 56. 3% 54. 3% 81. 5% 42. 2% 54. 9% Surface Form 49. 3% 62. 9% 54. 3% 81. 5% 44. 2% 64. 0% Unigram 55. 3% 66. 9% 74. 5% 83. 4% 73. 3% 75. 7% Bigram 55. 1% 66. 9% 72. 9% 89. 5% 72. 1% 79. 9% 27

Individual Word POS (Senseval-1) All Nouns Verbs Adj. Majority 56. 3% 57. 2% 56. 9% 64. 3% P-2 57. 5% 58. 2% 58. 6% 64. 0 P-1 59. 2% 62. 2% 58. 2% 64. 3% P 0 60. 3% 62. 5% 58. 2% 64. 3% P 1 63. 9% 65. 4% 64. 4% 66. 2% P-2 59. 9% 60. 0% 60. 8% 65. 2% 28

Individual Word POS (Senseval-2) All Nouns Verbs Adj. Majority 47. 7% 51. 0% 39. 7% 59. 0% P-2 47. 1% 51. 9% 38. 0% 57. 9% P-1 49. 6% 55. 2% 40. 2% 59. 0% P 0 49. 9% 55. 7% 40. 6% 58. 2% P 1 53. 1% 53. 8% 49. 1% 61. 0% P-2 48. 9% 50. 2% 43. 2% 59. 4% 29

Combining POS Features Sval-2 Sval-1 line hard serve interest Majority 47. 7% 56. 3% 54. 3% 81. 5% 42. 2% 54. 9% P 0, P 1 54. 3% 66. 7% 54. 1% 81. 9% 60. 2% 70. 5% P-1, P 0, P 1 54. 6% 68. 0% 60. 4% 84. 8% 73. 0% 78. 8% P-2, P-1, 54. 6% P 0, P 1 , P 2 67. 8% 62. 3% 86. 2% 75. 7% 80. 6% 30

Effect Guaranteed Pre-tagging on WSD Senseval-1 Senseval-2 Guar. P. Reg. P P-1, P 0 62. 2% 62. 1% 50. 8% 50. 9% P 0, P 1 66. 7% 54. 3% 53. 8% P-1, P 0, P 1 68. 0% 67. 6% 54. 7% P-2, P-1, P 0, P 1 67. 8% , P 2 66. 1% 54. 6% 54. 1% 31

Parse Features (Senseval-1) All Nouns Verbs Adj. Majority 56. 3% 57. 2% 56. 9% 64. 3% Head 64. 3% 70. 9% 59. 8% 66. 9% Parent 60. 6% 62. 6% 60. 3% 65. 8% Phrase 58. 5% 57. 2% 66. 2% Par. Phr. 57. 9% 58. 1% 58. 3% 66. 2% 32

Parse Features (Senseval-2) All Nouns Verbs Adj. Majority 47. 7% 51. 0% 39. 7% 59. 0% Head 51. 7% 58. 5% 39. 8% 64. 0% Parent 50. 0% 56. 1% 40. 1% 59. 3% Phrase 48. 3% 51. 7% 40. 3% 59. 5% Par. Phr. 48. 5% 53. 0% 39. 1% 60. 3% 33

Thoughts… n n Both lexical and syntactic features perform comparably. But do they get the same instances right ? n n How much are the individual feature sets redundant. Are there instances correctly disambiguated by one feature set and not by the other ? n How much are the individual feature sets complementary. Is the effort to combine of lexical and syntactic features justified? 34

Measures n Baseline Ensemble: accuracy of a hypothetical ensemble which predicts the sense correctly only if both individual feature sets do so. n n Quantifies redundancy amongst feature sets. Optimal Ensemble: accuracy of a hypothetical ensemble which predicts the sense correctly if either of the individual feature sets do so. n Difference with individual accuracies quantifies complementarity. We used a simple ensemble which sums up the probabilities for each sense by the individual feature sets to decide the intended sense. 35

Best Combinations Data Set 1 Set 2 Base Ens. Opt. Best Sval-2 47. 7% Unigrams 55. 3% P-1, P 0, P 1 55. 3% 43. 6% 57. 0% 67. 9% 66. 7% Sval-1 56. 3% Unigrams 66. 9% P-1, P 0, P 1 68. 0% 57. 6% 71. 1% 78. 0% 81. 1% line 54. 3% Unigrams 74. 5% P-1, P 0, P 1 60. 4% 55. 1% 74. 2% 82. 0% 88. 0% hard 81. 5% Bigrams 89. 5% Head, Par 86. 1% 87. 7% 88. 9% 91. 3% 83. 0% serve 42. 2% Unigrams 73. 3% P-1, P 0, P 1 73. 0% 58. 4% 81. 6% 89. 9% 83. 0% interest 54. 9% Bigrams 79. 9% P-1, P 0, P 1 78. 8% 67. 6% 83. 2% 90. 1% 89. 0% 36

Path Map n Introduction n Background n Data n Experiments n Conclusions 37

Conclusions n Significant amount of complementarity across lexical and syntactic features. n Combination of the two justified. n We show that simple lexical and part of speech features can achieve state of the art results. n How best to capitalize on the complementarity still an open issue. 38

Conclusions (continued) n Part of speech of word immediately to the right of target word found most useful. n n Pos of words immediately to the right of target word best for verbs and adjectives. Nouns helped by tags on either side. (P 0, P 1) found to be most potent in case of small training data per instance (Sval data). Larger pos context size (P-2, P-1, P 0, P 1 , P 2) shown to be beneficial when training data per instance is large (line, hard, serve and interest data) n Head word of phrase particularly useful for adjectives n Nouns helped by both head and parent. 39

Other Contributions n Converted line, hard, serve and interest data into Senseval-2 data format. n Part of speech tagged and Parsed the Senseval 2, Senseval-1, line, hard, serve and interest data. n Developed the Guaranteed Pre-tagging mechanism to improve quality of pos tagging. n Showed that guaranteed pre-tagging improves WSD. The hard and serve data, part of speech tagged using Guaranteed Pre-tagging is part of NLTK data kit. 40

Code, Data & Resources n n n Synta. Lex : A system to do WSD using lexical and syntactic features. Weka’s decision tree learning algorithm is utilized. pos. Senseval : part of speech tags any data in Senseval-2 data format. Brill Tagger used. parse. Senseval : parses data in a format as output by the Brill Tagger. Output is in Senseval-2 data format with part of speech and parse information as xml tags. Uses Collins Parser. Packages to convert line hard, serve and interest data to Senseval-1 and Senseval-2 data formats. Brill. Patch : Patch to Brill Tagger to employ Guaranteed Pre-Tagging. http: //www. d. umn. edu/~tpederse/code. html http: //www. d. umn. edu/~tpederse/data. html 41

Documentation n “Combining Lexical and Syntactic Features for Supervised Word Sense Disambiguation”, Mohammad, S. and Pedersen, T. , To appear in the Proceedings of Eighth Conference on Natural Language Learning at HLT -NAACL, May 2004, Boston. n n “Guaranteed Pre-Tagging for the Brill Tagger”, Mohammad, S. and Pedersen, T. , In Proceedings of Fourth International Conference of Intelligent Systems and Text Processing, February 2003, Mexico. “Combining Lexical and Syntactic Features for Supervised Word Sense Disambiguation”, Mohammad, S. , Masters Thesis, August 2003, University of Minnesota, Duluth. http: //www. d. umn. edu/~tpederse/students. html 42

Senseval-3 (Mar-1 to April 15, 2004) Around 8000 training and 4000 test instances. Results expected shortly. Thank You 43