1. CS626 Assignments Abhirut Gupta Mandar Joshi Piyush Dungarwal (Group No. 6)

2. Trigram Implementation Details Trigram probabilities were smoothened using bigram and unigram probabilities. Lambdas were calculated using deleted interpolation. Outside Viterbi, Capitalization was used to improve the accuracy of NP0 tag.

3. Unknown word handling Unknown words were given small emission probabilities. Considerable experimentation was done to fix the value of this emission probability.

4. Trigram Accuracy Fold No.Total TagsCorrect TagsAccuracy 122504321451595.322 2715026821095.395 3967499259995.710 435925534226695.271 511023610540395.616

5. Confusion Matrix CJS : accuracy = 87.43 % Confused with : PRP = 9.35 % Eg. “As” CJS in “diagnosed as being HIV positive” PRP in “give AIDS greater recognition not as a disease” VVB : accuracy = 70.72 % Confused with : VVI = 12.97 % Eg. Forget, send, live, return, etc.

6. Confusion Matrix AV0 : accuracy = 87.27 % Confused with : PRP = 4.67 % VVD : accuracy = 88.98 % Confused with : VVN = 6.49 % Eg. Sent, lived, returned, etc.

7. Insight Nouns and verbs are often confused because, Most verbs can also be used as nouns. E.g. laugh, people, etc. Adverbs (which often precede verbs) can also be used as adjectives (which often precede nouns) E.g. fast, slow, etc.

8. Next word prediction (W3, T3)* = argmax (W3, T3) (P(W3, T3 | W2, W1, T2, T1)) = argmax (W3, T3) (P(T3| W2, W1, T2, T1) * P(W3 | W2, W1, T3, T2,T1)) = argmax (W3, T3) (P(T3| T2, T1) * P(W3 | W2, W1, T3))

9. Next word prediction (alternate model) W3* = argmax W3 P(W3 | W1, W2, T1, T2) = argmax W3 Σ T3 P(W3, T3| W1, W2, T1, T2) = argmax W3 Σ T3 P(T3| W1, W2, T1, T2).P(W3 | W1, W2, T1, T2, T3) =argmax W3 Σ T3 P(T3| T1, T2).P(W3 | W1, W2, T3) Drawbacks: Extremely slow without any considerable improvement in accuracy

10. Next word prediction ModelAccuracyPerplexity Tagged Model11.32 %11.83 Untagged Model4.67 %22.87

11. Generative vs. Discriminative (Bigram) FOLDDiscriminative ModelGenerative Model 179.90 %82.89 % 281.08 %80.79 % 381.91 %81.02 % 480.84 %83.45 % 581.80 %80.95 % Total80.85 %82.49 %

12. Generative model possibly gives better accuracy than Discriminative model because, we model Transition and Lexical probability separately in the former model. Discriminative approach: T* = argmax (T) Π P( ti | ti-1, wi-1) Generative vs. Discriminative

13. A* Results 1 Trigram Viterbi after applying Capitalization A* accuracy (With unknowns) A* accuracyViterbi 52.3 %94.06 %95.39 %

14. Node Expansions

15. In Trigram Viterbi implementation, Capitalization is applied outside the algorithm. This results in certain words being tagged differently by two algorithms. From the graph, A* expands fewer nodes than Viterbi. However statistics for node expansion do not take into account the comparisons made in the Viterbi algorithm

16. Beam search algorithm An upper bound is maintained on the size of Open list of nodes Results for size = 50:  Reduction in execution time : 56.62%  Accuracy : 78.28 % Results for size = 25:  Reduction in execution time : 76.50%  Accuracy : 63.24 %

17. Projection of Parse Trees

18. Our Approach From the English parse tree, we obtain the Noun Phrases. Using an English-Hindi dictionary and the given translation, find corresponding groups in the Hindi sentence, and replace the English groups by the Hindi ones.

19. Replace the remaining words by corresponding Hindi words from the sentence. Now, apply hard-coded inversion rules to the tree. Example – PP -> P NP in english is inverted as PP -> NP P in hindi

20. S PP NP VP INNP NN law By NP PP DT all NNS children INNP ofJJ NN compulsory school age MD VP must VBNP get a proper full time education

21. S PP NP VP INNP NN kaanoon By NP PP DT sabhi NNS bacchon INNP ofJJ NN anivarya School jaane ki umar MD VP must VBNP get pure samay ki shiksha

22. S PP NP VP INNP NN kaanoon ke anusaar NP PP DT sabhi NNS bacchon ko INNP keJJ NN anivarya school jaane ki umar MD VP honi chahiye VBNP praapt pure samay ki shiksha

23. S PP NP VP INNP NN kaanoon ke anusaar NP PP DT sabhi NNS bacchon ko IN NP ke JJNN anivarya school jaane ki umar MD VP honi chahiye VBNP praapt pure samay ki shiksha

24. Challenges Translated sentences may have a different structure altogether. It may not be possible to detect similar phrases. Example -  आपका बच्चा ड्रग्स या सॉल्वैंट्स लेने के चक्कर में क्यों पड़ गया है, यदि आप उसके कारणों को समझते हैं, तो इस विषय में अपने बच्चों से बात करना आपके लिए काफ़ी आसान होगा ।  If you understand the reasons why a child can get involved with drugs and solvents, it's much easier for you to talk to your children about it.

25. Drawbacks Parse trees produced may be “approximate” Due to morphological differences it may become difficult to produce tags for extra words created in transtaltion. Example – “ko”, “jaane ki” produced in the previous example.

26. Yago Yago fact table was queried to find relevant subject-predicate-object pairs. BFS was used to find path from start node to the goal node. This approach gives all paths of minimum length.

27. Yago Interesting relations: Sachin Tendulkar and Rahul Dravid have both won the Arjuna Award. Dev Anand was born in Mumbai, died in London. Rohit Sharma and Vikram Pandit were born in Nagpur.