![Text Mining Term Project 18
Modeling – Word2vec
Finally, Normalize the cosine similarity result for each document vectors, binarize the
answers(label), and evaluate the performance of the model
Test set
Ex)
𝑉′ 𝑘 = 0.05,0.001, 0.003 … . 0.002
𝐴 𝑘 = [1,0,0 …. 0]](https://image.slidesharecdn.com/tmfinalslideshareversion-180428060724/85/Context-based-movie-search-using-doc2vec-word2vec-18-320.jpg)
Ad
More Related Content
What's hot (20)
Similar to Context-based movie search using doc2vec, word2vec (20)
Ad
Recently uploaded (20)
Ad
Context-based movie search using doc2vec, word2vec
- 1. Context-based movie search for user questions that ask the title of the movie 장진규, 박정인 School of Business Administration 2018. 4. 18
- 2. Contents I. Introduction II. Preprocessing III. Analysis IV. Results
- 3. Contents I. Introduction II. Preprocessing III. Analysis IV. Results
- 4. Text Mining Term Project 4 Introduction People sometimes have a craving for find a movie that they once glimpsed. At that time, they used to ask the movie name through Q&A sites and get the result. Answerers often seems ‘god of movie’, so we want to imitate their prophecy. Question Examples
- 5. Text Mining Term Project 5 Data Gathering We chose one expert of this field and gather his answers. Q&A Site http://kin.naver.com Expert ID xedz**** Question & Answer data 39,758 Date 2012 December ~ 2018March Unique Movie 5,900 Gathered Data Information
- 6. Text Mining Term Project 6 2 Types of Text Representation There are 2 kinds of text representation: sparse and dense Sparse: One-Hot Encoding Dense: Word Embedding Sparse Dense Dimension ▪ As many as unique words • Autonomous setting • Usually 20~200 dimensions Information • Lots of 0 value • No Information • Every element has value • Abundant Information Comparison of Text Representations source: https://dreamgonfly.github.io/machine/learning,/natural/language/processing/2017/08/16/word2vec_explained.html
- 7. Text Mining Term Project 7 Main Idea of Word2Vec Word2Vec is one of the word embedding methods. Its main idea is “You shall know a word by the company it keeps.” Every word has friends around them
- 8. Text Mining Term Project 8 Algorithms of Word2Vec Word2vec has two model architectures: continuous-bag-of-words (CBOW), skip-gram. Diagrams of CBOW and Skip-gram source: https://aws.amazon.com/ko/blogs/korea/amazon-sagemaker-blazingtext-parallelizing-word2vec-on-multiple-cpus-or-gpus/
- 9. Text Mining Term Project 9 Algorithms of Doc2Vec Doc2vec has two model architectures: distributed memory model (PV-DM) and Distributed bag of words model(PV-DBOW). Diagrams of PV-DM and PV-DBOW source: Distributed Representations of Sentences and Documents The concatenation or average of vector with a context of three words is used to predict the fourth word. The paragraph vector represents the missing information from the current context Ignore the context words in the input, but force the model to predict words randomly sampled from the paragraph in the output. Similar to Skip-gram model PV-DM PV-DBOW
- 10. Contents I. Introduction II. Preprocessing III. Analysis IV. Results
- 11. Text Mining Term Project 11 ▪ Tokenize with KoNLPy • using Twitter package ▪ Pos-tagging • only get noun, verb, and adjective ▪ Remove Token which has only one character ▪ Remove Stop-words ▪ Delete questions of which token length are less than 10 ▪ Remove unnecessary words • URL, Special characters (!, ?, *, @, <. >), Emoticon(ㅋㅋ, ㅠㅠ), multispacer ▪ Stem words that dictionary cannot correct • (남주 → 남자주인공), (페북 → 페이스북), (영환 → 영화인데), (여자애 → 여자) ▪ Delete unnecessory phrase in question • 좀 옛날 영화인데 ~, 페북에서 봤는데, ~ 장면이 있었는데 기억이안나네요 ▪ Delete questions of which length are less than 30 Preprocessing We did preprocessing for better performance and it is processed by 2 steps: whole text data and tokenized data. Raw Preprocessing Tokenizing
- 12. Text Mining Term Project 12 Select Movies and Split dataset There are 5,900 movies in dataset, but many movies has few questions. So we remove certain movies that have questions below cutoff value. Then we split the dataset with 8:2 ratio to test the model. Movie Count 스파이더위크가의 비밀 259 캐빈 인 더 우즈 222 비밀의 숲 테라비시아 179 Cutoff 무서운 영화 2 1 전우 1 전우치 1 The number of question per movies Movie Train Test 스파이더위크가의 비밀 207 52 케빈 인 더 우즈 177 45 비밀의 숲 테라비시아 143 36 레모니 스니켓의 위험한 대결 142 36 플립 141 36 … … … Split Train and Test *Basic cutoff = 3 *Using stratified method
- 13. Contents I. Introduction II. Preprocessing III. Analysis IV. Results
- 14. Text Mining Term Project 14 Modeling – Word2vec To train word2vec model, we put the answers (label) between the tokenized words in the question. Using this corpus, we trained word2vec model. *put labels in every 5 words Q: question, A: answer(label), W: word The number of labels in train, test data : 2021 Train data set : 22620 ,Test data set: 5655 skip-gram is employed Dimensionality of the feature vectors - 300 Window size - 10 Hierarchical softmax used
- 15. Text Mining Term Project 15 Modeling – Word2vec
- 16. Text Mining Term Project 16 Modeling – Word2vec Each word in the test set is embedded into the model to obtain a word vector. Combine all the vectors into one vector on a question-by-question basis (Document vector)
- 17. Text Mining Term Project 17 Modeling – Word2vec Also embedding the unique answers (label) into the model to obtain label vector. After that, Calculate pairwise cosine similarity between the label vectors (𝑽 𝑨𝒏 ) and the document vectors (𝐕′ 𝒌) K: the number test set data n: the number of unique labels
- 18. Text Mining Term Project 18 Modeling – Word2vec Finally, Normalize the cosine similarity result for each document vectors, binarize the answers(label), and evaluate the performance of the model Test set Ex) 𝑉′ 𝑘 = 0.05,0.001, 0.003 … . 0.002 𝐴 𝑘 = [1,0,0 …. 0]
- 19. Text Mining Term Project 19 Modeling – Doc2vec In the Doc2vec model, we do not need to put the correct answer like in the word2vec model, because the answer (label) is also learned as a tag. PV distributed memory is employed. Dimensionality of the feature vectors. - 300 Window size - 3 hierarchical softmax used use the sum of the context word vectors. The paragraph vectors(label vectors) are asked to a prediction task about the next word in the sentence. Every paragraph is mapped to a unique vector. The paragraph vector and word vectors are averaged to predict the next word in a context.
- 20. Text Mining Term Project 20 Modeling – Doc2vec *Computes cosine similarity between a simple mean of the projection weight vectors of the given docs.
- 21. Contents I. Introduction II. Preprocessing III. Analysis IV. Results
- 22. Text Mining Term Project 22 Model evaluation – ROC curve The ROC curve results for each labels are evaluated by two methods: micro-averaging and macro-averaging. micro-averaging - considering each element of the label indicator matrix as a binary prediction macro-averaging - gives equal weight to the classification of each label. word2vec doc2vec AUC 0.78, 0.82 AUC 0.97, 0.97 (Micro, macro)(Micro, macro)
- 23. Text Mining Term Project 23 Model evaluation – Top-n Accuracy approach Top-n accuracy approach results for each labels are evaluated. For example, Top-5 accuracy means that any of our model 5 highest probability answers must match the expected answer. (n: 1~10) word2vec doc2vec Accuracy* : top 1 and top 10 accuracy Accuracy* 0.08 to 0.20 Accuracy* 0.49 to 0.73
- 24. Text Mining Term Project 24 Discussion • Conclusion ✓ Overall, doc2vec shows better performance than word2vec model ✓ Building a service by presenting n (at least 5) correct answer lists for new questions ✓ Application to speech recognition based movie recommendation service • Further study ✓ Problems that questions about untrained movies - complementing through learning synopsis of the movies ✓ A method for dealing with imbalanced movie data is needed
- 25. Text Mining Term Project 25 Thank you
- 26. Text Mining Term Project 26 APPENDIX We drew graphs to find which movies and which genres are highly asked. We could find that people wanted to find mysterious and thrilling movies 259 222 179 178 177 166 151 147 143 131 스파이더위크가의 비밀 캐빈 인 더 우즈 비밀의 숲 테라비시아 레모니 스니켓의 위험한… 플립 트루먼 쇼 다이버전트 스플라이스 아바타 업사이드 다운 7447 5001 4585 2560 784 706 501 공포, 스릴러 SF, 판타지 로맨스, 멜로 액션, 무협 코미디 드라마 애니메이션 Asked Movie Ranking Asked Movie Genre
- 27. Text Mining Term Project 27 APPENDIX Delete unnecessory phrase in question At the beginning of the question and at the end, remove all phrases before and after the word. If the word in the check words list (within 20% of the length of the question)