Into the Wild - wilth Natural Language Processing and Text Classification - Data Natives Conference 2015

Talk from Data Natives Conference 2015 about a experimental project for Natural Language Processing.

Transcript

1. …with 
 Natural Language Processing and Text Classification Data Natives 2015 19.11.2015 - Peter Grosskopf
2. Hey, I’m Peter. Developer (mostly Ruby), Founder (of Zweitag) Chief Development Officer @ HitFox Group Department „Tech & Development“ (TechDev)
3. Company Builder with 500+ employees in AdTech, FinTech and Big Data
4. Company Builder = Ideas + People
5. How do we select the best people out of more than 1000 applications every month in a consistent way? ? ? ? Machine Learning ?
6. Yeah! I found a solution Not really
7. Our Goal Add a sort-by- relevance to lower the screening costs and invite people faster
8. Let’s Go!
9. Action Steps 1. Prepare the textual data 2. Build a model to classify the data 3. Run it! 4. Display and interpret 
 the results
10. 1. Prepare Load data Kick out outlier Clean out stopwords (language detection + stemming with NLTK) Define classes for workflow states Link data
11. 2. Build a model tf-idf / bag of words !: term-frequency idf: inverse document frequency
12. Transform / Quantization from a textual shape to a numerical vector-form I am a nice little text -> v(i, am, a, nice, little, text) -> v(tf*idf, tf*idf, tf*idf, tf*idf, tf*idf, tf*idf)
13. term-frequency (tf) Count occurrences in document I am a nice little text -> v(i, am, a, nice, little, text) -> v(1*idf, 1*idf, 1*idf, 1*idf, 1*idf, 1*idf)
14. inverse document frequency (idf) Count how often a term occurs in the whole document set and invert with the logarithm d1(I play a fun game) -> v1(i, play, a, fun, game) d2(I am a nice little text) -> v2(i, am, a, nice, little, text) -> v2(1*log(2/2), 1*log(2/1), 1*log(2/2), …) -> v2(0, 0.3, 0, 0.3, 0.3, 0.3)
15. bag of words Simple approach to calculate the frequency of relevant terms Ignores contextual information better: n-grams
16. n-grams Generate new tokens by concatenating neighboured tokens example (1 and 2-grams): (nice, little, text) -> (nice, nice_little, little, little_text, text) -> From three tokens we just generated 5 tokens. example2 (1 and 2-grams): (new, york, is, a, nice, city) -> (new, new_york, york, york_is, is, is_a, a, a_nice, nice, nice_city, city)
17. vectorize the resumes build 1 to 4 n_grams with Scikit (sklearn) TdIdf-Vectorizer
18. Define runtime Train-test-split by date (80/20) Approach: Pick randomly CVs out of the test group Count how many CVs have to be screened to find all the good CVs
19. 3. run it! After the resumes are transformed to vector form, the classification gets done with a classical statistical machine learning model 
 
 (e.g. multinominal-naive-bayes, stochastic-gradient-descent- classifier, logistic-regression and random-forest)
20. 4. Results Generated with a combination of stochastic-gradient-descent- classifier and logistic-regression with the python machine-learning library scikit-learn AUC: 73.0615 %
21. Wrap Up 1. Prepare 2. Build Model 3. Run 4. Interpret import data vectorize the CVs with 1 to 4 n_grams choose Machine Learning model visualize results clean data define train-test- split run it! Area under curve (AUC)
22. Conclusion After trying many different approaches (doc2vec, Recurrent Neuronal Networks, Feature Hashing)- bag of words still the best Explana<on: CV documents do not contain too many semantics
23. Outlook Build a better database Experiment with new approaches and tune models Build a continuous learning model
24. Happy End. Thanks :-)