Breaking Down NLP for SEOs - SMX Advanced Europe 2019 - Paul Shapiro
•Download as PPTX, PDF•
6 likes•2,108 views
Interested in learning about Natural Language Processing (NLP)? Are you using NLP for your SEO already and want to step it up a level? Join this session to get a crash course in NLP. From stemming and lemmatization to word embeddings and its applications for SEO. Paul Shapiro will break down NLP to explain how NLP technology uses machine learning to decipher and analyze our human languages in a way that is highly valuable for marketers and SEOs. Paul will also share specific examples using the Python programming language along the way so you can either start using NLP right away for SEO or find new and more effective ways to use NLP.
Recommended
More Related Content
What's hot
What's hot (20)
Similar to Breaking Down NLP for SEOs - SMX Advanced Europe 2019 - Paul Shapiro
Similar to Breaking Down NLP for SEOs - SMX Advanced Europe 2019 - Paul Shapiro (20)
More from Paul Shapiro
More from Paul Shapiro (8)
Recently uploaded
Recently uploaded (20)
Breaking Down NLP for SEOs - SMX Advanced Europe 2019 - Paul Shapiro
- 1. @fighto BREAKING DOWN NLP FOR SEO @fighto
- 2. Paul Shapiro Senior Partner, Head of SEO @ Catalyst, a GroupM Agency
- 3. @fighto Assumptions & Prerequisites • Familiarity with Python • Familiarity with common data science libraries such as pandas and NumPy • Familiarity with Jupyter Notebooks (optional) • But no prior knowledge of NLP
- 4. @fighto Libraries Used in Examples
- 5. @fighto KNIME as an Alternative https://www.knime.com
- 6. @fighto What is Natural Language Processing (NLP)?
- 7. @fighto “NLP is a way for computers to analyze, understand, and derive meaning from human language in a smart and useful way. By utilizing NLP, developers can organize and structure knowledge to perform tasks such as automatic summarization, translation, named entity recognition, relationship extraction, sentiment analysis, speech recognition, and topic segmentation.” https://blog.algorithmia.com/introduction-natural-language-processing-nlp/
- 9. @fighto Parse Semi/Unstructured Text Data https://github.com/niderhoff/nlp-datasets Input:
- 10. @fighto Example Data Sources • (Digital) Books • CSVs, Excel, JSON, XML, etc. • Word Docs/PDFs • Web Pages (most relevant to SEO)
- 11. @fighto
- 12. @fighto Text Pre-Processing • Tokenization • Text must be broken into units aka tokens • (Usually individual words)
- 13. @fighto Text Pre-Processing We need to parse, clean, and prepare text data for both analysis and conversion into a machine interpretable formats.
- 15. @fighto Text Pre-Processing • Noise and Junk Removal/Cleanup • Punctuation and Special Characters • Stop Words • Common Abbreviations • Common Character Cases • Etc.
- 16. @fighto Lowercase + Remove Punctuation
- 17. @fighto Tokenize & Remove Stop Words
- 19. @fighto Text Pre-Processing • Normalization and Standardization • Stemming • Lemmatization
- 20. @fighto Why Normalization, Text Analytics Ex • Speeds up machine learning analysis • Disambiguation • Say there are 500 jokes in our corpus that mention “Donald Trump” • 25 of those jokes include the word “economy, 15 include the word “economic” and 10 mention “world economies”. • All of these jokes have to do with both “economics” and “Donald Trump” but would turn up as 3 distinct co-occurences.
- 21. @fighto Why Stemming and Pitfalls • More basic method of reducing different forms of the same word to a common base • Stemming chops off the end of the word to accomplish this • Faster method • Results in terms that are not real words:
- 22. @fighto Stemming
- 23. @fighto Why Lemmatization and Pitfalls • More sophisticated method of reducing different forms of the same word to a common base • Lemmatizations leverages vocabulary and grammar to infer the root of a word • Requires Parts of Speech tagging • Slower but more accurate method
- 25. @fighto Information Extraction & Grouping • Getting more context • N-Grams • Parts of Speech Tagging • Chunking/Chinking • Named Entity Recognition • Word Embeddings
- 26. @fighto N-Grams
- 27. @fighto Parts of Speech Tagging
- 28. @fighto
- 30. @fighto Word Embeddings: word2vec, GloVe
- 31. @fighto Word Embeddings: word2vec, GloVe
- 32. @fighto Statistical Feature Creation • Leverage personal heuristics to create customized numeric representations that you think could be used by a machine learning model to make predictions
- 33. @fighto Example: Joke Lines & Length
- 35. @fighto Example: Number of Profane Words
- 36. @fighto Feature Normalization • Box-Cox Power Transformations • “A Box Cox transformation is a way to transform non- normal dependent variables into a normal shape. Normality is an important assumption for many statistical techniques; if your data isn’t normal, applying a Box-Cox means that you are able to run a broader number of tests.” https://www.statisticshowto.datasciencecentral.com/box-cox-transformation/
- 38. @fighto Check Distribution with Histogram
- 39. @fighto Check Distribution with Histogram
- 40. @fighto Box-Cox Transformation & Apply
- 41. @fighto Vectorization • Count Vectorizer • N-Gram Vectorizer • TF-IDF Vectorizer
- 42. @fighto Count Vectorizer – Cleaning Function
- 45. @fighto Let’s Talk About TF-IDF for a Moment • Count Vectorizer looked at how many times a term or n-gram appeared in a joke and represents as positive integer • TF-IDF would create a score that considers how many time a term appears in a joke as well as how many times it appears in the entire corpus of jokes. • Rarer words are deemed to more important because they can be used distinguish one joke from another. • Higher TF-IDF value = more uncommon • Lower TF-IDF value = less common
- 47. @fighto Decision Trees Will [Sports Team] win? Players statistics are favorable? Is the team their playing historically better? Yes No? Yes No
- 48. @fighto Random Forest Will [Sports Team] win? Players statistics are favorable? Is the team their playing historically better? Yes No? Yes No Will [Sports Team] win? Players statistics are favorable? Is the team their playing historically better? Yes No? Yes No
- 52. @fighto Having Done This Better • Reduce overfitting • Standardize features (mixing sparse and non-sparse data) • Word embeddings for more context • More sophisticated models
- 53. @fighto More Applications for SEO • Creating performant content (joke example extrapolated) • Predicting natural link earning potential • Natural language generation, writing bits of content • Semantic content optimization • Site architecture design and taxonomy • User flow creation • Keyword research • Etc.
- 54. @fighto How to Learn More, Resources • https://web.stanford.edu/~jurafsky/slp3/ • https://www.kaggle.com/learn/overview • https://towardsdatascience.com • https://github.com/keon/awesome-nlp
- 55. CatalystDigital.com | info@CatalystDigital.com Organic Search | Paid Search | Paid Social Advertising Ecommerce Marketing & Amazon Advertising Thank You!
Editor's Notes
- Let’s take a trip back to SEO grade school…
- We’re switching to spaCy for this because NLTK is slower and produces worse results
- Skip chunking