Modern text mining – understanding a million comments in 60 minutes
0 likes288 views
This document describes how Datanizing uses natural language processing and machine learning techniques to derive data-driven insights from large amounts of user-generated content. They automatically gather and clean relevant content, rank it by relevance, and calculate key insights. This includes detecting topics, creating data-driven personas, understanding semantic context, and predicting changing interests in real-time to help customers better align marketing messages. The process involves text analysis, word embeddings, topic modeling, and classification techniques.
Modern text mining – understanding a million comments in 60 minutes
- 2. How to derive data-driven insights … from user-generated content https://datanizing.com/dts
- 3. Automatically gather relevant content 1 Cleaning & Linguistics 2 Relevanceranking of gathered content 3 Data-driven calculationof main insights 4 Text analysis - automatically,regularly & based on large amount of data. Visualizationin dashboard& reports 5
- 4. What is a trending hotel and what about bars? Are there unmet needs I could address? What do people care about in NYC? Products for new target groups Marketing campaigns in new places Data-driven category management …
- 6. OurBig Datamarketresearch approachexplained on a common situation: Visiting foreign places.
- 7. Theproblem: Too little time, too much going on.
- 8. Opinion of 1 author
- 10. 14 years of NYC TripAdvisor Forum Jan 2005 – Apr 2019 Making sense of …
- 11. Much content! 1.6 mio posts 89.815 users Nobody can read that.
- 13. Solution?
- 14. Automatically gather relevant content 1 Cleaning & Linguistics 2 Relevanceranking of gathered content 3 Data-driven calculationof main insights 4 Displayin dashboard& reports 5 Text processing pipeline Language detection Synonyms Outlier detection Featureextraction Clustering Wordcombinations Categories Wordfrequencies Uniqueness Domainsimilarity Inverteddomain frequency
- 15. Perform quality assurance with the whole content Getoverviewof texts,data qualityand recognizepossible biasin data
- 16. Typical questions answered with statistics Do frequent authorswrite shorter posts? How does the number of articles change over time? How does the article length change over time? What is the length distributionof articles? Which are the most frequent words? How are keywords used over time?
- 17. Python (pandas) Content spidering Data extraction SQL database Jupyter Notebook Spidering, Database, Pandas, Jupyter
- 18. Question Answering Translation / Dialogue Summarization / Topic Mining Classification / Retrieval strong weak "ShallowNLP" • Simple language models with many simplifications (Bag-of-Words, n-grams) • Keywords, phrases • Robust algorithms "DeepNLP" • Complex language models necessary for deep understanding • Statements spanning sentences • Fragile algorithms Text prepration with NLP
- 19. Typical questions answered with statistics How does the number of articles change over time?
- 20. Typical questions answered with statistics How does the article length change over time?
- 21. Typical questions answered with statistics Do frequent authorswrite shorter posts?
- 22. Typical questions answered with statistics Which are the most frequent words?
- 23. How would you rate the quality? How are keywords used over time?
- 24. Summary quality assurance Be surethe textmassesfuture data-drivendecisionswillbe based on has a good-enoughquality Recognizepossible biasin data Getoverviewof texts& dataquality take-away value of text statistics
- 25. Create data-driven insights from post of NYC travelers Data-drivenfocuspointsfordigital marketing, category management,productdesign, personalization
- 26. Topic 1 Topic 2 Topic 3 Topic 4 Topic 5 Topic 6 80% 20% Kim User #1 Hank User #2 Marty # ... Olivia #68,641 How to create data-driven personas
- 27. Looking for hidden / latent structure 1) Which could be candidates for topics? 2) How are they distributed in document space? Basic idea of topic modelling Topic 1 Topic 2 Topic 3 topicsdocuments ... Topic k doc 1 doc 2 doc n...
- 28. Only use word frequencies • Term frequency (TF) • Very simple, but robust • Basis for many algorithms (retrieval, classification) Disadvantages • Very simplified model of language • No syntactical or relational information kept Improvements • TF/IDF, n-grams Need to vectorize data (BoW) Documents D1: „Steffi likes London." D2: „Steffi does not like London." D3: „Steffi likes London, but not Paris." D1 1 1 1 D2 1 1 1 1 1 D3 1 1 1 1 1 1
- 29. Most ML is boring maths x11 x1n ... ... ... xm1 ... xmn m documents with n features (words) • Use a matrix representation • m x n Matrix can become very large • 1.3 million rows, 500.000 columns • Matrix is sparse: Most documents containonly a few words Matrix can be simplified • Only keep certain number of features • Only keep features which occur more than x times features documents
- 30. How Topic Modeling works Adopted from http://topicmodels.west.uni-koblenz.de/ckling/tmt/svd_ap.html Topic modelling transforms the matrix • Re-arrange features (words) and documents • Find blocks • Word in blocks constitute topics • Documents in blocks belong to topic
- 31. Topic1: Transportation jfk train airport flight car taxi bus way fly arrive Topic4: Happiness just great good time like place think really people food Topic2: Newbie thank nyc hi look help good suggestion appreciateadvice visit Topic5: Organization day tour walk museum brooklyn square island central park park plan Topic3: Accommodation hotel stay room night look area square bed price times Topic6: Discounts ticket buy game seat book purchase discount just sell website Result: 6 data-driven topics
- 32. Summary topic modeling Decisionsbacked by data - fordigitalmarketing,category management,productdesign, personalization, … Detectdistincttopicsuserstalkabout Detecthiddencustomersegmentsbasedon interest take-away business value of data-driven topics
- 33. Can Tripadivor help in predicting popular posts?
- 34. Tripadvisor Posts Training Predictions Number of replies as metric for popularity <4: unpopular >15: popular • Use the “labels” • Training with 40,000posts • Classify text of possible offers • Find out which content appeals to travelers Example: Predict popularity
- 35. Example: Predict popular posts Meet the locals at Times Square Brooklyn at night Walk the high line Visit the One World Observatory Take a boat trip on the Hudson river
- 36. Summary classification Classifytextin customerservice,findhatespeech, find news categories, separateEnglish fromGermantext,… Classificationand prediction withcategorized texts take-away: construct labels from unstructured text and use as categories
- 37. Detect what people are interested in when talking about NYC Alignyour messagesto whatpeople actuallylikeabout you. Detectchanginginterestsin real-time.
- 38. Search result for “airport” in the TripAdvisor forum
- 39. Analysis of words in text • Order not used • Relations between words neglected • Lost semantics Analyze n-grams • Order taken into account • Static relations via tuples • Abstraction to semantics missing So far in text analytics… Each word is a single entity Context decides about semantics!
- 40. Analysis of words in text • Order not used • Relations between words neglected • Lost semantics Analyze n-grams • Order taken into account • Static relations via tuples • Abstraction to semantics missing So far in text analytics… Each word is a single entity Context decides about semantics!
- 41. Aim: Find contextinformation of words CBOW model model • Predict word from context Skip-gram model • Determine contextfrom word • Slower, more precise with infrequent words Training word vectors
- 42. Word2vec similarities in detail
- 43. Search result for “airport” after training word embeddings There are three airports in New York: JFK EWR LGA John F. Kennedy NewarkLiberty International Airport LaGuardia
- 44. Brooklyn Bridge goes to Brooklyn Where does Lincoln Tunnel go to? LincolnTunnel goes to: Hoboken Queens Jersey City
- 45. Summary word embeddings • Benefitfromchangingtrends • Createsemanticallyawaresearchresults Detectchanginginterestsin real-time Detectrelevant contextof a topic take-away business value of semantic context
- 46. To wrap it up insights & business value from UGC analysis Data-driven personas Semantic context Changing interests Decisionsbacked by relevantdata for marketing, categorymanagement, productdesign, … Alignyour messages to whatpeople really likeabout you & adjust over time Insights Business value
- 47. Looking beyond UGC derive business value from other text sources Technical documentation Data-drivenapproachto derivingdiverse,un-biased insight from… Company wikis Change requests Scientific publications … Future cost-drivers Knowledge bottlenecks Emerging competing technologies Technical debts
- 48. Data-driven travel recommendation from 1.6 mio NYC TripAdvisor posts