Ontos NLP Stack, Sep. 2016
2 likes857 views
The document discusses a project on streaming-based text mining using deep learning and semantic technologies, focusing on its use cases, market situation, and implementation details. Key areas include content augmentation in e-commerce, crime investigation assistance, and entity recognition from various data sources. The report also highlights lessons learned, including the importance of deep learning skills and the integration of semantic web technologies for data processing.
Ontos NLP Stack, Sep. 2016
- 1. WWW.LEDS-PROJEKT.DE STREAMING-BASED TEXT MINING USING DEEP LEARNING AND SEMANTICS MARTIN VOIGT / ONTOS
- 2. AGENDA Use Cases Lessons Learned WILDHORNMINER Market Situation What is required from project partners and customers? What is good, what not and what are the next steps? What are the others doing? How looks Ontos’ approach of a flexible text mining by using Deep Learning? How to analyze texts from various sources and interlink with existing knowledge graphs? 12. September 20162
- 3. USE CASES What is required from project partners and customers? 12. September 20163
- 4. USE CASES FROM LEDS • Content Augmentation in E-Commerce • Product descriptions imported from various sources as usually unstructured text • Much manual work is resource consuming and expensive http://www.walmart.com/ip/The-Revenant-Blu-ray-Digital-HD/50129277 12. September 20164
- 5. USE CASES FROM CUSTOMERS • Brand or competitor monitoring • What are the people or journalists writing about my brand? • What are my competitors doing? How is the market changing? Monitoring of web sites news feeds social media channels … Link with reports CRM Open data … 12. September 20166
- 6. USE CASES FROM CUSTOMERS • Assist criminal investigations in the (dark) web • More and more organized crime in the web • Detect entities with their relation and additional facts But Neumann and McKelvey eventually sold the business to their landlord Joshua Guttman. Neumann McKelvey ? Joshua Guttman Business sold_what sold_to a 12. September 20167
- 7. MAIN REQUIREMENTS • Detection and classification of entities, relations and facts with good F1 • Long, high quality texts vs. short texts with bad / missing grammar in multiple languages • Training not by linguists but domain experts • Flexible adaption to new domains and contexts • Many different data sources 12. September 20168
- 8. MARKET SITUATION What are the others doing? 12. September 20169
- 9. MARKET • MarketsAndMarkets: NLP (all) for 2020 • Market: $ 13,4 bn • CAGR: ~ 18,4% • ResearchMOZ: Text Analytics 2015-2019 • CAGR: ~ 16,1% • Transparency Market Research: Text Analytics 2016–2024 • CAGR: ~ 17,6% • Driving element: NLP-as-a-Service with CAGR of ~20,2% http://www.marketsandmarkets.com/PressReleases/natural-language-processing-nlp.asp http://www.researchmoz.us/global-natural-language-processing-market-2015-2019-report.html http://www.transparencymarketresearch.com/pressrelease/global-text-analytics-market.htm 12. September 201610
- 10. MARKET • ResearchAndMarkets: 2016 Top 10 Information and Communication Technologies • 7. Natural Language Processing (NLP) • Increasing Demand for Automation Drives Growth • Funding on the Rise - Start-ups Driving Key Innovations for NLP Applications • High Acceptance of Technology enables the US to Remain on the Top • Gartner‘s Top 10 Strategic Technology Trends für 2016 • 4 - Information of Everything: • Access to heterogeneous data sources • (Semantic) Linking of data items • 5 - Advanced Machine Learning: • Deep Learning and Neural Network for NLP http://www.researchandmarkets.com/research/vd8fr9/2016_top_10 12. September 201611
- 11. GETTING AN OVERVIEW • Understood domain by reviewing 30+ tools and frameworks • Checked scientific prototypes as well as open source tools • Investigated what other enterprise, companies, and startups are doing • Defined a set of 15+ criteria to compare and understand 12. September 201613
- 12. GETTING AN OVERVIEW • Understood domain by reviewing 30+ tools and frameworks • Findings • Startups and huge enterprises usually stick to Deep Learning (DL) high flexibility • Some existing companies loss market share, e.g., Attensity • Market saturation in U.S., only less companies in Europe (5) • More and more NLP-as-a-Service • Only a few use RDF data, e.g., for output or disambiguation (8) • Only 8 tools could extract relations and facts • Only 9 open source tools with available benchmarks 12. September 201614
- 13. GETTING AN OVERVIEW • Rough comparison of the main concepts • Conclusions for Ontos • Use DL as foundation for text mining tasks • Combine it with semantic technologies, e.g. for disambiguation • External view: NLP-as-a-Service • Internal view: NLP pipelines in order to combine different tools / technologies based on http://www.deeplearningbook.org/contents/intro.html Fig 1.5 12. September 201616
- 14. MINER How looks Ontos’ approach of a flexible text mining by using Deep Learning? 12. September 201617
- 15. MINER - OVERVIEW Language model generated from Corpora Model as input for supervised step Supervised Model Domain or task specific model for sequence labeling Trained on specific texts by domain experts Large collection of texts in a given language or “dialect”, e.g. Social Media / Twitter Large text corpora Unsupervised Model 12. September 201618
- 16. MINER – CORPUS CREATION • 1 corpus per language / dialect required • the larger and more heterogeneous the better get and model more contexts for words • Sources: Common Crawl, Wikipedia, news feeds, domain specific texts, Twitter, … 12. September 201619
- 17. MINER – UNSUPERVISED MODEL • language model that could be reused in multiple domains / contexts • reuse the concept of word2vec to create word embeddings Language Dictionary Word2Vec Model Word2Vec Dictionary Numeric Corpus 12. September 201620
- 18. MINER – SUPERVISED MODEL • Sequence labeling task • use a bi-directional LSTM (BLSTM) • Gated Recurrent Unit (GRU) as specialized LSTM • Output layer • join and normalize • classification LEDS is a research project Word Embeddings GRU GRU GRU GRU GRU GRU GRU GRU GRU GRU Forward GRU Backward GRU OUT OUT OUT OUT OUTOutput Layers 0 0 Output Textual Input 12. September 201621
- 19. MINER – IMPLEMENTATION • Torch framework • matured, efficient GPU support, good packages for neural networks, … • Scripting with Lua language • Service implementation with Go because of great integration with Lua • Integrated in Ontos Eiger workbench 12. September 201622
- 20. • Create new “projects” ad-hoc • Reuse pre-labeled corpora • Free definition of entity types as required for domain MINER – IMPLEMENTATION 12. September 201623
- 21. • List and manage documents of project corpora MINER – IMPLEMENTATION 12. September 201624
- 22. • Double-check and annotate new labels MINER – IMPLEMENTATION 12. September 201625
- 23. MINER – IMPLEMENTATION • English News • ~ 400 English texts from CoNLL 2003 and news feeds • 5 entity types defined and manually annotated by 2 experts • 1st value: correct start of entity • 2nd value: correct end of entity Class: Person Organization Product Location Event Per class F1:....... 0.976 | 0.990 0.932 | 0.962 0.873 | 0.942 0.958 | 0.976 0.891 | 0.928 Per class Recall:... 0.974 | 0.988 0.936 | 0.953 0.845 | 0.926 0.965 | 0.980 0.870 | 0.913 Per class Precision: 0.978 | 0.992 0.928 | 0.971 0.902 | 0.959 0.951 | 0.973 0.914 | 0.944 http://www.cnts.ua.ac.be/conll2003/ner/ 12. September 201626
- 24. MINER – IMPLEMENTATION • English Twitter • ~ 2400 tweets from Twitter NLP tools / A. Ritter 2011 • 5 entity types defined, some are combined from original source Class Person Organisation Place Product Thing Per class F1:....... 0.797 | 0.883 0.548 | 0.141 0.725 | 0.669 0.195 | 0.207 0.409 | 0.551 Per class Recall:... 0.769 | 0.852 0.487 | 0.085 0.696 | 0.647 0.120 | 0.129 0.337 | 0.529 Per class Precision: 0.828 | 0.918 0.627 | 0.414 0.756 | 0.693 0.516 | 0.514 0.520 | 0.574 https://github.com/aritter/twitter_nlp 12. September 201627
- 25. WILDHORN How to analyze texts from various sources and interlink with existing knowledge graphs? 12. September 201628
- 26. NLP PIPELINES • Problem: How to efficiently connect various tools in the data stream? • Sources • RSS feed reader, crawler, Twitter, FTP servers, … • Analytics • 1 MINER instance per language • Disambiguate and link to Knowledge graphs or taxonomies • Sinks • RDF stores, Apache Solr, HDFS, Apache Cassandra, … http://www.computernewsme.com/wp-content/uploads/2011/06/Cable-clutter.jpg12. September 201629
- 27. NLP PIPELINES • Best practices – by LinkedIn • http://www.confluent.io/blog/stream-data-platform-1/ • Apache Kafka! 12. September 201630
- 28. WILDHORN NLP PIPELINE Message Broker Apache Spark Apache Kafka / Confluent Schema Registry REST APITechnical Foundation: Twitter Extractor RSS Feed Crawler REST Proxy Log File Extractor Web / Darknet Crawler Taxo Extractor MINER AGDISTIS Named Entity Extraction & Disambiguation ENS RDF Serializer QUAD NoSQL Serializer Apache Cassandra Solr Serializer Apache Solr 12. September 201632 Document Classification Sentiment Analysis
- 29. WILDHORN NLP PIPELINE • Implementation • Apache Spark 1.6 & Confluent 3.0 • Use of Spark Jobserver to manage Apache Spark applications • Ontos Eiger backend ready • First tests • Simple Mesos cluster with 3 servers • ~ 1000 message / sec with 1 broker without MINER analytics: no data problem • Problem: 1 MINER instance currently scales up to 100 message / sec Kafka Pipeline Part Producers Consumers Topic Schema Part Producers Consumers Topic Schema ...https://github.com/spark-jobserver/spark-jobserver 12. September 201633
- 30. WILDHORN NLP PIPELINE https://github.com/spark-jobserver/spark-jobserver 12. September 201634
- 31. LESSONS LEARNED What is good, what not and what are the next steps? 12. September 201635
- 32. CURRENT STATUS • Detection and classification of entities, relations and facts with good F1 • Long, high quality texts vs. short texts with bad / missing grammar in multiple languages • Training not by linguists but domain experts • Flexible adaption to new domains and contexts • Many different data sources 12. September 201636
- 33. LESSONS LEARNED • Neural networks / deep learning provide great concepts & frameworks for flexible, high quality NLP tasks • Apache Kafka / Confluent Platform in combination with Apache Spark good foundation for data streaming and processing • Disambiguation and linking of entities to taxonomies and Knowledge Graphs via Semantic Web technologies is a core contribution for data integration • Hard to find employees with Deep Learning skills 12. September 201637
- 34. NEXT STEPS • Try letter-trigram word hashing to overcome out of dictionary problem of word2vec algorithm • Relation and fact extraction in MINER Adam Neumann is the CEO of super-hot office rental company WeWork, the most valuable startup in New York City. PersonFirst Name Sentiment Entity Type Adam super-hot Entity Adam Neumann Legend Relation Relation Type firstName1 firstName WhoHow positive1 positiveSentiment WeWork Organization ThingWhatHow a has A1 A2 A3 A4 Annotation https://arxiv.org/abs/1608.06757 12. September 201638
- 35. NEXT STEPS • Define NLP pipelines in frontend • Make MINER scalable • Usable search interface • Benchmark with GERBIL 12. September 201639
- 36. Q & A Dr. Martin Voigt Managing Director Ontos GmbH D-04319 Leipzig M: +49 178 40 222 58 E: martin.voigt@ontos.com Twitter: m_a_r_t_i_n 12. September 201640