Upgrade to Pro — share decks privately, control downloads, hide ads and more …

NLP A to Z: Using NLP to Enhance Broker Product...

Sponsored · SiteGround - Reliable hosting with speed, security, and support you can count on.
Avatar for B Johnson B Johnson
September 25, 2019

NLP A to Z: Using NLP to Enhance Broker Productivity in Deal IQ

Presented at the Digital & Technology Unity Conference in Dallas, TX

Avatar for B Johnson

B Johnson

September 25, 2019
Tweet

Other Decks in Programming

Transcript

  1. NLP A to Z Using NLP to Enhance Broker Productivity

    in Deal IQ Britt Johnson, CBRE Build
  2. Schedule AI Deploy Working Model to the ShipIt API Frontend

    Development Future Work & Features 5 min Collect, Organize, Analyze Collect & Organize User Data: pawprint & Event Horizon Analyze: EDA (Problem & Goal) 15 min Learn (ML) Clean Data for NLP Label Training Data Train Embeddings Proof of Concept (BoW & Logistic Regression) Find & Train the Most Performant RNN 15 min
  3. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA
  4. Analyze Learn AI Organize Allocate Tasks Lease Expiration Deal Stage

    Client Communications Cold Calls & Voicemails Follow Up with Client Client Information ToDo List Analyze Collect
  5. Analyze Learn AI Organize Allocate Tasks Lease Expiration Deal Stage

    Client Communications Cold Calls & Voicemails Follow Up with Client Client Information ToDo List Analyze Collect
  6. Analyze Learn AI Collect Organize "called for Maxwell. call him

    Max. He was quick and said he is already represented. Try to learn more about him on the next call." • "Left Voicemail. Not in office." • "LM with assistant for Charles. Own/occupy building next to 5-Star." • "Followed up with jacqueline...next is to drop by a timeline to help kick start a conversation" • "Tried to call today and could not connect. Need to do further research on correct number and contact." • *Identifying information has been changed for privacy & confidentiality reasons
  7. Organize Analyze Learn AI Signal? Clean & Label Presence Presence

    → Notes EDA Collect Collect Organize Collect
  8. Analyze Learn AI Collect Organize Manually Label the Training Dataset

    0: Cold Outreach (left vm, cold call, etc.) 1: Too Early (contact later) 2: Call to Action (set up a meeting, etc.) 3: Contact or Company Info (contact details, company description, etc.) 4: Deal, Lease, Building Info (requirements for a future lease, info about an existing lease) 5: Hot Communication (back and forth with a client, expect a response) 6: Upcoming Meeting (details about meeting)
  9. Analyze Learn AI Collect Organize Manually label the training dataset

    0: Cold Outreach (left vm, cold call, etc.) 1: Too Early (contact later) 2: Call to Action (set up a meeting, etc.) 3: Contact or Company Info (contact details, company description, etc.) 4: Deal, Lease, Building Info (requirements for a future lease, info about an existing lease) 5: Hot Communication (back and forth with a client, expect a response) 6: Upcoming Meeting (details about meeting)
  10. Analyze Learn AI Collect Organize Manually label the training dataset

    0: Cold Outreach (left vm, cold call, etc.) 1: Too Early (contact later) 2: Call to Action (set up a meeting, etc.) 3: Contact or Company Info (contact details, company description, etc.) 4: Deal, Lease, Building Info (requirements for a future lease, info about an existing lease) 5: Hot Communication (back and forth with a client, expect a response) 6: Upcoming Meeting (details about meeting)
  11. Analyze Learn AI Collect Organize Text Preprocessing "Good afternoon, Jack,

    As you can see from the e-mail trail below, the larger space at Harper really makes the annual rent a bit out-of-reach. Our district manager has requested that we look into the Westlake space, which is also new construction. With a potential March 2019 delivery date, we’ll be forced to get a short extension at the current location. Please get the particulars and submit an LOI for this space. With 65’ depth, we would have to have a 30’ width. Thank you, April From: John Smith Sent: Tuesday, October 27, 2019 4:13 PM To: April Smith <[email protected]> Cc: Jessie Brown <[email protected]> Subject: RE: Relocation, Westlake, WA 5673 April, Our next selection is the Makati Pkwy which is yet to begin construction. Can we get a plan that shows where we could be (spaces in our 2,700 square foot footprint) in that development and what else business will be in that development? Thanks, BC From: April Smith Sent: Monday, May 26, 2019 4:23 PM To: Bill Stevens <[email protected]>; Mark Williams <[email protected]>; Subject: FW: Relocation, WA 5673 Hi Everyone, Per the e-mail below, the landlord isn’t able to accommodate our request for a 2,700 sq. ft. space. As you know, we learned that the depth of the building is only 50’ so that would require a 25’ width. The landlord has only 1600 (too small) or two 1500 sq. ft. bays; therefore, we would have to take 2900 sq. ft. With the asking rent of $28/sq. ft., we would have a larger Base rent, plus triple net. This does not include our buildout costs/furniture, etc. Before we proceed, I wanted to ensure that we are comfortable with 2900 sq. ft. Thanks, April" • *Identifying information has been changed for privacy & confidentiality reasons
  12. Analyze Learn AI Collect Organize Text Preprocessing "NO SPOC contact

    - sdf,ds,fdsl;sdd md;s,d;sa,d'ad,a'sd, mda;ls,d,;sal,dsal;d,as alsdmas;ld,sa;l,sdl;mvsdl; a;sldmas;ldmsa;dlmsa;dlsad ;ald;as,dsa;ld,sa;dl,sad a;sldmma;sld,,sa;dlmas;ldmas d;asld,s;ld,s;dl,f;dslms;lfmsdl;fmd" • *Identifying information has been changed for privacy & confidentiality reasons
  13. Collect Organize Analyze MVP Learn AI Signal? Clean & Label

    Presence Presence → Notes EDA Collect Learn AI
  14. Analyze Learn AI Collect Organize Lease = [1, 0, 0,

    0, 0, 0, 0, …, 0] Call = [0, 1, 0, 0, 0, 0, 0, …, 0] Follow = [0, 0, 1, 0, 0, 0, 0, …, 0] Lease Call
  15. Analyze Learn AI Collect Organize Model Selection RandomizedSearchCV, GridSearchCV Vectorizer

    (Bag of Words) CountVectorizer, TfidfVectorizer Decomposer (Dimensionality Reduction) TruncatedSVD, Latent Dirichlet Allocation Classifier Logistic Regression (L1 & L2) 86.6% accuracy
  16. Collect Organize Analyze MVP Learn AI Signal? Clean & Label

    Presence Presence → Notes EDA Collect Learn AI
  17. Collect Organize Analyze MVP Learn AI Signal? Clean & Label

    Presence Presence → Notes EDA Collect AI
  18. Collect Organize Analyze MVP Learn AI Signal? Clean & Label

    Presence Presence → Notes EDA Collect AI
  19. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA AI Collect
  20. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA AI Collect
  21. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA AI Collect
  22. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA AI Collect
  23. Collect Organize Analyze MVP Learn AI RNN Signal? Clean &

    Label Presence Presence → Notes EDA AI Collect
  24. Future Improvements Integrating real broker labeled data Attention More cleaning

    Automate deployment & retraining Train Own Embeddings
  25. Taking What We Learned Use broker text fields to inform

    a hot prospective leads score Automatically update deal details from email communications Match the best broker to suit a client’s needs via the client’s activity (embeddings) Recommend relevant news and research articles to brokers Recommend similar listings from personalized searches (embeddings) Create broker embeddings to match orphan deals to the best broker
  26. Analyze Learn AI Collect Organize Manually label the training dataset

    0: Cold Outreach (left vm, cold call, etc.) 1: Too Early (contact later) 2: Call to Action (set up a meeting, etc.) 3: Contact or Company Info (contact details, company description, etc.) 4: Deal, Lease, Building Info (requirements for a future lease, info about an existing lease) 5: Hot Communication (back and forth with a client, expect a response) 6: Upcoming Meeting (details about meeting)