NLG - Natural Language Generation

NLG - Natural Language Generation Enter Narratives Trivadis triCast 29.05.2018, 16:00 bis 17:00 Uhr Isabela Anciutti Stefan Bartram

@trivadis

BASEL BERN BRUGG DÜSSELDORF HAMBURG KOPENHAGEN LAUSANNE

tricast

FRANKFURT A.M. FREIBURG I.BR. GENF MÜNCHEN STUTTGART WIEN ZÜRICH

Speaker Isabela Anciutti Consultant [email protected]

Stefan Bartram Senior Partner Manager

[email protected]

2

11.06.2018

Trivadis triCast

Trivadis triCast Format Tuesday at 4 p.m.

Talk, questions and answers

Current IT topics, Trivadis point of view webcast will be recorded

Short surveys Option for questions via question menu

3

11.06.2018

Trivadis triCast

Training with quality & success guarantee

4

11.06.2018

https://www.trivadis.com/en/training Trivadis triCast https://www.trivadis.com/en/training

NLG - Natural Language Generation

5

11.06.2018

Trivadis triCast

NLG – Natural Language Generation Enter Narratives

Isabela Anciutti

BASEL BERN BRUGG DÜSSELDORF HAMBURG COPENHAGEN LAUSANNE

FRANKFURT A.M. FREIBURG I.BR. GENEVA MUNICH STUTTGART VIENNA ZURICH

7


AGENDA 1. Introduction 2. Trivadis POC

3. Narratives

8


Introduction

9


Definition NLG, Narratives or aka Data Story-Telling: Data in, Language out! “It’s like a translator that converts a computer-based representation into a natural language representation”

Computational Linguistics + Language-oriented Artificial Intelligence Expertise in: – Linguistics – Psychology – Engineering – Computer Science 10


11


Advantages

Conclusion – Speeds the data understanding, reduces manual analysis, accelerates decisionmaking and makes insights more accessible

12


NLG Package What do you need to pack for NLG? – An NLG Engine – Data Sets – Generation Grammar

For example, using Template Method: – Raw data – Definition of labels – Definition of KPIs

– Thresholds 13


Trivadis POC

14


CV as first exposure to customer Consultant’s Business Card – Our customer receives Resumes per e-mail – New applications at customers every month

Winning impression • Professional, updated, straight to business • Better first impression of company and of employee • Wins in a comparison with other candidates’ Resumes

15


Paid alternative Writing or rewriting of summary by consultants • Expected little compliance or no significant improvement • Consultants missing Resume in a given language • New consultants start from scratch • Final text still needs review

Cost of writing and/or reviewing the Resume Summaries by an agency • 450+ CVs: 1 hour/CV • Recurrent cost based on continuous hiring and career progress

16


Target format Summary made of 4-5 statements containing • Relevant experience • Unique skills/qualities • Highlights/accomplishments • Languages What a Resume’s summary should avoid • Simple repetition of Resume’s contents • List of values, like IT tools or job roles • Chronological descriptions: since then, after that, beginning in 2005, etc. • Direct references implying gender and names (these shouldn’t be relevant) 17


CV Data – Valued and categorized variables Variable course_teach_cnt exp_year_cnt industry_cnt language_cnt last_role_title last_role_years_cnt presentation_publication_cnt project_cnt role_cnt skill_cnt skill_expert_cnt skill_guru_cnt technology_cnt

18

Value 2 23 8 4 Senior Consultant 13 10 69 7 184 48 2 29

Category CAT_RSH_1 CAT_EXP_3 CAT_IND_2 CAT_LNG_3 CAT_EXP_2 CAT_RSH_3 CAT_PRO_3 CAT_ROL_2 CAT_SKL_3 CAT_SKL_3 CAT_SKL_3 CAT_TEK_3


Variable Value course_teach_cnt 4 exp_year_cnt 16 industry_cnt 8 language_cnt 3 last_role_title Principal Consultant last_role_years_cnt 3 presentation_publication_cnt 24 project_cnt 139 role_cnt 7 skill_cnt 105 skill_expert_cnt 34 skill_guru_cnt 4 technology_cnt 38

Category CAT_RSH_2 CAT_EXP_3 CAT_IND_2 CAT_LNG_3 CAT_EXP_1 CAT_RSH_3 CAT_PRO_3 CAT_ROL_2 CAT_SKL_2 CAT_SKL_3 CAT_SKL_3 CAT_TEK_3

Results Target descriptive CV: – “Accomplished retail manager with more than five years of extensive retail, sales & hospitality background. First sales consultant to reach 5,000 clients for XYZ agency with generated sales of over $500,000. Quality writing and communication skills, multilingual (Spanish/Portuguese/German) and a strong passion for the hospitality/hotel industry.”

Resulting descriptive CV: – “Dynamic Senior Consultant applying know-how and leading business for more than two decades. Determined guru with miscellaneous publications and while effectively producing a multifold of vibrant projects. Enthusiastic and communicative. Efficiently engaging no less than 29 different technologies, specially succeeding at plenty of industry sectors and an authority in many roles.” – “Communicative Principal Consultant contributing on strong enterprises for close to twenty years. With many optimal skills and specially excelling in a lot of sectors. Self-confident and ambitious. Responsibly mastering more than 38 distinct applications, handling plenty of roles and with more than 139 projects.”

19


Narratives

20


NLG Engines:

21


NLG Engines:

22


Partners

NLG Engines:

23


Examples

NLG Engines:

24


Partners

NLG Engines:

Example Tableau Narratives for Tableau™ is an extension for Google Chrome that automatically generates insightful stories about visualizations created in Tableau Server 10.0, Tableau Public, and Tableau Online.

25


NLG Engines:

Example Qlik Sense® Qlik Sense® analytics software is a business intelligence tool designed for the enterprise. Narratives for Qlik is an extension for Qlik Sense.

26


NLG Engines: Power BI

Example Microsoft Narratives for Power BI enables Power BI users to immediately gain insight from all of their data by transforming it into intelligent narratives.

27


NLG Engines: Lumira

28


Example SAP BO

By 2019, natural-language generation will be a standard feature of 90% of modern BI and analytics platforms.

29


Q&A Tel. +41 79 909 7217 [email protected]

BASEL BERN BRUGG DÜSSELDORF HAMBURG COPENHAGEN LAUSANNE

FRANKFURT A.M. FREIBURG I.BR. GENEVA MUNICH STUTTGART VIENNA ZURICH

Output Results Turnover ($ thousand) Mobile Car Cosmetics Food Travel

Jan

Feb

Mar

Apr

May

Jun

120

125

130

121

111

160

212

250

165

150

160

120

56

76

60

65

29

18

345

467

289

565

530

480

2065

0

0

0

0

4533

Turnover report for the month of June The total turnover for all divisions was $5.3 Million. There was an extreme rise (640%) in relation to the previous gain on the month before (92%). 31


Output Results Turnover in thousand $ Mobile Car Cosmetics Food Travel

Jan

Feb

Mar

Apr

May

Jun

120

125

130

121

111

160

212

250

165

150

160

120

56

76

60

65

29

18

345

467

289

565

530

480

2065

0

0

0

0

4533

Turnover report for the month of June In two divisions only, the turnover increased during the month of June compared to the previous months (Travel and Mobile). 32


Output Results Turnover in thousand $ Mobile Car Cosmetics Food Travel

Jan

Feb

Mar

Apr

May

Jun

120

125

130

121

111

160

212

250

165

150

160

120

56

76

60

65

29

18

345

467

289

565

530

480

2065

0

0

0

0

4533

Turnover report for the month of June Caveat: The Cosmetics division needs closer surveillance: in June the turnover for Cosmetics fell below the minimum turnover goal set. 33


POC targets Targets ✓

Creation of support tables: verbs, adjectives and terms

• 180 adjectives, 46 pronouns, 41 adverbs, 109 verbs ✓

Categorization of subject’s skills for the proper use of adjectives

• For each topic a category and a level from 1-3 ✓

Calculation of topic’s weight based on the individual content

• Estimation can be optimized ✓

Randomization of synonyms usage

• 71 statement templates

34

6/11/2018


Results (cont.) Resulting templates: • Adjective job title Experience. Accomplishment. Accomplishment. Adjective and adjective. Accomplishment, Accomplishment and Accomplishment.

35

6/11/2018


SQL Server

36

6/11/2018


BI Challenges (cont.) Translation of data into sentences • 23 years of experience, 69 projects, 7 different roles:  with more than 20 years of success  projects  occupying roles

Wide range of vocabulary to maintain diversity across pool ✓ Text templates using statement templates using term’s lists

37

6/11/2018


Wordsmith and Alexa

38

6/11/2018


NLG Techniques Canned Text ▪ Simplest ▪ Single and Multi Phrase ▪ Trivial to create

▪ Very inflexible

39

Template

Phrase-based

▪ Pre-defined templates

▪ Generalized templates

▪ Unique set of features

▪ Flexible alterations

▪ Recursive phrasal patterns at different levels

▪ Each possible expression alternative

▪ Multisentence ▪ For regular texts


▪ Complex

Feature-based

▪ Singlesentence

Fragen und Antworten Isabela Anciutti Consultant [email protected]

Stefan Bartram Senior Partner Manager [email protected]

40

11.06.2018

Trivadis triCast

triCast & other events Upcoming triCast events Thursdays, at 4 p.m.

Trivadis LIVE events (https://www.trivadis.com/de/events)

(http://m.trivadis.com/tricast-about)

▪

41

▪

Europäische TDWI Konferenz, 25.-27. Juni 2018 in München

▪

WX Developer Week, 25.-28. Juni 2018 in Nürnberg

▪

JAVA Forum Stuttgart 2018 - SAVE the DATE, 05.07.2018, Stuttgart

31.07.2018 - Data Analytics on top of Azure Big Data Sources

11.06.2018

Trivadis triCast

Thanks for joining! Stefan Bartram Senior Partner Manager Tel. +49 89 99 27 59 322 [email protected]

@trivadis

tricast