natural language generation
play

Natural Language Generation . .. . . .. .. . .. . . . .. - PowerPoint PPT Presentation

Feb 09, 2024 •228 likes •1.71k views

. .. . . .. . . .. . . .. . . .. . . . . Ondej Duek Ondej Duek 1/ 40 May 11 th , 2016 Charles University in Prague Faculty of Mathematics and Physics Institute of Formal and Applied Linguistics (mostly) for Spoken

  1. . .. .. . .. . . .. . . .. . . .. . . . . Standard NLG Pipeline ( Textbook ) Ondřej Dušek 4/ 40 [Text] [Sentence plan(s)] [Content plan] [Inputs] Textbook NLG Pipeline . Introduction to NLG . .. . . .. .. . . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . ↓ Content/text planning (“what to say”) • Content selection, basic ordering ↓ Sentence planning/microplanning (“middle ground”) • aggregation, lexical choice, referring… ↓ Surface realization (“how to say it”) • linearization, conforming to rules of the target language

  2. • Content selection according to communication goal • Basic structuring (ordering) . . . .. . . .. .. . .. . . .. . . . . . . .. . . .. . Introduction to NLG Textbook NLG Pipeline Standard NLG Pipeline ( Textbook ) Inputs weather report numbers etc.) Content planning 5/ 40 Ondřej Dušek .. .. . . . . .. . . .. . . .. . . .. . . .. .. . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Communication goal (e.g. “inform user about search results”) • Knowledge base (e.g. list of matching entries in database, • User model (constraints, e.g. user wants short answers) • Dialogue history (referring expressions, repetition)

  3. . . .. . . .. . . .. . . .. . . .. . .. . Textbook NLG Pipeline Ondřej Dušek 5/ 40 Content planning weather report numbers etc.) Inputs Standard NLG Pipeline ( Textbook ) Introduction to NLG . . .. . . .. . . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . . .. .. . . .. . . .. Natural Language Generation • Communication goal (e.g. “inform user about search results”) • Knowledge base (e.g. list of matching entries in database, • User model (constraints, e.g. user wants short answers) • Dialogue history (referring expressions, repetition) • Content selection according to communication goal • Basic structuring (ordering)

  4. • Creating linear text from (typically) structured input • Ensuring grammatical correctness . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. . Introduction to NLG Textbook NLG Pipeline Standard NLG Pipeline ( Textbook ) Sentence planning (micro-planning) Surface realization 6/ 40 Ondřej Dušek .. .. . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Word and syntax selection (e.g. choose templates) • Dividing content into sentences • Aggregation (merging simple sentences) • Lexicalization • Referring expressions

  5. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . Introduction to NLG Textbook NLG Pipeline Standard NLG Pipeline ( Textbook ) Sentence planning (micro-planning) Surface realization 6/ 40 Ondřej Dušek . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . .. . . .. . . .. . . .. Natural Language Generation • Word and syntax selection (e.g. choose templates) • Dividing content into sentences • Aggregation (merging simple sentences) • Lexicalization • Referring expressions • Creating linear text from (typically) structured input • Ensuring grammatical correctness

  6. • Templates, grammars, rules, statistics, or a mix thereof • Varied, custom-tailored, non-compatible . .. . . .. . . . . . .. . . .. . .. .. . .. . . .. . Introduction to NLG Real NLG Systems Real NLG Systems Few systems implement the whole pipeline realization Approaches Data representations 7/ 40 Ondřej Dušek .. . . . . . .. . . .. . . .. . . .. . . . .. . .. .. . . .. . . .. . . .. . . .. . . Natural Language Generation • Systems focused on content planning with trivial surface • Surface-realization-only, word-order-only systems • One-step (holistic) approaches • SDS: content planning done by dialogue manager → only sentence planning and realization here

  7. • Varied, custom-tailored, non-compatible . . . .. . . .. .. . .. . . .. . .. . . . Real NLG Systems Ondřej Dušek 7/ 40 Data representations Approaches realization Few systems implement the whole pipeline Real NLG Systems . Introduction to NLG . .. . . .. .. . . . .. . .. . . .. . . .. . . .. . . .. . . . . . .. . . .. .. . .. . . .. . . Natural Language Generation • Systems focused on content planning with trivial surface • Surface-realization-only, word-order-only systems • One-step (holistic) approaches • SDS: content planning done by dialogue manager → only sentence planning and realization here • Templates, grammars, rules, statistics, or a mix thereof

  8. . .. .. . .. . . .. . . .. . . .. . . . . Real NLG Systems Ondřej Dušek 7/ 40 Data representations Approaches realization Few systems implement the whole pipeline Real NLG Systems . Introduction to NLG . .. . . .. .. . . . . .. .. . . .. . . .. . . .. . . .. . . . . . .. . .. .. . . .. . . .. . . Natural Language Generation • Systems focused on content planning with trivial surface • Surface-realization-only, word-order-only systems • One-step (holistic) approaches • SDS: content planning done by dialogue manager → only sentence planning and realization here • Templates, grammars, rules, statistics, or a mix thereof • Varied, custom-tailored, non-compatible

  9. • Surface realization can be rule based • you can hardcode the grammar, it is more straightforward to fix • Surface realizer is (relatively) easy to implement • and you can use a third-party one • Problem of all pipelines: error propagation • the more steps, the more chance to screw it up • Need to provide training sentence plans (statistical planners) • sometimes you may use existing analysis tools .. . . .. . . . . . . .. . . .. .. . .. . . .. . Introduction to NLG Two-step or One-step architecture? Two-step or one-step? Why go two-step Why go one-step 8/ 40 Ondřej Dušek .. . . .. .. . . .. . . . .. . .. . . .. . . . . .. .. . . .. . . .. . . . . . .. . . .. . Natural Language Generation • Dividing makes the tasks simpler • no need to worry about morphology in sentence planning

  10. • Surface realizer is (relatively) easy to implement • and you can use a third-party one • Problem of all pipelines: error propagation • the more steps, the more chance to screw it up • Need to provide training sentence plans (statistical planners) • sometimes you may use existing analysis tools . . .. . . .. . . .. .. . . .. . . . . .. . . .. . Introduction to NLG Two-step or One-step architecture? Two-step or one-step? Why go two-step Why go one-step 8/ 40 Ondřej Dušek .. . .. .. .. . . .. . . . . . .. . . .. . . . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . • Dividing makes the tasks simpler • no need to worry about morphology in sentence planning • Surface realization can be rule based • you can hardcode the grammar, it is more straightforward to fix

  11. • Problem of all pipelines: error propagation • the more steps, the more chance to screw it up • Need to provide training sentence plans (statistical planners) • sometimes you may use existing analysis tools . . . .. . . .. . . .. .. . .. .. . .. . . .. . . .. . Introduction to NLG Two-step or One-step architecture? Two-step or one-step? Why go two-step Why go one-step 8/ 40 Ondřej Dušek . . . . . . .. . . .. . . .. . . .. . . .. . . .. .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Dividing makes the tasks simpler • no need to worry about morphology in sentence planning • Surface realization can be rule based • you can hardcode the grammar, it is more straightforward to fix • Surface realizer is (relatively) easy to implement • and you can use a third-party one

  12. • Need to provide training sentence plans (statistical planners) • sometimes you may use existing analysis tools . . .. . . .. . . .. . . .. .. . . . .. . . .. . . .. . Introduction to NLG Two-step or One-step architecture? Two-step or one-step? Why go two-step Why go one-step 8/ 40 Ondřej Dušek .. . . . . .. .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Dividing makes the tasks simpler • no need to worry about morphology in sentence planning • Surface realization can be rule based • you can hardcode the grammar, it is more straightforward to fix • Surface realizer is (relatively) easy to implement • and you can use a third-party one • Problem of all pipelines: error propagation • the more steps, the more chance to screw it up

  13. . . .. . .. .. . . .. . . .. . . .. . .. . . . .. . . .. . Introduction to NLG Two-step or One-step architecture? Two-step or one-step? Why go two-step Why go one-step 8/ 40 Ondřej Dušek . . .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . .. Natural Language Generation . .. . . .. .. . . • Dividing makes the tasks simpler • no need to worry about morphology in sentence planning • Surface realization can be rule based • you can hardcode the grammar, it is more straightforward to fix • Surface realizer is (relatively) easy to implement • and you can use a third-party one • Problem of all pipelines: error propagation • the more steps, the more chance to screw it up • Need to provide training sentence plans (statistical planners) • sometimes you may use existing analysis tools

  14. • Each stage: . . . .. . . .. .. . .. . . .. . . . . . 1. Sentence planning Ondřej Dušek 9/ 40 2. Current state-of-the art / our works 1. History 3. One-step approaches to NLG 2. Surface realization NLG systems examples .. NLG Systems Examples . .. . . .. .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. . .. .. . . .. . . Natural Language Generation • Divided by NLG stage:

  15. • Each stage: . . . .. . . .. .. . .. . . .. . . . . . 1. Sentence planning Ondřej Dušek 9/ 40 2. Current state-of-the art / our works 1. History 3. One-step approaches to NLG 2. Surface realization NLG systems examples .. NLG Systems Examples . .. . . .. .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. . .. .. . . .. . . Natural Language Generation • Divided by NLG stage:

  16. • Each stage: . . . .. . . .. .. . .. . . .. . . . . . 1. Sentence planning Ondřej Dušek 9/ 40 2. Current state-of-the art / our works 1. History 3. One-step approaches to NLG 2. Surface realization NLG systems examples .. NLG Systems Examples . .. . . .. .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. . .. .. . . .. . . Natural Language Generation • Divided by NLG stage:

  17. • Each stage: . . . .. . . .. .. . .. . . .. . . . . . 1. Sentence planning Ondřej Dušek 9/ 40 2. Current state-of-the art / our works 1. History 3. One-step approaches to NLG 2. Surface realization NLG systems examples .. NLG Systems Examples . .. . . .. .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. . .. .. . . .. . . Natural Language Generation • Divided by NLG stage:

  18. . .. . . .. . . .. . . .. . . .. . . . .. 1. Sentence planning Ondřej Dušek 9/ 40 2. Current state-of-the art / our works 1. History 3. One-step approaches to NLG 2. Surface realization NLG systems examples . NLG Systems Examples . .. . . .. .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. Natural Language Generation .. . .. . . .. . . • Divided by NLG stage: • Each stage:

  19. • Actually typically handcrafued or non-existent • One-step approaches or simplistic systems • Here we focus on trainable approaches • …and especially on our own ☺ . . . . .. . . .. . . .. . . .. . . .. . . .. . . .. . NLG Systems Examples Sentence planning Sentence Planning Examples 10/ 40 Ondřej Dušek .. . .. . . . .. . . .. . .. .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. . . Natural Language Generation • Various input/output formats, not very comparable

  20. • One-step approaches or simplistic systems • Here we focus on trainable approaches • …and especially on our own ☺ . .. . . .. . . . . . .. . . .. .. .. . .. . . .. . . .. . NLG Systems Examples Sentence planning Sentence Planning Examples 10/ 40 Ondřej Dušek . . .. . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. . . .. . . .. Natural Language Generation • Various input/output formats, not very comparable • Actually typically handcrafued or non-existent

  21. . .. . .. . . .. . . .. . . .. . . . .. . .. . . .. . . .. . NLG Systems Examples Sentence planning Sentence Planning Examples 10/ 40 Ondřej Dušek .. . . .. . . .. . . .. . . .. . . .. . . . . . .. . . .. . . .. . . .. . . .. Natural Language Generation • Various input/output formats, not very comparable • Actually typically handcrafued or non-existent • One-step approaches or simplistic systems • Here we focus on trainable approaches • …and especially on our own ☺

  22. . .. .. . . .. . . . . . .. . . .. . . .. .. domain Ondřej Dušek 11/ 40 plans hand-annotated sentence (RankBoost) trained on overgeneration the flight information . Trainable Sentence Planning: SPoT Sentence planning NLG Systems Examples . .. . . .. . . . . .. . . .. . . .. . . .. . . .. . . .. .. . . .. . . .. . . .. . . .. . . Natural Language Generation • Spoken Dialogue System in • Handcrafued generator + • Statistical reranker

  23. • Paiva&Evans : linguistic features annotated in corpus generated with • PERSONAGE-PE : personality traits connected to linguistic features via . . . .. . . .. .. . .. . . .. . . . . . . .. . . .. . NLG Systems Examples Sentence planning Trainable Sentence Planning: Parameter Optimization Examples many parameter settings, correlation analysis machine learning 12/ 40 Ondřej Dušek .. .. . . . . .. . . .. . . .. . . .. . . .. .. . .. . . .. . . .. . Natural Language Generation . .. . . .. . . • Requires a flexible handcrafed planner • No overgeneration • Adjusting its parameters “somehow”

  24. . . .. . . .. . . .. . . .. . . .. . .. . Sentence planning Ondřej Dušek 12/ 40 machine learning many parameter settings, correlation analysis Examples Trainable Sentence Planning: Parameter Optimization NLG Systems Examples . . .. . . .. . . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . .. . Natural Language Generation . .. . . .. . . .. • Requires a flexible handcrafed planner • No overgeneration • Adjusting its parameters “somehow” • Paiva&Evans : linguistic features annotated in corpus generated with • PERSONAGE-PE : personality traits connected to linguistic features via

  25. • Typical NLG training: • Our sentence planner learns alignments jointly • training from pairs: MR + sentence . . .. . . .. . . .. . . .. . . . . .. . .. .. . . .. . NLG Systems Examples Our Approach to Sentence Planning 2. Learns from unaligned data a) requires alignment of MR elements and words/phrases b) uses a separate alignment step 13/ 40 Ondřej Dušek .. . . . . . .. . . .. . .. .. . . .. . . .. . . . . . .. . . .. . Natural Language Generation .. .. . . .. . . Our A ∗ /Perceptron Sentence Planner ( TGEN1 ) 1. Requires no handcrafued module

  26. • Our sentence planner learns alignments jointly • training from pairs: MR + sentence . . .. . . .. . . .. . . .. .. . . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning 2. Learns from unaligned data a) requires alignment of MR elements and words/phrases b) uses a separate alignment step 13/ 40 Ondřej Dušek .. . . . . .. .. . . .. . . .. . . .. . . .. . . . . . .. . . .. .. Natural Language Generation . .. . . .. . . Our A ∗ /Perceptron Sentence Planner ( TGEN1 ) 1. Requires no handcrafued module • Typical NLG training: MR inform(name=X, type=placetoeat, eattype=restaurant, area=riverside, food=Italian) alignment X is an italian restaurant in the riverside area . text

  27. . . .. .. . .. . . .. . . .. . . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning 2. Learns from unaligned data a) requires alignment of MR elements and words/phrases b) uses a separate alignment step 13/ 40 Ondřej Dušek . . .. .. . . .. . . .. . . .. . . . . . .. . . . .. . . . . .. .. . .. . . Natural Language Generation .. . Our A ∗ /Perceptron Sentence Planner ( TGEN1 ) 1. Requires no handcrafued module • Typical NLG training: • Our sentence planner learns alignments jointly • training from pairs: MR + sentence MR inform(name=X, type=placetoeat, eattype=restaurant, area=riverside, food=Italian) X is an italian restaurant in the riverside area . text

  28. • Output : deep-syntax dependency trees • based on TectoMT 's t-layer, but very • two attributes per tree node: • using surface word order • Conversion to plain text sentences – • Treex / TectoMT English synthesis . .. . . .. . . . .. . . .. . . .. . Our Approach to Sentence Planning .. 14/ 40 X is an Italian restaurant area=riverside, food=Italian) eattype=restaurant, inform(name=X, type=placetoeat, Natural Language Generation Ondřej Dušek (rule-based, later) . surface realization t-lemma + formeme simplified I/O formats . NLG Systems Examples .. . .. .. . .. . . .. . . . . . .. . . .. . . . .. . . . .. . . .. . . .. . .. . . . .. . . .. in the riverside area. • Input : a MR • dialogue acts: “inform” + slot-value pairs • other formats possible

  29. • Conversion to plain text sentences – • Treex / TectoMT English synthesis . . . .. . . .. .. . . .. . . .. . . .. .. . 14/ 40 X is an Italian restaurant area=riverside, food=Italian) eattype=restaurant, inform(name=X, type=placetoeat, Natural Language Generation Ondřej Dušek (rule-based, later) . surface realization t-lemma + formeme simplified I/O formats Our Approach to Sentence Planning NLG Systems Examples . .. . . .. .. . .. .. . . . . . .. . . .. . . . . .. . . . .. . . . . . .. . .. in the riverside area. . . . .. .. • Input : a MR • dialogue acts: “inform” + slot-value pairs t-tree • other formats possible be • Output : deep-syntax dependency trees v:fin • based on TectoMT 's t-layer, but very X-name restaurant n:subj n:obj • two attributes per tree node: italian area adj:attr n:in+X • using surface word order riverside n:attr

  30. . .. . . .. . . .. . . .. . . .. . . .. .. . . 14/ 40 X is an Italian restaurant area=riverside, food=Italian) eattype=restaurant, inform(name=X, type=placetoeat, Natural Language Generation Ondřej Dušek (rule-based, later) .. surface realization t-lemma + formeme simplified I/O formats Our Approach to Sentence Planning NLG Systems Examples . . . .. .. . . .. . . . . . . .. . . .. . . .. in the riverside area. . . . .. . . .. .. . .. . . .. . . . .. . • Input : a MR • dialogue acts: “inform” + slot-value pairs t-tree • other formats possible be • Output : deep-syntax dependency trees v:fin • based on TectoMT 's t-layer, but very X-name restaurant n:subj n:obj • two attributes per tree node: italian area adj:attr n:in+X • using surface word order riverside n:attr • Conversion to plain text sentences – • Treex / TectoMT English synthesis

  31. • Using two subcomponents: • candidate generator • churning out candidate sentence plan trees • given an incomplete candidate tree, add node(s) • scorer /ranker for the candidates • influences which candidate trees will be . . .. . . .. . . .. .. . . .. . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Overall Structure of Our Sentence Planner candidate sentence plan expanded (selects the most promising) 15/ 40 Ondřej Dušek .. . .. .. .. . . .. . . . . . .. . . .. . . . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . • A*-style search – “finding the path” empty tree → full sentence plan tree • always expand the most promising • stop when candidates don't improve for a while

  32. • scorer /ranker for the candidates • influences which candidate trees will be . . . .. . . .. .. . .. . . .. . . . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Overall Structure of Our Sentence Planner candidate sentence plan expanded (selects the most promising) 15/ 40 Ondřej Dušek Natural Language Generation .. .. . . . . .. . . .. . . .. . . .. . . .. .. . . . . .. . .. . .. . .. . . .. . . MR • A*-style search – “finding the path” empty tree → full sentence plan tree Sentence planner • always expand the most promising candidate generator • stop when candidates don't improve for a while A* search • Using two subcomponents: • candidate generator scorer • churning out candidate sentence plan trees • given an incomplete candidate tree, add node(s) Sentence plan (deep syntax tree)

  33. . . .. . . .. . . .. . . .. . . .. . .. . Our Approach to Sentence Planning Natural Language Generation Ondřej Dušek 15/ 40 expanded (selects the most promising) candidate sentence plan Overall Structure of Our Sentence Planner NLG Systems Examples . . .. . . .. . . .. .. .. . . .. . . .. . . .. . . .. . . . . . .. . . . .. . . .. . .. . . .. . MR • A*-style search – “finding the path” empty tree → full sentence plan tree Sentence planner • always expand the most promising candidate generator • stop when candidates don't improve for a while A* search • Using two subcomponents: • candidate generator scorer • churning out candidate sentence plan trees • given an incomplete candidate tree, add node(s) • scorer /ranker for the candidates • influences which candidate trees will be Sentence plan (deep syntax tree)

  34. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . . .. . . .. . . .. . . .. . . .. .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek . . .. . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. Natural Language Generation .. . .. . . .. . . • Given a candidate plan tree, generate its successors

  35. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . . . .. . . .. . . .. .. .. .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . . . . . . . . . .. . . .. . . .. . . .. .. . . .. . . .. . . .. .. . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree t-tree be recommend serve v:fin v:fin v:fin

  36. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . . . .. . . .. . . .. .. .. .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . . . . . . . . . .. . . .. . . .. . . .. .. . . .. . . .. . . .. .. . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree t-tree be recommend serve v:fin v:fin v:fin

  37. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . .. . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . . . . . . .. . . . . . .. . . .. . . .. .. . . .. .. . . . . .. .. . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree t-tree be be be v:fin v:fin v:fin be v:fin restaurant X-name restaurant n:obj n:subj n:subj

  38. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . .. . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . . . . . . .. . . . . . .. . . .. . . .. .. . . .. .. . . . . .. .. . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree t-tree be be be v:fin v:fin v:fin be v:fin restaurant X-name restaurant n:obj n:subj n:subj

  39. • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . . . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . .. . . . . .. . . . . . .. . . .. . . .. .. . .. .. .. . . .. . . . . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree be be be v:fin v:fin v:fin X-name restaurant X-name bar X-name n:subj n:obj n:subj n:obj n:subj

  40. • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size … . . . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . .. . . . . .. . . .. . . .. . . .. . . .. . . .. .. .. . . .. . . . . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree be be be v:fin v:fin v:fin X-name restaurant X-name bar X-name n:subj n:obj n:subj n:obj n:subj • “possible places” must be limited in practice

  41. • parent–child • t-lemma + formeme • number of children, tree size … . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation .. .. . . . . .. . . .. .. . .. . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. • Given a candidate plan tree, generate its successors t-tree t-tree t-tree be be be v:fin v:fin v:fin X-name restaurant X-name bar X-name n:subj n:obj n:subj n:obj n:subj • “possible places” must be limited in practice • using combination of things seen in training data

  42. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Candidate generator by adding 1 node (at every possible place) 16/ 40 Ondřej Dušek Natural Language Generation . .. .. . . . . . .. . . .. . . .. . . .. . .. .. . . .. . . .. . . .. . . .. . . • Given a candidate plan tree, generate its successors t-tree t-tree t-tree be be be v:fin v:fin v:fin X-name restaurant X-name bar X-name n:subj n:obj n:subj n:obj n:subj • “possible places” must be limited in practice • using combination of things seen in training data • parent–child • t-lemma + formeme • number of children, tree size …

  43. • occurence of input DA slots + t-lemmas / formemes • tree shape • tree edges (parent-child) • … . . . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek . .. . . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. Natural Language Generation .. . .. . . .. . . • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m

  44. • occurence of input DA slots + t-lemmas / formemes • tree shape • tree edges (parent-child) • … . . . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek . .. . . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. Natural Language Generation .. . .. . . .. . . • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m

  45. • occurence of input DA slots + t-lemmas / formemes • tree shape • tree edges (parent-child) • … . . . .. . . .. . . .. .. . .. . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek . .. . . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. Natural Language Generation .. . .. . . .. . . • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m

  46. • tree shape • tree edges (parent-child) • … . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek .. .. . . . . .. . . .. . . .. . . .. . . .. .. . .. . . .. . . .. . Natural Language Generation . .. . . .. . . • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m • occurence of input DA slots + t-lemmas / formemes

  47. • tree edges (parent-child) • … . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek .. .. . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m • occurence of input DA slots + t-lemmas / formemes • tree shape

  48. • … . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . .. . Natural Language Generation . .. . . .. . . .. • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m • occurence of input DA slots + t-lemmas / formemes • tree shape • tree edges (parent-child)

  49. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Scorer How to describe the fitness? Features: 17/ 40 Ondřej Dušek . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . .. Natural Language Generation . .. . . .. . . .. • a function: sentence plan tree t , MR m → real-valued score • describes the fitness of t for m • occurence of input DA slots + t-lemmas / formemes • tree shape • tree edges (parent-child) • …

  50. • Training: • given m , generate the best tree t top with current weights • update weights if t top t gold (gold-standard) • Update: w feat t gold m feat t top m • Updates based on partial trees • Estimating future value of the trees .. . . .. . . . . .. . . .. . . NLG Systems Examples .. . . .. . .. Our Approach to Sentence Planning Perceptron scorer Basic form w Our improvements Trying to guide the search on incomplete trees 18/ 40 Ondřej Dušek . .. . .. . .. . . .. . . . . . .. . . .. . . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation • score = w ⊤ · feat ( t , m )

  51. • Update: w feat t gold m feat t top m • Updates based on partial trees • Estimating future value of the trees . .. . . .. . . . . . .. . . .. .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Perceptron scorer Basic form w Our improvements Trying to guide the search on incomplete trees 18/ 40 Ondřej Dušek .. . .. .. .. . . .. . . . . . .. . . .. . . . . .. . . .. . . .. . . .. . . .. . . .. . Natural Language Generation • score = w ⊤ · feat ( t , m ) • Training: • given m , generate the best tree t top with current weights • update weights if t top ̸ = t gold (gold-standard)

  52. • Updates based on partial trees • Estimating future value of the trees . . . .. . . .. .. . .. . . .. . .. . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Perceptron scorer Basic form Our improvements Trying to guide the search on incomplete trees 18/ 40 Ondřej Dušek .. . . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • score = w ⊤ · feat ( t , m ) • Training: • given m , generate the best tree t top with current weights • update weights if t top ̸ = t gold (gold-standard) • Update: w = w + α · ( feat ( t gold , m ) − feat ( t top , m ))

  53. • Updates based on partial trees • Estimating future value of the trees . . . .. . . .. .. . .. . . .. . .. . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Perceptron scorer Basic form Our improvements Trying to guide the search on incomplete trees 18/ 40 Ondřej Dušek .. . . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • score = w ⊤ · feat ( t , m ) • Training: • given m , generate the best tree t top with current weights • update weights if t top ̸ = t gold (gold-standard) • Update: w = w + α · ( feat ( t gold , m ) − feat ( t top , m ))

  54. • Estimating future value of the trees . . . .. . . .. .. . .. . . .. . .. . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Perceptron scorer Basic form Our improvements Trying to guide the search on incomplete trees 18/ 40 Ondřej Dušek .. . . . .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • score = w ⊤ · feat ( t , m ) • Training: • given m , generate the best tree t top with current weights • update weights if t top ̸ = t gold (gold-standard) • Update: w = w + α · ( feat ( t gold , m ) − feat ( t top , m )) • Updates based on partial trees

  55. . . .. .. . .. . . .. . . .. . . .. . .. . Our Approach to Sentence Planning Ondřej Dušek 18/ 40 Trying to guide the search on incomplete trees Our improvements Basic form Perceptron scorer NLG Systems Examples . . .. . . .. . . . .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . .. Natural Language Generation . .. . . .. . . .. • score = w ⊤ · feat ( t , m ) • Training: • given m , generate the best tree t top with current weights • update weights if t top ̸ = t gold (gold-standard) • Update: w = w + α · ( feat ( t gold , m ) − feat ( t top , m )) • Updates based on partial trees • Estimating future value of the trees

  56. • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well .. . . .. . . . . . .. . . . .. .. . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek .. . . .. .. . . .. . . . .. . .. . . .. . . . . .. .. . . .. . . .. . . . . . .. . . .. . Natural Language Generation • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs

  57. • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well . . .. . . .. . . .. . .. . .. . . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek .. . .. .. .. . . .. . . . . . .. . . .. . . . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89%

  58. • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well . .. . . .. . . . . . .. . . .. .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results but: 19/ 40 Ondřej Dušek . .. .. .. . . . .. . . . . . .. . . .. . . . .. .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU…

  59. • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well . . . .. . . .. . . .. .. . .. .. . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek . . . . . . .. . . .. . . .. . . .. . . .. . . .. .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include

  60. • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well . . .. . . .. . . .. . . .. .. . . . .. . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek .. . . . . .. .. . . .. . . .. . . .. . . .. . . .. . . .. . . .. Natural Language Generation . . .. . . .. . . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful

  61. • slooooow, doesn't scale very well . . .. . . .. . . .. . . .. .. . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek . . .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . .. Natural Language Generation . .. . . .. .. . . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output

  62. . . .. . .. .. . . .. . . .. . . .. . .. . . . .. . . .. . NLG Systems Examples Our Approach to Sentence Planning Evaluation of Our Sentence Planner Data Results 19/ 40 Ondřej Dušek . . .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . . .. Natural Language Generation . .. . . .. .. . . • Restaurant recommendations from the BAGEL generator • restaurant location, food type, etc. • just 404 utterances for 202 DAs • basic setup 54.24% BLEU, best version 59.89% • less than BAGEL 's ~ 67% BLEU… but: • we do not use alignments • our generator decides itself what to include • outputs mostly fluent and meaningful • problems: • repeated/missing/irrelevant information on the output • slooooow, doesn't scale very well

  63. . NLG Systems Examples X is a restaurant in the riverside area near X. Generated X restaurant is near X on the riverside. Reference eattype=restaurant) inform(name=X-name, type=placetoeat, area=riverside, near=X-near, Input DA Example Outputs Our Approach to Sentence Planning . inform(name=X-name, type=placetoeat, area=X-area, .. . . .. . . .. . . Input DA pricerange=moderate, eattype=restaurant) . Input DA Ondřej Dušek 20/ 40 X is a Japanese restaurant in the centre of town near X and X. Generated X is a Chinese takeaway and Japanese restaurant in the city centre near X. Reference food=Japanese) area=citycentre, near=X-near, food=“Chinese takeaway”, inform(name=X-name, type=placetoeat, eattype=restaurant, X is a French restaurant in the riverside area which serves French food. Reference Generated X is a French restaurant on the riverside. Reference area=riverside, food=French) inform(name=X-name, type=placetoeat, eattype=restaurant, Input DA X is a restaurant in the X area. Generated X is a moderately priced restaurant in X. .. . .. . . . .. . . .. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . . .. . .. .. . . .. . . .. . . Natural Language Generation

  64. . NLG Systems Examples X is a restaurant in the riverside area near X. Generated X restaurant is near X on the riverside. Reference eattype=restaurant) inform(name=X-name, type=placetoeat, area=riverside, near=X-near, Input DA Example Outputs Our Approach to Sentence Planning . inform(name=X-name, type=placetoeat, area=X-area, .. . . .. . . .. . . Input DA pricerange=moderate, eattype=restaurant) . Input DA Ondřej Dušek 20/ 40 X is a Japanese restaurant in the centre of town near X and X. Generated X is a Chinese takeaway and Japanese restaurant in the city centre near X. Reference food=Japanese) area=citycentre, near=X-near, food=“Chinese takeaway”, inform(name=X-name, type=placetoeat, eattype=restaurant, X is a French restaurant in the riverside area which serves French food. Reference Generated X is a French restaurant on the riverside. Reference area=riverside, food=French) inform(name=X-name, type=placetoeat, eattype=restaurant, Input DA X is a restaurant in the X area. Generated X is a moderately priced restaurant in X. .. . .. . . . .. . . .. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . . .. . .. .. . . .. . . .. . . Natural Language Generation

  65. . NLG Systems Examples X is a restaurant in the riverside area near X. Generated X restaurant is near X on the riverside. Reference eattype=restaurant) inform(name=X-name, type=placetoeat, area=riverside, near=X-near, Input DA Example Outputs Our Approach to Sentence Planning . inform(name=X-name, type=placetoeat, area=X-area, .. . . .. . . .. . . Input DA pricerange=moderate, eattype=restaurant) . Input DA Ondřej Dušek 20/ 40 X is a Japanese restaurant in the centre of town near X and X. Generated X is a Chinese takeaway and Japanese restaurant in the city centre near X. Reference food=Japanese) area=citycentre, near=X-near, food=“Chinese takeaway”, inform(name=X-name, type=placetoeat, eattype=restaurant, X is a French restaurant in the riverside area which serves French food. Reference Generated X is a French restaurant on the riverside. Reference area=riverside, food=French) inform(name=X-name, type=placetoeat, eattype=restaurant, Input DA X is a restaurant in the X area. Generated X is a moderately priced restaurant in X. .. . .. . . . .. . . .. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . . .. . .. .. . . .. . . .. . . Natural Language Generation

  66. . NLG Systems Examples X is a restaurant in the riverside area near X. Generated X restaurant is near X on the riverside. Reference eattype=restaurant) inform(name=X-name, type=placetoeat, area=riverside, near=X-near, Input DA Example Outputs Our Approach to Sentence Planning . inform(name=X-name, type=placetoeat, area=X-area, .. . . .. . . .. . . Input DA pricerange=moderate, eattype=restaurant) . Input DA Ondřej Dušek 20/ 40 X is a Japanese restaurant in the centre of town near X and X. Generated X is a Chinese takeaway and Japanese restaurant in the city centre near X. Reference food=Japanese) area=citycentre, near=X-near, food=“Chinese takeaway”, inform(name=X-name, type=placetoeat, eattype=restaurant, X is a French restaurant in the riverside area which serves French food. Reference Generated X is a French restaurant on the riverside. Reference area=riverside, food=French) inform(name=X-name, type=placetoeat, eattype=restaurant, Input DA X is a restaurant in the X area. Generated X is a moderately priced restaurant in X. .. . .. . . . .. . . .. . . .. . . .. . . .. . . .. . . .. . .. . . . .. . . .. . . .. . .. .. . . .. . . .. . . Natural Language Generation

  67. . . . .. . . .. . . .. . . .. . . .. . .. .. . . .. . . .. . NLG Systems Examples Surface Realization Surface Realization Examples Treex / TectoMT realizer 21/ 40 Ondřej Dušek .. . . .. . . .. . . .. . . .. . . .. . . . . . .. . . .. . . .. . . .. . . .. Natural Language Generation • Also various input formats, at least output is always text • From handcrafued to different trainable realizers • Also including our own (developed here at ÚFAL): • actually handcrafued for the most part

  68. • General purpose • Functional Unification . . . .. . . .. . .. Surface Realization . . .. . . .. . . NLG Systems Examples KPML Grammar-based Realizers (90's): KPML , FUF/SURGE "It is raining cats and dogs." Ondřej Dušek 22/ 40 ) (C / OBJECT :LEX CATS-AND-DOGS :NUMBER MASS)) :TENSE PRESENT-CONTINUOUS :ACTEE (A / AMBIENT-PROCESS :LEX RAIN :LOGICALFORM :TARGETFORM .. EX-SET-1 :NAME (EXAMPLE Grammar FUF/SURGE Grammar multilingual .. . . . . . .. . . .. . .. .. . . .. . . .. . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. Natural Language Generation • General purpose, • Systemic Functional

  69. . . .. . . .. . . .. . NLG Systems Examples .. . . .. . . .. . Surface Realization . "It is raining cats and dogs." Ondřej Dušek 22/ 40 ) (C / OBJECT :LEX CATS-AND-DOGS :NUMBER MASS)) :TENSE PRESENT-CONTINUOUS :ACTEE (A / AMBIENT-PROCESS :LEX RAIN :LOGICALFORM :TARGETFORM Grammar-based Realizers (90's): KPML , FUF/SURGE EX-SET-1 :NAME (EXAMPLE Grammar FUF/SURGE Grammar multilingual KPML .. . .. .. . .. . . .. . . . .. . .. . . .. . . . . . . . .. . . .. . . .. . .. . . . .. . . .. Natural Language Generation • General purpose, • Systemic Functional • General purpose • Functional Unification

  70. . . .. . . .. . . .. . . .. . . .. . .. . Surface Realization Ondřej Dušek 23/ 40 enhancements Grammar multi-lingual Grammar-based Realizer: OpenCCG NLG Systems Examples . . .. . . .. . .. . .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . .. . . .. . . .. . . .. Natural Language Generation • General purpose, • Combinatory Categorial • Used in several projects • With statistical

  71. . . . . .. . . .. . .. . . . .. . . .. . .. NLG Systems Examples .. p.setVerb("chase"); Ondřej Dušek 24/ 40 >>> Mary chased the monkey. System.out.println(output); String output = realiser.realiseSentence(p); p.setFeature(Feature.TENSE, Tense.PAST); p.setObject("the monkey"); p.setSubject("Mary"); Surface Realization SPhraseSpec p = nlgFactory.createClause(); Realiser realiser = new Realiser(lexicon); NLGFactory nlgFactory = new NLGFactory(lexicon); Lexicon lexicon = new XMLLexicon("my-lexicon.xml"); (procedural) other languages Procedural Realizer: SimpleNLG .. . . .. . .. . . .. . . . .. . .. . . .. . . . . . . .. . . .. . . .. . .. . . . .. . . .. Natural Language Generation • General purpose • English, adapted to several • Java implementation

  72. • Ranking according to: • n -gram models ( NITROGEN, HALOGEN ) • Tree models (XTAG grammar – FERGUS ) • Predicted Text-to-Speech quality ( Nakatsu and White ) • Personality traits (extraversion, agreeableness… – CRAG ) • Provides variance, but at a greater computational cost . .. . . .. . . .. . . . .. . . .. .. . . .. . . .. . NLG Systems Examples Surface Realization Trainable Realizers: Overgenerate and Rank + alignment (repeating words uttered by dialogue counterpart) 25/ 40 Ondřej Dušek .. . . . . . .. . . .. . . .. . . .. . . . .. . . .. . . .. . . .. . Natural Language Generation .. .. . . .. . . • Require a handcrafued realizer, e.g. CCG realizer • Input underspecified → more outputs possible • Overgenerate • Then use a statistical reranker

  73. • Provides variance, but at a greater computational cost . . .. .. . .. . . .. . . .. . . .. .. . . . . .. . . .. . NLG Systems Examples Surface Realization Trainable Realizers: Overgenerate and Rank + alignment (repeating words uttered by dialogue counterpart) 25/ 40 Ondřej Dušek . . .. .. . . .. . . .. . . .. . . . . . .. . . . .. . . . .. Natural Language Generation . .. . . .. . . .. • Require a handcrafued realizer, e.g. CCG realizer • Input underspecified → more outputs possible • Overgenerate • Then use a statistical reranker • Ranking according to: • n -gram models ( NITROGEN, HALOGEN ) • Tree models (XTAG grammar – FERGUS ) • Predicted Text-to-Speech quality ( Nakatsu and White ) • Personality traits (extraversion, agreeableness… – CRAG )

  74. . . . .. .. . .. . . .. . . .. . . .. . .. .. . . .. . . .. . NLG Systems Examples Surface Realization Trainable Realizers: Overgenerate and Rank + alignment (repeating words uttered by dialogue counterpart) 25/ 40 Ondřej Dušek . . . .. . . .. . . .. . . .. . . .. . . . . . .. . . . .. . Natural Language Generation .. . . .. .. . . • Require a handcrafued realizer, e.g. CCG realizer • Input underspecified → more outputs possible • Overgenerate • Then use a statistical reranker • Ranking according to: • n -gram models ( NITROGEN, HALOGEN ) • Tree models (XTAG grammar – FERGUS ) • Predicted Text-to-Speech quality ( Nakatsu and White ) • Personality traits (extraversion, agreeableness… – CRAG ) • Provides variance, but at a greater computational cost

  75. . .. .. .. . . .. . . .. . . .. . . . .. . .. . . .. . . .. . NLG Systems Examples Surface Realization Trainable Realizers: Syntax-Based 26/ 40 Ondřej Dušek . . . .. . . .. . . .. . . .. . . .. . . . . . .. . . . . .. .. . . .. . . .. Natural Language Generation • StuMaBa : general realizer based on SVMs • Pipeline: ↓ Deep syntax/semantics ↓ surface syntax ↓ linearization ↓ morphologization

  76. • We use it for our experiments ( TGEN1 ) • analysis • Pipeline approach • Mostly simple, single-purpose, rule-based modules (blocks) • Word inflection: statistical ( Flect ) • Gradual transformation of deep trees into surface dependency • Surface trees are then simply linearized . .. . .. . . .. . . .. . . .. . . . . .. . .. .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realizer synthesis on BAGEL data = 89.79% BLEU trees 27/ 40 Ondřej Dušek . . .. .. .. . . .. . . . . . .. . . .. . . . . .. .. . .. . . .. . . . . . .. . . .. . Natural Language Generation • Domain-independent

  77. • Pipeline approach • Mostly simple, single-purpose, rule-based modules (blocks) • Word inflection: statistical ( Flect ) • Gradual transformation of deep trees into surface dependency • Surface trees are then simply linearized . .. . . .. . . .. . . . .. . . .. .. . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realizer trees 27/ 40 Ondřej Dušek . .. . . . . .. . . .. . . .. . . .. . . .. . . . .. . . .. . . .. Natural Language Generation .. . .. . . .. . . • Domain-independent • We use it for our experiments ( TGEN1 ) • analysis → synthesis on BAGEL data = 89.79% BLEU

  78. • Gradual transformation of deep trees into surface dependency • Surface trees are then simply linearized . .. . . .. . . . . . .. . . .. .. .. . . . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realizer trees 27/ 40 Ondřej Dušek . .. .. .. . . .. . . .. . . .. . . .. . . . . . . .. . . .. . Natural Language Generation . .. . . .. . . .. • Domain-independent • We use it for our experiments ( TGEN1 ) • analysis → synthesis on BAGEL data = 89.79% BLEU • Pipeline approach • Mostly simple, single-purpose, rule-based modules (blocks) • Word inflection: statistical ( Flect )

  79. . . . .. . . .. . . .. . . .. . . .. . .. .. . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realizer trees 27/ 40 Ondřej Dušek . .. . .. . . .. . . .. . . .. . . . . . .. . . .. . . . .. . Natural Language Generation .. . . .. . . .. • Domain-independent • We use it for our experiments ( TGEN1 ) • analysis → synthesis on BAGEL data = 89.79% BLEU • Pipeline approach • Mostly simple, single-purpose, rule-based modules (blocks) • Word inflection: statistical ( Flect ) • Gradual transformation of deep trees into surface dependency • Surface trees are then simply linearized

  80. • Copy the deep tree (sentence plan) • Determine morphological agreement • Add prepositions and conjunctions • Add articles • Compound verb forms (add auxiliaries) • Punctuation • Word inflection • Capitalization t-tree zone=en_gen jump v:fin cat window n:subj n:through+X . .. . . . . .. . . .. . Treex / TectoMT Surface Realization Example .. . . .. . NLG Systems Examples Our Surface realizer . 28/ 40 Ondřej Dušek Natural Language Generation .. . .. . . . .. . . .. . .. .. . . .. . . .. . . .. . . .. . . .. . . .. . . . . .. . . .. . . .. . • Realizer steps (simplified):

  81. • Determine morphological agreement • Add prepositions and conjunctions • Add articles • Compound verb forms (add auxiliaries) • Punctuation • Word inflection • Capitalization . . .. . . .. . . .. .. . .. .. . . . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realization Example 28/ 40 Ondřej Dušek Natural Language Generation . . .. .. .. . . .. . . . . . .. . . .. . . . . .. .. . .. . . .. . . . . . . .. . . .. • Realizer steps (simplified): • Copy the deep tree (sentence plan) t-tree zone=en_gen jump v:fin cat window n:subj n:through+X

  82. • Add prepositions and conjunctions • Add articles • Compound verb forms (add auxiliaries) • Punctuation • Word inflection • Capitalization . .. . . .. . . .. . . .. .. . . .. .. . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realization Example 28/ 40 Ondřej Dušek Natural Language Generation . . . . . . . . . .. . . .. . . .. . . .. .. . . .. . . .. . .. .. . . . .. . . .. . • Realizer steps (simplified): • Copy the deep tree (sentence plan) • Determine morphological agreement t-tree zone=en_gen jump v:fin cat window n:subj n:through+X

  83. • Add articles • Compound verb forms (add auxiliaries) • Punctuation • Word inflection • Capitalization . .. . . .. . . .. . . .. .. . . .. .. . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realization Example 28/ 40 Ondřej Dušek Natural Language Generation . . . . . . . . . .. . . .. . . .. . . .. .. . . .. . . .. . . .. .. . .. . . .. . . • Realizer steps (simplified): • Copy the deep tree (sentence plan) • Determine morphological agreement • Add prepositions and conjunctions t-tree zone=en_gen jump v:fin cat window n:subj n:through+X

  84. • Compound verb forms (add auxiliaries) • Punctuation • Word inflection • Capitalization . . .. . . .. . .. .. . . .. . . . . .. . . .. . . .. . NLG Systems Examples Our Surface realizer Treex / TectoMT Surface Realization Example 28/ 40 Ondřej Dušek Natural Language Generation .. . . . . . .. .. . .. . . .. . . .. . . .. . . .. . . .. . . .. . . . . .. . . .. • Realizer steps (simplified): • Copy the deep tree (sentence plan) • Determine morphological agreement • Add prepositions and conjunctions • Add articles t-tree zone=en_gen jump v:fin cat window n:subj n:through+X

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend

Outline of todays lecture Overview of Natural Language Generation Components of Natural

Outline of todays lecture Overview of Natural Language Generation Components of Natural

Natural Language Processing Outline of todays lecture Overview of Natural Language Generation Components of Natural Language Generation systems Data for NNs via classical realization Referring expressions Natural Language Processing

586 views • 37 slides
Natural Language Generation Survey in the State of the Art of Natural Topic Coverage

Natural Language Generation Survey in the State of the Art of Natural Topic Coverage

Natural Language Generation Survey in the State of the Art of Natural Topic Coverage Language Generation by Gatt and Krahmer Intro and NLG tasks ->Tianqi Wu NLG Architecture and Approaches ->Jianing Zhou Style Variation

1.02k views • 99 slides
Natural Language Processing 1 Lecture 11: Language generation and summarisation Katia Shutova

Natural Language Processing 1 Lecture 11: Language generation and summarisation Katia Shutova

Natural Language Processing 1 Natural Language Processing 1 Lecture 11: Language generation and summarisation Katia Shutova ILLC University of Amsterdam 6 December 2018 Natural Language Processing 1 Language generation Language generation

828 views • 50 slides
Let the AI do the Talk Adventures with Natural Language Generation @MarcoBonzanini PyParis 2018

Let the AI do the Talk Adventures with Natural Language Generation @MarcoBonzanini PyParis 2018

Let the AI do the Talk Adventures with Natural Language Generation @MarcoBonzanini PyParis 2018 PyData London Conference 12-14 July 2019 @PyDataLondon NATURAL LANGUAGE GENERATION Natural Language Processing 5 Natural Language Processing

1.43k views • 84 slides
STS for NLG Christian Chiarcos chiarcos@uni-potsdam.de Natural Language Generation Natural

STS for NLG Christian Chiarcos chiarcos@uni-potsdam.de Natural Language Generation Natural

STS for NLG Christian Chiarcos chiarcos@uni-potsdam.de Natural Language Generation Natural Language Generation (NLG) (...) is a subfield of artificial intelligence and computational linguistics that is concerned with building computer

557 views • 10 slides
Natural Language Processing 1 Lecture 10: Language generation and summarisation Katia Shutova

Natural Language Processing 1 Lecture 10: Language generation and summarisation Katia Shutova

Natural Language Processing 1 Natural Language Processing 1 Lecture 10: Language generation and summarisation Katia Shutova ILLC University of Amsterdam 2 December 2019 1 / 51 Natural Language Processing 1 Language generation Language

830 views • 54 slides
Natural Language Generation Andrea Zugarini SAILab December 5th, 2019 LabMeeting, December 5th

Natural Language Generation Andrea Zugarini SAILab December 5th, 2019 LabMeeting, December 5th

Natural Language Generation Andrea Zugarini SAILab December 5th, 2019 LabMeeting, December 5th - 2019 Natural Language Generation Natural Language Generation Natural Language Generation is the problem of generating text automatically.

514 views • 16 slides
Natural Language Generation AN OVERVIEW What is NL Generation? a definition, the roots, and

Natural Language Generation AN OVERVIEW What is NL Generation? a definition, the roots, and

Natural Language Generation An Overview Stephan Busemann DFKI GmbH Saarbrcken, Germany Stephan.Busemann@dfki.de Acknowledgement: Part of this presentation was inspired by Robert Dales and Ehud Reiters tutorial on Applied NL Generation at

430 views • 31 slides
Natural Language Generation and Dialog System Evaluation EE596/LING580 -- Conversational

Natural Language Generation and Dialog System Evaluation EE596/LING580 -- Conversational

Natural Language Generation and Dialog System Evaluation EE596/LING580 -- Conversational Artificial Intelligence Hao Cheng University of Washington Conv AI System Diagram 1 Natural Language Generation 2 NLG Approaches Template

782 views • 43 slides
Building Adaptable and Scalable Natural Language Generation Systems Yannis Konstas Natural

Building Adaptable and Scalable Natural Language Generation Systems Yannis Konstas Natural

Building Adaptable and Scalable Natural Language Generation Systems Yannis Konstas Natural Language Generation is everywhere (Machine Translation)

1.78k views • 174 slides
Natural Language Generation (Not Only) in Dialogue Systems Ond rej Du sek Institute of

Natural Language Generation (Not Only) in Dialogue Systems Ond rej Du sek Institute of

Overview Content Planning Surface Realization Natural Language Generation (Not Only) in Dialogue Systems Ond rej Du sek Institute of Formal and Applied Linguistics Charles University in Prague May 22, 2013 Ond rej Du sek

366 views • 24 slides
Natural Language Processing with Deep Learning CS224N/Ling284 Lecture 15: Natural Language

Natural Language Processing with Deep Learning CS224N/Ling284 Lecture 15: Natural Language

Natural Language Processing with Deep Learning CS224N/Ling284 Lecture 15: Natural Language Generation Christopher Manning Announcements Thank you for all your hard work! We know Assignment 5 was tough and a real challenge to do

1.24k views • 87 slides
Intro to Natural Language Generation Ehud Reiter (Abdn Uni and Arria/Data2text) Background read:

Intro to Natural Language Generation Ehud Reiter (Abdn Uni and Arria/Data2text) Background read:

Intro to Natural Language Generation Ehud Reiter (Abdn Uni and Arria/Data2text) Background read: Reiter and Dale, Building Natural Language Generation Systems Ehud Reiter, Computing Science, University of Aberdeen 1 What is NLG? NLG

873 views • 56 slides
Roadmapping for Natural Language Generation Robert Dale rdale@ ics.mq.edu.au www.clt.mq.edu.au

Roadmapping for Natural Language Generation Robert Dale rdale@ ics.mq.edu.au www.clt.mq.edu.au

Roadmapping for Natural Language Generation Robert Dale rdale@ ics.mq.edu.au www.clt.mq.edu.au LREC 2004-05-29 1 Underlying Premise The problem: Current NLG research delivers solutions that are looking for problems The disconnect:

406 views • 14 slides
Language Technology II: Natural Language Dialogue Verbal Output Generation in Dialogue

Language Technology II: Natural Language Dialogue Verbal Output Generation in Dialogue

Language Technology II: Natural Language Dialogue Verbal Output Generation in Dialogue Systems Ivana Kruijff-Korbayov ivana.kruijff@dfki.de Dialog System: Basic Architecture Input ASR Interpretation Dialogue

581 views • 39 slides
Natural Language Understanding We want to communicate with computers using natural language

Natural Language Understanding We want to communicate with computers using natural language

Natural Language Understanding We want to communicate with computers using natural language (spoken and written). unstructured natural language allow any statements, but make mistakes or failure. controlled natural language only

698 views • 20 slides
Content / Mastery of Basic Concepts Place Values of Whole

Content / Mastery of Basic Concepts Place Values of Whole

Content / Mastery of Basic Concepts Place Values of Whole Numbers up to Ten Thousands Addition / Subtraction / Multiplication / Division Algorithms Factors and Multiples Mixed Numbers and

525 views • 10 slides
Technical Aspects of the Paper: Improving Code Readability Models with Textual Features

Technical Aspects of the Paper: Improving Code Readability Models with Textual Features

Technical Aspects of the Paper: Improving Code Readability Models with Textual Features Deeksha Arya COMP762 Key Concepts Previous work as mentioned in the paper: QALP tool (to compute similarity between comment and code) Entropy

585 views • 16 slides
1 Annoucnements Homework 1 due Wednesday 2 Survey of Computer Science Decision Trees,

1 Annoucnements Homework 1 due Wednesday 2 Survey of Computer Science Decision Trees,

xkcd.com/518 1 Annoucnements Homework 1 due Wednesday 2 Survey of Computer Science Decision Trees, Conditionals Kevin Walsh kwalsh@holycross.edu Readings: CSI 9.4, Py 1.6 Example Program >>> lbs = 5 lbs = 5 >>> kg =

569 views • 24 slides
CSE 142 Chapter 4 Programming I l Read Sections 4.14.5, 4.74.9 l The book assumes that

CSE 142 Chapter 4 Programming I l Read Sections 4.14.5, 4.74.9 l The book assumes that

CSE 142 Chapter 4 Programming I l Read Sections 4.14.5, 4.74.9 l The book assumes that youve read chapter 3 on functions Read it anyway, you should do fine Conditionals Their order is a little screwy... 28. June, 2000 F-2 CSE

417 views • 8 slides
A Monte Carlo approach to a divergence minimization problem (work in progress) IGAIA IV, June

A Monte Carlo approach to a divergence minimization problem (work in progress) IGAIA IV, June

A Monte Carlo approach to a divergence minimization problem (work in progress) IGAIA IV, June 12-17, 2016, Liblice Michel Broniatowski Universit Pierre et Marie Curie, Paris, France June 13, 2016 Michel Broniatowski (Institute) Monte Carlo

1.85k views • 40 slides
M ODELLING AND A NALYSIS OF B IOCHEMICAL N ETWORKS WITH T IME P ETRI N ETS Louchka Popova-Zeugmann

M ODELLING AND A NALYSIS OF B IOCHEMICAL N ETWORKS WITH T IME P ETRI N ETS Louchka Popova-Zeugmann

CSP, C APUTH 2004 PN & Systems Biology M ODELLING AND A NALYSIS OF B IOCHEMICAL N ETWORKS WITH T IME P ETRI N ETS Louchka Popova-Zeugmann Humboldt University Berlin, Dep. of CS Monika Heiner Brandenburg University of Technology Cottbus,

794 views • 44 slides
A New Fractional Process: A Fractional Non-homogeneous Poisson Process Enrico Scalas University

A New Fractional Process: A Fractional Non-homogeneous Poisson Process Enrico Scalas University

Definitions Limit theorems Application to the CTRW Summary and Outlook References A New Fractional Process: A Fractional Non-homogeneous Poisson Process Enrico Scalas University of Sussex Joint work with Nikolai Leonenko (Cardiff

721 views • 37 slides
18.175: Lecture 20 Infinite divisibility and L evy processes Scott Sheffield MIT 18.175 Lecture

18.175: Lecture 20 Infinite divisibility and L evy processes Scott Sheffield MIT 18.175 Lecture

18.175: Lecture 20 Infinite divisibility and L evy processes Scott Sheffield MIT 18.175 Lecture 20 1 Outline Infinite divisibility Higher dimensional CFs and CLTs 18.175 Lecture 20 2 Outline Infinite divisibility Higher dimensional CFs and

567 views • 8 slides

More recommend