Situierte Generierung

Transcription

1 Situierte Generierung Grundlagen der Generierung (Reiter & Dale, 1997) Konstantina Garoufi 1. November 2011 basierend auf Folien von Alexander Koller

2 What is natural language generation (NLG)? non-linguistic representation of information text

3 Example: weather reports Reiter et al. (2005)

4 Example: museum This exhibit is a lekythos, created during the archaic period. It dates from circa 500 BC. It was painted by Amasis with the red figure technique and it originates from Attica.... Unlike the previous exhibit, it originates from Attica. Questo reperto è una lekythos. Αυτοʹ το εʹ κθεµα ειʹ ναι µια ληʹ κυθος. Isard et al. (2003)

5 Example: games ( Dwarf Fortress, generated in-game text, see

6 Related: summarization Apple profit: Apple fourth-quarter profit surges 70%, again exceeding expectations Propelled by huge demand for the latest iphone model, Apple Inc. on Monday posted tremendous increases in earnings and sales as it overtook BlackBerry's maker in the rankings of smart-phone manufacturers. (article 1) The tech juggernaut once again trounced analysts' expectations for earnings as well as sales, but its stock slumped in after-hours trading because of an unexpected drop in a key measure of profitability. (article 1) Waterloo, Ont.-based BlackBerry maker Research in Motion fired back Tuesday in a war of words with Apple Inc. Co-CEO Jim Balsillie sliced apart Apple's criticisms of his company, issuing a lengthy statement aimed at Apple CEO Steve Jobs. (article 3) (Dave Chidley/Canadian Press) A day earlier, Jobs bragged that Apple sold more iphones than RIM sold BlackBerrys in the most recent financial quarter. (article 3) The head of Android posted a passive-aggressive tweet on the subject, and Android developer TweetDeck made a more direct correction of Chief Executive Steve Jobs' claims about its development process. (article 4) Earlier in the day, the stock hit an all-time high of $319 a share, representing a 50% gain since January. (article 2) Apple surpassed longtime rival Microsoft in market capitalization, making the Cupertino, California, company the most valuable technology company in the world, for the moment, at least. (article 5) (Columbia Newsblaster)

7 Traditional structure of an NLG system what to say? = strategic NLG text planning how to say it? = tactical NLG realization sentence planning Pipeline model; Reiter (1994)

8 Text planning Content determination: what objects do we want to talk about? what do we want to say about them? if necessary, interpret domain data Discourse planning: how to arrange information into a discourse? how to break information up into sentence-sized chunks? Both are typically domain-specific.

9 Content determination Reiter & Dale (1997)

10 Content determination: how? Deep reasoning approaches: Determine what to say by reasoning about what information user needs to accomplish their goal Plan recognition paradigm (e.g. Allen & Perrault, 1980) Highly complex Domain-specific approaches: Analyze text in corpus Work out rules for what to say in each situation Easier to implement but less flexible

11 Discourse planning Reiter & Dale (1997)

12 Discourse planning: how? Planning-based approaches: Use AI-planning operators to represent discourse relations Preconditions: what needs to hold before a relation can be used Effects: impact on the hearer s beliefs of using the relation Schema-based approaches: Schemas (patterns) specify constituents and their discourse relations

13 Example: dialog systems naive system Demberg & Moore (2006)

14 Content determination depart arrive airline business direct price :00 13:20 BA :27 13:50 KLM :00 15:10 Lufthansa :00 15:30 BA :10 13:10 BMI :30 17:19 KLM :40 17:30 BMI User model: + direct, + business, + KLM Input: arrive around 3pm Demberg & Moore (2006)

15 Discourse planning 1. Flights in business class: 4 a) KLM: none b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

16 Linguistic (surface) realization Take specification of each sentence and translate it into the actual sentence in the output language. Input: specification of sentence (e.g., semantic representation) a grammar or something similar Output: one sentence

17 Realization 1. Flights in business class: 4 a) KLM: none b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

18 Template-based realization Hard-coded rule: flights in business class: X There are X flights in business class. Flights in business class: 4 There are four flights in business class.

19 Realization 1. There are four flights with availability in business class. a) KLM: none b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

20 Realization 1. There are four flights with availability in business class. a) None are on KLM. b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

21 Realization 1. There are four flights with availability in business class. a) None are on KLM. b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct don t want separate template for each combination

22 Realization with TAG semantics syntax S direct(f7) NP there is VP NP : f7 NP S VP a direct N : f7 flight there is a NP : f7 NP : f7 arrives(f7, 17:30) N* that N : f7 arrives RC at VP PP 17:30 direct N : f7 flight that N : f7 arrives RC at VP PP which is too late for you 17:30 NP : f7 There is a direct flight too_late(f7) NP * that arrives at 17:30, which is too late for you which is too late for you.

23 Realization 1. There are four flights with availability in business class. a) None are on KLM. b) describe f7: +direct, arrives 17:30, too late 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

24 Realization 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. Flights in economy class a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

25 Realization 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) describe f5: BMI, arrives 13:10 b)describe f2: KLM, arrives 13:50, but -direct

26 Realization 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) There is a direct flight on BMI, arriving at 13:10. b)describe f2: KLM, arrives 13:50, but -direct

27 Realization 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) There is a direct flight on BMI, arriving at 13:10. b)there s also a KLM flight arriving at 13:50, but it requires a connection in Amsterdam.

28 Where are we now? text planning surface realization? sentence planning Pipeline model; Reiter (1994)

29 Sentence planning Output of content determination may be not quite suitable as input of surface realizer. Sentence planning: Everything that needs to happen to map content determination output to surface realization input. Typically: sentence aggregation lexicalization referring expression generation

30 Sentence aggregation 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) There is a direct flight on BMI, arriving at 13:10. b)there s also a KLM flight arriving at 13:50, but it requires a connection in Amsterdam.

31 Sentence aggregation 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) There is a direct flight on BMI, arriving at 13:10. b)there s also a KLM flight arriving at 13:50, but it requires a connection in Amsterdam. Merge these two sentences

32 Sentence aggregation 1. There are four flights with availability in business class. a) None are on KLM. b)there is a direct flight that arrives at 17:30, which is too late for you. 2. If you re willing to travel economy, a) there is a direct flight on BMI, arriving at 13:10. b)there s also a KLM flight arriving at 13:50, but it requires a connection in Amsterdam.

33 Types of sentence aggregation Conjunction: apples and oranges Ellipsis: John went to the bank and ε deposited 50 euros. Set formation: every day of the week Embedding:..., which leaves at 10am.

34 Lexicalization Mapping semantic concepts to content words. Not as trivial as it sounds at first glance. How to realize generic concepts based on what they apply to? The temperature rose. The rain got heavier. The revenue increased.

35 Lexicalization How to map real-world values to words? what RGB values accepted as red car? as red wine? as red hair? How to distribute bits of meaning over different words? swim across the lake traverser le lac à la nage

36 Referring expressions (REs) Knowledge base: a b c d We want to say that this guy sleeps. The white rabbit sleeps.

37 Incremental algorithm for RE generation First, add a base noun (taxonomic property) Then, add modifiers (other properties) incrementally until the description has become distinguishing Fixed order for adding properties (e.g. type > color > size >...) Dale & Reiter (1995)

38 RE generation Universe: List of properties: - rabbit - polar bear - broccoli - white... the white rabbit distractors target referent Incremental algorithm: Dale & Reiter (1995)

39 RE generation Very active research area! Traditionally: focus on logical properties of REs (e.g. minimality) Recently: focus shift towards cognitive models of good REs (e.g. humanlikeness, understandability)

40 NLG tasks overview Content determination: deciding what to say Discourse planning: imposing order and structure Sentence aggregation: grouping messages together Lexicalization: picking the right words Referring expression generation: describing entities Linguistic realization: producing grammatical text

41 Summary What to say = content determination How to say it = discourse planning + sentence planning + linguistic realization Particular problem areas: aggregation, lexical choice, referring expressions

42 Next week Topic: Situated noun phrase generation Check the website for the relevant papers and read them me your questions by midnight before class Presenters: Arrange to meet me for feedback at the latest one week before your talk!