Presentation on Performing aggregation and ellipsis using discourse structures Mariet Theune, Feikje

1
Presentation onPerforming aggregation and
ellipsis using discourse structuresMariet
Theune, Feikje Hielkema, Petra Hendriks

• Presented by Pranesh Bhargava

2

• The article is about how an orphan clause

3

• The article is about how an orphan clause which
  is a result of determiner movement

4

• The article is about how an orphan clause which
  is a result of determiner movement inside the
  head phrase is tackled

5

• The article is about how an orphan clause which
  is a result of determiner movement inside the
  head phrase is tackled by flouting the
  pied-piping constraint,

6

• The article is about how an orphan clause which
  is a result of determiner movement inside the
  head phrase is tackled by flouting the
  pied-piping constraint, which is a member of the
  across-the-border movement, aka the island
  constraints family

7

• The article is about how an orphan clause which
  is a result of determiner movement inside the
  head phrase is tackled by flouting the
  pied-piping constraint, which is a member of the
  across-the-border movement, aka the island
  constraints family, notwithstanding the movement
  due to the verbal clause under the aegis of
  immediate schema number 3 of the Head Driven
  Phrase Structure Grammar

8
The Tacit Revenge of the Linguist on the Verbose
of AI )

• The article is about how an orphan clause which
  is a result of determiner movement inside the
  head phrase is tackled by flouting the
  pied-piping constraint, which is a member of the
  across-the-border movement, aka the island
  constraints family, notwithstanding the movement
  due to the verbal clause under the aegis of
  immediate schema number 3 of the Head Driven
  Phrase Structure Grammar

9

• Real Abstract of the Paper
• The generation of aggregated and elliptical
  sentences, using Dependency trees connected by
  rhetorical relations as input.

10

• Explanation of the Terminology of the Paper

11
Coordination and Subordination

• Used by language to combine smaller units to make
  larger units.
• Participating units called conjuncts.
• Coordination Each conjunct has same function.
  Also called Paratactic.
• Subordination One conjunct is main while the
  other is dependent. Also called hypotactic.
• Theoretical limit of conjunction infinite
• Ex
• Coordinators and, but, or
• Subordinators although, because, however
• John loves Mary and Mary loves John.
• John loves Mary, however she does not love him.

12
Aggregation

• Combining two or more linguistic structures into
  one linguistic structure. Reape Mellish (1999)
• Many types syntactic, semantic, lexical,
  conceptual, discourse.
• No example or definition of any types.
• Paper focuses on syntactic aggregation.
• Syntactic Aggregation Removing redundant
  information at the syntactic level, while leaving
  at least one item in the text to carry the
  meaning explicitly. Dalianis(1999)
• Example Coordination, Subordination and Ellipsis

13
Ellipsis

• Deleting redundant material from an aggregated
  sentence.
• Types of Ellipsis
• 1. Conjunction Reduction Subject of the second
  clause is deleted.
• Diana entered the desert and saw Brutus
• 2. Right Node Raising Rightmost string of the
  first clause is deleted.
• Diana kicks and the prince hits Brutus

14

• 3. Gapping Main verb of the second clause
  deleted.
• Diana left the desert and Brutus the forest.
• 4. Stripping Only one constituent of the second
  clause retained, and rest replaced by place
  holder words.
• Diana entered the desert and so did Brutus.
• Why Need Aggregation/Ellipsis
• Study by Callaway and Lester (2001).
• Subjects showed clear preference for narratives
  with aggregation.
• Help maintain coherence in the text.

15
Rhetorical Dependency Graph Dependency Trees

• RDG Graph with DTs expressing simple
  propositions as nodes, connected by rhetorical
  relations. (defined)
• Contains
• 1. DTs as its leaves
• 2. Relations as non terminal nodes.
• Assume DTs to be sentences.
• Assume RDG to be the text or a group of sentence.

16
Rhetorical Dependency Graph

• Nucleus Important clause
• Satellite Supporting clause

17
Dependency Tree

• Constructed on the basis of predicates and
  arguments.
• Connected to each other through rhetorical
  relations.
• No dependence on linear word order. Translate
  well over different languages.
• No limit on any number of children of any node.
• Alpino parser format used for dependency trees.
• Minor changes
• - A tag for morphology added
• - Tags indicating the position of a word left out
  initially.

18
Dependency Tree
19

• Abstract (again)
• The generation of aggregated and elliptical
  sentences, using Dependency trees connected by
  rhetorical relations as input.
• Accept simple input sentences from a source.
• Aggregate the sentences using dependency trees
  and rhetorical relations.
• Elide the superfluous material.
• Produce the complex sentence as output.

20
Input sentences

• Input sentences came from the NLG system called
  the virtual story teller.
• Plot contains two characters, several events.
• The Narrator Agent transforms plot into text.
• This input text is unaggregated, monotonous and
  redundant.
• Example text generated by the Narrator
• Diana afraid of Brutus. Diana fled to the forest.
  Brutus went to the forest.

21
Design of Narrator

• Content (Document) Planner Receives List of
  propositions, adds rhetorical relations based on
  cue words.
• Sentence (Micro) Planner Maps propositions to
  DTs
• Surface Realizer Performs syntactic aggregation
  and generates surface form.

22
Cue Words

• No definition provided.
• Rhetorical relations are indicated through cue
  words (normally, adverbs).
• Example,
• Tom quit his job, because he was tired of long
  hours. (Cause relation)
• Tom quit his job, he was tired of long hours
  anyway. (Not causal)
• Cue words determine how the trees are aggregated.
• Example, whereas, then again, because, etc.

23
Cue Word Taxonomy

• Cue-word taxonomy Classification of cue-words
  on the basis of what context they exist in.
  (Substitutability tests)
• Ex.
• Kate and Sam are like chalk and cheese. Sam lives
  for his books whereas/on the other hand/then
  again Kate is only interested in martial arts.
• I dont know where to eat tonight. The Indian
  restaurant is always good then again/on the
  other hand/whereas we had curry just the other
  night.

24
Cue Word Taxonomy

• Four classes of cue words were found. Each
  indicates a type of relation.
• 1. Cause relation (one sentence is the cause and
  other is effect)
• 2. Temporal relation (chronological event in
  sentences)
• 3. Contrast relation (not defined)
• 4. Additive relation (not defined)
• Each class may have smaller subsets.

25
Cue Word Taxonomy (Dutch)
26
Cue Word Taxonomy

• Empty boxes Missing cue words (no cue words
  were found)
• Example
• Features of temporal relation The order of the
  events expressed by the clauses.
• Before, after, sequence, finally, during, etc.

27
Performing Aggregation and Ellipsis

• Three steps
• 1. Select a cue word expressing rhetorical
  relation between two DTs.
• 2. Join the DTs depending on the properties of
  the cue word.
• 3. Check the joined DTs for any repeated elements
  that can be elided.

28
Selecting DTs for Aggregation

• The RD Graph is traversed depth-first.
• Relations having two DTs as children are
  searched.
• Appropriate cue word is selected to express this
  relation.
• The cue word becomes the node in the graph.
• Aggregation takes place The nucleus and
  satellite DTs combine to form a complex DT.
• The complex DT replaces the original relation.
• Algorithm continues looking for relations to
  transform.

29
(No Transcript)
30
Selecting DTs for Aggregation

• To avoid high complexity, maximum 3 DTs allowed
  to combine.
• If DTs cannot combine naturally, aggregation may
  be forced using adverbs like then, however, etc.
  to the second DT.
• Such non aggregated sentences get combined in
  subsequent traversal.

31
Aggregation

• Structure of resulting aggregated DT depends on
  properties of selected cue word
• Cue word is coordinator Paratactic structure
  (root node of new DT is CONJ)
• Cue word is subordinator Hypotactic structure
• Cue word is adverb It gets added to nucleus or
  satellite, and separate trees get added to the
  result
• The new tree is passed to the next stage,
  ellipsis.

32
Ellipsis

• Removal of superfluous material from a DT.
• Applies only to paratactic constructions.
• Both conjuncts of the aggregated DT searched for
  identical nodes/branches.
• Identifiers used to mark the identical nodes.
• If no nodes are marked, no ellipsis takes place.
• Depending on the node, one of the kind of
  ellipsis takes place
• Identical subjects Conjunction Reduction
• Identical verbs Gapping
• Identical objects Right Node Raising

33
Ellipsis

• An example of the procedure
• Input sentences
• Diana goes to the forest. Brutus goes to the
  forest.
• Aggregated with cue word and
• Diana goes to the forest and Brutus goes to the
  forest.
• Performing ellipsis (stripping)
• Diana goes to the forest and so does Brutus
• More than one type of Ellipsis may apply on the
  same DTs. Applying coordination
• Diana and Brutus go to the forest.

34
Ellipsis

• Thus, superfluous node is deleted.
• Parent node receives connection to the remaining
  twin.
• Connection is marked borrowed to show that the
  node does not appear on surface.

35
Results

• Here are the two versions of the output of the
  virtual storyteller
• Without aggregation
• Diana goes to the desert. Brutus goes to the
  desert. Diana is afraid of Brutus. Diana goes to
  the forest. Brutus goes to the forest.
• With aggregation
• Diana and Brutus go to the desert. Diana goes to
  the forest, because she is afraid of Brutus.
  Therefore, Brutus goes to the forest too.

36
Scope of Improvement

• - Addition of information about personal goals.
  e.g. why does Brutus follow Diana? Why does Diana
  go to the forest? Etc.
• - Scope of improvement on choice of ellipsis.
• Diana went to the desert and Brutus went to the
  desert too. v/s Diana went to the desert and so
  did Brutus.
• (Need better definition of where to use which
  type of ellipsis).
• - Distributed v/s collective reading.
• Diana and Brutus went to the desert. Separately
  or together?
• (Encoding of this information needed in the DTs).

37
References

• Callaway, G. and J. Lester (2001) Evaluating the
  effects of natural language generation. In
  Proceedings of the 23rd Annual Conference of the
  Cognitive Science Society (CogSci 2001), pages
  164-169, August 2001
• Dalianis, H. (1999) Aggregation in natural
  language generation. Computational Intelligence,
  15(4)384-413
• Reape, M. C. Mellish (1999) Just what is
  aggregation anyway? In Proceedings of the 7th
  European Workshop on Natural Language Generation,
  pages 20-29