Sentence Ordering and Topic Segmentation

1
Sentence Ordering and Topic Segmentation

• Vasileios Hatzivassiloglou
• University of Texas at Dallas

2
Ordering of sentences

• In certain applications, the context planning
  module is weak, but we receive (more or less
  fully formed) a number of sentences that need to
  be put together
• Examples
• Summarization of a single document
• Summarization of multiple documents
• Generating long answers for questions (e.g.,
  definitions)

3
Deciding the order

• It may be possible to use certain clues
• Presence of a certain features, such as
  genus-species information in a definitional
  sentence
• More generally, classifying the sentences in
  functional categories (e.g., genus-species,
  historical, example, etc.) and following a
  partial order between categories
• Using known order, e.g., in single-document
  summarization

4
Cohesion

• A major element in producing good text is
  cohesion, the semantic relatedness between the
  components (sentences)
• Cohesive text flows naturally
• How to estimate cohesion?
• shared words between adjacent sentences,
  especially important words for the topic
• co-reference (hard to recognize)
• semantically related words between adjacent
  sentences

5
Lexical chains

• Groups of related words and compounds
• Built using
• a source of semantic relations such as a
  thesaurus or WordNet
• adjacency information from known, ordered sets of
  sentences
• Can be constructed at various levels of
  granularity
• Can be topic-dependent

6
WordNet

• Developed at Princeton over the last 15 years
• Contains more than 120,000 words and compounds
  organized into synsets, sets of synonyms that
  define a word sense
• Hierarchically organized to indicate hyponymy
  (e.g., sugar maple ? maple ? tree ? plant)
• Limited antonymy (opposition) and meronymy
  (part-whole) links

7
Sample lexical chains

• proportion 1, face mask 1, equipment 2, pump 2,
  device 1, computer 3, machine 8
• Glasgow Royal infirmary 1, HCI 5, hospital 4
• drug 1, chloroform 1, anaesthetic 4
• surgery 2, operation 2
• Tory government 1, government 1
• anaesthetist 1, doctor 1, Scottish doctor 1,
  surgeon 1

8
Example definitional output

• Who is Sonia Gandhi?
• Congress President Sonia Gandhi, who married into
  what was once Indias most powerful political
  family, is the first non-Indian since
  independence 50 years ago to lead the Congress.
  After Prime Minister Rajiv Gandhi was
  assassinated in 1991, Gandhi was persuaded by the
  Congress to succeed her husband to continue
  leading the party as the chief, but she refused.
  The BJP had shrugged off the influence of the
  51-year-old Sonia Gandhi when she stepped into
  politics early this year, dismissing her as a
  foreigner. Sonia Gandhi is now an Indian
  citizen. Gandhi, who is 51, met her husband when
  she was an 18-year old student at Cambridge in
  London, the first time she was away from her
  native Italy.

9
Example definitional output

• Who is Sonia Gandhi?
• Congress President Sonia Gandhi, who married into
  what was once Indias most powerful political
  family, is the first non-Indian since
  independence 50 years ago to lead the Congress.
  After Prime Minister Rajiv Gandhi was
  assassinated in 1991, Gandhi was persuaded by the
  Congress to succeed her husband to continue
  leading the party as the chief, but she refused.
  The BJP had shrugged off the influence of the
  51-year-old Sonia Gandhi when she stepped into
  politics early this year, dismissing her as a
  foreigner. Sonia Gandhi is now an Indian
  citizen. Gandhi, who is 51, met her husband when
  she was an 18-year old student at Cambridge in
  London, the first time she was away from her
  native Italy.

10
Example definitional output

• Who is Sonia Gandhi?
• Congress President Sonia Gandhi, who married into
  what was once Indias most powerful political
  family, is the first non-Indian since
  independence 50 years ago to lead the Congress.
  After Prime Minister Rajiv Gandhi was
  assassinated in 1991, Gandhi was persuaded by the
  Congress to succeed her husband to continue
  leading the party as the chief, but she refused.
  The BJP had shrugged off the influence of the
  51-year-old Sonia Gandhi when she stepped into
  politics early this year, dismissing her as a
  foreigner. Sonia Gandhi is now an Indian
  citizen. Gandhi, who is 51, met her husband when
  she was an 18-year old student at Cambridge in
  London, the first time she was away from her
  native Italy.

11
Example definitional output

• Who is Sonia Gandhi?
• Congress President Sonia Gandhi, who married into
  what was once Indias most powerful political
  family, is the first non-Indian since
  independence 50 years ago to lead the Congress.
  After Prime Minister Rajiv Gandhi was
  assassinated in 1991, Gandhi was persuaded by the
  Congress to succeed her husband to continue
  leading the party as the chief, but she refused.
  The BJP had shrugged off the influence of the
  51-year-old Sonia Gandhi when she stepped into
  politics early this year, dismissing her as a
  foreigner. Sonia Gandhi is now an Indian
  citizen. Gandhi, who is 51, met her husband when
  she was an 18-year old student at Cambridge in
  London, the first time she was away from her
  native Italy.

12
Applications of lexical chains

• Ordering, as discussed
• Text classification
• Text segmentation and filtering
• focused information retrieval

13
TextTiling

• Step one Estimate the cohesion score between
  left and right blocks at each potential gap
• The cohesion score can be based on shared words,
  shared lexical chain elements, or new words
• One option is to treat the blocks as vectors, and
  then calculate their inner product

14
Depth scores

• Cohesion scores are converted to depth scores
• Intuitively, the depth score corresponds to
• accounting for the variability in cohesion within
  a segment (e.g., introduction vs. later detailed
  discussion)
• One formula is
• di (si-1 si) (si1 si)
• We are only interested at points where siltsi-1
  and siltsi1

15
Smoothing cohesion scores

• We need to detect gradual gaps and conversely,
  stronger nearby gaps when evaluating a particular
  gap
• This can be achieved by replacing si with
  (si-1sisi1)/3, which smoothes the cohesion
  score across successive gaps

16
Final gap determination

• The appropriate strength of the depth score
  varies according to the text
• A self-adjusting approach
• Calculate the mean depth score µ and its standard
  deviation s
• Choose gaps with di gt µcs, where c is a constant

17
Reading

• Section 15.5 on discourse segmentation and text
  tiling