Morphology and Sentences

1
Morphology and Sentences

• Vasileios Hatzivassiloglou
• University of Texas at Dallas

2
Matching words

• The same word can have multiple forms
• The most common case is sentence-initial
  capitalization we need to preserve this
  information while matching the forms
• Important to recognize proper nouns
• Another cause is alternative spellings
• labour / labor, labelled / labeled
• multiple formats for dates, phone numbers, etc.

3
Morphology

• Many languages have multiple word forms related
  to a single base form (root form)
• A lexeme is the representative base form from
  which other closely related forms are produced
• Three productive processes
• Inflection
• Derivation
• Compounding

4
Inflection

• Addition of prefixes and suffixes that change
  meaning only slightly
• Lexeme and syntactic category stays the same
• Features such as number and tense are modified
• Examples
• -s to form plural nouns
• -ed to form past tense

5
Derivation

• More radical change in meaning
• Possible change in syntactic category
• Examples
• -ly transforms adjectives to adverbs (wide-ly)
• -en transforms adjectives to verbs (weak-en)
• -able transforms verbs into adjectives
  (accept-able)
• -er transforms verbs into nouns (teach-er)

6
Compounding

• Combination of two words into a new word
• Result can be written as one word, as one
  hyphenated word, or as two words, but is
  pronounced as a single word and has distinct,
  recognizable meaning
• Most unrestricted process (noun-noun most common
  in English)
• Examples
• disk drive, mad cow disease

7
Irregularities in morphology

• Compounding unpredictable
• Exceptions in both inflectional and derivational
  morpology
• irregular plurals and verb tenses
• word formation affixes do not always apply, e.g.,
  difficultly and oldly
• sometimes other processes compete, e.g., hardly

8
Reverse morphology engineering

• Recall our goal to conflate similar word forms
• Approach recognize morphological inflation
  and/or derivation and combine the word types
• Is it advisable to do so?

9
When to remove morphology

• Depends on the task
• Removing morphology loses some syntactic
  information
• Process is inexact, so it will sometime lose
  semantic information as well
• Usually applied for retrieval purposes (e.g., in
  information retrieval) but not in most text
  analysis tasks
• Regardless, it is important to recognize
  morphology

10
Stemming

• Rule-based approach that removes common suffixes
  such as er and ing
• Most commonly used algorithm is Porters (1980)
• Result conflates sometimes unrelated word forms
  and will not work well in morphologically rich
  languages

11
Handling compounds

• By default, we will recognize members of many
  compounds as separate words
• It is often beneficial to add commonly used
  compounds into the lexicon
• This increases the lexicon size substantially,
  but also increases the accuracy of our
  representation and further processing

12
Sentences

• Definition
• A span of text that is syntactically complete
• Observable through the use of punctuation that
  separates sentences (period but also question
  mark, exclamation mark, semicolon, and maybe
  colon and long dash)
• Irregularities include quoted text, and
  typographic conventions in the US.
• You remind me, she said, of your mother.

13
Length of sentences

• Usually, we study short examples
• In real life, sentences in prepared text are
  longer
• News text
• Mode of 23 words
• 75 of sentences are longer than 15 words
• 58 of sentences are longer than 20 words
• 26 of sentences are longer than 30 words

14
Recognizing sentences

• Most significant problem is the ambiguous use of
  punctuation for the period in abbreviations and
  numbers as well
• Basic algorithm Treat all periods and other
  punctuation marks above as sentence separators
• Accuracy
• approximately 90

15
Improving sentence recognition

• Use information from
• capitalization (observable)
• knowledge of proper names
• knowledge of abbreviations
• knowledge of possible abbreviation positions
  (e.g., vs. and Dr. cannot be sentence final)

16
A heuristic algorithm

• Label every punctuation mark (not inside a word)
  as sentence separators
• Adjust for quote position in American English
• Revise this if the next word is not capitalized
• For periods, also revise this if the last word is
  an abbreviation that cannot occur at the end of
  the sentence

17
Sentence recognition is classification

• Decide the role of each punctuation mark
• Using observed data from labeled training cases
  to predict new occurrences
• Possible features
• length and case of words before/after
• probability of each word to be sentence
  final/initial
• probability of parts of speech to be sentence
  final/initial (class-based smoothing)

18
Reading

• Section 3.1 (introductory part, before 3.1.1)
• Review material on different word classes
  Sections 3.1.1 3.1.4
• Sections 4.2.3 4.2.4