Ryan Shaw
Applied Natural Language Processing
Assignment #2
September 29, 2004


I wanted to be able to correctly chunk noun phrases such as "rising
labor costs" in sentences like "Rising labor costs are forcing
manufacturers to make some difficult decisions." I wrote a parser (based
on the one you gave us) that would make a first pass at identifying noun
phrases, and then wrote a second parser that, among other things, would
look for present participles ("rising") preceding noun phrases ("labor
costs").

The problem with this was that my rule also chunked present participles
followed by their direct objects, like "forcing manufacturers." I
thought I could avoid this by looking for a preceding VBP (like "are")
and not chunking in that case.

Looking through the Python regex documentation, I saw that I could do
this with a "negative lookbehind assertion," which would match a VBG
followed by a NP only if it were *not* preceded by a match for VBP.

Unfortunately, the syntax for negative lookbehind assertions (?<!...)
includes the '<' character, which is treated specially by the NLTK
ChunkRule class in order to make writing chunking rules easier. The
ChunkRule constructor converts from a simplified regex syntax which
treats angle brackets like parentheses to normal Python regex syntax.
The HackedChunkRule eliminates this step, allowing rule
definitions to use normal Python regex syntax directly.

So, HackedChunkRule is basically just a way to work around a (IMO)
shortsighted design decision by the NLTK developers, in order to make
the full power of Python regular expressions available.

As for the <-NONE-> tag: this was just a lame ad hoc attempt to get
better results. I noticed that some of my VP chunking rules were
breaking because of <-NONE-> tags seemingly randomly sprinkled
throughout the treebank corpus. I couldn't find any info about these
tags, what they meant or why they were in there. So I decided to allow
optional <-NONE-> tags in verb phrases. I also noticed that the
construction "$X a share" (for stock prices) always had an extra
<-NONE-> tag between the dollar value and the determiner "a," so I took
advantage of this fact to write a rule for correctly chunking those
constructions as NPs, since none of my other rules handled those correctly.


These are my chunking parsers. They are meant to be run in the order presented.

-- Parser 1 -------------------------------------------------------
This parser operates on unchunked word tokens and produces NP chunks.
    rules = [
        ChunkRule(r'<\$><CD><-NONE-><DT><NN>',
                  'Share prices (ex: $18 a share)'),
        ChunkRule(r'<DT>?(<JJ.*>|<CD>|<VBN>|<NNP>)*<NN.*>+',
                  'Chunk nouns and their modifiers'),
        ChunkRule(r'<DT>?<RB>?(<JJ.*>|<CD>)+<VBG>+',
                  'Chunk gerunds and their modifiers'),
        ChunkRule(r'<\$><CD>+',
                  'Chunk currency values'),
        ChunkRule(r'<CD>+',
                  'Chunk numbers (like years)'),
        ChunkRule(r'<PRP>',
                  'Chunk personal pronouns')]

-- Parser 2 -------------------------------------------------------
This parser operates on NP chunks and produces higher-level NP
chunks. May be run more than once.
    rules = [
        HackedChunkRule(r'(?<!<VBP>)<VBG><NP>',
                  'Chunk present participles modifying NPs'),
        ChunkRule(r'<NP>((<,>|<CC>|<,><CC>)<NP>)*<CC><NP>',
                  'Chunk NPs joined by conjunctions'),
        ChunkRule(r'<NP><POS><NP>',
                  'Chunk possessives'),
        ChunkRule(r'<PRP\$><NP>',
                  'Chunk possessive pronouns')]

-- Parser 3 -------------------------------------------------------
This parser operates on NP chunks and produces PP chunks.
    rules = [
        ChunkRule(r'(<IN>|<TO>)+<NP>',
                  'Chunk prepositions followed by NPs')]

-- Parser 4 -------------------------------------------------------
This parser operates on NP and PP chunks and produces higher-level NP
chunks like "the first man on the moon."
    rules = [
        ChunkRule(r'<NP><PP>',
                  'Chunk NP-PP combos into NPs')]

-- Parser 5 -------------------------------------------------------
This parser operates on NP and PP chunks and produces VP chunks.
    rules = [
        ChunkRule(r'<VB.*><JJ.*>',
                  'Chunk verb-adjective combos (ex: rising higher)'),
        ChunkRule(r'(<MD>|<TO>)?<RB>?(<VB.*>)+<-NONE->?<RB>?<RP>?(<NP>|<PP>)*<RB>?',
                  'Chunk VPs with all their optional attachments'),
        ChunkRule(r'<VB.*>',
                  'Chunk lone verbs')]

-- Parser 6 -------------------------------------------------------
This parser operates on VP chunks and produces higher-level VP chunks
like "rose quickly and levelled off."
    rules = [
        ChunkRule(r'<VP><CC><VP>',
                  'Chunk VPs joined by conjunctions')]


Highlights of chunking performance:

-- Sentence 1 -----------------------------------------------------
Shorter maturities are considered a sign of rising rates because
portfolio managers can capture higher rates sooner.

-- Original chunking rules ----------------------------------------
(S:
  <Shorter/JJR>
  (NP: <maturities/NNS>)
  (VP: <are/VBP> <considered/VBN> (NP: <a/DT> <sign/NN>))
  <of/IN>
  (VP:
    <rising/VBG>
    (NP: <rates/NNS>)
    (PP: <because/IN> (NP: <portfolio/NN> <managers/NNS>)))
  <can/MD>
  <capture/VB>
  <higher/JJR>
  (NP: <rates/NNS>)
  <sooner/RB>
  <./.>)

-- Improved chunking rules ----------------------------------------
(S:
  (NP: <Shorter/JJR> <maturities/NNS>)
  (VP:
    <are/VBP>
    <considered/VBN>
    (NP:
      (NP: <a/DT> <sign/NN>)
      (PP: <of/IN> (NP: <rising/VBG> (NP: <rates/NNS>))))
    (PP: <because/IN> (NP: <portfolio/NN> <managers/NNS>)))
  (VP:
    <can/MD>
    <capture/VB>
    (NP: <higher/JJR> <rates/NNS>)
    <sooner/RB>)
  <./.>)

-- Comments -------------------------------------------------------
The original chunker incorrectly tagged "rising rates because
portfolio managers" as a VP. My chunker correctly identified "rising
rates" as a NP.


-- Sentence 2 -----------------------------------------------------
Japan's domestic sales of cars, trucks and buses in October rose 18%
from a year earlier to 500,004 units, a record for the month, the
Japan Automobile Dealers' Association said .

-- Original chunking rules ----------------------------------------
(S:
  (NP: <Japan/NNP>)
  <'s/POS>
  (NP: <domestic/JJ> <sales/NNS>)
  (PP: <of/IN> (NP: <cars/NNS>))
  <,/,>
  (NP: <trucks/NNS>)
  <and/CC>
  (NP: <buses/NNS>)
  (PP: <in/IN> (NP: <October/NNP>))
  (VP:
    <rose/VBD>
    (NP: <18/CD> <%/NN>)
    (PP: <from/IN> (NP: <a/DT> <year/NN>)))
  <earlier/JJR>
  <to/TO>
  (NP: <500,004/CD> <units/NNS>)
  <,/,>
  (NP: <a/DT> <record/NN>)
  (PP: <for/IN> (NP: <the/DT> <month/NN>))
  <,/,>
  (NP: <the/DT> <Japan/NNP> <Automobile/NNP> <Dealers/NNPS>)
  <'/POS>
  (NP: <Association/NNP>)
  <said/VBD>
  <./.>)

-- Improved chunking rules ----------------------------------------
(S:
  (NP:
    (NP: (NP: <Japan/NNP>) <'s/POS> (NP: <domestic/JJ> <sales/NNS>))
    (PP:
      <of/IN>
      (NP:
        (NP: <cars/NNS>)
        <,/,>
        (NP: <trucks/NNS>)
        <and/CC>
        (NP: <buses/NNS>))))
  (PP: <in/IN> (NP: <October/NNP>))
  (VP:
    <rose/VBD>
    (NP: (NP: <18/CD> <%/NN>) (PP: <from/IN> (NP: <a/DT> <year/NN>))))
  <earlier/JJR>
  (PP: <to/TO> (NP: <500,004/CD> <units/NNS>))
  <,/,>
  (NP:
    (NP: <a/DT> <record/NN>)
    (PP: <for/IN> (NP: <the/DT> <month/NN>)))
  <,/,>
  (NP:
    (NP: <the/DT> <Japan/NNP> <Automobile/NNP> <Dealers/NNPS>)
    <'/POS>
    (NP: <Association/NNP>))
  (VP: <said/VBD>)
  <./.>)

-- Comments -------------------------------------------------------
My chunker correctly identified "Japan's domestic sales of cars,
trucks and buses" as a single NP, whereas the original chunker did
not.


-- Sentence 3 -----------------------------------------------------
In the 1990s, spurred by rising labor costs and the strong yen, these
companies will increasingly turn themselves into multinationals with
plants around the world.

-- Original chunking rules ----------------------------------------
(S:
  (PP: <In/IN> (NP: <the/DT> <1990s/NNS>))
  <,/,>
  <spurred/VBN>
  <by/IN>
  (VP: <rising/VBG> (NP: <labor/NN> <costs/NNS>))
  <and/CC>
  (NP: <the/DT> <strong/JJ> <yen/NNS>)
  <,/,>
  (NP: <these/DT> <companies/NNS>)
  <will/MD>
  <increasingly/RB>
  <turn/VB>
  <themselves/PRP>
  (PP: <into/IN> (NP: <multinationals/NNS>))
  (PP: <with/IN> (NP: <plants/NNS>))
  (PP: <around/IN> (NP: <the/DT> <world/NN>))
  <./.>)

-- Improved chunking rules ----------------------------------------
(S:
  (PP: <In/IN> (NP: <the/DT> <1990s/NNS>))
  <,/,>
  (VP:
    <spurred/VBN>
    (PP:
      <by/IN>
      (NP:
        (NP: <rising/VBG> (NP: <labor/NN> <costs/NNS>))
        <and/CC>
        (NP: <the/DT> <strong/JJ> <yen/NNS>))))
  <,/,>
  (NP: <these/DT> <companies/NNS>)
  (VP:
    <will/MD>
    <increasingly/RB>
    <turn/VB>
    (NP:
      (NP: <themselves/PRP>)
      (PP: <into/IN> (NP: <multinationals/NNS>)))
    (PP: <with/IN> (NP: <plants/NNS>))
    (PP: <around/IN> (NP: <the/DT> <world/NN>)))
  <./.>)

-- Comments -------------------------------------------------------
The original chunker did not correctly identify either of the two VPs
in this sentence, and falsely identified a NP as a VP. My chunker
handled all three phrases (pretty much) correctly.


-- Sentence 4 -----------------------------------------------------
The index of the 100 largest Nasdaq financial stocks rose modestly as
well, gaining 1.28 to 449.04.

-- Original chunking rules ----------------------------------------
(S:
  (NP: <The/DT> <index/NN>)
  <of/IN>
  <the/DT>
  <100/CD>
  <largest/JJS>
  (NP: <Nasdaq/NNP>)
  (NP: <financial/JJ> <stocks/NNS>)
  <rose/VBD>
  <modestly/RB>
  <as/IN>
  <well/RB>
  <,/,>
  <gaining/VBG>
  <1.28/CD>
  <to/TO>
  <449.04/CD>
  <./.>)

-- Improved chunking rules ----------------------------------------
(S:
  (NP:
    (NP: <The/DT> <index/NN>)
    (PP:
      <of/IN>
      (NP:
        <the/DT>
        <100/CD>
        <largest/JJS>
        <Nasdaq/NNP>
        <financial/JJ>
        <stocks/NNS>)))
  (VP: <rose/VBD> <modestly/RB>)
  <as/IN>
  <well/RB>
  <,/,>
  (NP:
    (NP: <gaining/VBG> (NP: <1.28/CD>))
    (PP: <to/TO> (NP: <449.04/CD>)))
  <./.>)

-- Comments ------------------------------------------------------- 
My chunker correctly tagged "The index of the 100 largest Nasdaq
financial stocks" as a single NP, where the original chunker only
tagged "The index." The original chunker also missed the VP "rose
modestly;" mine did not.


-- Verb context analysis ------------------------------------------ 
I selected the verb "to rise," various forms of which appeared 30
times in 29 sentences in the subset of the Penn Treebank that I
used. Given that most of the articles in this corpus came from the
Wall Street Journal, it is perhaps unsurprising that the contexts in
which it appeared were fairly similar.

Subjects of "rise" were always some quantitative value like "sales,"
"prices," "stock," "ratings," etc. In about 2/3 of the sentences this
value was a NNS (plural common noun).

As "rise" is an intransitive verb, it did not take any direct objects,
but there were some common contructions following it. In the majority
of sentences it was followed by a dollar amount or a percentage, a PP
beginning with "to" and sometimes a PP beginning with "from," as the
following template shows:

         rise [AMOUNT] [to AMOUNT] [from AMOUNT]

It was also common to see the VP include temporal specifiers such as
"in September" or "last year," and to have adverbs such as "quickly"
or "sharply" modifying the verb.


-- Similar verbs --------------------------------------------------
Though I did not have time to do an analysis to find verbs similar to
"rise," I am fairly confident that an analysis using the same corpus
of synonyms such as "increase" or antonyms such as "fall" would find
similar patterns to the ones described above.