Lecture 11: Parts of speech - UMass Amherst

21 Pages·2014·1.09 MB·English

Lecture 11:
Parts of speech
Intro to NLP, CS585, Fall 2014
http://people.cs.umass.edu/~brenocon/inlp2014/
Brendan O’Connor
(http://brenocon.com)
Some material borrowed from Chris Manning,
Jurafsky&Martin, and student exercises
1
Tuesday, October 7, 14
•
Review Next Tues. Which?
•
12-1:30pm ?
•
5:30-7pm ?
2
Tuesday, October 7, 14
What’s a part-of-speech (POS)?
•
Syntactic categories / word classes
•
You could substitute words within a class and have
a syntactically valid sentence.
•
Give information how words can combine.
•
I saw the dog
•
I saw the cat
•
I saw the {table, sky, dream, school, anger, ...}
3
Tuesday, October 7, 14
POS is an old idea
•
Dionysius Thrax of Alexandria (100 BCE):
8 parts of speech
•
Common in grammar classes today:
noun, verb, adjective, preposition, conjunction,
pronoun, interjection
•
Many other more fine-grained possibilities
https://www.youtube.com/watch?
v=ODGA7ssL-6g&index=1&list=PL6795522EAD6CE2F7
4
Tuesday, October 7, 14
Open class (lexical) words
Nouns Verbs Adjectives old older oldest
Proper Common Main
Adverbs slowly
IBM cat / cats see
Italy snow registered
Numbers
… more
122,312
one
Closed class (functional)
Modals
Determiners the some can Prepositions to with
had
Conjunctions and or Particles off up … more
Pronouns he its Interjections Ow Eh
5
Tuesday, October 7, 14
Open vs closed classes
•
Closed
•
Determiners: a, an, the
•
Pronouns: he, she, it, they ...
•
Prepositions: on, over, under, of, ...
•
Why “closed”?
•
Many are “grammatical function words.”
•
Open
•
Nouns, verbs, adjectives, adverbs
6
Tuesday, October 7, 14
Many tagging standards
•
Brown corpus (85 tags)
•
Penn Treebank (45 tags) ... the most common one
•
Coarse tagsets
•
Petrov et al. “Universal” tagset (12 tags)
•
http://code.google.com/p/universal-pos-tags/
•
Motivation: cross-linguistic regularities
•
e.g. adposition: pre- and postpositions
•
For English, collapsing of PTB tags
•
Gimpel et al. tagset for Twitter (25 tags)
•
Motivation: easier for humans to annotate
•
We collapsed PTB, added new things that were necessary
for Twitter
7
Tuesday, October 7, 14
Coarse tags, Twitter
D: It's
D: a
A: great
N: show
V: catch
O: it
P: on
D: the
Proper or common?
^: Sundance
Does it matter?
N: channel
#: #SMSAUDIO
Grammatical category??
U: http://instagram.com/p/trHejUML3X/
Not really a grammatical
category, but perhaps an
important word class
8
Tuesday, October 7, 14
Why do we want POS?
•
Useful for many syntactic and other NLP tasks.
•
Phrase identification (“chunking”)
•
Named entity recognition
•
Full parsing
•
Sentiment
9
Tuesday, October 7, 14
same 80 TOEFL questions (Landauer & Dumais, some minor algebraic manipulation, if co-
1997). occurrence is interpreted as NEAR:
The Pointwise Mutual Information (PMI) be-
SO(phrase) =
tween two words, word and word , is defined as
1 2
mantic orientation of a given phrase is calculated & Wiebe, f2o0ll0o0w; sW (Ciehbuer,c 2h0 &00 H; aWnkiesb, e1 9e8t 9a)l:. , 2001). hits(phrase NEAR “excellent”) hits(“poor”) (3)
by comparing its similarity to a positive reference However, although an isolated adjective may indi- log
2
p(word & word )
1 2 hits(phrase NEAR “poor”) hits(“excellent”)
word (“excellent”) with its similarity to a negative cate subjectivity, there may be insufficient context
PMI(word , word ) = log (1)
1 2 2
reference word (“poor”). More specifically, a to determine semantic orientation. For example,
p(word ) p(word )
1 2 Equation (3) is a log-odds ratio (Agresti, 1996).
phrase is assigned a numerical rating by taking the the adjective “unpredictable” may have a negative
To avoid division by zero, I added 0.01 to the hits.
Here, p(word & word ) is the probability that
mutual information between the given phrase and orientation in an autom1otive revi2ew, in a phrase
I also skipped phrase when both hits(phrase
word and word co-occur. If the words are statisti-
the word “excellent” and subtracting the mutual such as “unpred1 ictable ste2e ring”, but it could have
NEAR “excellent”) and hits(phrase NEAR
cally independent, then the probability that they
information between the given phrase and the word a positive orientation in a movie review, in a
“poor”) were (simultaneously) less than four.
co-occur is given by the product p(word )
“poor”. In addition to determining the direction of phrase such as “unpredictable plot”. Therefore the 1
These numbers (0.01 and 4) were arbitrarily cho-
p(word ). The ratio between p(word & word ) and
the phrase’s semantic orientation (positive or nega- algorithm extract2s two consecutive words, w1here 2
sen. To eliminate any possible influence from the
p(word ) p(word ) is thus a measure of the degree
tive, based on the sign of the rating), this numerical one member of th1e pair is a2n adjective or an adverb
testing data, I added “AND (NOT host:epinions)”
of statistical dependence between the words. The
rating also indicates the strength of the semantic and the second provides context.
to every query, which tells AltaVista not to include
log of this ratio is the amount of information that
orientation (based on the magnitude of the num- First a part-of-speech tagger is applied to the
the Epinions Web site in its searches.
we acquire about the presence of one of the words
3
ber). The algorithm is presented in Section 2. review (Brill, 1994). Two consecutive words are
The third step is to calculate the average seman-
when we observe the other.
Hatzivassiloglou and McKeown (1997) have extracted from the review if their tags conform to
tic orientation of the phrases in the given review
POS patterns: sentiment
The Semantic Orientation (SO) of a phrase,
also developed an algorithm for predicting seman- any of the patterns in Table 1. The JJ tags indicate
and classify the review as recommended if the av-
phrase, is calculated here as follows:
tic orientation. Their algorithm performs well, but adjectives, the NN tags are nouns, the RB tags are
erage is positive and otherwise not recommended.
4
it is designed for isolated adjectives, rather than adverbs, an d tShOe (pVhBra stea)g =s PaMreI (vpehrrbasse., “Texhcee lsleencto”n) d Table 2 shows an example for a recommended
(2)
phrases containing adjectives or adverbs. This is pattern, for e x a m p l e , m e a-n PsM tIh(apth rtawseo, “cpoonosre”c) utive review and Table 3 shows an example for a not
discussed in more detail in Section 3, along with words are ex•tracted if the first word is an adverb recommended review. Both are reviews of the
The reference words “excellent” and “poor” were
Turney (2002): identify bigram phrases useful for sentiment
other related work. and the second word is an adjective, but the third Bank of America. Both are in the collection of 410
chosen because, in the five star review rating sys-
The classification algorithm is evaluated on 410 word (which is naotn eaxlytrsacisted) cannot be a noun. reviews from Epinions that are used in the experi-
tem, it is common to define one star as “poor” and
2
reviews from Epinions , randomly sampled from NNP and NNPS (singular and plural proper nouns) ments in Section 4.
five stars as “excellent”. SO is positive when
four different domains: reviews of automobiles, are avoided, so that the names of the items in the
phrase is more strongly associated with “excellent”
Table 2. An example of the processing of a review that
banks, movies, and travel destinations. Reviews at review cannot influence the classification.
and negative when phrase is more strongly associ- 6
the author has classified as recommended.
Epinions are not written by professional writers;
ated with “poor”.
Table 1. Patterns of tags for extracting two-word
Extracted Phrase Part-of-Speech Semantic
any person with a Web browser can become a
PMI-IR estimates PMI by issuing queries to a
phrases from reviews.
Tags Orientation
member of Epinions and contribute a review. Each search engine (hence the IR in PMI-IR) and noting
online experience JJ NN 2.253
First Word Second Word Third Word
of these 410 reviews was written by a different au- the number of hits (matching documents). The fol-
low fees JJ NNS 0.333
(Not Extracted)
thor. Of these reviews, 170 are not recommended lowing experiments use the AltaVista Advanced
local branch JJ NN 0.421
1. JJ NN or NNS anything
and the remaining 240 are recommended (these Search engine5, which indexes approximately 350
small part JJ NN 0.053
2. RB, RBR, or JJ not NN nor NNS
classifications are given by the authors). Always million web pages (counting only those pages that online service JJ NN 2.780
RBS
guessing the majority class would yield an accu- are in English). I chose AltaVista because it has a printable version JJ NN -0.705
3. JJ JJ not NN nor NNS
racy of 59%. The algorithm achieves an average NEAR operator. The AltaVista NEAR operator direct deposit JJ NN 1.288
4. NN or NNS JJ not NN nor NNS
well other RB JJ 0.237
accuracy of 74%, ranging from 84% for automo- constrains the search to documents that contain the
5. RB, RBR, or VB, VBD, anything
inconveniently RB VBN -1.541
bile reviews to 66% for movie reviews. The ex- RBS words wViBthNin, otre Vn BwGo rds of one another, in either
located
perimental results are given in Section 4. order. Previous work has shown that NEAR per-
The second step is to estimate the semantic ori- other bank JJ NN -0.850
The interpretation of the experimental results, forms better than AND when measuring the
entation of the extracted phrases, using the PMI-IR true service JJ NN -0.732
the limitations of this work, and future work are strength of semantic association between words
Average Semantic Orientation 0.322
algorithm. This algorithm uses mutual information
discussed in Section 5. Potential applications are (Turney, 2001).
as a measure of the strength of semantic associa-

outlined in Section 6. Finally, conclusions are pre- Let hits(query) be the number of hits returned,
(plus other sentiment stuff)
tion between two words (Church & Hanks, 1989).
sented in Section 7. given the query query. The following estimate of
PMI-IR has been empirically evaluated using 80
SO can be derived from equations (1) and (2) with
6
The semantic orientation in the following tables is calculated
synonym test questions from the Test of English as
2 Classifying Reviews 10 using the natural logarithm (base e), rather than base 2. The
a Foreign Language (TOEFL), obtaining a score of
natural log is more common in the literature on log-odds ratio.
The first step of the algorithm is to extract phrases Tu7es4d%ay, O(cTtobuerr n7 ,e 1 y4 , 2 0 0 1 ) . F o r c o m p a r i s o n , L a tent Se- Since all logs are equivalent up to a constant factor, it makes
5
mantic Ana hlyttspi:/s/ w(wLwS.Aal)ta, vaisntao.tchoemr/ ssitteast/issetairccah/la dmve asure no difference for the algorithm.
containing adjectives or adverbs. Past work has
of word association, attains a score of 64% on the
demonstrated that adjectives are good indicators of
subjective, evaluative sentences (Hatzivassiloglou

3 http://www.cs.jhu.edu/~brill/RBT1_14.tar.Z
2 4
http://www.epinions.com See Santorini (1995) for a complete description of the tags.