Proceedings of the CoNLL Shared Task Session of EMNLP-CoNLL 2007, pp. 1051–1055,
Prague, June 2007. c ?2007 Association for Computational Linguistics
Frustratingly Hard Domain Adaptation for Dependency Parsing
Mark Dredze1and John Blitzer1and Partha Pratim Talukdar1and
Kuzman Ganchev1and Jo˜ ao V. Grac ¸a2and Fernando Pereira1
1CIS Dept., University of Pennsylvania, Philadelphia, PA 19104
2L2F – INESC-ID Lisboa/IST, Rua Alves Redol 9, 1000-029, Lisboa, Portugal
We describe some challenges of adaptation
in the 2007 CoNLL Shared Task on Domain
Adaptation. Our error analysis for this task
suggests that a primary source of error is
differences in annotation guidelines between
treebanks. Our suspicions are supported by
the observation that no team was able to im-
prove target domain performance substan-
tially over a state of the art baseline.
Dependency parsing, an important NLP task, can be
done with high levels of accuracy. However, adapt-
ing parsers to new domains without target domain
labeled training data remains an open problem. This
paper outlines our participation in the 2007 CoNLL
Shared Task on Domain Adaptation (Nivre et al.,
2007). The goal was to adapt a parser trained on
a single source domain to a new target domain us-
ing only unlabeled data.
15K sentences of labeled text from the Wall Street
Journal (WSJ) (Marcus et al., 1993; Johansson and
Nugues, 2007) as well as 200K unlabeled sentences.
The development data was 200 sentences of labeled
biomedical oncology text (BIO, the ONCO portion
of the Penn Biomedical Treebank), as well as 200K
unlabeled sentences (Kulick et al., 2004). The two
test domains were a collection of medline chem-
istry abstracts (pchem, the CYP portion of the Penn
Biomedical Treebank) and the Child Language Data
Exchange System corpus (CHILDES) (MacWhin-
ney, 2000; Brown, 1973). We used the second or-
der two stage parser and edge labeler of McDonald
et al. (2006), which achieved top results in the 2006
We were given around
CoNLL-X shared task. Preliminary experiments in-
dicated that the edge labeler was fairly robust to do-
main adaptation, lowering accuracy by 3% in the de-
velopment domain as opposed to 2% in the source,
so we focused on unlabeled dependency parsing.
Our system did well, officially coming in 3rd
place out of 12 teams and within 1% of the top sys-
tem (Table 1).1In unlabeled parsing, we scored
1st and 2nd on CHILDES and pchem respectively.
However, our results were obtained without adap-
tation. Given our position in the ranking, this sug-
gests that no team was able to significantly improve
performance on either test domain beyond that of a
After much effort in developing adaptation meth-
ods, it is critical to understand the causes of these
negative results. In what follows, we provide an er-
ror analysis that attributes domain loss for this task
to a difference in annotation guidelines between do-
mains. We then overview our attempts to improve
adaptation. While we were able to show limited
adaptation on reduced training data or with first-
order features, no modifications improved parsing
with all the training data and second-order features.
2 Parsing Challenges
We begin with an error analysis for adaptation be-
tween WSJ and BIO. We divided the available WSJ
data into a train and test set, trained a parser on
the train set and compared errors on the test set
and BIO. Accuracy dropped from 90% on WSJ to
84% on BIO. We then computed the fraction of er-
rors involving each POS tag. For the most common
1While only 8 teams participated in the closed track with us,
our score beat all of the teams in the open track.
Table 1: Official labeled (l) and other unlabeled (ul)
submitted results for the two test domains (pchem
and childes) and development data accuracy (bio).
The parser was trained on the provided WSJ data.
POS types, the loss (difference in source and tar-
get error) was: verbs (2%), conjunctions (5%), dig-
its (23%), prepositions (4%), adjectives (3%), de-
terminers (4%) and nouns (9%).2Two POS types
stand out: digits and nouns. Digits are less than
4% of the tokens in BIO. Errors result from the BIO
annotations for long sequences of digits which do
not appear in WSJ. Since these annotations are new
with respect to the WSJ guidelines, it is impossi-
ble to parse these without injecting knowledge of
the annotation guidelines.3
common, comprising 33% of BIO and 30% of WSJ
tokens, the most popular POS tag by far. Addi-
tionally, other POS types listed above (adjectives,
prepositions, determiners, conjunctions)oftenattach
to nouns. To confirm that nouns were problem-
atic, we modified a first-order parser (no second or-
der features) by adding a feature indicating correct
noun-noun edges, forcing the parser to predict these
edges correctly. Adaptation performance rose on
BIO from 78% without the feature to 87% with the
feature. This indicates that most of the loss comes
from missing these edges.
The primary problem for nouns is the difference
between structures in each domain.
tion guidelines for the Penn Treebank flattened noun
phrases to simplify annotation (Marcus et al., 1993),
so there is no complex structure to NPs. K¨ ubler
(2006) showed that it is difficult to compare the
Penn Treebank to other treebanks with more com-
phrase “the New York State Insurance Department”.
The annotation indicates a flat structure, where ev-
Nouns are far more
2We measured these drops on several other dependency
parsers and found similar results.
3Forexample, thephrase“(R=28%(10/26); K=10%(3/29);
chi2 test: p=0.014).”
ery token is headed by “Department”. In contrast,
a similar BIO phrase has a very different structure,
pursuant to the BIO guidelines. For “the detoxi-
cation enzyme glutathione transferase P1-1”, “en-
zyme” is the head of the NP, “P1-1” is the head of
“transferase”, and “transferase” is the head of “glu-
tathione”. Since the guidelines differ, we observe no
corresponding structure in the WSJ. It is telling that
the parser labels this BIO example by attaching ev-
ery token to the final proper noun “P1-1”, exactly as
the WSJ guidelines indicate. Unlabeled data cannot
indicate that BIO uses a different standard.
Another problem concerns appositives. For ex-
ample, the phrase “Howard Mosher, president and
chief executive officer,” has “Mosher” as the head
of “Howard” and of the appositive NP delimited by
commas. While similar constructions occur in BIO,
there are no commas to indicate this. An example is
the above BIO NP, in which the phrase “glutathione
transferase P1-1” is an appositive indicating which
“enzyme” is meant. However, since there are no
commas, the parser thinks “P1-1” is the head. How-
ever, there are not many right to left attaching nouns.
In addition to a change in the annotation guide-
lines for NPs, we observed an important difference
in the distribution of POS tags. NN tags were almost
twice as likely in the BIO domain (14% in WSJ and
25% in BIO). NNP tags, which are close to 10% of
the tags in WSJ, are nonexistent in BIO (.24%). The
cause for this is clear when the annotation guide-
lines are considered. The proper nouns in WSJ are
names of companies, people and places, while in
BIO they are names of genes, proteins and chemi-
cals. However, for BIO these nouns are labeled NN
instead of NNP. This decision effectively removes
NNP from the BIO domain and renders all features
that depend on the NNP tag ineffective. In our above
BIO NP example, all nouns are labeled NN, whereas
the WSJ example contains NNP tags. The largest
tri-gram differences involve nouns, such as NN-NN-
NN, NNP-NNP-NNP, NN-IN-NN, and IN-NN-NN.
However, when we examine the coarse POS tags,
which do not distinguish between nouns, these dif-
ferences disappear.This indicates that while the
overall distribution of POS tags is similar between
the domains, the fine grained tags differ. These fine
grained tags provide more information than coarse
tags; experiments that removed fine grained tags
hurt WSJ performance but did not affect BIO.
Finally, we examined the effect of unknown
words. Not surprisingly, the most significant dif-
ferences in error rates concerned dependencies be-
tween words of which one or both were unknown
to the parser. For two words that were seen in the
training data loss was 4%, for a single unknown
word loss was 15%, and 26% when both words were
unknown. Both words were unknown only 5% of
the time in BIO, while one of the words being un-
known was more common, reflecting 27% of deci-
sions. Upon further investigation, the majority of
unknown words were nouns, which indicates that
unknown word errors were caused by the problems
Recent theoretical work on domain adapta-
tion (Ben-David et al., 2006) attributes adaptation
loss to two sources: the difference in the distribu-
tion between domains and the difference in label-
ing functions. Adaptation techniques focus on the
former since it is impossible to determine the lat-
ter without knowledge of the labeling function. In
parsing adaptation, the former corresponds to a dif-
ference between the features seen in each domain,
such as new words in the target domain. The de-
cision function corresponds to differences between
annotation guidelines between two domains. Our er-
ror analysis suggests that the primary cause of loss
from adaptation is from differences in the annotation
guidelines themselves. Therefore, significant im-
provements cannot be made without specific knowl-
amount of source training data can help if no rele-
vant structure exists in the data. Given the results
for the domain adaptation track, it appears no team
successfully adapted a state-of-the-art parser.
3 Adaptation Approaches
We survey the main approaches we explored for this
task. While some of these approaches provided a
modest performance boost to a simple parser (lim-
ited data and first-order features), no method added
any performance to our best parser (all data and
A natural approach to improving parsing is to mod-
ify the feature set, both by removing features less
likely to transfer and by adding features that are
more likely to transfer. We began with the first ap-
proach and removed a large number of features that
we believed transfered poorly, such as most features
for noun-noun edges. We obtained a small improve-
ment in BIO performance on limited data only. We
ically designed to improve noun phrase construc-
tions, such as features based on the lexical position
of nouns (common position in NPs), frequency of
occurrence, and NP chunking information. For ex-
ample, trained on in-domain data, nouns that occur
more often tend to be heads. However, none of these
features transfered between domains.
A final type of feature we added was based on
the behavior of nouns, adjectives and verbs in each
domain. We constructed a feature representation
of words based on adjacent POS and words and
clustered words using an algorithm similar to that
of Saul and Pereira (1997). For example, our clus-
tering algorithm grouped first names in one group
and measurements in another. We then added the
cluster membership as a lexical feature to the parser.
None of the resulting features helped adaptation.
Training diversity may be an effective source for
adaptation. We began by adding information from
multiple different parsers, which has been shown
to improve in-domain parsing. We added features
indicating when an edge was predicted by another
parser and if an edge crossed a predicted edge, as
well as conjunctions with edge types. This failed
to improve BIO accuracy since these features were
less reliable at test time. Next, we tried instance
bagging (Breiman, 1996) to generate some diversity
among parsers. We selected with replacement 2000
training examples from the training data and trained
three parsers. Each parser then tagged the remain-
ing 13K sentences, yielding 39K parsed sentences.
We then shuffled these sentences and trained a final
parser. This failed to improve performance, possibly
because of conflicting annotations or because of lack
of sufficient diversity. To address conflicting annota-
tions, we added slack variables to the MIRA learn-
ing algorithm (Crammer et al., 2006) used to train
the parsers, without success. We measured diversity
by comparing the parses of each model. The dif-
ference in annotation agreement between the three
instance bagging parsers was about half the differ-
ence between these parsers and the gold annotations.
While we believe this is not enough diversity, it was
not feasible to repeat our experiment with a large
number of parsers.
Another approach to adaptation is to favor training
examples that are similar to the target. We first mod-
ified the weight given by the parser to each training
sentence based on the similarity of the sentence to
target domain sentences. This can be done by mod-
ifying the loss to limit updates in cases where the
sentence does not reflect the target domain. We tried
a number of criteria to weigh sentences without suc-
cess, including sentence length and number of verbs.
Next, we trained a discriminative model on the pro-
vided unlabeled data to predict the domain of each
sentence based on POS n-grams in the sentence.
Training sentences with a higher probability of be-
ing in the target domain received higher weights,
also without success. Further experiments showed
that any decrease in training data hurt parser perfor-
mance. It would seem that the parser has no dif-
ficulty learning important training sentences in the
presence of unimportant training examples.
A related idea focused on words, weighing highly
tokens that appeared frequently in the target domain.
We scaled the loss associated with a token by a fac-
tor proportional to its frequency in the target do-
main. We found certain scaling techniques obtained
tiny improvements on the target domain that, while
significant compared to competition results, are not
statistically significant. We also attempted a sim-
ilar approach on the feature level. A very predic-
tive source domain feature is not useful if it does
not appear in the target domain. However, limiting
the feature space to target domain features had no
effect. Instead, we scaled each feature’s value by a
factor proportional to its frequency in the target do-
main and trained the parser on these scaled feature
values. We obtained small improvements on small
amounts of training data.
Target Focused Learning
4 Future Directions
Given our pessimistic analysis and the long list of
failed methods, one may wonder if parser adapta-
tion is possible at all. We believe that it is. First,
there may be room for adaptation with our domains
if a common annotation scheme is used. Second,
we have stressed that typical adaptation, modifying
a model trained on the source domain, will fail but
there may be unsupervised parsing techniques that
improve performance after adaptation, such as a rule
based NP parser for BIO based on knowledge of the
annotations. However, this approach is unsatisfying
as it does not allow general purpose adaptation.
We thank Joel Wallenberg and Nikhil Dinesh for
their informative and helpful linguistic expertise,
Kevin Lerman for his edge labeler code, and Koby
Crammer for helpful conversations. Dredze is sup-
ported by a NDSEG fellowship; Ganchev and Taluk-
dar by NSF ITR EIA-0205448; and Blitzer by
DARPA under Contract No. NBCHD03001. Any
opinions, findings, and conclusions or recommen-
dations expressed in this material are those of the
author(s) and do not necessarily reflect the views of
the DARPA or the Department of Interior-National
Business Center (DOI-NBC).
Shai Ben-David, John Blitzer, Koby Crammer, and Fer-
nando Pereira. 2006. Analysis of representations for
domain adaptation. In NIPS.
1996. Bagging predictors.Machine
R. Brown. 1973. A First Language: The Early Stages.
Harvard University Press.
Koby Crammer, Ofer Dekel, Joseph Keshet, Shai Shalev-
Shwartz, and Yoram Singer. 2006. Online passive-
aggressive algorithms. Journal of Machine Learning
Research, 7:551–585, Mar.
R. Johansson and P. Nugues.
constituent-to-dependency conversion for English. In
Proc. of the 16th Nordic Conference on Computational
Sandra K¨ ubler.
schemes influence parsing results? or how not to com-
pare apples and oranges. In RANLP.
2006.How do treebank annotation
S. Kulick, A. Bies, M. Liberman, M. Mandel, R. Mc- Download full-text
Donald, M. Palmer, A. Schein, and L. Ungar. 2004.
Integrated annotation for biomedical information ex-
traction. In Proc. of the Human Language Technol-
ogy Conference and the Annual Meeting of the North
American Chapter of the Association for Computa-
tional Linguistics (HLT/NAACL).
B. MacWhinney. 2000. The CHILDES Project: Tools
for Analyzing Talk. Lawrence Erlbaum.
M. Marcus, B. Santorini, and M. Marcinkiewicz. 1993.
Building a large annotated corpus of English: the Penn
Treebank. Computational Linguistics, 19(2):313–330.
Ryan McDonald, Kevin Lerman, and Fernando Pereira.
2006. Multilingual dependency parsing with a two-
stage discriminative parser. In Conference on Natural
Language Learning (CoNLL).
J. Nivre, J. Hall, S. K¨ ubler, R. McDonald, J. Nils-
son, S. Riedel, and D. Yuret.
2007 shared task on dependency parsing.
of the CoNLL 2007 Shared Task. Joint Conf. on Em-
pirical Methods in Natural Language Processing and
Computational Natural Language Learning (EMNLP-
Lawrence Saul and Fernando Pereira.
gate and mixed-order markov models for statistical
language modeling. In EMNLP.