Question Answering Using Syntax-Based
Concept Nodes
Demetrios G. Glinos and Fernando Gomez
School of Electrical Engineering and Computer Science
University of Central Florida, Orlando, FL 32816
Abstract. This paper presents a syntax-based formalism for representing
atomic propositions extracted from textual documents. A method is described
for constructing a network of concept nodes for indexing such logical forms
based on the discourse entities they contain. The decomposition of meaningful
questions into Boolean combinations of question patterns using the same for-
malism is presented, in which free variables represent the desired answers. A
method is described for using this formalism for robust question answering us-
ing the concept network and WordNet synonym, hypernym, and antonym rela-
tionships. Finally, the encouraging performance of an implementation of this
formalism against the factoid questions from the 2005 Text Retrieval Confer-
ence, which operated upon the AQUAINT document corpus, is discussed.
1 Introduction
Question answering (QA) represents an intermediate stage along the path from docu-
ment retrieval to text understanding. As an area of research interest, it serves as a
proving ground where strategies for document processing, knowledge representation,
question analysis, and answer extraction may be evaluated in real world information
extraction contexts. Such contexts typically involve large text document collections,
web content, multimedia documents, or domain-specific databases. In this paper, we
present our formalism for processing text documents to produce a concept network
for persistent knowledge, for analyzing questions to produce question patterns em-
ploying the same formalism, and for robust question answering using the concept
network so obtained together with the WordNet lexical resource. We also discuss the
performance of an implementation of our methods against the factoid questions from
the 2005 Text Retrieval Conference (TREC) Question Answering Track.
2 Question Answering Systems
Current QA systems typically involve large text document collections and primarily
use keywords or other attributes extracted from the question to restrict the subset of
the document collection for consideration and also to extract an answer from the
documents and/or passages retrieved [10]. However, systems vary widely in their
approaches for achieving these ends.
G. Glinos D. and Gomez F. (2006).
Question Answering Using Syntax-Based Concept Nodes.
In Proceedings of the 3rd International Workshop on Natural Language Understanding and Cognitive Science, pages 81-90
DOI: 10.5220/0002477500810090
Copyright
c
SciTePress
Some systems, such as that described in [7], are built around a named entity recog-
nizer for selecting the relevant passages, to which surface patterns are applied to ex-
tract candidate answers. Other systems, such as [8], process documents to produce a
set of data structures which is mined to produce answers. In [6], the authors describe
a system in which the input text is processed to produce a set of “quasi-logical forms”
(QLFs), from which the answer in obtained by resolving the reference for the ques-
tion variable. And in [2], questions are resolved by Prolog unification against a set of
discourse representation structures.
Mixed strategies are also frequently employed, as in the three-stage approach de-
scribed in [5], which involved: (i) surface patterns learned from the Web; (ii) seman-
tic type extraction based on a predefined taxonomy; and (iii) searching the document
set for occurrences of hyponyms of a key word extracted from the question. A system
where different strategies were applied to different question types is described in [9].
By contrast, our approach involves logical forms that explicitly capture the n-ary
relationships among the syntactical components of atomic propositions, not just their
binary relationships, so that ancillary data structures are not needed in order to infer
higher order relationships. Moreover, question pattern matching is performed not
against the surface text itself, but against the logical forms that are extracted from the
text, and includes all features of the input question. Finally, our approach uses a net-
work of concept nodes that is unique in that it is used for: (i) indexing the logical
forms; (ii) capturing “is-a” relationships among concepts; and (iii) providing the entry
points for extending “is-a” relationships into the WordNet [4] lexical database.
3 Knowledge Representation
3.1 Proposition Tuples
We propose the notion of a “proposition tuple”, which we define as a logical proposi-
tion consisting of an atomic predicate and its arguments, as the basic unit of knowl-
edge that is mined for question answering. For a simple sentence, the main verb may
be taken as the predicate, with the remaining syntactical components as its arguments.
Thus, for “Peter threw the ball,” we have the predicate threw(Peter,ball). This is not
generally the case for more complex sentences. Consider, for example, the following
sentence from the AQUAINT corpus used in TREC: “Tourism is one of the major
industries in Port Arthur, a town at the southern tip of the island.” The implementa-
tion described in Section 6 of this paper successfully splits off the apposition, produc-
ing two simpler sentences: “Port Arthur is a town at the southern tip of the island,”
and “Tourism is one of the major industries in Port Arthur.” In this form, each of
these simpler sentences can now be expressed as a logical proposition from which the
predicate and its arguments may be easily discerned. In our implementation, sentence
splitting is performed by heuristics for recognizing appositions, subordinate clauses,
non-defining relative clauses, coordinations, and similar syntactical constructs, from
which separate sentences are constructed, although it should be noted that our algo-
rithms do not depend upon the particular method for performing such splitting.
Given that the input text can be decomposed into distinct propositions through sen-
tence splitting as defined above, we define the proposition tuple for a given proposi-
tion as the 6-tuple
82
<subject,verb,gerinf,modifiers,indirect,direct>
where “subject” refers to the noun phrase representing the subject of the sentence,
“verb” refers to the main verb phrase, “gerinf” represents any gerund or infinitive
form, “modifiers” refers to adverbials and adverbial complements, typically preposi-
tional phrases, “indirect” refers to the indirect object, if any, and “direct” refers to the
direct object, if any. The purpose of separating the gerund/infinitive from the main
verb is to facilitate question answering.
Proposition tuples capture the n-ary relationships among all of the non-null syntac-
tical components of the input proposition. Thus, a single tuple may encode the binary
relationships between subject and verb, object and verb, subject and object, and be-
tween any of the modifiers and other sentence components or even other modifiers. It
also encodes higher order relationships, such as a ternary relationship between a
given subject, object, and verb, and so on. As such an encoding mechanism, the tuple
represents an efficient means of representing all such relationships.
As an example, consider the following sentence, which is also taken from the
AQUAINT corpus: “As a professor at the University of Chicago in the early 1940s,
Fermi designed and built the first nuclear reactor that later was put into use for re-
search into nuclear weapons.” There are two propositions contained in this sentence,
so that there are, correspondingly, two proposition tuples, the first of which is:
Proposition tuple #1:
subject: Fermi,
verb: designed
gerinf: null
modifiers: as a professor; at the University of Chicago; in the early 1940s
indirect: null
direct: the first nuclear reactor that later was put into use for research
into nuclear weapons
where the second proposition tuple is identical to the first, except that the verb is
“built”. The three verb modifiers in this example, all of which are prepositional
phrases, are separated by semicolons in the modifier field, indicating that they are
separate and distinct objects. This supports question answering using any combina-
tion of these modifiers, from none to all, in any order. Also noteworthy is the capture
of an entire phrasal concept as the direct object. Phrasal concepts form the basis of
the concept network discussed in the next subsection.
3.2 Concept Network
Given that a proposition tuple represents the interrelationships among the non-null
syntactical components of the associated proposition, the tuple is relevant for answer-
ing questions that pertain to any of the discourse entities contained in any of these
components. Accordingly, as an indexing mechanism, we create “concept nodes” for
all such entities and associate the tuple with each such node. For these purposes, a
discourse entity is taken to be each noun phrase contained in a subject, indirect ob-
ject, and direct object, and the noun phrase objects of prepositional modifiers. Direct
quotes, subordinate clauses, and other phrasal concepts are not further decomposed at
this stage. For example, given the sentence “Peter enjoyed reading the book that
Mary recommended,” we generate the simple proposition tuple <'Pe-
83
ter','enjoyed','reading',null,null,'the book that Mary recommended'>, from which we
create concept nodes for 'Peter', 'Mary', and 'the book that Mary recommended'.
To construct a concept set that will support complex question answering, we con-
sider, for example, a document containing the sentences:
The Russian submarine Kursk sank in the Barents Sea.
The Russian submarine Kursk sank in deep water.
If the question were “Where did the submarine sink?” both “in the Barents Sea” and
“in deep water” could arguably be acceptable answers. However, if the question
were “In what sea did the submarine sink?” then only the first would be acceptable.
Since the propositions represented by these sentences are syntactically identical, we
distinguish them by postulating a mechanism by which we recognize “the Barents
Sea” as an instance of the “sea” category contained in the question.
The method for distinguishing these cases is to encode all “is-a” relationships as
explicit parent-child relationships among the corresponding concept nodes. Thus, if
the document also contained a sentence stating, essentially, that the Barents Sea is a
sea, this would be sufficient. However, we observe anecdotally that neither ordinary
conversation nor newswire text typically contains such explicit categorizations.
Therefore, as a corollary to the rule above, we postulate the need for deriving parent-
child node relationships from individual concept nodes so that, in the example above,
“the Barents Sea” is a “sea” because the lowercase head noun “sea” can be found to
be a common noun, of which “the Barents Sea” is therefore an instance.
Table 1 list a number of parent-child derivations that the system described in Sec-
tion 6 has implemented. The arrows in the “Example” column of the table run from
the child concept to the parent concept. No doubt additional derivations may be
found; however, the set presented serves to illustrate the mechanism.
Table 1. Parent-child concept derivations.
____________________________________________________________________
Derivation Example
Common noun-proper noun “space shuttle Atlantis”-->“space shuttle”
Common noun-common noun “oil tanker” --> “tanker”
Proper noun with preposition “King of England” --> “king”
Common noun parent of proper “Nobel Prize” --> “prize”
Proper noun-common noun “Cadillac sedan” --> “Cadillac”, ”sedan”
Adjective-common noun “Russian submarine” --> “submarine”
Multiple proper names “Peter and Paul” --> “Peter”, “Paul”
NP coordination “fish and chips” --> “fish”, “chips”
NP following preposition “jar of beans” --> “jar”
NP following adjective/adverb “fast cars” --> “cars”
NP following prep in proper noun “Nobel Prize for Physics”>“Nobel Prize”
Possessive form “Mary's car” --> “car”
Business entity suffix “IBM Corp.” --> “IBM”
Comma-separated location “Normandy, France” --> “France”
Concept prefixed by ordinal “49
th
pageant” --> “pageant”
Title before proper name “Miss Lara Dutta” --> “Lara Dutta”
Cardinal number prefix “five coins” --> “coins”
Concept begins with dollar sign “$200 Million” --> “dollar”
Subordinate clause without “that” “the book Mary read” --> “book”
84
To support these capabilities, we define a “concept node” to be the 4-tuple:
<name,{parents},{children},{tuples}>, where “name” refers to the noun
phrase for the discourse entity for which the node is constructed, “{parents}” and
“{children}” refer to the (possibly empty) sets of parent and children nodes for the
concept, and “{tuples}” refers to the tuples in which the concept (discourse entity)
appears.
These derivations are repeatedly applied to a concept extracted from a tuple com-
ponent until it cannot be decomposed further. Thus, for the concept “space shuttle
Discovery”, application of the last derivation followed by the one above it, produces
the links that establish that “space shuttle Discovery “ is-a “space shuttle” is-a “shut-
tle”, which has the beneficial effect of rooting a set of concepts in a common noun
which, in this case, can be found in WordNet. This supports using synonyms when
answering queries concerning the nodes so rooted, as described in Section 5.
A node may have more than one parent. For example, a node for “John Hancock”
may have parent links to both “insurance company” and “revolutionary war hero.”
Similarly, a node may have more than one child. For example, the concept “subma-
rine” may have children “Russian submarine” and “attack submarine”. Thus, concept
nodes constructed in this manner are organized into a network of one or more disjoint
sub-networks, each containing one or more related concept nodes.
When concepts are indexed in the network, “nicknames” are generated for proper
nouns so that only one node is created for each proper noun concept, and each of its
nicknames is mapped as an alias to the concept. Thus, for example, the phrase “John
Kennedy” will be mapped to the concept node “John F. Kennedy”. The concept addi-
tion logic is so constructed that for such proper noun associations, the longer concept
is maintained as the concept node. If a shorter concept is encountered first, then when
the related longer concept phrase is added to the network, it subsumes the shorter one
by adjusting all necessary parent and child links and adding an alias for the shorter
one linking it to the master concept node.
4 Question Analysis
Questions are decomposed into proposition tuples in the same manner as text, with
the addition of free variables for the desired answer. These variables may take the
form of a general directive, such as “*who”, “*what”, *when”, “*where”, and
“*why”, or a target preposition type, such as “*in”, when the answer is expected to be
modified by a preposition. We also define the answer variable “*ans” to serve as a
referent to a candidate answer obtained in response to a previous tuple, so that we can
construct question patterns as boolean combinations of separate tuple patterns.
For simple questions, a single tuple pattern may suffice. For example, for the ques-
tion “What did Peter eat?” the corresponding question tuple is <'Peter','did eat',_,
_,_,”*what”>, where we have indicated the free variable with a leading asterisk and
null values with underscores for clarity of display.
More complex questions require Boolean combinations of question pattern tuples.
For example, the question “What kind of car does Peter drive?” the question decom-
poses into the conjunction:
<'Peter','does drive',_,_,_,”*what”> <and> <'*ans','is',_,_,_,”car”>
85
Peter may drive his mother crazy or a nail with a hammer, but neither is a “car” and
will not be returned as an answer because the second pattern would not match.
In a similar fashion, some questions require disjunctive combinations of question
pattern tuples to accommodate alternative syntactic realizations of the expected an-
swer. For the question “What is the title of the book?” the question decomposes into:
<'*what','is',_,_,_,'the title of the book'> <or>
<'the title of the book','is',_,_,_,'*what'>
Still other questions may require both disjunctive and conjunctive patterns. For ex-
ample, the question “What kind of book did Peter give to Mary?” the question de-
composes into: ( <'Peter',’did give',__,’to Mary,’,__,*what’'> <or>
'Peter',’did give',__,_,Mary,’*what’'> ) <and>
<'*ans','is',_,_,_,'book'>
which is needed to find answers where the document base contains sentences of the
form “Peter gave Mary the novel,” and “Peter gave the novel to Mary.”
The final step in question analysis involves creating question tuples for the passive
forms of active question constructions, and vice versa, and for possessive forms. Pat-
terns produced at this step are added disjunctively to the existing question tuple set.
5 Question Answering
Question answering is a three-stage process in which: (a) the question is analyzed as
described in the preceding section; (b) the tuples of interest are retrieved from the
concept network; and (c) the tuples retrieved then examined to search for candidate
answers. The second and third stages are discussed separately in this section.
5.1 Tuples of Interest
Since the concept network is, by construction, indexed by the concept names it in-
cludes, it is not necessary to search the network for candidate propositions. Thus, the
network supports incremental growth as a body of knowledge is acquired, with mini-
mal performance penalty for such growth. The tuples of interest are retrieved as fol-
lows. First, noun phrases are extracted from the various components of the question
tuple or tuples, and for each such phrase, the concept network is checked for the pres-
ence of a concept node by that phrase, or the presence of a node to which such phrase
is mapped as an alias. When a relevant concept node is found, all tuples associated
with that node are included in the return set. Thus, for a question inquiring about
“John Kennedy”, the tuples indexed by the “John F. Kennedy” node would be re-
turned. The child nodes of the concept node are also examined recursively, and any
new tuples found are also added to the return set. Once all noun phrases are checked
against the network, the resultant tuples are returned as the set of tuples of interest.
By construction, every tuple in this set contains one or more of the phrases of interest
from the question, and no tuples in the concept network that contain one of the de-
sired phrases are omitted.
86
5.2 Question Pattern Matching
Question pattern matching proceeds by a straightforward unification algorithm in
which the entire boolean combination of question tuples is applied to each tuple until
an answer.
The examination of a single proposition tuple proceeds by checking its tuple com-
ponents against the non-null components of the question pattern. For each such com-
ponent that does not involve a free (answer) variable, a matching algorithm is exe-
cuted. If any such component fails to match, the proposition tuple is rejected and ex-
amination proceeds to the next tuple in line. For subjects and direct objects, an ex-
amination set is created consisting of all system aliases for a proper noun phrase, if
any, otherwise the common noun phrase itself, augmented by all aliases for the target
if the phrase contains a member of the target synset. For example, if the target is
“Russian submarine Kursk” and the proposition tuple contains the word “submarine”,
then all target aliases, including the word “Kursk” are included, so that a match will
be found against a question pattern in which “Kursk” is specified.
For verbs, the main verb from the proposition tuple is extracted and WordNet
methods are used to obtain all verb roots. Each root is then compared against the
WordNet root of the main verb in the question tuple and a match is found if any root
of one is found in any synset of the WordNet hypernym expansion of the other. This
algorithm will find a match between the verbs “give” and “transfer”, for example,
since “transfer” appears in the hypernym tree for “give.”
Gerunds and infinitives are matched in the same manner as main verbs, except that
the infinitive “to do” in the question pattern is considered to match any infinitive that
is present in the proposition tuple.
Indirect objects receive special treatment, since a nominal indirect is semantically
equivalent to a similarly structured sentence in which the same noun phrase occurs as
the object of a “to” or “for” preposition. Accordingly, the proposition tuple's nominal
indirect is checked, if any, and if none, then its modifiers commencing with “in” or
“for”, if any, are examined.
Modifiers also receive special treatment since there can be more than one in either
the question pattern or the proposition tuple. Where the question pattern contains
more than one modifier, a match is recorded only if all modifiers are matched inde-
pendently. However, so long as all question modifiers are matched, it does not matter
that a proposition tuple may have additional, unmatched modifiers, since these are, by
construction, irrelevant to the question. We seek to achieve robustness in matching
through examinations of WordNet synonyms for a noun phrase head, whenever pos-
sible, which obtains for most common nouns and some proper nouns.
Now, where the question pattern component contains the answer variable “*ans”,
an answer retrieval algorithm is executed according to the component type. For sub-
jects and direct objects, the algorithm returns the corresponding component of the
proposition tuple if the expected answer type is confirmed through WordNet. For
example, a location type is confirmed if “structure” or “location” appears in the head
noun's hypernym tree. Similarly, verbs and other tuple components are retrieved after
appropriate type checks by algorithms that parallel their corresponding match meth-
ods.
If a match is not obtained for a given proposition tuple, then the reciprocal predi-
cate is checked if the subject and direct object are both nonempty, there is no ger-
und/infinitive, and the main verb is not copular, not the verb “do”, and possesses an
87
antonym in WordNet. This construction increases recall by supporting a match for the
question “What did Mary receive?” where the database contains, for example, “Peter
gave Mary the book.” Once the reciprocal tuple is formed, it is matched in the same
manner described above for the direct pattern.
Different processing is performed for question patterns that seek to confirm “is-a”
relationships for previously found candidate answers. Where, for example, we seek to
confirm whether a candidate answer for the location of the sinking of the submarine
Kursk is in fact an instance of the “sea” class, the parents of the candidate answer are
searched recursively to the head of their equivalence network for the desired class. If
none is found, then the WordNet hypernym tree for each root of the network is
checked. We note that there can be more than one such root, since a concept node can
have more than one parent. If the desired class name is found in the hypernym tree for
any node examined, then the candidate concept is deemed a member of the class and
a match is recorded.
6 A Test Implementation
The algorithms described above were implemented in our Semantic Extractor (SE-
MEX) tool, a Java application for testing semantic extraction algorithms. SEMEX
incorporates the Brill tagger [3], the Cass partial parser [1], and a comprehensive set
of empirically derived grouping and sentence splitting heuristics, plus rudimentary
pronomial coreference resolution, to provides a graphical user interface for viewing
intermediate results at each stage of processing, from POS tagging, through parsing,
sentence splitting, syntactic role assignment and resolution, and concept extraction.
SEMEX was configured to exercise the TREC 2005 QA questions against the top-
50 documents for each target returned by NIST's generic IR engine from the
AQUAINT newswire collection. SEMEX output for the first 200 factoid questions
was analyzed in detail and scored manually using the TREC-furnished factoid an-
swer patterns, adjusted to eliminate the limitations of the tool and document set. Thus,
a question was discarded from the analysis if: (a) its answer was not in the top-50
documents; (b) the document containing the answer was not readable by SEMEX; (c)
the answer was in a direct quote, since SEMEX was not configured to look within
direct quotes; (d) the answer was in the dateline or headline; (e) the question form
was not one of the types that was implemented in SEMEX; or (f) the answer pattern
provided by TREC was a wrong answer, or no answer, to the question. Similarly, an
answer was considered found successfully if only a minor manual correction to the
parsed output of text or question using the SEMEX GUI was sufficient to allow SE-
MEX to run on the question, as this were considered a near-term programming im-
provement for the tool. Importantly, however, questions whose answers required rea-
soning over the concept set or more than minor syntactical parsing adjustments were
retained as failures, as these indicated areas of further inquiry.
On this basis, SEMEX was able to answer correctly 66 out of the 135 non-
discarded questions, for an adjusted score of 48%. Of the 66 correct answers, 22 were
found without manual intervention, for a raw score of 16%. Both of these scores com-
pare favorably with other systems, as [10] reports factoid answer accuracies in a
range from 21.3% to 77.0% for the top 10 runs submitted, with only the top 3 systems
achieving factoid accuracy scores greater than 35%. And with sixty-three runs sub-
88
mitted to the QA track, it is clear that most of the systems achieved factoid accuracies
below the lower bound of this range.
Of the 65 discards, 29 involved questions that had no answers in the documents,
14 involved the headline or dateline, 9 involved documents that could not be read, 6
involved question types that were not implemented, and 2 had incorrect answers.
Generally, the newswire articles that comprised the document sets provided chal-
lenges to our implementation, as they were characterized by numerous spurious
HTML codes and meta-data within the text, such as identification of sources, sports
scores, and multiple datelines within the text. Where acceptable parses were obtained,
however, the algorithms proposed operated well. Algorithm success and limitations
are well illustrated in a separate test where for the target “Russian submarine Kursk
sinks” TREC question 66.7 asked the list question, “Which U.S. submarines were
reportedly in the area?” For this question, SEMEX generated the question pattern:
<or> [S:*which][V:were][GI:][M:reportedly;in the area][IO:][DO:]
<and> [S:*ans][V:is][GI:][M:][IO:][DO:U.S. submarine]
and returned the answer “the Toledo”. What is significant about this result is that the
document text consisted of three sentences, the first two of which were:
<P> The second U.S. submarine in the Barents Sea when the Kursk sank
was the Toledo, a Russian news agency reported Thursday. </P> <P>
The agency, Interfax, said the Toledo was in the area along with another
U.S. submarine, the Memphis, during the Russian naval exercises in mid-
August, when the Kursk sank, with the loss of 118 lives. </P>
In this example, the fact that the Toledo was a U.S. submarine is established in the
first sentence of the document, while the fact that it was reportedly in the area when
the Kursk sank was furnished by the second sentence. The connection between these
two facts is provided by the concept network, in which the “Toledo” concept has a
parent, “second U.S. submarine in the Barents Sea when the Kursk”, which by opera-
tion of the derivations of Table 1 in Section 3.3, has a parent, “second U.S. subma-
rine” which, by a further derivation has a parent, “U.S. submarine”, which is the de-
sired class. We note that SEMEX was unable to find the second valid answer, “the
Memphis”, since the SEMEX did not understand that vessel to have been in the area,
although the network structure shows that it was recognized as a U.S. submarine.
Similarly, TREC question 87.3 asked the factoid question, “What Nobel Prize was
Fermi awarded in 1938?” for which SEMEX correctly returned “the Nobel Prize for
Physics”, which was made possible by the parent-child relationship present in the
concept network between “Nobel Prize” and the answer returned. Indeed, a separate
test has shown that SEMEX would still return the same result if the question had
asked what “prize” was Fermi awarded, again as a consequence of the concept net-
work, and also if the question had asked about an “award”, which though not in the
network was nevertheless a WordNet hypernym of “Nobel Prize” found by SEMEX.
7 Conclusions
The test implementation of the algorithms proposed demonstrated the value of a con-
cept network of parent-child concept relationships and indexing proposition tuples
that encode the relationships among the syntactical components of atomic sentences.
The results also establish the robustness that may be achieved through the use of
89
WordNet. We note, however, that the current implementation does not explicitly dis-
ambiguate words but relies instead on the constellation of syntactical components in a
proposition tuple to serve as selectional restrictions when answering questions. It re-
mains a question for further investigation whether a separate disambiguation module
is necessary. Further research is also needed to establish the extent to which a concept
network of the type proposed can support reasoning and advanced question answer-
ing.
References
1. Abney, S.: Partial Parsing via Finite-State Cascades. In: Proceedings of Workshop on Ro-
bust Parsing, 8
th
European Summer School in Logic, Language and Information. Prague,
Czech Republic (1996) 8-15.
2. Ahn, K., Bos, J., Clark, S., Curran, J., Dalmas, T., Leidner, J., Smillie, M., Webber, B.:
Question Answering with QED and Wee at TREC-2004. In: Voorhees, E.M., Buckland,
L.P. (eds.): Proceedings of the Thirteenth Text REtrieval Conference (TREC 2004), NIST
Special Publication 500-261. Gaithersburg, MD (2005)
3. Brill, E.: Some Advances in Part of Speech Tagging. In: Proceedings of the Twelfth Na-
tional Conference on Artificial Intelligence (AAAI-94). Seattle, Washington (1994)
4. Fellbaum, C. (ed.): WordNet: An Electronic Lexical Database. The MIT Press, Cambridge
London (1998)
5. Gaisauskas, R., Greenwood, M., Hepple, M., Roberts, I., Saggion, H.: The University of
Sheffield's TREC 2004 Q&A Experiments. In: Voorhees, E.M., Buckland, L.P. (eds.): Pro-
ceedings of the Thirteenth Text REtrieval Conference (TREC 2004), NIST Special Publica-
tion 500-261. Gaithersburg, MD (2005)
6. Greenwood, M., Roberts, I., Gaizauskas, R.: The University of Sheffield TREC 2002 Q&A
System. n: Voorhees, E.M., Buckland, L.P. (eds.): Proceedings of the Eleventh Text RE-
trieval Conference (TREC 2002), NIST Special Publication 500-251. Gaithersburg, MD
(2003)
7. Harabagiu, S., Moldovan, D., Clark, C., Bowden, M., Williams, J., Bensley, J.: Answer
Mining by Combining Extraction Techniques with Abductive Reasoning. In: Voorhees,
E.M., Buckland, L.P. (eds.): Proceedings of the Twelfth Text REtrieval Conference (TREC
2003), NIST Special Publication 500-255. Gaithersburg, MD (2004) 375-382
8. Litkowski, K.: Question Answering Using XMS-Tagged Documents. In: Voorhees, E.M.,
Buckland, L.P. (eds.): Proceedings of the Eleventh Text REtrieval Conference (TREC
2002), NIST Special Publication 500-251. Gaithersburg, MD (2003) 156-165
9. Schone, P., Ciany, G., McNamee, P., Mayfield, J., Bassi, T. Kulman, A.: Question Answer-
ing with QACTIS at TREC-2004. In: Voorhees, E.M., Buckland, L.P. (eds.): Proceedings
of the Thirteenth Text REtrieval Conference (TREC 2004), NIST Special Publication 500-
261. Gaithersburg, MD (2005)
10. Voorhees, E.M.: Overview of the TREC 2004 Question Answering Track. In: Voorhees,
E.M., Buckland, L.P. (eds.): Proceedings of the Thirteenth Text REtrieval Conference
(TREC 2004), NIST Special Publication 500-261. Gaithersburg, MD (2005)
90