Proceedings of HLT/EMNLP 2005 Demonstration Abstracts, pages 8–9,
Vancouver, October 2005.
MindNet: an automatically-created lexical resource 
 
Lucy Vanderwende, Gary Kacmarcik, Hisami Suzuki, Arul Menezes 
Microsoft Research 
Redmond, WA 98052, USA 
{lucyv, garykac, hisamis, arulm}@microsoft.com 
 
Abstract 
We will demonstrate MindNet, a lexical resource 
built automatically by processing text.  We will 
present two forms of MindNet: as a static lexical 
resource, and, as a toolkit which allows MindNets 
to be built from arbitrary text.  We will also intro-
duce a web-based interface to MindNet lexicons 
(MNEX) that is intended to make the data con-
tained within MindNets more accessible for explo-
ration.  Both English and Japanese MindNets will 
be shown and will be made available, through 
MNEX, for research purposes. 
1 MindNet 
A MindNet is a collection of semantic relations 
that is automatically extracted from text data using 
a broad coverage parser. Previous publications on 
MindNet (Suzuki et al., 2005, Richardson et al., 
1998, Vanderwende 1995) have focused on the 
effort required to build a MindNet from the data 
contained in Japanese and English lexicons. 
Semantic Relations 
The semantic relations that are stored in MindNet 
are directed, labeled relationships between two 
words; see Table 1:  
Attributive Manner Source 
Cause Means Synonym 
Goal Part Time 
Hypernym Possessor TypicalObject 
Location Result TypicalSubject 
Table 1: A sampling of the semantic relations stored in 
MindNet 
 
These semantic relations are obtained from the 
Logical Form analysis of our broad coverage 
parser NLPwin (Heidorn, 2000).  The Logical 
Form is a labeled dependency analysis with func-
tion words removed.  We have not completed an 
evaluation of the quality of the extracted semantic 
relations.  Anecdotally, however, the quality varies 
according to the relation type, with Hypernym and 
grammatical relations TypicalSubject and Typi-
calObj being reliable, while relations such as Part 
and Purpose are less reliable. By making MindNet 
available, we solicit feedback on the utility of these 
labeled relationships, especially in contrast to sim-
ple co-occurrence statistics and to the heavily used 
hypernymy and synonymy links. Furthermore, we 
solicit feedback on the level of accuracy which is 
tolerable for specific applications. 
Semantic Relation Structures 
We refer to the hierarchical collection of semantic 
relations (semrels) that are automatically extracted 
from a source sentence as a semrel structure. Each 
semrel structure contains all of the semrels ex-
tracted from a single source sentence.  A semrel 
structure can be viewed from the perspective of 
each unique word that occurs in the structure; we 
call these inverted structures.  They contain the 
same information as the original, but with a differ-
ent word placed at the root of the structure. An ex-
ample semrel structure for the definition of 
swallow is given in Figure 1a, and its inversion, 
from the perspective of wing is given in Figure 1b: 
 
swallow           wing 
 Hyp bird           PartOf bird 
       Part wing             Attrib small 
       Attrib small          HypOf swallow 
 
Figure 1a and b: Figure 1a is the semrel structure for the 
definition of swallow1, Figure 1b the inversion on wing. 
2 MNEX 
MNEX (MindNet Explorer) is the web-based inter-
face to MindNet that is designed to facilitate 
browsing MindNet structure and relations. MNEX 
displays paths based on the word or words that the 
                                                           
1
 Swallow: a small bird with wings (LDOCE).  Definition 
abbreviated for purposes of exposition.   
8
user enters. A path is a set of links that connect one 
word to another within either a single semrel struc-
ture or by combining fragments from multiple 
semrel structures.  Paths are weighted for compari-
son (Richardson, 1997). Currently, either one or 
two words can be specified and we allow some 
restrictions to refine the path search.  A user can 
restrict the intended part of speech of the words 
entered, and/or the user can restrict the paths to 
include only the specified relation. When two 
words are provided, the UI returns a list of the 
highest ranked paths between those two words. 
When only one word is given, then all paths from 
that word are ranked and displayed.  Figure 2 
shows the MNEX interface, and a query requesting 
all paths from the word bird, restricted to Noun 
part of speech, through the Part relation:  
 
 
Figure 2: MNEX output for “bird (Noun) Part” query 
3 Relation to other work 
For English, WordNet is the most widely used 
knowledgebase. Aside from being English-only, 
this database was hand-coded and significant effort 
is required to create similar databases for different 
domains and languages. Projects like EuroWord-
Net address the monolingual aspect of WordNet, 
but these databases are still labor intensive to cre-
ate.  On the other hand, the quality of the informa-
tion contained in a WordNet (Fellbaum et al., 
1998) is very reliable, exactly because it was 
manually created.  FrameNet (Baker et al., 1998) 
and OpenCyc are other valuable resources for Eng-
lish, also hand-created, that contain a rich set of 
relations between words and concepts. Their use is 
still being explored as they have been made avail-
able only recently. For Japanese, there are also 
concept dictionaries providing semantic relations, 
similarly hand-created, e.g., EDR and Nihongo 
Goi-taikei (NTT). 
The demonstration of MindNet will highlight 
that this resource is automatically created, allowing 
domain lexical resources to be built quickly, albeit 
with lesser accuracy.  We are confident that this is 
a trade-off worth making in many cases, and en-
courage experimentation in this area.  MNEX al-
lows the exploration of the rich set of relations 
through which paths connecting words are linked. 
4 References 
Baker, Collin F., Fillmore, Charles J., and Lowe, John 
B. (1998): The Berkeley FrameNet project. in Pro-
ceedings of the COLING-ACL, Montreal, Canada. 
Fellbaum, C. (ed). 1998. WordNet: An Electronic Lexi-
cal Database. MIT Press. 
Heidorn, G. 2000. Intelligent writing assistance. in 
R.Dale, H.Moisl and H.Somers (eds.), A Handbook 
of Natural Langauge Processing: Techniques and 
Applications for the Processing of Language as Text. 
New York: Marcel Dekker. 
National Institute of Information and Communications 
Technology. 2001. EDR Electronic Dictionary Ver-
sion 2.0 Technical Guide. 
NTT Communications Science Laboratories. 1999. Goi-
Taikei - A Japanese Lexicon. Iwanami Shoten. 
OpenCyc. Available at: http://www.cyc.com/opencyc. 
Richardson, S.D. 1997, Determining Similarity and In-
ferring Relations in a Lexical Knowledge Base. PhD. 
dissertation, City University of New York. 
Richardson, S.D., W. B. Dolan, and L. Vanderwende. 
1998. MindNet: Acquiring and Structuring Semantic 
Information from Text, In Proceedings of ACL-
COLING. Montreal, pp. 1098-1102. 
Suzuki, H., G. Kacmarcik, L. Vanderwende and A. 
Menezes. 2005. Mindnet and mnex. In Proceedings 
of the 11th Annual meeting of the Society of Natural 
Language Processing (in Japanese).  
Vanderwende, L. 1995. Ambiguity in the acquisition of 
lexical information. In Proceedings of the AAAI 
1995 Spring Symposium Series, symposium on rep-
resentation and acquisition of lexical knowledge, 
174-179. 
9
