Dependency Based Logical Form Transformations 
Stephen Anthony and Jon Patrick  
School of Inform ation Technologies 
The University of Sydney 
Sydney, Australia 2006 
{stephen,jonpat}@it.usyd.edu.au 
 
Abstract 
This paper describes a system developed for the 
transformation of English sentences into a first 
order logical form representation. The metho d-
ology is centered on the use of a dependency 
grammar based parser . We demonstrate the suit-
ability of applying a dependency parser based 
solution to the given task and in turn explain 
some of the limitations and challenges involved 
when using such an approach. The efficiencies 
and deficiencies of our approach are discussed 
as well as considerations for further enhanc e-
ments.  
1 Introduction 
In addition to the well-known all words and lexi-
cal sample tasks deployed in previous Senseval 
workshops a number of new tasks have been in-
cluded in this sense evaluation. These new tasks in-
clude identification of semantic roles as in FrameNet 
(Gildea and Jurafsky 2002), disambiguation of 
WordNet glosses  (Miller 1990; Fellbaum 1998; 
Harabagiu, Miller et al. 1999), automatic acquisitio n 
of subcategorisation frames (Korhonen 2002; Preiss 
and Korhonen 2002), and Logic al Form Identific a-
tion (LFI) (Rus 2002; Rus and Moldovan 2002). 
This paper discusses a solution developed for the 
LFI task. The approach used here employs a func-
tional dependency parser (Järvinen and Tapanainen 
1997; Tapanainen and Järvinen 1997) and uses a 
limited number  of additional resources. This contri-
bution is intended to demonstrate the suitability of a 
dependency parser to the given task and also explain 
some of the limitations and challenges  involved 
when using such an approach. 
1.1 Motivation 
Part of the initial step towards the interpretation of 
a sentence as postulated by Hobbs et al. (1993) in-
volves the proof of the logical form of a sentence. 
This statement entails the transformation of a sen-
tence into a logical form as a fundamental building 
block towards sentence interpretation.  
 
Advantages specifically related to the utilis ation 
of logical forms in language processing include a 
simplified interface between syntax and semantics, a 
natural and easily exploitable representation of syn-
tactic arguments, and the potential for formation of 
conceptual predicates (Rus 2002) if predicates are 
disambiguated with respect to a general ontology 
such as WordNet.  
1.2 Task Description 
The Logic al Form (LF) employed in this task is a 
flat, scope-free first order logic representation that 
embeds lexical and syntactic information. A predi-
cate is generated for every nominal, verbal, adjecti-
val, and adverbial content word. The name of the 
predicate is a concatenation of the lemmatised word 
form and part-of-speech category. The sentence be-
low is followed by its corresponding target logical 
form representation. 
 
Some students like to study in the mornings. 
 
student:n_ (x3) like:v_ (e4, x3, e6) to (e5, e6) 
study:v_ (e6, x3, x9) in (e7, x9) morning:n_ (x9). 
 
Relationships between predicates are shared 
through their arguments. The two types of argu-
ments used are events (e) and entities (x). Using the 
transformation shown above as an example, the 
event predicate ‘like’ is labeled as e4 and has sub ject 
argument x3 which corresponds to ‘student’ and 
grammatical object argument e6 which corresponds 
to the ‘study’ event predicate.  
 
The remainder of the argument slots are reserved 
for indirect and prepositional objects.  Determiners, 
plurals, negation, auxiliaries, verb tenses, and punc-
tuation are excluded from the final representation.  
2 Methodology 
The system is built using a highly modular design 
and is intended to be as generic and reusable as pos-
sible. The basic data structure is a flat list-like repre-
sentation with generic property slots attached to each 
element. This structure maximises compatibility 
with the final representation and allows for greater 
flexibility in the types of information that may be 
                                             Association for Computational Linguistics
                        for the Semantic Analysis of Text, Barcelona, Spain, July 2004
                 SENSEVAL-3: Third International Workshop on the Evaluation of Systems
associated with each predicate. Figure 1 illustrates 
the major proces s ing modules available and the 
work flow. 
 
Figure 1: Logical form identification work flow 
A syntactic parse including functional dependen-
cies is produced on a per sentence basis. Definitions 
of the properties associated with each token are pre-
sented in Table 1.  
 
Attribute Value 
Word ID Integer sentence position 
Head ID Integer position of head dependency 
Text The original word form 
Lemma Lemmatised word form 
Morpho Morphological function tags. Parts 
of speech and sub-features  
Syntax Surface syntactic tags 
Depend Dependency functions 
MAIN Main element 
SUBJ  Position of syntactic subject 
OBJ Syntactic object position 
I-OBJ  Indirect object position 
COMP Position of syntactic complement 
PCOMP Prepositional complement position 
DET Determiner dependent 
ATTR Attributive nominal 
CC Coordinating conjunction 
GOAL  Position of goal 
OC Object complement 
Table 1: Linguistic information stored for each  
token 
The resultant parse is transformed into a linear 
data structure indexed by word position. This is il-
lus trated in Table 2 using the example sentence 
‘Some students like to study in the mornings’. The 
original token text is stored, as is the lemmatised 
form.  
Word
ID
 
Lemma
 
Head
ID
 
De
pend
 
Text
 
Mo
rpho
 
Syntax
 
2 some 3  det Some DET  >N 
3  stu-
dent 
4  subj stu-
dents 
N NOM 
PL 
NH 
4  like 1 main  like V PRES  VA 
5  to 6  pm  to INF 
MARK 
AUX 
6 study 4  obj study V INF VA 
7  in 6  tmp  in  PREP EH 
8  the 9  det  the DET  >N 
9 morn
ing 
7 pcomp morn-
ings 
N NOM 
PL 
NH 
Table 2: Example syntactic parse 
Head and dependency type are the most important 
class of information used by the system. The de-
pendency type and head of the token is often di-
rectly, if not indirectly, translatable into a predicate 
argument. Examples of the types of dependency 
functions employed include subject, object, preposi-
tional complement, agent, subject and object com-
plements, indirect object, goal, and coordinating 
conjunctions. Determiner and negator functions are 
also of interest because they are excluded from the 
final represent ation.  
 
The filter module moderates the presence or ab-
sence of tokens using stop lists or pass lists or a 
combination of both. Stop lists are used to specify 
content to be excluded from the token stream and 
pass lists specify elements that should remain. To-
kens may be filtered from the stream based on any 
attribute type and value listed in Table 1. This in-
formation is provided in the filter set. The principal 
types of information filtered in this system are de-
terminers based on morpholog ical tags and auxilia-
ries based on syntactic tag inform ation. For example 
‘some’ and ‘the’ are filtered as a cons equence of a 
morpho property equals ‘DET’ stop list rule . 
 
When the token stream has been annotated with 
the necessary information and has passed through 
the filter, the tokens that remain are passed through 
the logical form processor (LFP). The main function 
of the LFP is to build an inverted index identifying 
all dependent tokens. Once grammatical dependen-
cies  are assigned and the inverted index is built the 
logical form representation may be constructed. 
Each predicate is constructed from the token stream 
in turn based on the part-of-speech category of the 
token. The base form of the token is concatenated 
with the part-of-speech tag. A mapping table is used 
to transform the part-of-speech information pro-
     Filters 
Logical Form 
processor 
Functional 
dependency 
parser 
Natural lan-
guage sen-
tences  
Target logical 
form 
Filter 
set 
duced by the parse into the coarser grained Wor dNet 
tags. 
 
Entities are the simplest type of predicate to con-
struct as they contain only a single argument, for 
which the word identifier attribute value is used. 
Noun tokens ‘student’ and ‘morning’ from the ex-
ample are transformed into the predicates             
student :n_(x3) and morning:n_(x9). Pronouns, 
prepositional complements, and coordinating con-
junctions are dealt with individually using their re-
spective dependency function values. 
 
Adjectives are constructed using the head depend-
ency value as the argument unless the dependent is 
marked with a subject. In this case the argument be-
comes the head of the subject. Adverbs are created 
primarily using the dependency function alone. 
 
Verbal predicates are constructed using SUBJ, 
OBJ, GOAL, OC, I-OBJ, COMP, and PCOMP de-
pendencies in the specified order. A special case 
exists for verbs that have object complement de-
pendencies. In these cases attributive nominals are 
identified and assigned as arguments independently.  
 
The main verb ‘like’ in our example is trans-
formed into the pred icate like:v_(e4, x3, e6) as a 
result of subject (SUBJ) and object (OBJ) dependen-
cies found in ‘student’ and ‘study’ respectively. 
Given the fact that we are dealing with the main 
verb, the LFP inverts the subject and object depend-
encies, inserts them into the head verb token prop-
erty slot and assigns their respective word identifier 
values. The inverted properties augment the token 
slot for ‘like’ which has word identifier four in Table 
2. The additional elements of the inverted index used 
to build the predicate are listed in Table 3. 
 
Attribute Value 
OBJ 6 
SUBJ  3 
depend main 
head 1 
lemma like 
morpho V PRES  
syntax VA 
text like 
Table 3: Augmented token slot for ‘like’ 
Verbal predicates which also serve as grammatical 
objects also warrant special treatment. The token 
‘study’ is an example of this as it serves as the object 
of the head verb ‘like’.  A cache is used to store the 
sentential head, prepositional complements, subjects, 
and coordinating conjunctions. The cache is used in 
this instance to assign the subject and prepositional 
complement arguments in order to form the predi-
cate study:v_(e6, x3, x9). Notice from Table 2 word 
identifier three matches the grammatical subject to-
ken ‘students’ and word identifier nine matches the 
head of the prepositional phrase ‘ in the mornings’. 
 
Once all tokens are processed the logical form 
transformation is complete and the final representa-
tion is presented in the aforementioned notation. 
3 Evaluation 
Argument, predicate, and sentence level precision 
and recall measures are used to evaluate perform-
ance of the system as compared to a gold-standard. 
The system was trained on a set of 50 sentences with 
corresponding logical forms. Final testing was per-
formed on a set of 300 LF -sentence pairs.  
3.1 Argument Level 
Precision at the argument level is defined to be the 
number of correctly identified arguments divided by 
the number of all identified arguments. Recall is de-
fined to be the number of correctly identified argu-
ments divided by the real number of arguments that 
should be present in the target transformation. 
3.2 Predicate Level 
Predicates must identify all arguments correctly to 
be counted as a correct predicate. Precision is de-
fined to be the number of correctly ident ified predi-
cates divided by the number of all attempted 
predicates. Recall is defined as the number of cor-
rectly identified predicates divided by the real num-
ber of predicates that were supposed to be identified 
in the target transformation.  
3.3 Sentence Level 
Various oth er sentence level measures are also 
used. Sentence-argument is defined as the number of 
sentences that have all arguments correctly identi-
fied divided by the number of sentences attempted. 
Sentence-predicate is similar except conditioned on 
predicates. Sentence-argument-predicate is defined 
to be the number of sentences that have all argu-
ments correctly identified divided by the number of 
sentences which have all predicates correctly identi-
fied. Sentence-argument-predicate-sentences refers 
to the number of sentences that have all arguments 
and all pred icates correctly identified divided by the 
number of sentences attempted. 
4 Results 
As stated earlier the final evaluation was con-
ducted on a set of 300 sentence-LF pairs. Table 4 
lists the evaluation precision and recall results using 
the measures discussed in section 3 which have been 
converted into percentages.  
 
Evaluation Measure Score 
Argument Precision 76.4 
Argument Recall 65.6 
Predicate Precision 84.0 
Predicate Recall 85.0 
Sentence-Argument 16.0 
Sentence-Predicate 35.3 
Sentence-Argument-Predicate 38.7 
Sentence-Argument-Predicate-Sentences  13.7 
Table 4: Evaluation results as percentages  
The major source of error in terms of arguments 
originated from the parser’s inappropriate ha ndling 
of coordinating conjunctions. Another common 
source of error arose from poor handling of nominal 
group complexes. With regard to predicate perform-
ance, the decision to forfeit the use of the available 
multi-word item list proved costly.  
5 Future Work  
Harabagiu et al. (1999) proposed a scheme for at-
taching sense tags to predicates within the frame-
work of transforming WordNet glosses into a logical 
form. In this way conceptual predicates may be 
formed to manipulate a meaning representation in 
more significant ways. Naturally the sense inventory 
must be sensitive enough to allow for a meanin gful 
and representative mutation to be applied to the 
meaning representation.  
6 Conclusions 
Dependency grammars provide a natural and in-
tuitive solution to the task of logical form identific a-
tion. We have managed to demonstrate relatively 
good overall performance on the given task with 
minimal additional processing and a very small 
amount of training data. 
 
It is argued that a dependency grammar based 
parse provides a rich source of knowledge that is 
suitable for the transformation of English sentences 
into a logical form. It would appear that there is to a 
large extent enough information embedded within 
the parser’s output to achieve the desired outcome.  It 
is however apparent that other types of information 
could further improve the solution. These types of 
information include named entity recognition and 
multi-word phrase detection.  

References  
Fellbaum, C. (1998). WordNet : An Electronic Lexi-
cal Database. Cambridge, Massachusetts; 
London, MIT Press. 
Gildea, D. and D. Jurafsky (2002). "Automatic La-
beling of Semantic Roles." Computational 
Linguistics 28(3): 245-288.  
Harabagiu, S. M., G. A. Miller, et al. (1999). Word-
Net 2 - A Morphologically and Semantically 
Enhanced Resource. SIGLEX. 
Hobbs, J. R., M. E. Stickel, et al. (1993). "Interpreta-
tion as Abduction." Artificial Intelligence 
63: 69-142.  
Järvinen, T. and P. Tapanainen (1997). Dependency 
Parser for English. Helsinki, University of 
Helsinki, Department of General Linguis-
tics. 
Korhonen, A. (2002). Subcategorization Acquisition. 
Ph.D. Dissertation. Computer Laboratory, 
University of Cambridge. 
Miller, G. (1990). "WordNet: An Online Lexical 
Database." International Journal of Lexi-
cography 3(4). 
Preiss, J. and A. Korhonen (2002). Improving Sub-
categorization Acquisition with WSD . In 
Proceedings of the ACL Workshop on Word 
Sense Disambigu ation: Recent Successes 
and Future Directions. 
Rus, V. (2002). Logic Form For WordNet Glosses. 
Ph.D. Dissertation. Computer Science De-
partment,  School of Engineering, Southern 
Methodist University.  
Rus, V. and D. I. Moldovan (2002). "High Precision 
Logic Form Transformation." International 
Journal on Artificial Intelligence Tools 
11(3). 
Tapanainen, P. and T. Järvinen (1997). A Non-
Projective Dependency Parser. In Proceed-
ings of the 5th Conference on Applied Natu-
ral Language Processing, Association for 
Computational Linguistics, Washington 
D.C. 
