The FrameNet tagset for frame-semantic and syntactic coding of 
predicate-argument structure 
Christopher JOHNSON 
FrameNet Project, 
International Computer Science Institute 
currently: Soliloquy, Inc. 
255 Park Ave S 
New York NY 10010 
c johnson@soliloquy.corn 
Charles J. FILLMORE 
FrameNet Project, 
International Computer Science Institute 
1947 Center St., Suite 600 
Berkeley, CA, USA 94704 
fillmore@icsi.berkeley.edu 
Abstract 
This paper presents the syntactic and semantic 
tags used to annotate predicate-argument 
structure in the Berkeley FrameNet Project. It 
briefly explains the theory of frame semantics 
on which semantic annotation is based, 
discusses possible applications of FrameNet 
annotation, and compares FrameNet to other 
prominent iexical resources. 
Introduction 
This paper presents the tagset used to annotate 
the predicate-argument structures of English 
verbs, adjectives, and nouns in the Berkeley 
FrameNet Project (NSF IR\]-9618838, "Tools for 
Lexicon Building"), a corpus-based 
computational lexicography project based on the 
theory of frame semantics (see Fillmore 1982). 
It briefly explains the theoretical background and 
shows how frame-semantic annotation creates 
lexicographic generalizations that are not 
possible with more traditional linguistic 
approaches to argument structure based on 
thematic roles. 
1 The FrameNet Project 
1.1 What is frame semantics? 
Frame semantics characterizes the semantic and 
syntactic properties of predicating words by 
relating them to semantic frames. These are 
schematic representations of situations involving 
various participants, props, and other conceptual 
roles, each of which is a frame element (FE). 
The semantic arguments of a predicating word 
correspond to the FEs of the frame or frames 
associated with that word. 
Frames are organized in a structure that can 
be modeled by an inheritance lattice. They range 
from being very general, like case frames 
(Fillmore 1968) or other simple event schemas 
underlying thematic roles, to being lexically 
specific. The most interesting frames are those at 
an intermediate level of specificity which 
encapsulate generalizations about the semantic 
and syntactic properties of word classes that are 
overlooked by thematic role analyses. 
1.2 An example 
One example is the commercial transaction 
frame, which includes the following FEs: Buyer, 
Seller, Goods, and Money. The following 
sentence schemas show how these FEs are 
expressed differently by different Commercial 
Transaction words: 
Buyer bought Goods from Seller for Money 
Buyer paid Seller Money for Goods 
Buyer paid Money to Seller for Goods 
Seller sold goods to Buyer for Money 
Seller sold Buyer Goods for Money 
Buyer spent Money on Goods 
(Seller not expressed) 
Goods cost Buyer Money 
(Seller not expressed) 
Different words assign the Commercial 
Transaction FEs to different Phrase Types (PTs) 
and Grammatical Functions (GFs). For example, 
buy treats the Buyer as an NP subject (i.e. 
  
  56
External Argument) and the Seller as a PP 
complement headed by from, while sell treats the 
Buyer as a direct object or a PP complement 
headed by to and the Seller as a subject. The 
purpose of FrameNet annotation is to gather 
information like this about the grammatical 
realization of FEs for various frames. 
1.2 Project goals 
The goals of the project are to create a database 
of information about English words and the 
frames they inherit, provide annotated corpus 
examples that illustrate how information about 
FEs is expressed by complements and modifiers 
of these words in attested sentences, and 
contribute to a suite of software tools to support 
annotation, database building, and database 
interface. 
An important part of FrameNet work is the 
annotation of corpus sentences with frame- 
semantic information. We use the British 
National Corpus (BNC), because no equally 
comprehensive corpus exists for American 
English (though efforts are underway to create a 
comparable American National Corpus--see 
Fillmore et al. 1998). Each annotated example 
sentence shows argument-structure properties of 
one target verb, adjective or noun. The main 
task of annotation is to tag the arguments (and 
occasionally modifiers) of the target with the 
names of the FEs that they express. A secondary 
task is to mark other lexicographically relevant 
elements, such as support verbs of target nouns, 
and certain non-meaningful elements that 
indicate lexicographically relevant grammatical 
constructions (e.g. extraposition and the 
existential construction). (See Fillmore & Atkins 
1998 on lexicographic relevance.) 
Here is an example sentence from the BNC 
showing the annotation properties of the 
complements of the target verb tell: 
(1) \[Maltravers (Speaker, NP, Ext)\] decided not 
to tell 
\[Stephen (Addressee, NP, Obj)\] 
\[about the inscription in the Le Carr6 book 
(Topic, PPabout, Comp)\]. 
The annotated constituents appear in brackets. 
Following each constituent is a set of 
parentheses containing the FE, PT and GF 
associated with that constituent (actual 
annotations consist of XML markup created 
using the Alembic Workbench software from the 
Mitre Corporation). PT and GF information is 
added by an automatic phrase labeler developed 
by the technical team. 
2 The FrameNet tagset 
2.1 Phrase Types 
Below are the FrameNet Phrase Types. This is 
intended to be a comprehensive list of the types 
of syntactic constituent that can express FEs of 
major predicating words of English (nouns, 
verbs and adjectives). These constituents occur 
either as arguments or as lexicographically 
relevant modifiers. In constructing this list, we 
made extensive reference to the Comlex syntax 
(Meyers et al. 1995). 
2.1.1 Noun phrase types 
NP Noun phrase (the witness) 
N Non-maximal nominal (personal chat) 
Poss Possessive NP (the child's decision) 
There Expletive there (there was a fight) 
It Expletive it (it's nice that you came) 
2.1.2 Prepositional phrase types 
PP Prepositional phrase (look at me) 
Ping PP with gerundive object (keep from 
laughing) 
Part Particle(look it up) 
2.1.3 Verb phrase types 
VPfin Finite verb phrase (we atefish) 
VPbrst Bare stem VP (let us eatfish) 
VPto To-marked infinitive VP (we want to eat 
 sh) 
VPwh WH-VP (we know how to win) 
VPing Gerundive VP (we like winning) 
2.1.4 Complement clause types 
Sfin 
Swh 
Sif 
Sing 
Sto 
Sforto 
Sbrst 
Finite clause (it's nice thatyou came) 
WH-clause (ask who won) 
lflwhether clause (ask if we won) 
Gerundive clause (we saw them running) 
To-marked clause (we want them to win) 
For-to-marked clause (we would like for 
them to win) 
Bare stem clause (we insist that they 
win) 
  
  57
In certain cases FrameNet marks as two 
constituents what are treated as "small clauses" 
in some analyses. For example, in the sentence I 
consider Pat a genius, Pat and a genius would 
be tagged separately. 
2.1.50therphrase types 
AjP Adjective phrase (an interesting idea) 
AdvP Adverb phrase (you put that nicely) 
Quo Quote ("Indeed, "she said) 
2.2 Grammatical Functions 
Below is a list of the FrameNet GFs. This list is 
intended to characterize all the grammatical 
contexts relative to English verbs, adjectives and 
nouns that are regularly occupied by FE- 
expressing constituents. In this list, we do not 
make the traditional distinction between obliques 
and arguments/complements (the former are 
simply PP complements). 
2.2.1 FrameNet grammatical functions 
Ext External Argument 
(Argument outside phrase headed by target verb, 
adjective or noun) 
Comp Complement 
(Argument inside phrase headed by target verb, 
adjective or noun) 
Mod Modifier 
(Non-argument expressing FE of target verb, 
adjective or noun) 
Xtrap Extraposed 
(Verbal or clausal complement extraposed to end 
of VP) 
Obj Object 
(Post-verbal argument; passivizable or does not 
alternate with PP) 
Pred Predicate 
(Secondary predicate complement of target verb 
or adjective) 
Head Head 
(Head nominal in attributive use of target 
adjective) 
Gen Genitive Determiner 
(Genitive Determiner of nominal headed by 
target) 
2.3 Nonexpression of FEs 
Besides being expressed by the PTs and GFs 
listed above, FEs may remain unexpressed under 
the different conditions discussed below (see 
Fillmore 1986). 
2.3.1 Indefinite Null lnstantiation 
In sentence (1), the FEs Message, Medium and 
Code are not expressed. It can however be 
inferred that there must be a Message, Medium 
and Code in a communicative event of the type 
described by this sentence. These FEs, while 
conceptually present in this sentence, are 
optionally expressed, and there are no particular 
restrictions on their nonexpression. This is called 
Indefinite Null Instantiation (INI). 
2.3.2 Definite Null Instantiation 
With some words, an FE may be unexpressed in 
a sentence only if it assumed that the person to 
whom the sentence is addressed has specific 
information about the frame element in question. 
This kind of non-expression is called Definite 
Null Instantiation (DNI). For example, the 
words tell, inform and notify allow a Message to 
be omitted when it is clear what the Message is, 
e.g. How did I know you won? Because Pat 
already tom me. 
2.3.3 Constructionally licensed null 
instantiation 
Certain grammatical constructions, such as 
Passive and Imperative, allow an External 
Argument FE to be unexpressed, e.g. Harsh 
things were said, Tell me about yourself. This is 
Constructionally Licensed Null Instantiation 
(CNI). 
FrameNet annotation marks prominent 
unexpressed FEs as well as FEs that are 
expressed overtly. In order to achieve this, we 
place the symbols INI, DNI and CNI 
immediately after the target word in every 
annotated sentence, and place the appropriate FE 
tag on the appropriate symbol. For example, the 
case of DNI discussed above would be annotated 
as follows: 
(2) \[Pat (Speaker)\] already told INI \[DNI 
(Message)\] CNI \[me (Addressee)\]. 
  
  58
2.4 Frame Elements 
Frames are organized into the following 
domains: Body, Chance, Cognition, 
Communication, Emotion, Health, Life Stages, 
Motion, Perception, Society, Space, Time, 
Transaction, and a General domain. Each 
domain contains several frames that characterize 
different word classes. Because there are many 
frames, it is not possible to give the complete list 
of FEs. The next section discusses FEs from the 
Communication domain. 
3 Communication frames 
It is typically the case that different frames in the 
same domain share FEs. For that reason, each 
domain can be characterized by a basic frame 
that defines its FEs in general terms, and more 
specific frames, corresponding to word classes, 
that are based on this basic frame through 
inheritance or some other principled relation. Let 
us consider the basic frame of verbal 
communication. 
3.1 Basic Verbal Communication frame 
The following FEs consistently appear in frames 
relating to verbal communication: 
Speaker 
(A person who performs an act of verbal 
communication) 
Addressee 
(An actual or intended recipient of a verbal 
message) 
Message 
(A communicated proposition) 
Topic 
(The subject matter of a message) 
Medium 
(A physical channel of communication) 
Code 
(The  or other code used to 
communicate) 
These FEs all derive their meaning from the 
concept of a basic communicative event. Clearly 
a true frame representation cannot just consist of 
a list of role names, but must characterize such 
events. Currently FrameNet frame descriptions 
exist only in text form, but the ultimate aim is to 
express them in a machine-readable format. An 
important component of such a representation 
will be feature structures that express relations 
between frames symbolically. These might be 
combined with computational event models that 
are able to generate inferences, such as the x- 
schemas developed by Bailey et al. (1997) and 
Narayanan (1997). 
The FEs of the Basic Verbal Communication 
frame are relevant to sentence (1) above, in 
which the Speaker is expressed as an NP Ext, the 
Addressee is expressed as an NP Obj, and the 
Topic is expressed as a PP Comp headed by 
about. The other FEs are not expressed. 
We will examine how the FEs above are 
realized in different Communication frames, and 
in the process, will see some of the kinds of 
generalizations that can be expressed through 
frame-semantic lexical analysis. 
3.2 Other frames in the Communication domain 
The Basic Verbal Communication frame 
characterizes events of verbal communication in 
the most general terms. Different 
Communication words represent different types 
of communicative event and different ways of 
construing such events. Generalizations over 
these words are captured by different frames in 
the domain. Some focus on, or profile, in 
Langacker's (1987) terminology, the relation 
between a Speaker and a propositional Message. 
These are grouped into frames characterizing 
different speech acts, e.g. asking (Questioning 
frame), requesting (Request frame), asserting 
(Statement frame), and promising (Commitment 
frame). A few words profile the relation between 
a Speaker and an act of speaking, but not the 
propositional Message communicated (talk, 
speak). Some denote events of reciprocal 
communication (e.g. discuss, argue, 
conversation, etc.). And so on. 
The following sections summarize properties 
of specific frames, and discuss some 
assumptions about the ontology of the 
Communication domain that might account for 
these properties. 
3.3 The Statement frame 
Verbs and nouns in the Statement frame profile a 
relation between a Speaker and a propositional 
  
  59
Message that has the speech act status of an 
assertion. Because of the importance of the 
Message FE, these words frequently occur with 
Sfin Comp (finite clausal complements): 
(3) \[Others (Speaker, NP, Ext)\] assert \[that 
anthropology is the tree and sociology the branch 
(Message, Sfin, Comp)\]. 
(4) \[Managers (Speaker, NP, Ext)\] claim \[there 
was no radiological hazzard to staff or the public 
(Message, S fin, Comp)\]. 
(5) \[His (Speaker, Poss, Gen)\] claims \[to have 
more energy (Message, VPto, Comp)\] are simply 
laughable. 
Message can be expressed with different PPs as 
well. For example, with the target noun claim, it 
can be expressed as a to-marked infinitive VP, as 
in example (5). 
3.4 The Speaking frame 
Verbs and nouns in the Speaking frame profile a 
relation between a Speaker and an act of 
speaking, but do no allow a Message to be 
expressed: 
(6) \[She (Speaker, NP, Ext)\] never spoke \[about 
her feelings (Topic, PPabout, Comp)\]. 
This fact can be explained if we analyze the 
basic meaning of Speaking words as being 
something like 'X say something'. The Message 
role can be thought of as being incorporated 
into this meaning, with incorporation in this case 
being equivalent to obligatory INI. 
3.5 The Request frame 
Like words in the Statement frame, these words 
frequently occur with clausal Complements. 
However, because these Complements express 
requests rather that assertions, they often occur 
as bare stem clauses: 
(7) In all cases \[the respondent (Speaker, NP, 
Ext)\] may request \[in writing (Medium, PPin, 
Comp)\] \[that the disciplinary findings be 
published (Message, Sbrst, Comp)\]. 
3.6 The Conversation flame 
One class of words in the Communication 
domain does not use the FEs Speaker and 
Addressee as they occur in other frames. These 
are nouns and verbs or reciprocal 
communication, which are treated in the 
Conversation frame. In this frame, the human 
interlocutors can be expressed in separate 
constituents, assigned the roles Protagonist-I 
(Prot-1) and Protagonist-2 (Prot-2), or they can 
be expressed in a single conjoined or plural 
constituent, assigned the role Protagonists 
(Prots). 
This class of words demonstrates the 
complex interaction of frames. The FEs Prot-1, 
Prot-2 and Prots are not equivalent to any of the 
Basic Verbal Communication FEs, but do relate 
to them in a regular way. Each of these roles 
must be thought of as relating to two or more 
Communicative subevents, and as corresponding 
to a Speaker in some and an Addressee in others. 
This basic structure is taken not from any frame 
in the domain of Communication, but from a 
Reciprocity frame, which is in the General 
domain. The Reciprocity frame may be thought 
of as an Aktionsart frame that structures events 
and relations from other frames in a particular 
way, such that there are multiple subeventualities 
of the same type as that of the input frame, and 
the bindings or fillers of the roles are reversed 
from one subeventuality to the another. 
The complexity introduced by frames such 
as these points to the need to distinguish between 
conceptual roles and FEs in any given frame. 
While the Communication roles Speaker and 
Addressee are not FEs in the Conversation 
frame, they are conceptual roles, because the 
conceptual representation of Conversation makes 
reference to them. (In the Speaking frame 
discussed above, we can also think of the 
incorporated Message role as being a conceptual 
role rather than an FE.) FEs in any given frame 
should therefore be defined as those roles for 
which the frame specifies conventional means of 
syntactic expression, even if these means are not 
employed in all sentences. 
3.6.1 Disjoint expression: Prot-1 and Prot-2 
Here is a BNC example of argue with a disjoint 
expression of interlocutors: 
(8) "\[You (Prot-1, NP, Ext)\] can't argue 
\[politics (Topic, NP, Comp)\] \[with foreigners 
(Prot-2, PPwith, Comp)\]", sighed the policeman. 
Prot-2, the less prominent interlocutor, is 
regularly expressed in this frame by a PP Comp 
  
  60
headed by with. This is consistent with the 
general behavior of the Reciprocity frame as it 
occurs in combination with other frames and 
domains (e.g. Pat had a 
collision/relationship/agreement with Kim). 
3. 6.2 Joint expression: Protagonists 
Here are BNC examples of the joint expression 
of interlocuters with argue: 
(9) \[They (Prots, NP, Ext)\] argued \[angrily 
(Manner, AdvP, Mod)\] \[over who was the real 
"Prince of Sleaze" (Topic, PPover, Comp)\]. 
(10) \[Mr. and Mrs. Popple (Prots, NP, Ext)\] 
always argued \[INI (Topic)\] at least once a 
week. 
3.7 FrameNet complements existing lexical 
resources 
FrameNet annotations provide more detail than 
existing lexical resources about the way in which 
particular semantic roles (i.e. FEs) are linked 
with particular means of syntactic expression. 
Since different senses of ambiguous words are 
defined relative to different frames, this linking 
information could potentially be used for lexical 
disambiguation. 
For example, consider the verb argue as it is 
treated in the WordNet database (Fellbaum 
1998). Below are the three WordNet senses of 
argue, and the sentence frames that are 
associated with each sense: 
Sense 1: argue, reason -- (present reasons and 
arguments) 
EX: Sam and Sue argue 
EX: Sam wants to argue with Sue 
Sense 2: argue, contend, debate, fence -- (have 
an argument about something) 
EX: Sam and Sue argue 
EX; Sam wants to argue with Sue 
Sense 3: argue, indicate -- (give evidence of; 
"The evidence argues for your claim"; "The 
results indicate the need for more work") 
The three senses listed above correspond to three 
different frames: Sense 1 corresponds to the 
Statement frame, Sense 2 to the Conversation 
frame, and Sense 3 to the Evidence frame in the 
Cognition domain. However, the information 
about sentence frames provided in WordNet 
does not correspond to the generalizations that 
are apparent in FrameNet. For example, 
WordNet gives the same sentence frames for 
Senses 1 and 2, while in the FrameNet database, 
the senses of argue defined relative to the 
Statement and Conversation frames are 
characterized by different argument structures: 
only Statement argue allows finite clausal 
complements expressing Message; Conversation 
argue has the properties of other reciprocal 
communication words, which Statement argue 
lacks, and does not allow clausal Complements 
(or any other expression of Message). 
COMLEX (Meyers et al. 1995) recognizes 
the syntactic frames in which argue occurs, but 
does not provide information about the linking of 
syntactic constituents with semantic roles, or 
about the different complementation properties 
of different senses of ambiguous words. 
Conclusion 
FrameNet semantic annotation captures human 
knowledge about the ways in which semantic 
roles (FEs) are conventionally expressed by 
different words in various word classes and 
domains. The kind of information in the 
FrameNet database is not expressed in the same 
level of depth in any existing print dictionary or 
computational lexical resource. While WordNet 
describes semantic relations between words, it 
does not recognize the conceptual schemas, i.e. 
frames, that mediate in these relations, and 
therefore does not have the means to link 
arguments of predicating words with the 
semantic roles they express. COMLEX and 
NOMLEX provide detailed information about 
the syntactic frames in which verbs and nouns 
occur, but also lack a means to link syntactic 
arguments with semantic roles. FrameNet 
therefore provides information that complements 
major existing lexical resources. 
Acknowledgements 
The authors would like to thank the National 
Science Foundation for supporting this project, 
and the International Computer Science Institute 
in Berkeley for giving it a home. Thanks also to 
Oxford University Press and Sue Atkins for 
making it possible for us to use the British 
National Corpus, and to IMS-Stuttgart and Uli 
Heid for providing us with the tagged version. 
We are also grateful to Catherine Macleod and 
Ralph Grishman of the NYU Proteus Project for 
allowing one of use to be a visitor, and various 
people in the Berkeley linguistics department 
and in the Neural Theory of Language project at 
ICSI for enlightening discussions. 

References 

Bailey, D., J. Feidman, S. Narayanan and G. Lakoff. 
(1997) Modeling embodied lexical semantics. In 
Proceedings of the 19th Annual Meeting of the 
Cognitive Science Society COGSCI-97. Stanford, 
CA: Stanford University Press. 

Baker, Collin F., Charles J. Fillmore, and John B. 
Lowe. (1998) The Berkeley FrameNet Project. In 
COLING-ACL '98 Proceedings of the Conference 
held August 10-14, 1998, in Montreal, Canada, 
pp.86-90. 

Fellbaum, Christiane (1987) WordNet: An electronic 
lexical database. Cambridge, MA: MIT Press. 

Fillmore, Charles J. (1968) The case for case. In E. 
Bach and R. Harms (eds.), Universals in Linguistic 
Theory. New York: Holt, Rinehart & Winston. 

Fillmore, Charles J. (1982) Frame semantics. In 
Linguistics in the Morning Calm, pp. 111-137. 
Seoul, South Korea: Hanshin Publishing Co. 

Fillmore, Charles J. (1986) Pragmatically Controlled 
Zero Anaphora. In Proceedings of the 12th annual 
meeting of the Berkeley Linguistics Society, pp. 
95-107. 

Fillmore, Charles J. and B. T. S. Atkins. (1999) 
Describing polysemy: the case of 'crawl'. In 
Polysemy: Linguistic and Computational 
Approaches, (eds.) Yael Ravin and Claudia 
Leacock. Oxford: Oxford University Press. 

Fillmore, Charles J. and B.T.S. Atkins. (1998) 
FrameNet and lexicographic relevance. In 
Proceedings of the First International Conference 
on Language Resources and Evaluation, Granada, 
Spain, 28-30 May 1998. 

Fillmore, Charles J., Nancy Ide, Daniel Jurafsky, and 
Catherine Macleod (May 1998) An American 
National Corpus: A Proposal. In Proceedings of the 
First International Conference on Language 
Resources and Evaluation (LREC); Granada, Spain 
28-30 May 1998. 

Gahl, Susanne (1998) Automatic extraction of 
Subcorpora based on Subcategorization Frames 
from a Part-of-Speech tagged Corpus. In COLING ACL '98 Proceedings of the Conference held 
August 10-14, 1998, at the University of Montr6al, 
Canada, pp.428-32. 

Gahl, Susanne (1998) Automatic extraction of 
Subcorpora for Corpus-based Dictionary-building. 
In Thierry Fontenelle et al. (eds.), EURALEX '98 
Proceedings: Papers submitted to the Eighth 
EURALEX Conference, at the University of Li6ge, 
Belgium, pp.445-452. 

J.B. Lowe, C.F. Baker, and C.J. Fillmore. (1997) A 
frame-semantic approach to semantic annotation. 
In Proceedings of the SIGLEX workshop "Tagging 
Text with Lexical Semantics: Why, What, and 
How?" held April 4-5, in Washington, D.C., USA 
in conjunction with ANLP-97. 

Johnson, Christopher (1999) Multiple frame 
inheritance in lexical descriptions. Annual 
Meeting of the Linguistic Society of America. Los 
Angeles, January 9. 

Johnson, Christopher. (1999) Syntactic and semantic 
principles of FrameNet annotation. TC-99-018, 
International Computer Science Institute, Berkeley, 
CA. 

Langacker, Ronald. (1987) Foundations of Cognit&e 
Grammar, Vol. 1: Theoretical Perspectives. 
Stanford, CA: Stanford University Press. 

Levin, Beth. (1993) English Verb Classes and 
Alternations. Chicago: University of Chicago Press. 
Meyers, Adam, Catherine Macleod, and Ralph 

Grishman. (1995) Comlex syntax 2.0 manual for 
tagged entries. Technical report, Proteus Project, 
New York University. 

Narayanan, S. (1997) KARMA: Knowledge-Based 
Active Representations for Metaphor and Aspect. 
Ph.D. Dissertation, Computer Science Division, 
University of California, Berkeley. 
