FrameNet is a lexico-semantic resource developed at the University of Berkeley (have a look at the
It uses so called frames for annotating English texts with discourse semantic action scenarios. Each frame consists of ...
the frame itself, e.g. Giving, which is attached to the target word,
a set of roles, such as Donor (the person who gives), Recipient
(the person who receives the given object) and Theme (the transferred entity) for Giving;
FrameNet distinguishes between core roles (important roles that are frequently realized in the English
samples) and non-core roles
and a set of lexical units by which the frame is invoked (ex.: English to give or
Sanskrit pradā for the frame Giving)
Extensive definitions of these components can be found in the
frame index of FrameNet.
Given these components, the task of Semantic Role Labeling (SRL) consists in assigning the correct
frames and roles to words in a sentence. Consider the Sanskrit sentence
sa bhīmasenāya gadām adāt. The FrameNet annotation of this sentence would look
The same colour scheme (red = target, green = core roles, grey = non-core roles) is used on the annotation page,
where users can select a frame in the listbox at the top of the page to display its current state of annotation.
To my knowledge, frame annotations have not yet been used in Indological studies, although they offer interesting
new perspectives for philological research. Just a few ideas:
Content based search and "automatic text understanding": Because frames constitute a level of linguistic
information that strongly abstracts from the linguistic surface structure, they can be used to trace identical
content in passages that have a completely different wording. Such approaches can be used for content based and
even inter-corpus queries, for text summarization and for other advanced aspects of computational text
Studying the argument structure of Sanskrit words: Which persons and objects realize the different roles of a
frame? Can temporal, religious or regional differences be detected?
Lexicographic studies based on the topics just mentioned
To annotate a sentence, proceed as follows:
Click on the FN link that shows up to the right of sentences on the text page and on the references page.
You are redirected to the annotation page.
Select the frame you want to annotate in the selected sentence from the listbox Frame at
the top of the annotation page. The roles of the frame are displayed in a table right below this listbox.
Drag the roles from the role table into the correct fields of the table that records the lexical composition of the
Sanskrit sentence. Please make sure to annotate the target word, e.g., pradā
in the case of Giving. The annotation interface will not store your annotation if you
do not mark the target word.
Press the button Save to store the annotation. Please press the button only once, and
be patient if storing takes a few seconds.
After the annotation has been stored, you may proceed to another randomly selected sentence from the DCS
that contains the same lexical unit (e.g., pradā).
Frames can be annotated directly by every user of the DCS. It is recommended that anyone who plans to annotate a
larger number of frames gets a user name (contact me!). In this way, individual contributions are easy to track, and registered
users are able to delete the frames they added to the database [editing is planned for a future version of the
It is strongly recommended to annotate on a framewise basis for the beginning. This means that you should select
one frame you are interested in, retrieve the respective lexical units from the query page (use the
flag "Search for meanings"), and annotate only this frame. Full annotation of text passages may
be frustrating at the beginning, because it requires intensive search in the FrameNet index.
To generate a well balanced corpus of frame semantic annotations, it is recommended to use the random approach
Every user of the DCS gets full access to all FrameNet annotations. To display and download the data in XML
format, please follow the link "Generate XML data" on the display page.
Data is rebuilt once per day on user request.