| | | Annotation Basics |
FrameNet is a computational lexicography project based on the principles of frame semantics, a theory developed by Charles Fillmore, the Principal Investigator. Frame semantics characterizes the semantic and syntactic properties of 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. The most interesting frames do not simply characterize individual words; rather, they are generalizations over classes of words that share semantic and syntactic properties.
A word whose meaning is defined relative to a given frame is said to inherit the frame. The frame elements in that frame, or some subset of them, constitute the semantic arguments of the words inheriting the frame. Complements and, occasionally, modifiers of these word in actual sentences provide information about the frame elements. Facts about the syntactic complements of a word and the frame elements they express are referred to as valence properties of the word.
Grammatical constructions associated with a frame express generalizations about the syntactic realization of frame elements. For example, it is a general property of frames in the communication domain that a Topic may be expressed by an about-PP.
They asked me [about my family].The President made a statement [about the economy].
We had a conversation [about our favorite music].
The words inheriting a frame may specify their own idiosyncratic constraints on the expression of frame elements, and these lexically specific constraints may override the generalizations made at the frame level. For example, with the verb discuss, the frame element Topic is expressed as a direct object rather than an about-PP:
We discussed [our favorite music].* We discussed about our favorite music.
Words may also place semantic constraints on frame elements that are more specific than those specified at the frame level. For example, the word slither inherits a general frame pertaining to the directed motion of a living being under its own power, but places its own special constraints on the nature of the living thing that moves, as well as on the manner in which it moves.
An example of a semantic frame is the Commercial Transaction frame. This frame characterizes simple events of buying and selling, as such events are linguistically encoded in English. At the minimum, it has the following frame elements: the Buyer, the Seller, the Money, and the Goods. Different words associated with this frame are characterized by the different phrase types and grammatical functions they use to provide information about these frame elements. Limiting ourselves to the use of these words in the active voice, we can say that the word buy, for example, expresses the Buyer as a subject NP and the Goods as a direct object NP. The Seller can optionally be expressed as a PP complement headed by from, and the Money as a PP complement headed by for. The word charge expresses the same roles with its complements, but assigns each role to a different phrase type-grammatical function pair: the Buyer is expressed as a direct object NP, the Goods as a PP Complement headed by for, the Seller as a subject NP, and the Money as a non-object NP complement. Other verbs defined relative to the same frame include sell, pay, spend, and cost. Below, the different syntactic complementation patterns associated with these words are summarized using frame element names. An example sentence follows each pattern.
Buyer bought Goods from Seller for Money.Al bought one golf club from Bill for $1000.
Buyer paid Seller Money for Goods.
Al paid Bill $1000 for one golf club.
Buyer paid Money to Seller for Goods.
Al paid $1000 to Bill for one golf club.
Seller sold Goods to Buyer for Money.
Bill sold one golf club to Al for $1000.
Seller sold Buyer Goods for Money.
Bill sold Al one golf club for $1000.
Buyer spent Money on Goods. (Seller not expressed)
Al spent $1000 on one golf club.
Goods cost Buyer Money. (Seller not expressed)
One golf club cost Al $1000.
The purpose of FrameNet annotation is to mark the complements and, occasionally, modifiers of predicating words with the names of the frame elements that they express.
The goal of the FrameNet project is to construct a computer-readable database of information about English words and the frames they inherit, together with attested examples that illustrate the way frame elements are expressed by complements and modifiers of these words in real sentences.
An important part of this work is the annotation of corpus sentences with frame semantic information. Each example sentence shows valence properties of one predicating word--typically a verb, adjective or noun. In the context of a given example sentence, the word whose semantic and syntactic properties are of interest is called the target word, or just the target.
The main task of annotation is to tag to the important syntactic constituents in example sentences with the names of the frame elements that they express. Generally speaking, these are the constituents that bear grammatical functions with respect to the target word, but they may sometimes be modifiers of the target. A secondary task of annotation is to mark certain lexico-syntactically relevant elements in sentences, such as support verbs of target nouns.
In order to annotate a collection of sentences for a given target word, it is necessary for annotators to understand the frame associated with that word. Frame descriptions are provided by the Vanguard, and are read carefully before annotation begins in any frame.
Since the goal of the project is to annotate the best and simplest example sentences for a target word in a particular frame, we only use the sentence-level Metaphor tag to indicate metaphoric use of a target word when there is a systematic transfer across frames. The transfer must be one that permits the use of the same frame elements for all parts of a sentence in two different frames.
Each time a sentence is tagged with the sentence-level Metaphor tag, an annotator also begins a discussion about a new frame.
Use INC on an FE when a target word incorporates an FE. For example, in Scalar_Position2, the noun increase incorporates the FE Path. INC may be used on any part of speech, but will most often be needed with verbs. Consider the following sentence.
He boxed the toy for shipment.
Every occurance of box.v incorporates the FE Goal, which must be so indicated. Unlike INI, DNI,and CNI, INC may be used even when the FE is instantiated. For example,
He boxed the toy [in a red container GOAL.]
Annotate on an additional FE layer when more than one FE is expressed by a single constituent. For example, in the bodily_injury frame,
Primary FE Layer: He [hit TARGET][my head BP]. Secondary FE Layer: [my VICTIM]
Note that there is no PT or GF indicated for the FE on the secondary layer.
In the following sentence these FEs do not overlap. He [hit TARGET] [me VICTIM] [on the head BP].
One of the two FEs may be expressed twice, in which case, double tag that FE, once on the primary layer and once on the secondary layer, as shown here.
Primary Layer: He [hit TARGET] [me VICTIM] [on my head BP]. Secondary Layer: [my VICTIM]
| | | Annotation Basics |