NLP research on Alzheimer Disease
Chapter 0 --- Introduction
Chapter 1 --- Entity Type Annotation Instructions
Chapter 2 --- Entity Attribute Annotation Instructions
Chapter 3 --- Relation Annotation Instructions
3.1 Core Argument
agent, theme, dative
3.2 Thematic Roles
attribute 3.4 Disclosure
Chapter 4 --- Special Case Handling.
Chapter 0 Introduction
We let potential patients with Alzheimer’s disease depict a picture called ‘circus procession’ and record their answers. The annotation of the transcripts creates a valuable corpus, which can be used as training data for natural language processing research on diagnosis of the disease. In our annotation, a sentence is annotated to identify mentions of some real-world entities (things) and their types, and a relation between two. Named entity recognition and information extraction tasks are accomplished. The main tasks of this annotation are: entity type labeling, entity attribute labeling and relation labeling. Each aspect is discussed in detail below.
Chapter 1 Entity Type Annotation Instructions
Entities are real life objects in the picture, commonly noun phrases. We omit articles in our annotation. Entities are classified into two general types---known and unknown. Known entities are things that constantly appear in each transcript. For example, elephants, clown and tricycle.
Modifiers are said to modify entities and can be removed without affecting the grammar of the sentence. Nmod is the class of adjective modifiers. Most common types of adjectives appeared are Color / Order / Size / Quantity.
Examples of the ‘Order’ adjectives: first, the other, another….
Note: sometimes compound nouns can function as modifier as well. Ex: a polka dot dress.
Predicate is the part of a sentence that tells what the subject does. We only annotate one word which is the verb. We omit the auxiliary verbs(am, is are). One type of predicates we pay special attention to is Motion. For motion verbs, this paper is good reference. However, the most common ones in the texts are ‘march, ride, walk’.
Attention: We treat have/hold/dressed up as normal predicates.
Xmod is the class of adverbial modifiers
- Temporally related(modifiers of events): treasures are just lying around, waiting to be found
- Intentional (modifiers of propositions):
Probably, likely -> fuzzy
Must -> certain.
Very, clearly, really, definitely, absolutely -> emphasis
- Focus-sensitive: only, even
- Sentential (evaluative, attitudinal):fortunately, legally, frankly speaking, clauses beginning with given that, despite, except for or if.
Chapter 2 Entity Attribute Annotation Instruction
We define the following attribute types : Abstract, Disfluency, External, Opinion, Possessive, Subset
Commmonly appeared examples: something, one
Fragments of words, interruptions, incomplete sentences, filters and discourse markers.
- subjective adjectives (fancy, bored, beautiful, normal,etc)
- descriptive clauses that contain the word "like" (dressed up in a human -> human is opinion, dressed up like a millionaire)
- sentences with hint words at the front (In my experience, as far as i can see, it is obvious that, etc)
Show ownership by adding an apostrophe, an 's' or both.
Chapter 3 Relation Annotation Instructions
3.1 Core Argument
Entity – Predicate – Entity (- Entity)
-Agent and Theme Relation
Ex: the elephant has a hat
R1 agent Arg1:has Arg2:elephant R2 theme Arg1:has Arg2:hat
refers to indirect object of a verb
Ex: I gave him a book
R1 agent Arg1:gave Arg2:I R2 theme Arg1:gave Arg2:book R3 dative Arg1:I Arg2:him
3.2 Thematic Roles
Directional relations show motion along some path. For example: step forward, walk along the river.
Locative relations indicate where some action takes place. Both physical location and abstract locations are marked as locative.
Temporal words show when an action takes place. Also included in this category are adverbs of frequency: always, often, adverbs of duration: for a year.
Causal adverbials specify the reason for an action. Canonical cause clauses start with 'because'.
Manner relations indicate how an action is performed. For example: walk together, sing beautifully, ran quickly.
Explains the motivation for some action. Clauses beginning with 'in order to' and 'so that' are common purpose clause.
Usually between Xmod and predicates.
Nmod is Attribute of Entity
Ex: little parade
R1 attr Arg1:parade Arg2:little
! We treat “with” as attributes
Ex: a clown with makeup. (Makeup is attr of clown)
Chapter 4 Special Cases
- Treat compound nouns as a single entity
Ex: Polka dot, Straw hat
- One is holding a fan or umbrella -> treat as two entities.
- We do not annotate anything that is not related to the picture.