-
Notifications
You must be signed in to change notification settings - Fork 1
Tasks
In our project, we will focus on analogy and similarity tasks. A collection of tasks can be found in [4].
In order to understand the tasks, one shall first understand the vector offset method.
Each relationship is characterized by a relation-specific vector offset in the vector space word representations. all pairs of words sharing a particular relation are related by the same constant offset. [1]
Consider the questions of the form “a is to b as c is to ___”. Given a:b, we search for d in c:d. This is done by subtracting the vector of b from a and adding the resulting vector to c. Then you search the words in your embedding to find the word closest to the result in terms of cosine distance [1]
In a nutshell: Rating the relation similarity between one word pair with others.
Given two pairs of words, A:B and C:D, determine the degree to which the semantic relations between A and B are similar to those between C and D [2]. Here is why this task matters: Consider the following word pairs:
- dog:bark
- cat:meow
- car:vroom
These three X:Y pairs are all instances of the semantic relation ENTITY:SOUND. But notice that dog:bark is more similar to cat:meow than car:vroom. The discrete classification ENTITY:SOUND drops this information. With the similarity task, we focus on measuring the degree of relational similarity. [3]