Information Technology Reference
In-Depth Information
1
4
Preprocessing:
- remove noise data
- convert to simpler
sentences
Baseline method:
- deep linguistic
parser
- syntax patterns
- Google Map API
parsable
sentences
the other
sentences
2
5
Predict labels of
activity attributes
Learning model
Training data
Feature model
of training data
3
Add more training data
Fig. 2.
Approach of extracting activities from Twitter
the sentence “Tanaka is now taking refuge at Akihabara”,
actor
,
act
,
time
and
location
are “Tanaka”, “take refuge”, “now”, “Akihabara” respectively. Figure 2 explains the
key ideas of our previous work [7] for extracting activity attributes in each sentence
retrieved from Twitter, are summarized as follows:
-
We deploy self-supervised learning [8], and use the linear-chain conditional random
field (CRF) [9] as a learning model.
-
Using deep linguistic parser [10], 8 syntax patterns, and Google Map API to make
initial training data from parsable sentences.
-
Based on the initial training, we collect more trustworthy training from weblogs.
-
Since sentences retrieved from Twitter often contain noise data, we remove these
noise data before testing. Additionally, to avoid error when testing, we convert com-
plex sentences to simpler sentences by simplifying noun phrases and verb phrases.
-
We consider not only time stamp of tweets, but also time expression (now, this
evening, etc) to decide time of activities in these tweets.
3
Building Timeline Action Network
In this section, we first design TiAN. We then explain how to create semantic data for
TiAN.
3.1
Designing Timeline Action Network
To represent human activities in real-time, we add time information into TiAN. As
shown in Figure 3, TiAN is expressed as a directed graph whose nodes are activity
attributes, and whose edges are relations between these activity attributes. It is important
to help the computers understand the meaning of data, thus we design TiAN based on
OWL. Since N3 [11] is a compact and readable alternative to RDF's XML syntax, we
use N3 to describe data of TiAN.
To represent data of TiAN, we create classes and properties as below:
-
ActionClass, ActClass, WhereClass, and WhatClass are classes of activity, act, lo-
cation, object respectively.
