3
$\begingroup$

I have a classification problem to solve, that seems to be common but I am struggling to find the name of this task and the best way to model this problem.

Suppose I have a series of events that are sequential in time.

2 Jan - I matched with a nice girl on Tinder - ACTION_TYPE = SOCIAL_EVENT
5 Jan - I meet with her, it was nice - ACTION_TYPE = SOCIAL_EVENT
8 Jan - I just got accept to a new job. I will meet my boss tomorrow- ACTION_TYPE = PROFESSIONAL_EVENT
10 Jan - I meet with her, it was nice - ACTION_TYPE = PROFESSIONAL_EVENT

It is supervised learning, where I have correctly tagged timelines to train. But during prediction, I have to tag every single event.

I started with a text classification for the text, but I can not distinguish between the events on " 5 Jan" and " 10 Jan".

My instinct is to combine this problem with a sequence tagging, with a CRF layer at the end. But it would be nice if you could look at other possible solutions in the literature.

How would I model this problem? Is this problem known in the literature, and if so, how can I find it?

$\endgroup$
4
  • $\begingroup$ At first sight sequence labelling seems a good idea to me. What is strange in your data is that apparently all the conversations are mixed? Normally I would expect the first two to be part of one conversation and the last two to belong to another. $\endgroup$
    – Erwan
    Nov 9, 2021 at 23:14
  • $\begingroup$ I guess the pronouns are the problem. "Her" can refer to the boss or the girl. If the pronoun got replaced, the problem would be a lot easier. So maybe as a start, replace all pronouns by whatever noun preceeded, then try the prediction again. $\endgroup$ Nov 10, 2021 at 8:48
  • $\begingroup$ @Erwan and Eulenfuchswiesel this is just a example. I want to be able to model a classifier that takes the previous classification in consideration using something like a CRF or Beam Search. It must have something in the literature about it, but I can´t find $\endgroup$
    – bratao
    Nov 10, 2021 at 14:13
  • $\begingroup$ @bratao as far as I know this is exactly what sequence labeling does. You could also look into methods used for event detection, it looks similar. $\endgroup$
    – Erwan
    Nov 10, 2021 at 15:36

3 Answers 3

0
$\begingroup$

You can frame the problem as classification. The features are text and the day-months. The target is one of the discrete category labels (i.e., SOCIAL_EVENT or PROFESSIONAL_EVENT).

$\endgroup$
1
  • $\begingroup$ The day-months are random. What I want to model is the temporal event. One thing happen after another. $\endgroup$
    – bratao
    Nov 7, 2021 at 20:55
0
$\begingroup$

In time series you use data from the past to predict the future. So here your text at time t is the one you need to classify. But your data can have lagged data as input or even some aggregation function based on the past N points. In time series we use average and std over a moving window.

For example like including lags: Input data for your model:

[curent text, previous text, text before previous text]

Converted using bag of words:

[word 1 in text 1, ... ,word n in text 1, word 1 in text 2, ... , word n in text 2 ...] 

Here position matters, but neural networks can identify it.

$\endgroup$
0
$\begingroup$

Since your text data is sequential in nature best is to opt for sequential classification, below papers and tutorial for more information.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.