1
$\begingroup$

I extracted event triples from sentences using OpenIE. Can I concatenate the components in the event triple to make it a sentence and use Sentence-Bert to embed it? It seems no one has done this way before so I am questioning my idea.

I'm using news headlines to predict next day's stock movement. For example, there are two news headlines, the first is "U.S. stock index futures points to higher start", I used openIE to extract it and there are two event triples, [('U.S. stock index futures', 'points to', 'start'), ('U.S. stock index futures', 'points to', 'higher start')]. (There are repetition in the openIE extracted event triples and I don't know how to avoid it.) Since it contains events I'm interested in (stock index), I will embed these two events and take their mean as the the embedding.

The second headline is "STOCKS NEWS US- Economic and earnings diary for Jan 4", it contains no events as it is only contain nouns. So I will embed it as 0 vector in this case.

$\endgroup$
4
  • $\begingroup$ Why not using sentence-Bert in the first place? Could you give more information about your objective? $\endgroup$ Jun 12 at 20:17
  • $\begingroup$ @NicolasMartin Because at first I want to extract event triples to see if there are events I'm interested in, and I only care about the main parts of the sentence, which I think is the event triple(subject, predicate and object). $\endgroup$
    – user900476
    Jun 12 at 20:22
  • $\begingroup$ Alright. You should edit your question and clarify all those things because it is complex for me to give you a good recommendation. An example would also be helpful. $\endgroup$ Jun 12 at 21:02
  • $\begingroup$ @NicolasMartin Hi, I have updated the description of my question. $\endgroup$
    – user900476
    Jun 13 at 2:55

1 Answer 1

0
$\begingroup$

Using triples could lead to wrong results because some headlines could contain double negations or other complex structures that are difficult to classify with triples.

However, you can apply directly on the headlines Bert sentiment analysis instead, which can process complex semantics correctly.

Here is an example using Bert's twitter roberta sentiment analysis:

enter image description here

Note: in this specific case neutral and positive have almost the same value, and you will want to set some threshold to consider a headline as positive, like positive > 0.4. It could also require some fine tuning because tweets are a bit different from headlines.

You can even apply sentiment analysis levels (very negative, negative, neutral, positive, very positive) to get even better predictions.

$\endgroup$
5
  • $\begingroup$ Thank you very much! Actually I'm doing sentiment analysis for all news headlines (I wouldn't believe I can predict stock movement simply from text embeddings like some papers). My idea is that I use event embedding of every headline as a key (or query vector I don't know how to call it) to produce inner prodcut with a trainable vector v (query or key) to get a weight for every headline. $\endgroup$
    – user900476
    Jun 13 at 7:56
  • $\begingroup$ So in this case the second headline in my example contains no events, it has a huge neutral sentiment but taking the mean of the sentiments of this day would produce a huge neutral sentiment (there might not be much news one day) and makes no sense (cause it only contains some nouns, didn't indicate stock market events and movements). So its embedding vector will be 0, and thus the weight will be 0. $\endgroup$
    – user900476
    Jun 13 at 7:57
  • $\begingroup$ Why not detecting the event and the sentiment, then use a decision tree like random forest? Embedding is also possible, but it could require more training. $\endgroup$ Jun 13 at 8:11
  • $\begingroup$ I want to use deep learning models (actually I'm writing a thesis). Thank you, I will also look into decision tree models. $\endgroup$
    – user900476
    Jun 13 at 8:16
  • $\begingroup$ For sentiment analysis, I used to have better result with Bert, and for simple classifications, random forest used to work better than NN, but it depends on cases. A thesis is a great opportunity to compare such algorithms. $\endgroup$ Jun 13 at 9:14

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

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