Media analysis : negativity

Article on text polarity / media polarity

This is my second article. Since the first one took a long time to write, I will try to keep this one short.

It’s about detecting negativity inside a media (news article, video etc).

Since it can be subjective, it can be difficult to analyse.

There are different degrees of sentiment that can be analysed:

• Overall sentiment : it would be the general sentiment of the article

  • Even if the possible effects would be positive (denouncing something, for instance), I would classify it as negative if the topic is negative

  • There is not necessarily an overall sentiment (it can be neutral)

• Per section sentiment : you can divide the article into sections, and assess the sentiment of the section / paragraph / phrase

• Per topic sentiment : a topic can be engaged in multiple sections, it could makes sense to extract the sentiment per topic

After a bit a research, I discovered that this is named “text polarity” analysis.

Since it can be applied to any media, I would argue to call it “media polarity”.

But it would project political polarity, so I’ll keep it this way.

Actually, a lot of search engine research will come up with political media polarity results.

Assessing

It can be overwhelming.

What is negative, positive, about whom, to whom etc.

I’ll start with a simple topic, a usual news topic : someone murdered somebody else.

The gain represent the potential political gain for the groups that person B belongs too.

It might makes someone of a particular group feel positively, but it’s overally negative for the person A (and the groups he belongs too).

In some cases, it might be an incitement to attacks people of person B’s groups, but usually not.

In some cases, it might be neutral, if there is information about person B (negative information) in the article.

Common good

We can see that, the sentiment is mostly coming from the action, here.

I would use “common good” assessment, there are things that are positive, and other that are negative.

Even if someone could personally assess that killing is positive (a thief for instance), because of his personal bias (political, mental etc), under this analysis, we would classify it as negative.

I fetched some words from the news of today to help creating a simple list below.

It was actually tough to find any positive actions / words.

I did not find any, so I made them up.

The nouns used can affect the analysis, for instance “helped a murderer” would be negative, even though “help” is a positive word.

I was not sure if it should be indicated as negative, but after thinking about it, it is.

But “killed a thief” would still be negative.

It would be interesting to analyse this further, but it’s not the subject of this article.

I would say that any topic related to conflict, death / diseases are negative though.

Some of this topics could be debatable.

If you take “cure”, for instance, it could make you feel bad, but I would still stay it’s a positive topic.

Cats are definitively positives.

For processing

I’ll advise to start simple.

You would probably want to do it yourself, but you can find existing products (looking up “text polarity” for instance), free and open-source, or paid on-the-shelf.

To have a MVP, it would be sufficient.

I would start with the overall polarity, since it’s the most simple.

Then, I would create more precise analysis.

I am not sure if there are existing products that are more precise, actually.

Text comprehension is difficult.

This is a subject where you want to use neural-networks, since text comprehension can be hard to implement.

It’s one of the few tech subject where neural-networks actually make sense.

You would classify existing documents by hands, or use a premade dataset, and train the AI on it.

Relates to

• Sentiment analysis

• Text polarity

• Media analysis

To go further

I’ll try to make articles about these topics:

• Existing products implementing this (for the non-developers)

• Existing technologies (for developers)

• Existing datasets (for developers)

• Topic analysis / text comprehension

• Subjectivity

• Personal analysis