Is it possible to use Lambeq for keyword extractions in a text ( ~ 20 words) ?

[chin-jey]

Hello, I'm exploring Quantum NLP and I found this awesome work from you guys.

I was wondering if, apart from question answering and classification tasks showed in the documentation, I could use it for keyword extraction in short messages of 20-25 words.
Best regards

[dimkart]

Hi @chin-jey and thanks for the question. In general, lambeq works at the sentence level, and keyword extraction is a task that does not fit very well in this. In keyword extraction, each word needs to be evaluated individually, so compositionality here is not very helpful. The simplest way to do keyword extraction would be to compute TF-IDF values for each word in your dataset and use them to extract the important words, or use something more clever like a PageRank-style algorithm at the lever of words.

A suggestion for further reading: In this paper, Bob Coecke proposes a way to connect sentences into document-level circuits, and measure how each word is changed at the end of the interaction. Looks it might have a possible application to keyword extraction.

[chin-jey]

Hello @dimkart and thanks for you answer.

The simplest way to do keyword extraction would be to compute TF-IDF values for each word in your dataset and use them to extract the important words, or use something more clever like a PageRank-style algorithm at the lever of words.

here you mean "in general" or you mean that using lambeq (especially the quantum part) after doing what you suggest might work ?
thanks, chin.

[dimkart]

I mean "in general" -- as I mentioned, for this task it's easier to work at the word level (sometimes by just collecting statistics for individual words) rather than the sentence level.

[chin-jey]

Then you don't recommend using lambeq for this task ? I looked for something quantum-based to explore the stack

[dimkart]

If you want to use a compositional method for this task, you need to find a way to do predictions over words. One way is to set your task as a multi-label classification, making predictions over the vocabulary. Another is to encode your sentence into a tensor net that allows predictions for individual words, e.g. something like

<S>  John   loves   Mary   <E>
 o----o-------o-------o-----o
      |       |       |

where each word is a tensor of order 3. In the above diagram each word is associated with a wire that carries a prediction. You might want to try these approaches and compare with the simpler methods I mentioned above.