# How to approach for predicting semantic similarity between two phrases

I need pointers on the latest research, tools, and techniques for predicting semantic similarity between two phrases.

Problem Statement: Given two propositions A and B with A know to be true, predict if B agrees, contradicts, or is neutral.

Examples

A = X is greater than Y | B = Y is greater than X | Result = contradict

A = X is greater than Y | B = Y is less than X | Result = agree

A = X is greater than Y | B = P is less than X | Result = neutral

A = X increases with Y | B = X is directly proportional to Y | Result = agree

A = X is joining Y and Z | B = Y and Z are connected with X | Result = agree

What I have tried

I am looking into Google's Universal Sentence Encoder and alternatives for evaluating semantic similarities with mixed results. Most of the solutions available will give a result of exactly similar for the first example. I am also going through some research papers on grammar-based similarity techniques in natural language.

Any help or direction towards research papers, tools, or libraries is greatly appreciated.

• As an initial thought, it could be helpful if you fed in a model with word dependency relations in the sentence (i.e. is A is the subject or object of the sentence). You can get these relations via dependency parsing (web.stanford.edu/~jurafsky/slp3/15.pdf) Sep 14 '20 at 12:55

There is plenty of research in this kind of task, and its variants, like cross-lingual NLI (XNLI). I suggest you have a look at nlpprogress (link) or paperswithcode (link) to understand the landscape of benchmark datasets and SoTA models in NLI.