About This Project

We are providing an automated method to detect ambiguous sentences based on machine translation. The proposed method works regardless of the semantics and therefore, is highly generalizable.

Potential practical applications are automatic translation and detection of ambiguity of critical texts such as in law or politics. Furthermore, it can be applied in fine-tuning multilingual LLMs and development of a ambiguity-free global human language.

Ask the Scientists

What is the context of this research?

Detecting ambiguity in text prevents automatic machine translators from misleading the reader of the target translation in understanding the text. Furthermore, by discovering a function that detects ambiguity, we would be able to fine-tune LLM-based machine translators to preserve the ambiguity in ambiguous sentences in translation.

Apart from that, this project, is part of a bigger project at Synaptosearch which tries to optimize an AI-generated global human language. By finding the ambiguity detection function, we would be able to use it as part of the optimization criteria to achieve an ambiguity-free language.

What is the significance of this project?

Language ambiguity brings up misunderstandings and conflicts in real world interactions such as legal, political, commercial and cultural interactions. This misunderstanding can lead to either waste of huge amount of time in negotiation between the parties for conflict resolution or even in worse case results in conflicting actions.

Detecting ambiguous statements in global interactions can be a critical problem that requires scientific research in providing solutions. One major organization that can benefit from the proposed research is the United Nations (UN) where different countries with different languages interact with each other.

What are the goals of the project?

The goal of this project is to propose a generalizable method to detect ambiguity in sentences, regardless of the semantics. In order to do so, we consider machine translation, as potential actions performed on single text. Conflict in different actions, accordingly, specifies ambiguity in the text.

In order to provide a semantic-free approach, we propose investigating how a text is translated to many target language and then translated back to the source language. By mapping the two step translations together and studying the properties of this mapping, we would be able to detect ambiguity, regardless of the semantics of the sentence.