Possibly (but probably not)!

Translating between two languages with only monolingual corpora (unsupervised machine translation) is currently possible. It's an area of active research in NLP because current machine translation methods use large, parallel sentences which are expensive to create and don't exist between many language pairs.

The current state-of-the-art in unsupervised translation is Song et al. (2019), which reports BLEU of 37.5 on English-French. For reference, Google Translate, which uses parallel data, only scored about 35.7 as of 2017 (higher BLEU is better) (Johnson et al., 2017).

However, EN-FR is one of the easiest pairs because:

• There is a lot of high-quality parallel and non-parallel data since both are official languages of the U.N., the E.U., various countries, etc., are spoken by millions of people worldwide
• The languages have many cognates and some shared vocabulary
• The languages share a fairly simple writing system (esp. compared to hieroglyphics)

Lample et al. (2018) tested their system on Urdu->English, two unrelated languages with different writing systems and with (relatively) little available data, and obtain 12.3 BLEU. I don't have a reference point for how good that is, but it's definitely a start.

Finally, Zhang et al. (2019) train a translation system on Chinese -> Japanese, and show that it is possible to learn information about logographic writing systems, but Chinese and Japanese share cognates, and there's shared vocabulary between Chinese characters and kanji.

That being said, hieroglyphs are (IMO) a more complicated writing system than even Chinese. And even for English-Urdu, a "low-resource" language pair, Lample et al. use 5.5M sentences. I have no idea how much text exists in hieroglyphs, but suspect it's less than this. But in theory, if we dug up and digitized millions of tablets of an ancient, unknown language, then yes, we do have tools to translate it.

Possibly (but probably not)!

Translating between two languages with only monolingual corpora (unsupervised machine translation) is currently possible. It's an area of active research in NLP because current machine translation methods use large, parallel sentences which are expensive to create and don't exist between many language pairs.

The current state-of-the-art in unsupervised translation is Song et al. (2019), which reports BLEU of 37.5 on English-French. For reference, Google Translate, which uses parallel data, only scored about 35.7 as of 2017 (higher BLEU is better) (Johnson et al., 2017).

However, EN-FR is one of the easiest pairs because:

• There is a lot of high-quality parallel and non-parallel data since both are official languages of the U.N., the E.U., various countries, etc., are spoken by millions of people worldwide
• The languages have many cognates and some shared vocabulary
• The languages share a fairly simple writing system (esp. compared to hieroglyphics)

Lample et al. (2018) tested their system on Urdu->English, two unrelated languages with different writing systems and with (relatively) little available data, and obtain 12.3 BLEU. I don't have a reference point for how good that is, but it's definitely a start.

Finally, Zhang et al. (2019) train a translation system on Chinese -> Japanese, and show that it is possible to learn information about logographic writing systems, but Japanese kanji is borrowed from Chinese characters, so there's a lot of shared vocabulary.

That being said, hieroglyphs are (IMO) a more complicated writing system than even Chinese. And even for English-Urdu, a "low-resource" language pair, Lample et al. use 5.5M sentences. I have no idea how much text exists in hieroglyphs, but suspect it's less than this. But in theory, if we dug up and digitized millions of tablets of an ancient, unknown language, then yes, we do have tools to translate it.