HSK Manga: Turning Manga into Beginner-Level Chinese

2 points by briankung


elihunter173

Biasing individual tokens seems like a clever approach but I am suspicious that you can produce comprehensible graded readers just translating a single panel at a time.

As someone who has read a lot of Chinese graded readers, I can tell you that a lot of what makes a graded reader good is what they choose to repeat and what they choose to skip. Good graded readers repeatedly re-emphasize the same concepts, grammar patterns, and words consistently throughout the book to give you varied exposure to them, while at the same time skipping large, non-essential chunks of the story in order to make it not too boring.

I don't think an LLM translating panel by panel is capable of making these sort of decisions.

dutc

I think that these kinds of pet projects were more beneficial where the majority of the work was done by the author.

Obviously, having an LLM generate a bunch of code on your behalf makes it more likely you will be able to complete something that you can show others, but there seems to be a dilemma in that the better the model is at generating output, the less one is exposed to their own creation. As a result, it is possible to create things that are polished but ultimately not particularly meaningful or useful.

To be clear, even if the vast majority of the work was performed by the LLM, I think it's impressive for a high school student to have created anything at all. But, to be clear, I also am skeptical that the result, that this tool, is actually very useful or interesting for learning.

In my life, I have found that the process of building something by hand draws me into the creation, and it is often only after significant effort that I realise that I shouldn't be trying to build the thing in the first place. Were I to be able to build the thing without this effort and without this friction, it seems I would be robbed of this realisation.

In the case of this particular project, the home page claims “learning Chinese is hard,” and the linked page mentions that “writing [graded readers] by hand doesn't scale” and “engaging graded text is hard to come by.” None of this is true.

The post and the website seem to focus on preparing for the HSK 3 level. While I am not familiar with this exam, as I understand it, the level of proficiency it measures is incredibly low. I suspect that traditional (low-volume) instructional materials would be wholly adequate to achieve this level. In fact, I suspect that traditional materials would likely to be sufficient for the entire set of HSK examinations, and additional materials would be necessary only for developing proficiency beyond these. However, at that level, it's probably the case that one might be able to rely on materials designed for young age native speakers.

Again, it is impressive and laudable to see any kind of output generated by any means by a high school student, but I suspect a project like this is primarily valuable in what building it cultivates in the creator. If mostly everything was LLM-generated, what's left?