掲示板 Forums - Text Analyzer v2: feedback

I'd love any feedback on the visual update of the Text Analyzer, as well as the new data available. Specifically, here's what's changed:

1. It no longer runs each piece of text through the *fast* and then the *thorough* parsers - it just sticks to the more accurate one.

2. On tapping the name of the text, you get a new page with all the text info / reader, in addition to a new breakdown of grammar found in the text.

3. On the updated text list, you now have a Word and Grammar level. It's determined by the easiest (highest) level of the JLPT that covers 70% of the material.

Example: If you have N5: 80% of the words, it would be rated as N5

However, if you have 60% of N5 and 15% of N4 (=75% at N4 or easier), it would be rated N4.

Current concerns that I have:

1. Wondering if, in addition to overall grammar, if conjugation data should be split out as well (what percent is in polite, casual, etc.)

2. Should the 70% value be customizable?

3. While the JLPT level works fairly well on grammar, I feel like it works less so on vocab. When it comes to a lot of the text being posted in here (lyrics, novels, manga, anime subs, etc.), I feel like the JLPT (even with the expansions that renshuu is putting in with the revised Beginner/Basics lists) is narrow enough that more things are going to be marked as N1 (highest level), giving texts an impression of being overly difficult.

Anyway, I'd love any feedback or comments!

I think making the 70% value customizable would be good. I would probably want to use a much higher value (closer to 90%).

That being said, I'm not actively using the Text Analyzer, so my opinion on the matter should only hold a teeny tiny amount of weight.

I just noticed something else. The vocabulary percentages for the N Levels for my texts (including "not in JLPT") are not adding up to 100%. They all seem to give a total of around 80%.

The 100% issue does seem to be working ok now. And thanks for reminding me about how to view the "not in JLPT" words as well. That's very useful! But the overall word level score is still very misleading. Take for example one of my texts, where the vocab levels are:

• N5 - 52%
• N4 - 6%
• N3 - 6%
• N2 - 2%
• N1 - 0%
• Not - 34%

This text is rated as an N1 vocabulary, mostly because 34% are not in JLPT. But when I looked at those unlisted words I see:

• 十月
• ~ごろ
• 居酒屋
• 焼鳥
• ポテト
• トマト

Most of these words can be found in the most common word lists, or the most common news words list. So I assume that means that they are words that N5 or N4 learners would be able to recognize (at least the kana versions for words like izakaya and yakitori).

I understand that in an advanced text, the "not in JLPT" words would probably be advanced N1+ words. But in easy, graded reader style stuff like what I am looking at, the "not in JLPT" words are often common every day things that aren't "academic/business" enough to be on the JLPT lists, so it is skewing the results.

Potential solutions are (maybe?)

• an option to not consider the "not in JLPT" words and just use the cutoff percentage on what is left
• a histogram or other visual way to see the breakdown (echoing a previous post)
• add information to the analysis about how many words are in the "common" word lists
• allow users to set the cutoff (based on what I see in my data, 50% seems more accurate)

Oh, and by the way, most of the texts that I saved have been fairly easy, and the grammar level for each is mostly showing N4 or N5, so that seems to be working pretty well.

Most texts I've tried (easier and harder) came out as N1 regardless, which is misleading and makes the feature feel useless overall.

I do like the idea of using a frequency list, rather than JLPT level, to estimate text difficulty. But I agree that finding one that conforms to "common sense" wouldn't be easy.

I ran a quick search in some lists and got fairly different results: Wiktionary (https://en.wiktionary.org/wiki/Wiktionary:Frequency_lists/Japanese) puts 秋 at 794 and 遊ぶ at 3143, Leeds Corpus (http://corpus.leeds.ac.uk/frqc/internet-jp.num) puts 秋 at 1899 and 遊ぶ at 1593 and NINJAL's BCCWJ (https://clrd.ninjal.ac.jp/bccwj/en/freq-list.html) puts 秋 at 1357 and 遊ぶ at 995.

Bit of an update - since there is always some ambiguity for certain terms (usually ones without kanji), I can (currently) match up 25.5k out of 30k of the top words in the third frequency list.

As I continue to work on this, I'm wondering how to divide up the groupings. For example, if I set 5 difficulty levels (5 easiest, 1 hardest), what "categories" would line up with each of those?

Example:

70% in top 1,000 = Level 5

70% in top 2,500 = Level 4

70% in top 4,000 = Level 3

70% in top 8,000 = Level 4

otherwise

Level 5

Something is wrong. Everything looks skewed the other way now:

I would've expected a text on Kant's categorical imperative to rank slightly higher than level 4...

The word level definitely looks better now.

The grammar level percentages might need to be tweaked a little. Most texts I've tried are marked N4. Presumably because even texts geared towards N1 still use plenty of more common grammar too (e.g. から, ながら, だろう). A text comprised solely of N1 grammar examples still only ranked N2:

By the way, does the text analyzer parse grammar differently from the sentence lookup tab?

I tried checking the same sentence (#167770) in both. The lookup said 笑っているの = の、~こと (Changes A into a noun) while the text analyzer said 笑っているの = て、~で (Since A; As A) + いる (Present Casual).

I don't expect it to be perfect, but I did think it would be consistent...

On my smartphone the table is too wide to fit on the screen so you have to scroll to the right. This also seems to cause the pop-up dictionary to act word.

something else I would love to have would be some way to store additional notes about the text. I often paste stuff like nhk articles into it and would like to save the URL of the original article.

I found a minor but interesting bug.

I do the following:

Read a previously analyzed text.

close it again to get back to the menu.

try to add a new text.

the submit button is missing, only cancel is there.