PhrasePump algorithm

[jtsoftware]

I’ve been using Phrase Pump for about a month now. In general I find it very useful for listening practice, but I have some concerns about it.

It might be good for you to describe the algorithm used for selecting the words used to find the example sentences. (I’m a language tool software developer myself, so I like to understand the algorithms of the tools I use.) Your help pages are surprisingly sketchy.

My first observation is that although each time there are new words and phrases coming up, it seems it is staying on the same words for a long time. I haven’t measured it, but it seems a long time. Is there a way to manage what words are used? Is it pure spaced repetiion? In the settings, I see a “New” control, which defaults to 15 %. Is that for new sentences or new words from which the sentences are drawn?

The Phrase Pump front page shows I have 1630 words that I’ve marked to learn (I only recently started to accumulate them. Previously, before using Phrase Pump, I would accumulate them from 1 or more videos, color coded for multiple videos, download them for one video for study in my flash cards, and then delete them, so I can start fresh.) It also says I am learning 36 words (3 months), 2 words (week), 1 word (day). The “learning” seems to imply “in progress”. Isn’t that way too slow a rate?

For the listening practice, when I study a sentence, if I understand a sentence before exposing the written form and translation, it seems like there should be a “knew it” button that would tell the algorithm to bring it back less frequently. It seems it repeats sentences I understand too often.

I don’t understand how it picks a few sentences for filling blanks or translating mode at the beginning. I would prefer to have the option of all one mode at a time, or mixing them up. One bug is that if the autotranslation comes out as “-”, these sentences are not excluded.

That’s my feedback for now. But thanks for your efforts in developing such a great tool!

[Anxis]

Having used spaced repetition software for 20 years, I share your concerns and find the lack of engagement by LR staff on the matter very regrettable.

[Renn_Raven]

I think I have a similar problem. I have been learning words through the Phrasepump for over 4 months now. Not a single word has moved to learnt in the last 4 months. I have been looking at the ‘half life’ and each time a word is reviewed it adds only a small amount (often 2 days). This means it takes a very long time for a word to become known - I’m not even sure what ‘half life’ a word needs to get to, to turn into a ‘learnt’ word.

It is great to have lots of repetiton, but I’m finding that it doesn’t show progress. For example I had a word that was to be reviewed in 13 days, then after 13 days, it’s half life changed to 15 after I had selected known in the quiz.

[cb83]

I made a post about this a little while and agree with you completely.

I’m just adding my comment here too in the hope that Phrase Pump is the next part of LR to get an upgrade.

[joan_LanguageProcess]

I was able to find this Information related to the PhrasePump algorithm/process/method:

@oaprograms

The algorithms are relatively simple, based on spaced repetition, it is not secret, just we should write about them more and fine-tune them a bit still… We’re busy making Chat and ASR now… Almost everything on the website is in Beta still. Likely we should have had different priorities, but we will fix lots of things soon, I think.

@David_Wilkinson

Currently the flashcard algorithm adds words to the study pool by first taking all the ‘Marked to Learn’ (queued) words first, then falls back to suggestions (most frequent words that you don’t know). We plan to add a mode shortly (could do it today) that selects appropriately between the two categories… taking ‘Marked to Learn’ words first, as long as they are not too infrequent (still in the orange colour range). This would mean you could mark a word as ‘Marked to Learn’ when you start to recognise it in context, but it wouldn’t be jammed straight to the front of the flashcard queue, but would show as soon as it can justified to learn intensively.

@hawardbrain334

PhrasePump is a language model designed to assist users with generating natural language content and providing responses to various prompts. The “Learnt” number you mentioned refers to the number of times the model has been trained on different datasets and examples. The model’s training process involves exposure to a wide range of text and language patterns to improve its understanding and generation capabilities.

Regarding the lack of a “correct” or “mistake” button when reviewing, it’s important to note that PhrasePump, as an AI language model, doesn’t have a direct feedback mechanism for users to mark correct or incorrect responses. The model generates responses based on its training and the input it receives. It doesn’t have the ability to evaluate its own responses for correctness.

@David_Wilkinson

Flashcards

Now that Language Reactor knows which words you know, and which you don’t, we have been able to integrate a powerful flashcard system, with some interesting features.

• You can review saved words and phrases as flashcards in a ‘Cloze’ format. Item reviews are scheduled using the advanced library ‘Ebisu’.
• Language Reactor can schedule extra ‘Look and Listen’ cards (enabled by default) that provide an opportunity to become more familiar with how the word is used. The cards have many example sentences with audio that you can review.
• At the end of a study session, Language Reactor can suggest new words to add to your study routine. It examines your current vocabulary and identifies gaps.

@David_Wilkinson

PhrasePump

We added this just a few days ago. This feature finds sentences from the Tatoeba database that only contain words that have been marked as known, or that you have practiced in the flashcards at least once (familiar words, solid outline). The sentence is read (audio); the text and translation are hidden until you hover over them, or hit the ‘e’ key. The concept is simple, but powerful. Use it to improve your listening comprehension and strengthen your language knowledge. By setting a small ‘known’ vocabulary (50 words), it’s a great way to start learning a language from scratch. We will build on this feature soon and perhaps integrate it with the flashcard feature.

They merged the above “Flashcards” and “PhrasePump” at some point.

I also managed to find this on Reddit from 9ish months ago:

Sentences are read using the latest neural TTS. Items are scheduled using an advanced library, ‘Ebisu’.
— davidzweig

I hope some of this helps!

P.S. I will update this if I find out any more information (things just kind of “pop up” on this forum sometimes that reveal a bit more of what LR is and how it works— /worked, in the case of some posts—unfortunately).

[Renn_Raven]

Thank you, that looks like a lot of people have the same problem. I still am having issues. I think they have accidently put +2 each time you review instead of x2. None of my words have moved to learnt yet. The one I photographed as being 13, then 15, is now at an interval of 18. It’s ridiculious. I hope they fix it soon as I do pay for pro and it doesn’t support good learning if things don’t move to learnt.