Thanks Nora, I'll take a look at those links...

The reason for suggesting you disable this is because you have already complained about the way it works.  It does have flaws that need to be addressed and a quick search in here will show you that.  The way you describe it is exactly the way it works today if you enable it, but with a reasonable machine and no need to reduce slow look up times caused by poor connectivity for example there seems less value in using it.

Yes, I have quite a bit of coding knowledge. I have developed my own Lucene index applications for example.  People without coding knowledge are perhaps more likely than I to be impressed by what CAT Translation Memory tools are doing.  It ain't really rocket science: Levenshtein distances, all that mullarkey.

No doubt some people will claim that neural networks are involved.  I very much doubt that SDL 2019 really uses advanced things like that in any real sense of the term. 

I (obviously) don't know how familiar you are with Lucene indices (aka reverse indices)... they actually provide functionality which is more than powerful enough for TM apps.

Thanks. In my question I talked about phenomena A and B: B is the slowness I am finding with fuzzy matching with segments consisting of 30 or more words.  I should mention that I tend to translate with a colossal TM, built up over years, so maybe that is a factor here. 

If I turn off the "Match Repair" stuff maybe it'll become less sluggish but at the moment working *without* lookahead just on a normal translation job is painful in the extreme:  I am talking of waits of up to a minute. Insane.

Christopher Scully said:

No doubt some people will claim that neural networks are involved.

Where?

Christopher Scully said:

Yes, I have quite a bit of coding knowledge. I have developed my own Lucene index applications for example. 

Well, I'll look forward to the day you finish your new and improved toolset so I can see just how well things could work.

Christopher Scully said:

I should mention that I tend to translate with a colossal TM

What is a colossal TM?

Christopher Scully said:

I am talking of waits of up to a minute.

Clearly there is something wrong here.  But it would be good to know what size your resources are.  Most performance issues in Studio are related to influences such as these:

https://gateway.rws.com/csm?id=kb_article_view&sysparm_article=KB0032531

Hahaha.  Don't worry, it'll be FOSS.

Just let me ask you this: I do a lot of French legal contracts.  Very often French lawyers have a habit of capitalising certain terms in such documents: "La Partie qui dépose... ", "Le Bénéficiaire s'addressera au Vendeur ... ". Often these terms are repeated tens, even hundreds, of times in a given document. It's a right pain having to capitalise each English occurrence in the target.

So I've developed a little app which analyses the Word document for sequences and frequencies (there are Java classes which can analyse Word documents into paragraphs, words, etc.).  If it finds that the word "Bénéficiaire" crops up 100 times, with that capitalisation, the best thing is then not to bother with capitalising each instance , but to capitalise globally at the end.  My app mentions the number of lower-case ocurrences, which I can then search for if there are, say, only a couple in the whole document.

I go further: why bother even translating the word Bénéficiaire? Just leave it in the target until the very end of the job.  Then search-and-replace.

If SDL were FOSS someone somewhere would have had such thoughts, and a "case-sensitive sequence-matching"  function would have been built in years ago, to provide an analysis of things which will make your productivity (and consistency) better. 

Of course you can use such techniques on-the-fly, without prior analysis: I see this word "Vendeur" near the start of a document and assume it is going to crop up hundreds of times, with different casing variants, so I just leave it in the French until the end. Maybe many people already do this. But proper sequence analysis is better.

PS google | SDL "neural networks" |. There are tons of claims along these lines, e.g. www.sdl.com/.../what-linguistic-AI-has-learned.html

Thanks.  I'll also follow this up.