AI and descriptive notation

[Gerard Killoran]

I was intrigued by the possibilty that AI could translate games in newspapers from descriptive into algebraic notation, so I gave it a go with Copilot using the example below from the Field - Saturday 30 January 1915 p.43

test.png (322.35 KiB) Viewed 1324 times

Here are the results:

Copilot>
I can absolutely convert a chess position or game written in **descriptive notation** into **algebraic notation** — but I’ll need to *see the picture* first.

Upload the image of the chess game, and I’ll translate every move into clean, modern algebraic notation.

Having done so, I asked it to produce a pgn file

Copilot>
Absolutely, Gerard — here is the complete PGN file for the McCann–Bonwick French Defence game, fully reconstructed from the descriptive notation you provided.

You can copy‑paste this directly into any chess database or viewer.

With the following result

[Event "Lancashire vs Middlesex"]
[Site "?"]
[Date "????.??.??"]
[Round "?"]
[White "Dr. J. L. McCann"]
[Black "W. E. Bonwick"]
[Result "1-0"]
[Opening "French Defence"]

1. e4 e6 2. d4 d5 3. Nc3 dxe4 4. Nxe4 Nd7 5. Nf3 Ngf6 6. Bd3 Nxe4 7. Bxe4 Nf6 8. Qe2 Nxe4 9. Qxe4 Be7 10. O-O O-O 11. Bf4 Rb8 12. Rad1 a6 13. Rfe1 Bb7 14. d5 Bd6 15. Ng5 h6 16. Bxd6 Qxg5 17. Bxb8 Bxh2+ 18. Kxh2 Rxb8 19. Ba7 Qc7+ 20. Qe5 Rb7 21. Bb8 Qxe5+ 22. Bxe5 Rb4 23. g4 Rb6 24. Bc7 1-0

However this is the true score

[Event "Lancashire vs Middlesex"] [Site "?"] [Date "????.??.??"] [Round "?"] [White "Dr. J. L. McCann"] [Black "W. E. Bonwick"] [Result "1-0"] [Opening "French Defence"] 1. e4 e6 2. d4 d5 3. Nc3 dxe4 4. Nxe4 Nd7 5. Nf3 Ngf6 6. Bd3 Nxe4 7. Bxe4 Nf6 8. Qe2 Nxe4 9. Qxe4 Be7 10. O-O O-O 11. Bf4 Rb8 12. Rad1 b6 13. Rfe1 Bb7 14. d5 Bd6 15. Ng5 g6 16. Bxd6 Qxg5 17. Bxf8 Bxd5 18. Rxd5 exd5 19. Be7 Qd2 20. Qe5 Re8 21. Bb4 Rxe5 22. Bxd2 Rf5 23. g4 Rf3 24. Bh6 1-0

As you can see, Copilot goes wrong on move 12...a6 instead of ...b6 for no obvious reason. I wonder if other AI models can do a better job?

[Matt Mackenzie]

I see this is the really old fashioned form of Descriptive that uses (eg) "K sq" instead of K1.

[Ian Thompson]

As you can see, Copilot goes wrong on move 12...a6 instead of ...b6 for no obvious reason. I wonder if other AI models can do a better job?

If it was intelligent, it would, at least, realise that Black's move 13 was illegal as a result of its mistake on move 12, so the score it was offering could not be right. At the very least, it should tell you that move 13 in its game score is illegal.

When I tried this with Copilot on your gamescore image it came up with

Code: Select all

1. e4      e6  
2. d4      d5  
3. Nc3     Nf6  
4. Bg5     Bb4  
5. Bd3     Nc6  
6. Nf3     O-O  
7. O-O     h6  
8. Bh4     dxe4  
9. Nxe4    Be7  
10. Nxf6+  Bxf6  
11. Bxf6   Qxf6  
12. Qe2    Rd8  
13. Rad1   Bd7  
14. Qe4    g6  
15. Ne5    Nxe5  
16. dxe5   Qg5  
17. Qxb7   Qxe5  
18. Qxc7   Qxc7  
19. Rd2     Rd8  
20. f3      Qe7  
21. f4      Qd4+  
22. Kh1     Bc6  
23. b3      Bb7  
24. Rdf2    Rxf2  
25. Rxf2    Rd5  

0–1

so it's gone wrong on move 3 and is completely wrong after that. Black's move 19 is illegal.

It converted this to PGN as

[Event "?"] [Site "?"] [Date "????.??.??"] [Round "?"] [White "?"] [Black "?"] [Result "0-1"] 1. e4 e6 2. d4 d5 3. Nc3 Nf6 4. Bg5 Bb4 5. Bd3 Nc6 6. Nf3 O-O 7. O-O h6 8. Bh4 dxe4 9. Nxe4 Be7 10. Nxf6+ Bxf6 11. Bxf6 Qxf6 12. Qe2 Rd8 13. Rad1 Bd7 14. Qe4 g6 15. Ne5 Nxe5 16. dxe5 Qg5 17. Qxb7 Qxe5 18. Qxc7 Qxc7 19. Rd2 Rd8 20. f3 Qe7 21. f4 Qd4+22. Kh1 Bc6 23. b3 Bb7 24. Rdf2 Rxf2 25. Rxf2 Rd5 0-1

[Christopher Kreuzer]

If you pay to use better AI you likely get better results. If you keep training AIs with this sort of thing and tell them where the mistakes are, they will get better. But unless you have a large enough corpus to train it on, you will get errors. I am presuming these are all LLM AIs.

It is slightly depressing that AI gets used as a way to do OCR scanning (mainly because of the environmental considerations). But it is one of the applications. Some industry experts really fall in love with it. Others see some issues that are difficult to solve.

Smudged print, typos, idiosyncratic syntax and formats (e.g. the annotation numbering above). One of the major issues is if there is a missing word in a scanned text and it could be one of several plausible options. Would you trust the human who is not sure, or the AI who confidently insists it knows the answer? (The reason some AIs have been programmed to be confident is obvious but still worrying as humans all too often fall for confidence shysters.)