If you’ve ever wondered whether the same algorithms that power word embeddings in natural language processing can be applied to something completely non-linguistic – like chess moves – then this episode is for you. In the latest installment of In-Depth with Academia, we explore a fascinating paper by Boris Oricov titled You Shall Know a Piece by the Company It Keeps: Chess Plays as a Data for Word2vec Models. It’s a wild blend of linguistics, chess, and machine learning that asks a deceptively simple question: can we use distributional semantics – the idea that words mean something by the company they keep – to understand sequences of chess moves? And if so, what does that tell us about meaning, context, and structure in non-human systems? The paper doesn’t claim to revolutionize chess engines or offer practical training tips. Instead, it’s a purely academic investigation into whether the same contextual logic that gives words their meaning also applies to move sequences in millions of recorded chess games. Oricov treats chess moves as words and entire games as sentences, then applies Word2vec – a popular embedding model – to see if similar moves cluster together based purely on their surrounding context. Two main approaches are tested: one that looks only at sequences of moves, and another that incorporates actual board positions as part of the context. The findings are surprisingly elegant. In the simpler move-only models, the closest synonyms for any given move tend to be moves made by the same piece – which is remarkable because the model has no prior knowledge of what a rook or a pawn is. It reconstructs that structure purely from co-occurrence patterns. When board positions are added, the model picks up even finer distinctions, such as differentiating early-game king moves from endgame king moves, recognizing castling as a unique event, and even discovering a mirror symmetry in games without queens – where moves by Black and White cluster in perfect balance. The model also passes simple outlier tests, like spotting an endgame move sneaked into a set of typical opening moves. While Oricov himself is clear that he doesn’t see immediate practical applications for chess engines or training, the real value is conceptual: it shows that meaning – or at least meaningful structure – can emerge from context in domains far beyond language. This isn’t about building a better chess AI; it’s about testing whether distributional logic holds in any structured sequence of symbols. If you’re a chess player, you might start thinking about your own move sequences in a new light. If you’re a linguist or NLP practitioner, it’s a reminder that context-driven pattern recognition isn’t tied to words alone. And if you just love the messiness of academic exploration, it’s a great case study in how applying tools from one field to another can reveal unexpected connections – even when the results aren’t immediately “useful.”
Why this matters for developers:
the core technique is pure NLP – Word2vec embeddings – applied to a completely non-textual domain. That means the same pipeline (tokenization, context windows, vector training, similarity metrics) is reused with minimal changes. This opens up a broader question: what other structured sequences could you embed? Musical notes? DNA codons? Trading patterns? The paper also highlights the importance of context design – whether you use local move sequences or global board states dramatically changes what the model learns. For developers working with embeddings, this is a perfect reminder that the choice of context definition is not neutral; it shapes the semantic space you discover. The implementation itself is straightforward (treat moves as tokens, feed into Word2vec), which makes it reproducible and hackable. And the evaluation is refreshingly interpretable – not perplexity scores on held-out data, but simple odd-one-out tests and nearest-neighbor inspections that let you see exactly what the model has captured.
What you’ll take away:
a concrete example of how to reframe a non-linguistic problem as a text-analogy problem and process it with standard NLP tools. You’ll see why distributional semantics is more about structure than about language per se, and how careful context engineering (moves only vs. moves + positions) leads to qualitatively different insights. You’ll also pick up a healthy dose of academic humility: the author explicitly states that this work is about exploration, not application, which is a useful antidote to the constant pressure for immediate practical results. And you’ll leave with a new way to think about sequences – whether in chess, code, or any other domain where order and co-occurrence matter.
Watch the full episode and ask yourself: if a computer can learn that a bishop’s move to C4 is more similar to a bishop’s move to D3 based solely on surrounding moves, what structured patterns might your own data reveal when you stop treating it as just numbers or events and start treating it as a language?
Stay curious – and maybe run Word2vec on your Git commit history to see what kind of “developer style” emerges.