The Row Is Only The Starting Point
Last timeDirections That Mean Something
A single row cannot hold two senses of a word, so the layers above rebuild each position out of its neighbours until the same piece has different numbers in different sentences.
Consider the word bank. It has two senses that have nothing to do with each
other, and the table has one row for it. Not one row per sense. One row, consulted
by identifier, with no knowledge of which sentence the word turned up in.
This is a genuine limitation and not a technicality. The best a single row can do
is sit somewhere between the two senses, which puts it near neither set of related
words and makes it a poor representation of either use. For a long time this was
simply accepted as the cost of using a table.
- the numbers currently at position j in the passage
- how much position i draws from position j, computed from the content itself
- the new numbers at position i, built from the whole passage
It is worth being clear about what has and has not changed. The table is still
there, it is still consulted by identifier alone, and it still holds exactly one
row per entry. Nothing about the lookup became context-aware. What happened is
that the lookup was demoted from being the representation to being the first
guess, and the work of distinguishing one use from another was moved into the
layers above it, where the neighbours are available.
That reframing is the single most useful idea in this lesson. People often ask what
a model thinks a word means, expecting the answer to live in the table. The table
holds a starting point shared by every appearance of that piece. The meaning the
model acts on is assembled fresh for each occurrence and exists only while that
passage is being processed.
Watching one word diverge
The useful way to measure this is to take two appearances of the same word in
different sentences and ask how similar their representations are at each layer.
At the lookup, the answer is exactly one, because both are the same row. Going up,
the similarity falls.
The fall is not uniform. Words whose use genuinely varies across contexts diverge
more than words that behave the same everywhere, which is exactly what one would
hope. The upper layers are the most occurrence-specific, and the very top is pulled
back towards whatever shape the next-piece prediction job needs.
The lesson stops here
3 more paragraphs to go
You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.
The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.
See the planThe contentsThis is the reading half
Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.
The contents