An overview of the lexicon module's role and abilities.

§1. Prerequisites. The lexicon module is a part of the Inform compiler toolset. It is presented as a literate program or "web". Before diving in:

§2. A symbols table for natural language. This module provides an analogue to the "symbols table" used in a compiler for a conventional language. For example, in a C compiler, identifiers such as int, x or printf might all be entries in such a table, and any new name can rapidly be checked to see if it matches one already known.

In natural language we have "excerpts", that is, contiguous runs of words, rather than identifiers. But we must similarly remember their meanings. Examples might include:

american dialect, say close bracket, player's command, open, Hall of Mirrors

Conventional symbols table algorithms depend on the fact that identifiers are relatively long sequences of letters (often 8 or more units) drawn from a small alphabet (say, the 37 letters, digits and the underscore). But Inform has short symbols (typically 1 to 3 units) drawn from a huge alphabet (say, 5,000 different words found in the source text). Inform also allows for flexibility in matching: the excerpt meaning give # bone, for example, must match "give a dog a bone" or "give me the thigh bone".

We also need to parse in ways which a conventional compiler does not. If C has registered the identifier pink_martini, it never needs to notice martini as being related to it. But when Inform registers "pink martini" as the name of an instance, it then has to spot that either "pink" or "martini" alone might also refer to the same object.

Finally, we have to cope with ambiguities. An innocent word like "door" might have multiple meanings, and the more so once multi-word flexible patterns are involved.

§3. This is not a large module, but it contains tricky and speed-critical code. In compensation, it exposes a very simple API to the outside world, all of which is found in Lexicon (in lexicon).

The lexicon is stored using excerpt_meaning objects, in Excerpt Meanings. Entries are added with Lexicon::register and retrieved with Lexicon::retrieve.

In either case the user must supply a "meaning code", such as TABLE_MC, giving a very loose idea of the context; we will use that both to make lookups faster, to provide separate namespaces (one can search for just TABLE_MC meanings, for example), and to control the style of parsing done. See How To Include This Module (in lexicon).

§4. Optimisations. This is a speed-critical part of Inform and has been heavily optimised, at the cost of some complexity. There are two main ideas:

Firstly, each word in the vocabulary gathered up by the words module — i.e., each different word in the source text — has a vocabulary_lexicon_data object attached to it. This in turn contains lists of all known meanings starting with, ending with, or simply involving the word.

For example, if "great green dragon" is given a meaning, then this is added to the first-word list for "great", the last-word list for "dragon", and the middle-word list for "green".

In addition, every word in an excerpt which is not an article adds the meaning to its "subset list" — here, that would be all three words, but for "gandalf the grey", it would be entered onto the subset lists for "gandalf" and "grey". Subset lists tend to be longer and thus slower to deal with, and are used only in contexts where it is legal to use a subset of a name to refer to the meaning — for example, to say just "Gandalf" but mean the same wizard.

§5. Secondly, recall that each vocabulary entry has a field 32 bits wide for a bitmap, of which only 6 bits were used in the lexer. (See Vocabulary (in words).) For example, cardinal numbers had the NUMBER_MC bit set.

We're now going to use the other 26 bits. The idea is that if a meaning is registered for the name of, say, a table, then the TABLE_MC bit would be set for each of the words in that name. For example, if "table of tides" is such a name, then each if the words table, of and tides picks up the TABLE_MC bit.

What we gain by this is that if we are ever testing some words in the source text to see if they might be the name of a table, we can immediately reject, say, "green table" because the word green does not have the TABLE_MC bit.

For more on this, and for complications arising to do with case sensitivity, see ExcerptMeanings::hash_code_from_token_list.

§6. Performance in practice. The following statistics show how many times the lexicon was used during a typical Inform 7 compilation (the same one used to generate the data in Performance Metrics (in inform7)).

Optimisation is worthwhile if:

Optimisation is efficient if: 

Size of lexicon: 3169 excerpt meanings
  Stored among 859 words out of total vocabulary of 10634
  723 words have a start list: longest belongs to report (with 293 meanings)
  21 words have an end list: longest belongs to digits (with 7 meanings)
  29 words have a middle list: longest belongs to to (with 4 meanings)
  108 words have a subset list: longest belongs to street (with 4 meanings)

Number of attempts to retrieve: 110617
  of which unsuccessful: 92459
  of which successful: 18158

Total attempts to match against excerpt meanings: 279511
  of which, total with incorrect hash codes: 256747
  of which, total with correct hash codes: 22764
  of which, total which matched: 19929