Given a bilingual dictionary that translates lemmas of L₁ into L₂, its conversion is its transformation into an L₂–L₁ dictionary. If a dictionary were a glossary, this process would be trivial. However, as explained in the respective section, a bilingual dictionary is thoroughly asymmetrical. In practical linguistic work devoted to the description of L₂, the lexical database usually takes the perspective of an L₂–L₁ dictionary: the records of the database are constituted by L₂ lemmas, and linguistic information on them is provided in L₁. Typically, each L₂ lemma is provided with a set of L₁ equivalents. There is, thus, nothing in the microstructure of the database that could automatically provide L₁ lemmas.

There are a big solution and a small solution to the problem, both of which will only be sketched here. The small solution is roughly as follows:

Outputting a converse dictionary from a lexical database

If you plan to produce a converse dictionary from your lexical database, introduce additional fields (with respect to the microstructure set out in the respective section) into the database. At least the following are needed:

• L₁ lemma: The content of the field will mostly be one of the words appearing in the meaning definition of the L₂ lemma. More often than not, the L₂ lemma will correspond to more than one L₁ lemma. In that case, various instances of this field in a record are the solution in a free-field-structure database. In a relational database, a table of L₁ lemmas must be set up and a cross-table providing multiple linking of the L₂ lemma table and the L₁ lemma table.
• For each L₁ lemma, there is a set of fields specifying its grammatical properties. In a free-field database, this leads to a nesting of fields within a record. In a relational database, the L₁ lemma table is expanded by the appropriate number of columns.
• Finally, the set of examples associated with the L₂ entry must be assigned to the L₁ lemma fields, because some of them illustrate one L₁ sense, while others illustrate another L₁ sense. This is again done by nesting the fields of the record of the free-field database. In the relational database, there will be an additional cross-table linking the existent examples (so far linked only the L₂ entries) also to the L₁ lemma table.

Practically all of this is additional manual (or intellectual) work; very little of it can be automatized. Once the database has been prepared in such a way, it is then sorted according to the L₁ lemma field. If there are multiple occurrences of it within records, the same record must reappear in the list.

Then the L₁–L₂ text file can be output from the database in essentially the same way as the L₂–L₁ dictionary. Naturally, the fields have to be rearranged in the output; in particular, the order of the L₁ and the L₂ versions of each example must correspond to the direction of the dictionary.

Once the L₁–L₂ text file has been output, homonymy and polysemy must be distinguished manually: This involves spotting sequences of homonymous L₁ lemmas, deciding on homonymy vs. polysemy, grouping the senses of a polysemous item under one entry and numbering homonymous items.

Such a converse dictionary will be much thinner that its L₂–L₁ source, because instead of meaning definitions, it only contains copies of L₂ lemmas, and many of the fields of the database, including all of the diasystematic and encyclopedic information, will be missing.

Multilingual lexical database

The big solution to the bilingual dictionary is rather different. It is based on the idea of a multilingual lexicon, implemented as a relational database with the following structure:

1. There is a table of concepts that are language-independent.
2. There is a table of conceptual relations (like hyponymy, antonymy etc.), which is a kind of cross-table on the first table.
3. For each language L_i to be associated with the system, there is a table for its words. It contains all the relevant grammatical information for each word.
4. For each such L_i table, there is a cross-table associating each L_i item with one or more concepts of table #1. The association may be conditioned upon certain grammatical properties of the L_i item.
5. All pairings of an item of L_i with an item of L_j are then mediated by table #1. If there is, for a given L_i item, no L_j item linked to the same concept of table #1, a related concept – esp. a hyponym or a hyperonym – may be sought by using table #2.

Outputting a bilingual dictionary from such a database involves

• choosing any pair of L_i and L_j,
• for the L_i–L_j volume, associating the set of L_i items with their L_j equivalents by mechanism #5 above,
• sorting the database by L_i,
• outputting a text file in which all the information gathered from the various tables is assembled in L_i dictionary entries.