I used Antlr3 to create a NLP parser for entering a calendar event. For iOS, the C runtime's parse tree source was many megabytes and difficult to debug directly. For Android, the Java runtime parse tree ran out of memory and had to be broken into smaller pieces. Haven't tried Antlr4.
Current workaround is to parse a document directly into a token tree. Instead of separate lever + parser, merged a tweaked BNF with regex. Memory stays rather small and can download a revised grammar without recompiling a new binary.