An NLG pipeline for a legal expert system: a work in progress

First, we declare data types for the entities needed to represent the rules. A round of RPS is a Game with two Players, each throwing Sign. Then we introduce the three signs: rock, paper and scissors.

Next, we define four predicates, used later in our rules. The current syntax of L4 is adopted from functional programming languages. For example, the type signature for Win : Player $\rightarrow$ Game $\rightarrow$ Bool means “the function Win takes a Player and a Game, and returns a Boolean”.

For the purposes of constructing an interview, it is useful to have information on how to group the different predicates. L4 offers an alternative syntax for the predicates to be fields in the data types, as follows.

The class definition with the fields throw, win and participate is equivalent to the stand-alone predicates in capital letters. Note that capital letters have no significance in L4, we just use them in this example to distinguish between standalone predicates and those in a field.

As an alternative, participate and win can be defined as fields of the class Game. With this grouping, the type signature becomes Game $\rightarrow$ Player $\rightarrow$ Bool. We will use this grouping in the examples in Section 4.1, because that makes the best order for asking the questions.

The questions are embedded in a Docassemble interview. Docassemble is an open-source legal expert system, where users answer a set of questions through a browser interface. The responses are then compiled together into a document or processed further in other applications—in our case, an automated reasoner.

In order to let the user customise the interview—for instance, the NLG, question ordering, additional links to source material— our system outputs a LExSIS (Legal Expert System Interface Schema) file, which is short and declarative. The final interview, a long and imperative piece of Python code, is then constructed from the LExSIS file.

The interview includes three types of questions: a yes–no question (1), an open-ended wh-question (2) and two enumeration questions (3,4). These are determined by the L4 code.

Game is a class, so the first question we ask is whether a game exists. When we know that a game exist, we ask wh-questions about the predicates related to the game: who participates and which sign do they throw. The L4 code lists three signs, but no pre-existing players: that’s why the answer options for (2) is a slot for free text, and for (3,4) an enumerated list.

The choice of who in (2) is because player is a human noun. The GF-WordNet lexicon Angelov (2020) contains this information for about 6500 words, so we just make it as a rule in the grammar. If some human nouns are missing the animacy information, they will just be verbalised with the general strategy, which players.

The order of the questions is determined by Docassemble and the L4 encoding.

Docassemble has its internal logic: the goal of the interview is this case to find out the winner of a game of RPS, and thus it starts out from the first prerequisite, is there a game?

L4 too has its own logic: if we used the encoding where participate is a field of the class Game, then it will group the questions accordingly. Participation is a feature of a game, so after ensuring that a game exists, the user is asked about the game’s participants—game is now old information, so in English it gets a definite article and is placed later in the question. With a different grouping of predicates in L4, the questions will also be grouped and phrased differently: for instance, Are there any players, followed by Which game does Alice participate in.

Table 2 shows the GF expressions needed to construct different types of sentences. All of the functions in normal typeface come from the language-independent API of the GF RGL²²2 http://www.grammaticalframework.org/lib/doc/synopsis/. The variables in italics are generic: any noun phrase could be in place of np, any transitive verb for v2. For a verb like participate in, the preposition is part of the verb’s lexical entry—that’s why the expression to construct a VP is identical for participates in RPS and throws rock.

None of the expression templates which includes a noun phrase specifies a determiner. The NP can be constructed in different phases: it can be new or old information, or it may be a proper noun. Unfortunately, the lexicon doesn’t contain properties like mass noun vs. count noun, nor idiomatic expressions, like throw rock instead of throw a rock. Currently, we just accept that sometimes the determiner choice is unnatural, and leave it to the user to postprocess the output.

The reasoner we use is s(CASP) Arias et al. (2018): a logic programming language for Constraint Answer Set Programming.

The L4 code is translated into s(CASP) rules and statements. The query “who wins a game of RPS” is phrased as ?- win(Game,Player), where Player and Game are s(CASP) variables. The order of the arguments comes from the L4 encoding where win is a field of the class Game.

The L4 encoding doesn’t list any players or games, so s(CASP) needs more information before it can meaningfully run the query. As we have seen in 4.1, this information comes from the interview.

The answer consists of one or more sets of s(CASP) statements. These statements, just like their source in L4, are connected to the CNL descriptions, which are stored as GF abstract syntax trees. This makes it possible to aggregate them, such as grouping them by subject or predicate. In case of multiple sets for the same query, we group repeating statements in one block, grouped by and, and the rest under another, grouped by or. An example of this strategy is shown in Table 2, which shows an answer to a slightly different query: “list all ways that Alice can win a game”.

Currently, our order of aggregating predicates is intransitive before transitive, and nouns/adjectives before verbs. This results in sentences like are players and participants in RPS, are players and participate in RPS, play and participate in RPS, or less fluently, are participants in RPS and play.