Modularity in Generation Systems

It is widely accepted that one may divide the problem of natural language generation (NLG) into two subtasks:

• determination of the content of an utterance
• determination of its linguistic realization

This `divide and conquer' view of generation is the base of current architectures of systems. With few exceptions (e.g., [Appelt1985]) the following two components are assumed:

• `what to say' part (strategic component)
• `how to say it' part (tactical component)

Currently, in systems where the separation is advocated it is assumed that the strategic component is able to provide all information needed by the tactical component to make decisions about lexical and syntactic choices [McDonald1983,McKeown1985,Busemann1990,Horacek1990]. As a consequence, this implies that the input for tactical components is tailored to determine uniquely a good sentence, making the use of powerful grammatical processes redundant. In such approaches, tactical components are only front-ends while the strategic component needs detailed information about the language to use.

In [Appelt1989] it is shown that the use of a reversible grammar affects the modular status of a NLG in such a way that only the tactical component should be concerned with the specific details of a grammar and the strategic component should only perform general reasoning and planning tasks. In this view, the tactical component has the same status for the production process as the parser has for the understanding process.

Consequently, the strategic component cannot completely control the tactical component. For example, the following can happen. A message which is constructed precisely enough to satisfy the strategic component's goal can be under-specified from the tactical viewpoint. In particular, the generator can run into the risk of being misunderstood because of the produced utterance's ambiguity.

For example, if the strategic component specifies the following structure SEM as input to the tactical component:⁴

$$\left[\vphantom{ \begin{array}{l}sem: \left[ \begin{array}{l} con... ...y} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array} }\right.$$$$\begin{array}{l}sem: \left[ \begin{array}{l} cont: \left[ \begin{... ...\end{array} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array}$$ $$\left.\vphantom{ \begin{array}{l}sem: \left[ \begin{array}{l} con... ...y} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array} }\right]$$

then a possible utterance is `Remove the folder with the system tools' with the corresponding derived grammatical structure where the PP `with the system tools' is an adjunct to the VP:

$$\left[\vphantom{ \begin{array}{l} phon: \langle \mbox{\it remove... ...stem tools}\rangle\\ \end{array} \right] \vspace{0.5ex} \end{array} }\right.$$$$\begin{array}{l} phon: \langle \mbox{\it remove the folder with ... ...th the system tools}\rangle\\ \end{array} \right] \vspace{0.5ex} \end{array}$$ $$\left.\vphantom{ \begin{array}{l} phon: \langle \mbox{\it remove... ...stem tools}\rangle\\ \end{array} \right] \vspace{0.5ex} \end{array} }\right]$$

From the generator point of view this utterance is grammatical reflects exactly what the generator wants to express. For the hearer there also exists however the alternative grammatical structure where the PP `with the system tools' is a nominal adjunct:

$$\left[\vphantom{ \begin{array}{l} phon: \langle \mbox{\it remove... ...ex}\end{array}\rangle\\ \end{array} \right] \vspace{0.5ex}\end{array} }\right.$$$$\begin{array}{l} phon: \langle \mbox{\it remove the folder with ... ...space{0.5ex}\end{array}\rangle\\ \end{array} \right] \vspace{0.5ex}\end{array}$$ $$\left.\vphantom{ \begin{array}{l} phon: \langle \mbox{\it remove... ...ex}\end{array}\rangle\\ \end{array} \right] \vspace{0.5ex}\end{array} }\right]$$

with the semantic reading SEM':⁵

$$\left[\vphantom{ \begin{array}{l}sem: \left[ \begin{array}{l} con... ...y} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array} }\right.$$$$\begin{array}{l}sem: \left[ \begin{array}{l} cont: \left[ \begin{... ...\end{array} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array}$$ $$\left.\vphantom{ \begin{array}{l}sem: \left[ \begin{array}{l} con... ...y} \right] \vspace{0.5ex}\end{array} \right] \vspace{0.5ex}\end{array} }\right]$$

The whole situation can graphically be represented as follows:

The left triangle represents the domain of the derivation between the semantic structure SEM and the utterance `Remove the folder with the system tools' obtained during generation. Both triangles represent the domain of derivation between the utterance and the semantic structures SEM and SEM' computed during parsing.

Now the problem can be stated as follows. Since a tactical component is mainly guided by the compositional structure of the semantic input, it cannot control by itself those particular combinations of partial strings of the whole utterance which will lead to alternative derivations when the hearer is parsing this utterance. This means that possible ambiguities are out of the generator's view, and will only arise during parsing.

Of course, one could argue that if the generator had produced the utterance `Remove the folder by means of the system tools' instead of `Remove the folder with the system tools' then the kind of ambiguity exemplified above would not occur. Choosing the former instead of the latter in order to avoid ambiguity would mean that the strategic component is able to foresee that the generation process will run into the risk of generating an ambiguity, and hence of conveying misinformation. The problem here depends on the alternative possible realizations of the instrument role, namely `with' or `by means of'. The strategic component could have chosen `by means of' for some reasons internal to it (e.g., stylistic reasons, preferences, etc.) but not because it could foresee the ambiguity of `with'. In other words, given the modular design, the fact that at some point a potentially ambiguous LF surfaces as a un-ambiguous string cannot be assumed to be due to the fact that the ambiguity was foreseen, just other factors, independent from that, made the utterance unambiguous. If the strategic component chose `by means of' in order to restrict the set of possible derivations during parsing, this would mean that it is able to make decisions because of grammatical reasons.

The particular realization of the instrument role is not always relevant in order to avoid ambiguity. For example, in German (a language with relatively free word order) it would also be possible to utter:

`Mit den Systemwerkzeugen den Ordner löschen'
`[With the system tools] [the folder] remove'
(which can only mean Remove the folder by means of the system tools)

In this case the utterance is disambiguated by means of a specific ordering of the constituents. But now the same problem holds: Without detailed grammatical background the strategic component would not be able to specify the correct ordering in order to avoid ambiguity.

Summarising, it should be clear now that the strategic component cannot have this kind of control because otherwise this would blur the modular design of a generation system mentioned above. Fortunately, in many situations of communication a speaker need not worry about the possible ambiguity of what she is saying because she can assume that the hearer will be able to disambiguate the utterance by means of contextual information or that she would otherwise ask for clarification (Nevertheless, in the next section we show that the same problem mentioned above occurs also during clarification dialogs). However, an adequate generation system should also be able to avoid the generation of ambiguous utterances in some specific situations, e.g., when utterances refer to actions that have to be performed directly or in some specific dialog situations. As long as the strategic component has no detailed knowledge of a specific grammar it could not express `choose this particular form to avoid ambiguity'. Therefore it can happen that the intended message will not be conveyed.