Encoding LP constraints in a constraint-based framework without adding LP constraints to the formalism.

This sketch (initially written by AnnCopestake) illustrates how an encoding of a bit-vector style implementation of linear precedence could be done directly in a constraint-based formalism. The approach is inspired by Daniels and Meurers. The point is that the bit-vector approach could be directly implemented in the typed feature structure logic and this is an attempt to work out what this would entail. Unlike normal HPSG parsing, the conditions that apply to phrasal spanning are fully encoded in the feature structure constraints with no additional requirements.

(It is entirely possible that something similar has been done before - it would be good if pointers to such work could be added to this page.)

The encoding here requires:

  • a full append;
  • one additional type of recursive constraint (here called afterme), which is used as a linear precedence constraint although defined in terms of the vectors;
  • the ability to specify types on a per sentence basis (necessary because the length of the vector has to be fixed);
  • initial settings of the vector on lexical signs according to their position in the input;
  • the ability to define an indefinitely large set of constant types (cf instloc values in generation).

The encoding would allow a quick-check style mechanism to be used to enforce LP constraints. Constituents may be discontinuous.

The intuition behind the encoding is that all phrases contain a vector (encoded as a list) which has the same length as the number of signs in the initial input. For each phrase, the type of each element in the vector indicates whether it corresponds to an initial element before or after the phrase, or part of the phrase, or spanned by the phrase but not a constituent part. By comparing vectors for different phrases, we can enforce LP constraints.

A. Types encoding the vector

The general condition on phrase is:

[ phrase
  LP [ pos-constraint
       BEFORE [1]
       ME append ([2],span-list) unify append(span-list,[3])
       AFTER [4] ]
  DTR1.LP [ pos-constraint
            BEFORE [1] 
            ME [2] ]
  DTR2.LP [ pos-constraint
            ME [3]
            AFTER [4] ]  ]

where the general constraint on pos-constraint is:

[ pos-constraint
  POSVEC append([1],[2],[3])
  BEFORE [1] before-list
  ME [2] span-list
  AFTER [3] after-list ]

where before-list is defined as a list of zero or more elements of type before, after-list is a list of type after and span-list of type span. pos-constraint is the appropriate value of LP which is a feature found on all signs.

pos := top.
outside := pos.
after := outside.
before := outside.
span := pos.
me := span.
meA := me.
meB := me.

and so on.

If we are parsing a five word sentence, the type posvec which is the appropriate value of POSVEC is set to the list <pos,pos,pos,pos,pos>. Thus all signs have a 5-element posvec. The idea that types can be varied according to the input sentence is unfamiliar, but formally does not seem to be problematic, since we define parsing of a sentence with respect to a type system.

POSVEC values for the input signs are set according to their positions in the input (see below).

The root condition is that the POSVEC value is <me,me,me,me,me> (i.e., that the phrase spans all the input).

It follows logically from the append condition above, that before may not follow a me etc, but this would not follow in the feature structure formalism without some additional constraint.

B. An example of vectors in phrases

Assuming we are parsing the input A B C D E

initial POSVEC values are set as follows:

A  <meA, after, after, after, after>
B  <before, meB, after, after, after>
C  <before, before, meC, after, after>
D  <before, before, before, meD, after>
E  <before, before, before, before, meE>

The setting of the POSVEC values should not cause any formal worries - there has to be some notion of mapping between an input string (or list of token-strings) and signs: this just complicated that notion a little.

Suppose we have a parse tree such that:

S dominates X and Y
X dominates A and C
Y dominates B and Z
Z dominates D and E

phrase X has POSVEC < meA, span, meC, after, after> (indicating that it dominates A and C)

phrase Y has POSVEC < before, meB, span, meD, meE >

phrase Z has POSVEC < before, before, before, meD, meE >

S has < meA, meB, meC, meD, meE >

The phrasal POSVEC values arise from the initial values by virtue of the constraint on type phrase. To see this in a relatively complex case, consider the formation of S:

[ phrase
  LP [ POSVEC append([1],[3],[2])
       BEFORE [1] <>
       ME [3] append ([4],span-list) unify append(span-list,[7]) 
        ;;   < meA, span, meC, span, span> unify  <span, meB, span, meD, meE > 
        ;;   < meA, meB, meC, meD, meE >
       AFTER [2] <> ]
  DTR1.LP [ POSVEC append([1],[4],[5]) ;;; X
            BEFORE [1] <>
            ME [4] < meA, span, meC>
            AFTER [5] < after, after> ]
  DTR2.LP [ POSVEC append([6],[7],[2]) ;;; Y
            BEFORE [6] < before >
            ME [7] < meB, span, meD, meE > 
            AFTER [2] <> ]  ]

The trick here is the specification of ME on the phrase. This says that it is the unification of the filled-in values of the MEs of the daughters. The assumption is that DTR1 has a leftmost element which precedes the leftmost element of DTR2. In all cases, the length of ME of the phrase will be known, since the POSVEC is constrained to be the length of the sentence and the lengths of the BEFORE and AFTER lists are known (the BEFORE list coming from the BEFORE of DTR1 and the AFTER list from the AFTER of DTR2). Thus, although the append constraints on ME specify that the filler is a span-list of indefinite length (possibly zero) the length will actually be known when the rule is applied (assuming we’re operating bottom-up). In this case, DTR1 and DTR2 overlap, but the rule also allows for them to be contiguous (corresponds to the situation where ME of the phrase is the append of the ME of the DTRs) or to have some stuff in between (corresponds to an append of the ME of DTR1, some spanning elements, and the ME of DTR2). The use of the different me subtypes (meA etc) allows for failure of unification in cases where we attempt to combine two phrases which each contain the same constituent. Thus this formalisation encodes the no-overlap constraint reqired in parsing.

C. LP constraints

To allow lexical signs to specify that they select for signs that preceed/follow them, we use the type afterme (other types could be defined for before / immediately before / immediately after but these won’t be considered here).

For instance:

LP.POSVEC [1]
SUBJ.LP.POSVEC [2]
COMPS <LP.POSVEC [3], LP.POSVEC [4]>
LP.CONSTRAINTS < afterme([1],[3]),afterme([1],[4]),
                 afterme([3],[4]),afterme([2],[1])>

corresponds to the order: subj < sign < comp1 < comp2

To implement this, afterme is defined in terms of posvecs, so that any span values in the first posvec in an afterme are matched by before values in the second posvec. This can be done with a recursive type definition.

[ afterme
  PV1 posvec
  PV2 posvec ]

has two subtypes

[ afterme-elist
  PV1 elist
  PV2 elist ]

[ afterme-ne
  PV1 [ FIRST pos
        REST [1]]
  PV2 [ FIRST pos
        REST [2] ]
  JUNK [ afterme
         PV1 [1]
         PV2 [2]]]

where afterme-ne has two subtypes

[ afterme-me
  PV1.FIRST span
  PV2.FIRST before ]

and

[ afterme-other
  PV1.FIRST outside ]

Thus, if there is an afterme constraint between two posvecs, then any span values in the PV1 posvec have to be matched by before values in the PV2 posvec. This corresponds to the situation where the PV1 phrase has to be before the PV2 phrase in the input.

e.g.

PV1 < meA, span, meC, after, after > 
PV2 < before, before, before, meD, after >

satisfies afterme

PV1 < meA, span, meC, after, after > 
PV2 < before, meB, span, meD, after >

fails afterme.

Consider the A B C D E example as before, and assume that C somehow selects for Y and that we know afterme(C,Y)

X is

   [ LP [ POSVEC append([1],[4],[5]) 
           BEFORE [1] <>
           ME [4] < meA, span, meC>
           AFTER [5] < after, after>
           CONSTRAINTS afterme(< before, before, meC, after, after >, [8]) ]
     SELECTED.LP.POSVEC [8]]

Y has

   [ LP [ POSVEC append([6],[7],[2]) 
            BEFORE [6] < before >
            ME [7] < meB, span, meD, meE > 
            AFTER [2] <> ]  ]

so instantiating SELECTED with Y will give

afterme(< before, before, meC, after, after >, < before, meB, span, meD, meE >)

and this will fail.

The general requirement to implement this is that the constraints can be specified between two signs whose POSVECs can be accessed. This works in cases where the signs are selected for (including cases where the two signs in an afterme relationship are specified by a third sign, as in the two COMP example above), but does not allow for global constraints of the type considered by Reape, although it’s possible that these could be encoded somehow as constraints on phrases.

D. Implementation prospects

At one level, this is a theoretical exercise, since it requires recursive constraints. However, both the use of append and afterme could be implemented as special cases in a bottom-up parser. By keeping the vectors in the feature structure rather than having a separate LP layer, we might obtain the following advantages:

  1. formally, this seems straightforward and relatively clean
  2. existing techniques for efficient parsing can be reused. Specifically, the POSVEC constraints can be checked by the a variant of the quick-check mechanism and the rule filter might also operate reasonably directly. The parser itself is a simplification of existing algorithms. Basically, to implement this within the current LKB one would have to make use of a hacky way of setting the POSVECs instead of the append and also directly set the vectors for the afterme types (cf setting the ORTH value in morphology via evaluate-unifications).
  3. there is no need to expand the TDL formalism to include a new notation for LP constraints and there is some scope for experimentation by the grammar writer with different approaches to constraints, since they are part of the FS (provided the afterme hack is implemented sufficiently generically).
  4. constraint effects propagate directly. Consider this example again:
LP.POSVEC [1]
SUBJ.LP.POSVEC [2]
COMPS <LP.POSVEC [3], LP.POSVEC [4]>
LP.CONSTRAINTS < afterme([1],[3]),afterme([1],[4]),
                 afterme([3],[4]),afterme([2],[1])>

and assume that [1] is < before, me, after, after, after> then [3] is < before, before, pos, pos, pos >, [4] is < before, before, pos, pos, pos >, [2] is < me, after, after, after, after >. If we find [4] is < before, before, before, me, me >, then we get [3] is < before, before, me, after, after >

Bad things:

  1. POSVECs in feature structures could be confusing. This is a very low-level sort of thing to have in a grammar.
  2. The approach may not be expressive enough.

Last update: 2006-12-16 by AnnCopestake [edit]