alignment.md

Alignment between UMR graph and the sentence tokens

Alignment is an important part of UMR annotation that anchors UMR concept nodes to the sentence. Unlike AMR, alignments in UMR are supposed to be available for every annotated sentence and are directly stored in the same file as the graphs. However, at present they are not documented in the UMR annotation guidelines. We have to guess the rules from the released data or from what the UMR annotation tool does.

A UMR file has four annotation blocks for each sentence. Each block starts with a comment line (first character is #) and ends with an empty line. The alignment is described in the third block but the first block is important as well because it defines the tokens to which the alignment refers. The format of the first block is not unified and varies across languages in UMR 1.0. The important point is that the surface sentence must be presented as a sequence of tokens where neighboring tokens must be separated by a space character. So for example, we must insert a space between a word and a following comma.

The third block has as many non-comment lines as there are concept nodes in the sentence level graph. The order of the lines (nodes) is not significant. Each such line starts with a node id (variable), followed by a colon and a space, followed by numeric references to token ranges. For example, the following line says that node s16p is aligned to the first token of the sentence (whose index is 1, not 0):

s16p: 1-1

Note that even if the node is aligned to a single token, it is still presented as a range.

Discontinuous alignment

Sometimes a node corresponds to multiple tokens that are not adjacent in the sentence. It means that we need discontinuous alignments. Nothing similar occurs in the UMR 1.0 data, so we define the notation here. An alignment line may contain multiple ranges separated by a comma and a space. When this happens, the sub-ranges must be ordered by the token numbers and the first token number of a sub-range must be higher than the last token number of the previous sub-range + 1 (that is, there must be a gap containing at least one token, otherwise the sub-ranges could be merged).

s16p: 1-2, 4-4

Unaligned nodes

Abstract concepts (reifications, discourse relations etc.) that do not have a corresponding token on the surface still have an alignment line but their alignment is 0-0. Naturally, 0 cannot be combined with real token ranges, hence 0-1 would be illegal.

Certain special nodes (such as the name concept attached via a :name relation to a concept representing a named entity) are anchored to "-1" instead of 0, so their range is -1--1. See also issue #2 in the UMR annotation repository.

Unaligned tokens

??? – to investigate

Overlapping alignments?

??? – to investigate whether two nodes can map to the same token

What to align to what

These are our (ÚFAL) guidelines. They may be inspired by what we saw in UMR 1.0 but they do not attempt to mimic exactly the approach taken there.

• The easiest alignment is between a content word and the concept node that represents it: entities to nouns, states to adjectives or verbs, and processes to verbs.

• Overtly expressed discourse connectives often have their own nodes, too.

• Auxiliary verbs are aligned together with the main verb to the same event concept. The same holds for non-referential reflexive markers (smát se “to laugh”) and for verbal particles (come up).

• Some prepositions may have their own concept nodes. If they do not, then they should be aligned to the same node as their noun (they are like case markers in other languages). Note that this may lead to discontinuous alignment if there is an adjective between the preposition and the noun.

• Subordinating conjunctions are to clauses what prepositions are to nominals, so we might treat them accordingly and align them with verbs, unless they have their own node. This would be parallel to languages where subordination is marked morphologically on the verb.

• Numerical quantities do not have their own concept node because they are annotated as numerical :quant attributes, e.g. (s1h / house :quant 10). Therefore we should align s1h to the whole expression ten houses. (This is different from approximate quantities that have their own node and :quant is the relation that attaches them, e.g. (s1h / house :quant (s1s / several)) will have s1h attached to houses and s1s to several.)

• Punctuation tokens are normally not aligned with nodes. An exception would be that a node is aligned to a range of tokens, there is a punctuation symbol somewhere in the middle of the range and excluding it from the alignment would break the otherwise contiguous alignment into two sub-ranges.

• Reifications (the *-91 event concepts) are meant as abstract concepts, meaning that they typically do not have a corresponding token. However, if there is a token that is not aligned to anything else and that gave rise to the event, we should align it with the *-91 node. In particular, the copula (být “to be”) will often correspond to have-mod-91.

• The abstract concepts person, thing etc. may be aligned to overtly expressed pronouns. If the concept is only inferred from morphological agreement marked on the verb, it will stay unaligned.

• Somewhat schizophrenic situation arises with named entities. Typically there is an abstract concept (person, organization etc.) with a name child node. The abstract parent is aligned to the name tokens in the sentence. The name child stays unaligned, although it directly points to the orthographic words of the name via its :opN attributes. (This rule is inferred from the data relased in UMR 1.0.)

  • Nevertheless, there are situations when a name node is aligned to the name tokens. If the parent node has other children and they are aligned, then the parent node will not be aligned to the name tokens, hence the name node will align with them. For example, the Philippine island of Leyte is analyzed as (s4i2 / island :wiki "Leyte" :name (s4n2 / name :op1 "Leyte") :place (s4c / country :wiki "Philippines" :name (s4n3 / name :op1 "Philippine"))) where s4i2 is aligned to island, s4n2 to Leyte, s4c to Philippine, and s4n3 is unaligned.

• More generally, the approach in UMR 1.0 seems to be:

  • If the parent node covers the same tokens as its child node, the alignment will be assigned to the parent and the child will be formally unaligned. For example (english_umr-0003, snt8), doctor is represented as (s8p2 / person :ARG1-of (s8h / have-role-91 :ARG3 (s8d / doctor))); the token is aligned to s8p2, while s8h and s8d are unaligned.
  • If the parent node has multiple children that together completely cover the parent's span, the alignment will be assigned to the children and the parent will be formally unaligned. For example (english_umr-0003, snt6), next several days is represented as (s6t / temporal-quantity :quant (s6s2 / several) :unit (s6d / day) :mod (s6n / next)); here, s6s2, s6d and s6n are aligned to their respective tokens while s6t is unaligned.

• While the above rules strive to align as many non-punctuation tokens as possible, it is not required that all of them are aligned to concepts. There may be words that are not even distantly related to any individual node; such words will stay unaligned.