ACE lattice and its use with and without ubertagging

In the lexical parsing stage, ACE builds a lattice with lots of edges, so in the end, for the sentence Abrams heard barks, for example, we have:

lexical edge #99 le 0x7f7d67cbc40c vtx [3-4] lexeme 'period_pct' letype 'pt_-_period_le'
lexical edge #102 le 0x7f7d67cce55c vtx [1-2] lexeme 'heard_v3' letype 'v_vp_sorb-psv_le'
lexical edge #122 le 0x7f7d67b5b91c vtx [1-2] lexeme 'generic_quoted' letype 'n_-_gq_le'
lexical edge #128 le 0x7f7d67b7e640 vtx [2-3] lexeme 'generic_quoted' letype 'n_-_gq_le'
lexical edge #130 le 0x7f7d67b87144 vtx [0-1] lexeme 'abrams' letype 'n_-_pn_le'
lexical edge #132 le 0x7f7d67b8fb20 vtx [0-1] lexeme 'generic_quoted' letype 'n_-_gq_le'
lexical edge #236 le 0x7f7d67b9e060 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #237 le 0x7f7d67b9fe88 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #239 le 0x7f7d67ba5fe0 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #240 le 0x7f7d67ba7700 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
lexical edge #241 le 0x7f7d67ba9550 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
lexical edge #244 le 0x7f7d67bb0148 vtx [1-2] lexeme 'hear_v3' letype 'v_np-vp_bse_le'
lexical edge #245 le 0x7f7d67bb1fe8 vtx [1-2] lexeme 'hear_v3' letype 'v_np-vp_bse_le'
lexical edge #247 le 0x7f7d67bb5788 vtx [1-2] lexeme 'hear_v4' letype 'v_np-prd_oeq_le'
lexical edge #248 le 0x7f7d67bb7650 vtx [1-2] lexeme 'hear_v4' letype 'v_np-prd_oeq_le'
lexical edge #250 le 0x7f7d67bbda78 vtx [1-2] lexeme 'hear_v4' letype 'v_np-prd_oeq_le'
lexical edge #268 le 0x7f7d67be4590 vtx [2-3] lexeme 'bark_n1' letype 'n_-_mc_le'
lexical edge #270 le 0x7f7d67be84b0 vtx [2-3] lexeme 'bark_v1' letype 'v_-_le'
lexical edge #272 le 0x7f7d679ec2b8 vtx [0-1] lexeme 'abrams' letype 'n_-_pn_le'
lexical edge #273 le 0x7f7d679ef850 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #274 le 0x7f7d679f2de8 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #275 le 0x7f7d679f5570 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
lexical edge #277 le 0x7f7d679f9098 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
lexical edge #278 le 0x7f7d679fa5b0 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
lexical edge #284 le 0x7f7d67a08c88 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #285 le 0x7f7d67a0c6a8 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #286 le 0x7f7d67a0fa88 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #287 le 0x7f7d67a11ae0 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #288 le 0x7f7d67a15528 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #289 le 0x7f7d67a18930 vtx [1-2] lexeme 'hear_v1' letype 'v_np*_le'
lexical edge #290 le 0x7f7d67a1ad70 vtx [1-2] lexeme 'hear_v2' letype 'v_cp_le'
SENT: Abrams heard barks.

Note the multiple edges for hear_v1 (as opposed to just having an edge per lexical entry), etc.

My question is: is this lattice necessary in general (and if yes why) or is it necessary specifically for being able to do ubertagging later? @sweaglesw (if you answer, please remember we have issues with answering through email, sadly…)

My current understanding is: these edges correspond to the different possibilities of building the incoming orthography. For example, in English, hear can be the infinitive or any of the other forms of the verb.

so there will be lexical rules which correspond to the morphosyntactic feature bundles: the infinitive, I hear, you hear, they hear, we hear… – and each such option will be an edge.

Is this correct?

Right. For the case of hear_v1, you are seeing one generated by v_psp_olr, one by v_pst_olr, and one by v_pas_odlr, as is attested by the log messages explaining the generation of those edges (which would appear shortly before those “lexical edge” messages in the ACE log). I believe this is correct and expected behavior. The fact that it is potentially lattice shaped, meaning there can be entries that overlap in interesting ways, has to do with the possibility of tokenization ambiguity and also ambiguity about which MWEs apply. It is not specifically for ubertagging.

1 Like