nx_absa_lib.nx

nx_absa_lib.nx -- the SCORING MATH for SemEval-2014 Task 4 subtask 1 (aspect term extraction), as a shared lib so the CLI (nx_absa_bench) and the gate (nx_absa_bench_gate) share ONE definition of P, R and F1 and one gold parser. The metric is the paper's own (Pontiki et al. 2014, eq. 1-2): P=|S and G|/|S|, R=|S and G|/|G|, F1=2PR/(P+R), which for a set overlap is exactly 2*inter/(|S|+|G|) -- computed from counts so there is no double rounding. Everything is permil (1000 = 1.00). Nothing here is a threshold: the lib reports numbers, the gate says what a correct scorer must return, and the CLI prints the number beside the published winners. THE GOLD FORMAT (ABSA-PyTorch .seg mirror of the official test set): three lines per record -- line 0 mod 3: the sentence with the aspect term replaced by the placeholder $T$ line 1 mod 3: the aspect term (single or multi word, e.g. "hard disk", "delivery times") line 2 mod 3: the polarity (1 / 0 / -1) The gold aspect-term SET G is the distinct normalised terms across the term lines (footnote 3 of the paper counts occurrences; this is the distinct-term variant, DECLARED, a related but not identical metric). THE PREDICTOR is a UNIGRAM DOCUMENT-FREQUENCY baseline over the test sentences: it composes the ONE tokenizer (rm_next_token via rm_mine_doc) and the ONE frequency ranker (rm_mine_rank) from nx_reviewmine_lib, so no second tokenizer or ranker exists. It is a weak proxy for Hu-Liu (2004) frequent-noun-phrase mining -- our miner does not yet extract multi-word noun phrases (rung rm_aspects), so a multi-word gold term is UNMATCHABLE by a unigram prediction under exact-set match. That is the honest reason this baseline sits far below the supervised systems, and the gap IS the un-landed rung. license_tier: ORIGINAL No hw writes (Rule 26). LIB.