📄 Automatic Acquisition of English Topic Signatures Based on a Second Language

Summary

The paper addresses the knowledge acquisition bottleneck in word sense disambiguation (WSD): hand-built sense-tagged data is scarce. Wang proposes building English topic signatures, which are sets of terms that typically co-occur with a given word sense, using Chinese monolingual text instead of parallel corpora. Each sense of an ambiguous English word is mapped through an English-Chinese dictionary (Yahoo! Student) to a distinct Chinese word. Chinese text containing that word is retrieved from the Chinese Gigaword Corpus and People's Daily On-line, then segmented with ICTCLAS. The text is mapped back to English word by word with the LDC Chinese-English Translation Lexicon, so each context becomes a bag of English concepts labelled with a sense. The approach relies on English and Chinese being distant languages, which makes it likely that different senses receive different translations. The signatures were used to train an adaptation of Schütze's context-group discrimination, a second-order co-occurrence model. On six nouns from the TWA Sense Tagged Data Set, the method beat a most-frequent-sense baseline for five words, with no manual annotation at all.

Key points

• Pipeline: English-Chinese lexicon, then Chinese corpus or Web retrieval, then ICTCLAS segmentation and POS tagging, then LDC Chinese-English Translation Lexicon 3.0, then stop-word filtering.

• Data sources:

  • Mandarin Chinese Gigaword Corpus (CGC): 1.3GB of Xinhua newswire.
  • People's Daily On-line: added for bass, crane and palm, whose sense translations are rare in CGC.
  • Training data was capped at 6,000 instances per sense.

• WSD model:

  • The clustering step was dropped because the data arrives already sorted by sense.
  • The vector space is a "Concept Space" built from the 2,500 most frequent concepts, where a concept is the set of English translations of one Chinese word.
  • Each test instance is assigned to the sense vector with the highest cosine score.

• Results on TWA (baseline → system):

  Word	Baseline	System
  bass	90.7%	93.5%
  crane	74.7%	76.6%
  palm	71.1%	76.1%
  plant	54.3%	70.2%
  tank	62.7%	70.1%
  motion	70.1%	69.7%

  Motion was the only miss, blamed on its ambiguous Chinese translations.

• Main error sources:

  • an outdated lexicon, missing words such as "mobile phone";
  • segmentation mistakes that distort word distributions;
  • Chinese translations that are themselves polysemous.

Interesting ideas

• I1. A distant second language can act as a free sense annotator: if two senses get different translations, monolingual text in the other language supplies labelled contexts without any parallel corpus.
• I2. The sense inventory is defined only by distinctions that a bilingual dictionary actually lexicalises, which gives coarse, practically useful senses instead of WordNet's very fine granularity.
• I3. Each feature is a concept, meaning a set of synonyms from one foreign word, rather than a surface word, and a word maps to every concept that contains it.
• I4. Ambiguous pivot words can be repaired by expanding the query with monosemous synonyms, a fix the paper identifies after the motion failure.
• I5. Evaluation compares the unsupervised system against a "supervised" most-frequent-sense baseline taken from the test data, which is a stricter bar than random choice.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Silver sense or frame labels could be harvested this way and then cross-checked with our witness parsers before training.
• I2 (world): When compiling WordNet-based ontology entries, a "cross-lingually distinct" flag could mark which sense splits are worth keeping as separate world concepts.
• I3 (frames): Context-based sense choice for verbs could use synonym-set features, so that the VerbNet class selection rests on concepts rather than a raw bag of words.
• I4 (learn): The motion case works as a built-in negative control, so a repair rule could flag and expand ambiguous pivots before any silver data is accepted.
• I5 (answer): Reader components could report their scores against a majority-answer baseline, so that apparent gains are not just frequency effects.