We propose a novel two-level model, which computes similarities between word-level vectors that are biased by topic-level context representations.

Evaluations on a naturally-distributed dataset show that our model significantly outperforms prior word-level and topic-level models.

However, while DIRT computes sim(v, 21’) over vectors in the original word-level space, topic-level models compute sim(d, d’, w) by measuring similarity of vectors in a reduced-dimensionality latent space.

slots in the original word-level space while biasing the similarity measure through topic-level context models.

To address this hypothesized caveat of prior context-sensitive rule scoring methods, we propose a novel generic scheme that integrates word-level and topic-level representations.

Rather than computing a single context-insensitive rule score, we compute a distinct word-level similarity score for each topic in an LDA model.

Specifically, topics are leveraged for high-level domain disambiguation, while fine grained word-level distributional similarity is computed for each rule under each such domain.

This result more explicitly shows the advantages of integrating word-level and context-sensitive topic-level similarities for differentiating valid and invalid contexts for rule applications.

Thus, our model computes similarity over word-level (rather than topic-level) argument vectors, while biasing it according to the specific argument words in the given rule application context.

The core of our contribution is thus defining the context-sensitive word-level vector similarity measure sim(v, v’ , w), as described in the remainder of this section.

This way, rather than replacing altogether the word-level values v(w) by the topic probabilities p(t|dv, w), as done in the topic-level models, we use the latter to only bias the former while preserving fine-grained word-level representations.

Degorski (2011) uses concatenation of word-level base forms assigned by the tagger as a baseline.

According to the lemmatisation principles accompanying the NCP tagset, adjectives are lemmatised as masculine forms (gléwny), hence it is not sufficient to take word-level lemma nor the orthographic form to obtain phrase lemmatisation.

It is worth stressing that even the task of word-level lemmatisation is nontrivial for inflectional languages due to a large number of inflected forms and even larger number of syncretisms.

To show the real setting, this time we give full NCP tags and word-level lemmas assigned as a result of tagging.

The notation cas=n om means that to obtain the desired form (e. g. gléwne) you need to find an entry in a morphological dictionary that bears the same word-level lemma as the inflected form (gféwny) and a tag that results from taking the tag of the inflected form (adj : sgzinst :n:pos) and setting the value of the tagset attribute cas (grammatical case) to the value nom (nominative).

Our idea is simple: by expressing phrase lemmatisation in terms of word-level transformations we can reduce the task to tagging problem and apply well known Machine Learning techniques that have been devised for solving such problems (e. g. CRF).

The development set was enhanced with word-level transformations that were induced automatically in the following manner.

The dictionary is stored as a set of (orthographic form, word-level lemma, tag).

The task is to find a suitable transformation for the given inflected form from the original phrase, its tag and word-level lemma, but also given the desired form being part of human-assigned lemma.

Segmentation performance is measured using word-level precision (P), recall (R), and F-measure (F).

The best performance result achieved by G2 in our experiment is 81.7 in word-level F-measure, although this was obtained from search setting (c), using a heuristic p value 0.37.

It would be interesting to confirm this by studying the correlation between description length and word-level F-measure.