The most critical issue in generating and recognizing paraphrases is developing a wide-coverage paraphrase knowledge base. To attain the coverage of paraphrases that should not necessarily be represented at surface level, researchers have attempted to represent them with general transformation patterns. However, this approach does not prevent spurious paraphrases because there is no practical method to assess whether or not each instance of those patterns properly represents a pair of paraphrases. This paper argues on the measurement of the appropriateness of such automatically generated paraphrases, particularly targeting at morpho-syntactic paraphrases of predicate phrases. We first specify the criteria that a pair of expressions must satisfy to be regarded as paraphrases. On the basis of the criteria, we then examine several measures for quantifying the appropriateness of a given pair of expressions as paraphrases of each other. In addition to existing measures, a probabilistic model consisting of two distinct components is examined. The first component of the probabilistic model is a structured N-gram language model that quantifies the grammaticality of automatically generated expressions. The second component approximates the semantic equivalence and substitutability of the given pair of expressions on the basis of the distributional hypothesis. Through an empirical experiment, we found (i) the effectiveness of contextual similarity in combination with the constituent similarity of morpho-syntactic paraphrases and (ii) the versatility of the Web for representing the characteristics of predicate phrases.
An anaphoric relation can be either direct or indirect. In some cases, the antecedent being referred to lies outside of the discourse its anaphor belongs to. Therefore, an anaphora resolution model needs to consider the following two decisions in parallel: antecedent selection–selecting the antecedent itself, and anaphora type classification–classifying an anaphor into direct anaphora, indirect anaphora or exophora. However, there are non-trivial issues for taking these decisions into account in anaphora resolution models since the anaphora type classification has received little attention in the literature. In this paper, we address three non-trivial issues: (i) how the antecedent selection model should be designed, (ii) what information helps with anaphora type classification, (iii) how the antecedent selection and anaphora type classification should be carried out, taking Japanese as our target language. Our findings are: first, an antecedent selection model should be trained separately for each anaphora type using the information useful for identifying its antecedent. Second, the best candidate antecedent selected by an antecedent selection model provides contextual information useful for anaphora type classification. Finally, the antecedent selection should be carried out before anaphora type classification.