SPADE: Evaluation Dataset for Monolingual Phrase Alignment

SPADE: Evaluation Dataset for Monolingual Phrase Alignment Yuki Arase1⋆ and Junichi Tsujii⋆2 1Osaka University, Japan ⋆Artificial Intelligence Research Center (AIRC), AIST, Japan 2NaCTeM, School of Computer Science, University of Manchester, UK arase@ist.osaka-u.ac.jp, j-tsujii@aist.go.jp Abstract We create the SPADE (Syntactic Phrase Alignment Dataset for Evaluation) for systematic research on syntactic phrase alignment in paraphrasal sentences. This is the first dataset to shed lights on syntactic and phrasal paraphrases under linguistically motivated grammar. Existing datasets available for evaluation on phrasal paraphrase detection define the unit of phrase as simply sequence of words without syntactic structures due to difficulties caused by the non-homographic nature of phrase correspondences in sentential paraphrases. Different from these, the SPADE provides annotations of gold parse trees by a linguistic expert and gold phrase alignments identified by three annotators. Consequently, 20, 276 phrases are extracted from 201 sentential paraphrases, on which 15, 721 alignments are obtained that at least one annotator regarded as paraphrases. The SPADE is available at Linguistic Data Consortium for future research on paraphrases. In addition, two metrics are proposed to evaluate to what extent the automatic phrase alignment results agree with the ones identified by humans. These metrics allow objective comparison of performances of different methods evaluated on the SPADE. Benchmarks to show performances of humans and the state-of-the-art method are presented as a reference for future SPADE users. Keywords: phrase alignment, paraphrase detection 1. Introduction Paraphrases have been applied to various NLP applica- tions, and recently, they are recognized as a useful resource for natural language understanding, such as semantic pars- ing (Berant and Liang, 2014) and automatic question an- swering (Dong et al., 2017). While most previous studies focused on sentential para- phrase detection, e.g., (Dolan et al., 2004), finer grained paraphrases, i.e., phrasal paraphrases, are desired by the applications. In addition, syntactic structures are important in modeling sentences, e.g., their sentiments and semantic similarities (Socher et al., 2013; Tai et al., 2015). A few studies worked on phrasal paraphrase identification on sen- tential paraphrases (Yao et al., 2013); however, the units of correspondence in previous studies are defined as se- quences of words and not syntactic phrases due to difficul- ties caused by the non-homographic nature of phrase corre- spondences. To overcome these challenges, one promising approach is phrase alignment on paraphrasal sentence pairs based on their syntactic structures derived by linguistically motived grammar. A flexible mechanism to allow non- compositional phrase correspondences is also required. We have published our initial attempt on this direction (Arase and Tsujii, 2017). For systematic research on syntactic phrase alignment in paraphrases, an evaluation dataset as well as evalua- tion measures are essential. Hence, we constructed the SPADE (Syntactic Phrase Alignment Dataset for Evalua- tion) and released it through Linguistic Data Consortium1 (catalog ID: LDC2018T092). In the SPADE, 201 senten- tial paraphrases are annotated gold parse trees, on which 20, 276 phrases exist. Three annotators annotated align- 1https://www.ldc.upenn.edu/ 2https://catalog.ldc.upenn.edu/LDC2018T09 (will be effective since March 2018) ments among these phrases as shown in Figure 1, resulted in …

Content truncated to 500 words