Coordination Boundary Identification without Labeled Data for Compound Terms Disambiguation

Proceedings of the 28th International Conference on Computational Linguistics, pages 3043–3049 Barcelona, Spain (Online), December 8-13, 2020 3043 Coordination Boundary Identification without Labeled Data for Compound Terms Disambiguation Yuya Sawada ♠ Takashi Wada♣ Takayoshi Shibahara♠ Hiroki Teranishi♠ Shuhei Kondo♠ Hiroyuki Shindo♠ Taro Watanabe♠ Yuji Matsumoto♦ ♠ Nara Institute of Science and Technology ♣ School of Computing and Information Systems, The University of Melbourne ♦ RIKEN Center for Advanced Intelligence Project (AIP) {yuya.sawada.sr7,shibahara.takayoshi.sk4,teranishi.hiroki.sw5, shuhei-k,shindo,taro}@is.naist.jp twada@student.unimelb.edu.au yuji.matsumoto@riken.jp Abstract We propose a simple method for nominal coordination boundary identification. As the main strength of our method, it can identify the coordination boundaries without training on labeled data, and can be applied even if coordination structure annotations are not available. Our system employs pre-trained word embeddings to measure the similarities of words and detects the span of coordination, assuming that conjuncts share syntactic and semantic similarities. We demonstrate that our method yields good results in identifying coordinated noun phrases in the GENIA corpus and is comparable to a recent supervised method for the case when the coordinator conjoins simple noun phrases. 1 Introduction In the scientific literature, coordination is a common syntactic structure and is frequently used to describe technical terminologies. These coordinate structures often involve ellipsis, a linguistic phenomenon in which certain redundant words inferable from the context are omitted. For instance, the phrase “prostate cancer and breast cancer cells” conjoins two cell names, “prostate cancer cell” and “breast cancer cell,” with the token “cell” eliminated from the first conjunct. This phenomenon raises significant challenges in named entity recognition (NER) tasks, and most of the current NER models (Ma and Hovy, 2016) can identify only non-elliptical conjuncts, e.g., “breast cancer cells,” or incorrectly extract the whole coordinate phrases as single complex entities. Therefore, identifying coordinated noun phrases is crucial to improving the model performance in NLP, particularly in NER and relation extraction tasks within in scientific domains. In this paper, we propose a simple yet effective method for finding coordination with related compound nouns, such as technical terms. Compared to previous methods (Ficler and Goldberg, 2016b; Teranishi et al., 2017; Teranishi et al., 2019), our approach does not require any training on labeled data, and is applicable under the realistic conditions where annotations of coordinate structures are not readily avail- able. Our method employs recent pre-training language models such as ELMo (Peters et al., 2018) and BERT (Devlin et al., 2019) to measure the similarities of words, and identifies coordination boundaries based on the property in which conjuncts share syntactic and semantic similarities. This property has been exploited in traditional alignment-based methods (Kurohashi and Nagao, 1994; Shimbo and Hara, 2007; Hara et al., 2009), and our system extends and simplifies such methods by using neural embed- ding representations instead of the handcrafted features or heuristic rules used in their approaches. Our experiments show that, even without training, our method achieves good results in identifying nominal coordination boundaries in the GENIA corpus (Tateisi et al., 2005). When targeting only the coordina- tion of noun …

Content truncated to 500 words