Approaching textual coherence of machine translation with complex network

This study analyzes topological properties of complex networks of textual coherence, and investigates the textual coherence of machine translation by contrasting these properties in machine-translated texts with those in a human-translated text. The complex networks of textual coherence are built by drawing on the knowledge from Systemic Functional Linguistics, with Themes and Rhemes denoted as vertices and the semantic connections between them as edges. It is found that the coherence networks are small-world, assortatively mixed, scale-free with an exponential cut-off, and hub-dependent. The basic building blocks consist of fully-connected triads and fully-connected squares, with the latter playing a more significant role in the network construction. Compared with the complex network of human translation, the networks of machine translations have fewer vertices and edges, lower average degree, smaller network diameter, shorter average path length, larger cluster coefficient, bigger assortativeness coefficient and more types of motifs. Thus, we suggest that the machine-translated texts are sparsely, locally, unevenly and monotonously connected, which may account for why and how machine translation is weak in coherence. This study is the first effort ever to employ complex networks to explore textual coherence of machine translations. It may hopefully promote the cross-disciplinary interaction between linguistics, computer science and network science.