“…The multiclausal aspect of Topic Chains supports Figure 2: Analysis steps adopted in this study: (a) Grammatical subjects and objects of each main verb are identified via dependency parsing on the whole story discourse of The Little Prince (See a sentence example from Table 1, columns "S", "V", "O"); (b) Semantic role annotation: for all the subjects and objects, annotate their semantic roles as AGENT or PATIENT (See Table 1 column "V-agent" and "Vpatient"); (c) Character role annotation: assign story character roles to the entities, see character occurrences in Table A1, and Table 1 column "character"; (d) History verb retrieval for each story character: for each story character, tabulate the verbs that are its main verbs being used in the discourse (See example Table A3); (e) Relevance between history verbs and a current verb: for each current verb, calculate its relevance to the history verbs, and sum with or without their distance weight (See Table 2 and A5); (f) Salience of the correct character: for each verb, calculate how "salient" the correct character is compared to all other characters (See example Table A6); (g) Group test between pro-drop verbs vs. non-pro-drop verbs, and apply logistic regression to test predictability of character salience on dropping behavior (See group results in Table 3 and Figure 3). long-distance coreference (Sun, 2019). Taking a dynamic perspective, Pu (2019b) suggests that a topic chain "encodes a referent that is cognitively most accessible at the moment of discourse production, as enhanced by maximum discourse coherence of topic continuity and thematic coherence".…”