Article Preview
TopIntroduction
This section first introduces background research in CG. Next, to explain the problem to be solved in the evaluation experiments, the section deals with background research in WS anaphora resolution. Finally, in view of the background research the proposed implementation is briefly discussed and improvements over a former implementation are outlined.
The CG theory is based upon describing “a linguistic sign” (e.g., a word) by pairing its form (or sound structure) with the corresponding meaning, an idea found in early linguistics (De Saussure et al., 2011). Nowadays the label CG is used for various approaches—about seven research directions in total—to describing form-meaning pairings not only for morphemes and words but also for “idioms and abstract phrasal patterns” (Hoffmann and Trousdale, 2013). The CG has also been used to analyze speech discourse by chunks larger than a sentence (Antonopoulou and Nikiforidou, 2011). Moreover, an existing CG approach takes into account phenomena of breaking linguistic conventions to express new meanings i.e. the “fluidity” of the natural language (Steels, 2011a). Another direction in CG research examines how the language is acquired and used form the neurolinguistic perspective and models relations among concepts, i.e. among word meanings (Feldman, 2008). The CG finds its practical applications in robotics from experiments in robot linguistic communication (Steels, 2015) to experiments in entity recognition and acting on natural language commands (Khayrallah et al., 2015) to designing a humanoid robot with language acquisition capability (Hild et al., 2012).
The WS, on which the proposed implementation is evaluated, is “a small reading comprehension test” that can be used to determine if a machine is capable of acting similarly to people’s thinking (Levesque et al., 2011). A separate Schema typically consists of a descriptive part, a question and two possible answers. The descriptive part presents a reference (pronoun anaphora, etc.) resolution challenge such as that in the example: “The city councilmen refused the demonstrators a permit because they [feared/advocated] violence.” (Winograd, 1972). Depending on which alternative in the square brackets is used in the sentence, an adult proficient in English would easily understand that “they” refers to “the city councilmen” or “the demonstrators”. Understanding this, however, can pose serious problems for a machine as not only word matching and syntactic analysis but also examining “facts about the world” is needed (Winograd, 1980). Moreover, the WS is considered “Google-proof”, i.e. if the solution is not yet posted on-line, accessing a large text corpus is unlikely to help much (Levesque et al., 2011; Bailey et al., 2015).
Along the same lines, anaphora resolution research—dealing with what an ambiguous pronoun refers to in a sentence—demonstrates that the pronoun reference phenomenon should be explained from the semantic and pragmatic viewpoints (Bach, 1994; Peterson, 1997). Moreover, corpus-based statistical methods of anaphora resolution are considered shallow in semantics (Mitkov, 2014; Richard-Bollans et al., 2018). Most of the WS anaphora resolution approaches known to the author appear to rely on finding matches in large textual data, and show only modest resolution accuracy. Several such approaches (Kruengkrai et al., 2014; Sharma et al., 2015) generate Google queries for the WS, the solution is found by comparing query results returned by Google and the text that has been used for query generation. Similar works (Budukh, 2013; Peng et al., 2015; Emami et al., 2018) along with results returned by a search engine utilize other data such as the human knowledge base. A different study (Hofford, 2014) deals with parsing entities, such as people, and entity features, such as something people can do, in the WS. Other works (Kiselev, 2017; Raghuram et al., 2017) demonstrate that CG can be more resourceful in terms of the knowledge structure and reasoning, than the mentioned approaches, for the difficult task of the WS anaphora resolution.