The increasing cultural communication among countries brings more foreign works into China and in this process, translators play a vital role. At the same time, translation qualities are determined by various factors like translation theories, translators' education background and market requirement. Thus translation quality assessment is of great importance, which is especially true for literary translation. Among various assessment criteria, “faithfulness, expressiveness and elegance” put forward by Yan Fu in 1898 is regarded as an important principle. Owing to the ambiguity of its definition, controversies and queries have been raised; however, most scholars agree that “faithfulness” is a necessary and principal component in translation. The implications of “faithfulness” include faithfulness in word meanings, figure of speech, rhetorical devices, style, etc. In other words, any variation in those aspects can violate “faithfulness”. Traditional methods in faithfulness fall into two groups: analyzing typical expressions or translation strategies. The emergence of corpus brings a new method into translation studies, and stylistic variation has been found in translations with the help of corpora. But those studies usually use limited materials and improper reference corpora, so improvements are needed. Stray Birds is a collection of gnomic verses written by Tagore and was published in 1916. The first Chinese version was translated by Zheng Zhenduo in 1922, and since then several versions have come into being. The publication of Feng Tang's version in July, 2015 reheated the discussion about which translation version is the best. Various opinions are expressed on the basis of personal understandings and thus are lack of objectivity. Using FLOB and self-built English poetry corpus as English reference and also choosing LCMC and self-constructed Chinese poetry corpus as Chinese reference corpora, this paper aims to explore three questions: (1) whether the three translations follow the principle of “faithfulness” from the stylistic perspective? (2) if stylistic variation does exist, then what factors lead to the variation? (3) What is the relation between stylistic variation and readers' acceptance? The clustering analysis shows that the original text belongs to poetry and stylistic variation exists in all three translations: Zheng's translation shares similarities with romantic fictions and it is also similar to poetry; Feng and Lu's translation is similar to detective fictions. The principle of faithfulness has restrictions on the translators, while at the same time, the three translators have different understandings towards faithfulness. Reasons for stylistic variation are analyzed from three aspects: (1) Selecting thematic words with high frequencies in the original text, we find that Zheng tended to translate a word into one particular expression, while the other two translators chose several different expressions; (2) Representative linguistic features in each translation are extracted by correspondence analysis and analyzed in their contexts. The result shows that Zheng and Lu expressed the original meaning exactly on the whole and the style of word usage in their translation conforms to the original text. Moreover, Zheng and Lu did not use rhythm in their translations, just as the original text did. Feng also expressed the original meaning clearly on the whole; however, sometimes he emphasized on simplification and rhythm too much, so that he mistranslate some words and left some information untranslated; (3) Feng's vocabulary richness is the highest among all and Zheng's vocabulary richness is the closet to that of the original text. This research also indicates that the more similar the translation is to the original, the more acceptance it will get. Cluster analysis and correspondence analysis overcome subjective features of assessment and thus provide new methods in translation studies. On the basis of “faithfulness”, this study has shown their vigorousness in evaluating poetry. In future research, they can be used to measure different genres to achieve more scientific and reliable results.