| As the pinnacle of the art of Ci studies,Song Ci Poetry has unique artistic charm and ideological value and has always been an important research center of literature and linguistics.In recent years,with the development of artificial intelligence technology,the automatic analysis and generation of poetry have gradually developed into a new research hotspot in computational linguistics,and a number of databases and software systems such as poetry corpus,knowledge graphs,and automatic writing have been formed.At present,the research of the construction and quantification of ancient poetry annotation corpus,which is the basis of poetry computation,has made some achievements,forming some raw corpora such as the poems of Tang and Song dynasties in the level of one million words and Quan Song Ci(Complete Song poems),as well as some labeled corpora and knowledge bases such as some small-scale Tang poetry dependency treebanks,the knowledge ontology of Three Hundred Tang Poems,Quan Song Ci(Complete Song poems)corpus.Using these resources,a large number of string-based arrangements and analyses have been carried out,some of which have carried out the analysis of word segmentation,part of speech,prosody and syntax.However,there is still a lot of room for the annotation,analysis level and depth of the corpus,especially the image analysis of poetry based on the annotation data to improve.This paper mainly summarizes these problems into the following three points:Firstly,some measurement studies did not carry out word segmentation and part-of-speech tagging,which resulted in insufficient accuracy of statistics.For example,the twocharacter word“青天”(blue sky)was regarded as two words“青”(blue)and“天”(sky)for statistics;Secondly,there is less involvement in semantic labeling,which lacked semantic information,causing errors in semantic classification,such as classifying“三月”(March)and“明月”(bright moon)into one category of statistics.Thirdly,The labeling level stays at the word level,and less attention is paid to the distinctive image and emotional level in poetry,which leads to limited access to image and emotional information,and it is impossible to dig out deep-level poetry information.In response to the above three issues,this paper is based on Notes of Three Hundred Song Poems compiled by Tang Guizhang,and constructs a multi-level annotated corpus containing 282 Song Ci Poetry and more than 20,000 characters.Firstly,in view of the limitation of the labeling level,this paper breaks the previous single labeling level,tries to use words as the unit,and then designs a “word(word segmentation,part of speech)-image-emotion” multi-level labeling system.Secondly,this article has carried out a deep exhaustive annotation of words,including word segmentation,part-of-speech tagging and semantic tagging,and then we can see the semantic content of Song Ci Poetry.Finally,a more effective labeling system was designed for the image and emotional level,and the image was secondly labeled on the basis of word labeling,marking the internal part of the speech structure of the image,and clearly showing the relationship between the word and the image.This paper labelled the image semantic content such as the literal meaning,non-literal meaning and the mapping relationship between the two;Song Ci Poetry is also marked with a textual sentiment.Therefore,it is useful to carry out a thorough analysis of the imagery of Song Ci Poetry,and statistics such as metaphor patterns and emotional distribution.Based on the labelled information of the Three Hundred Song Poems annotated corpus,this paper conducts a multi-angle quantitative research and draws the following conclusions:(1)The use of characters,words,and images in Song Ci Poetry all present a longtailed distribution,which conforms to Zipf’s law.In the words of Song Ci Poetry,monosyllable words occupy the dominant position,among which the appearance frequency of“人”(people)is the highest;the proportion of content words is much higher than that of function words,but compared with Tang Poetry,function words are used more frequently in Song Ci Poetry.Among the images,static images and generic images occupy the dominant position,and the“东风”(east wind)image appears most frequently.Song Ci Poetry’s textual emotion is mainly negative,presents a pessimistic and negative emotional tone as a whole.(2)Based on the semantic statistics of words and images,the usage rules of Song Ci Poetry on the semantic level are obtained.In the semantic distribution of words,poets have different preferences for seasonal words and time-of-day words.The frequency order of using seasonal words is spring>autumn;the frequency order of using time-of-day words: year>day>month;what’s more,the time-of-day words mainly refer to evening or night semantics.In the distribution of the sense categories of imagery,the frequency order of the celestial sense categories is moon>sun>star;the frequency order of the earth sense categories is land>water area>airspace;the frequency order of the liquid sense categories is rain>snow>frost>dew;the frequency order of the gas sense categories is east wind> west wind,cloud> fog.(3)Through annotation and statistics,it is found that the “Three Hundred Song Poems” has 24,058 character tokens and 2,080 character types;there are 22,221 word tokens and 2,745 word types.At the literal level,these words can be reduced to 916 different semantic categories.There are 2,002 image tokens and 1,228 image types.At the literal level,these images can be reduced to 275 different sense categories,and at the non-literal level,these images can be reduced to 160 different sense categories.This further shows that the image semantics has multiple layers,and the literal meaning and non-literal meaning are not in a one-to-one correspondence.The semantic mapping of the two is mainly emotional metaphor,accounting for 30.47%.These data fully verify the necessity and feasibility of annotation in this paper and reveal the specific distribution of words,images and metaphorical patterns in Three Hundred Song Poems,which provides valuable basic resources for computer analysis and automatic generation of Song Ci Poetry. |
