Toward love and sex narrative generation using a noun conceptual dictionary

2.1.System structure

First, the story is input into the system. A story consists of surface textual representation and a conceptual structure. The conceptual structure has a so-called “structure tree” that represents the story content and a list of instances. The structure tree is a hierarchical structure of events in the story. An instance is a concrete object that exists in the real world. In this work, instances include characters, objects, and locations that appear in a story. Additionally, the attributes given to an instance by colouring are also considered instances themselves, assuming that they are unique information in a story. Section 2.3 provides the structure tree and the details of the list of instances.

The output of the system is a coloured story. We believe that the system can give the story an atmosphere of love and sex by sprinkling in noun concepts related to the two concepts. This assumption is based on the saying, “If you want to give a story a damp, watery feel, try placing water-related details everywhere,” initially stated by Shimada (2009).

Here, the system colours the story with love and sex by transforming the initial story. This transformation of the story replaces the values of the attribute frames of the instances in the story; this adds explanatory events to the story’s conceptual structure and changes the sentence expressions accordingly. An attribute frame is a structure that indicates the characteristics and state of an instance in a story.

The prototype of the system consists of a colouring module, a sentence generation module, a conceptual dictionary, an attribute information knowledge base, and a notation dictionary. Figure 1 presents the relationship of each mechanism.

Fig. 1.

Architecture of the proposed system.

2.2.Noun conceptual dictionary

While this section only describes the noun conceptual dictionary, Section 3.2 describes the verb conceptual dictionary only for the sentence patterns used in this study. A noun concept describes information about the meaning of a word, and the dictionary systematically stores about 12,000 noun concepts. The systematic structure is based on 5,809 categories (called intermediate concepts) and 13 hierarchies. The categories and noun concepts in the noun conceptual dictionary are all in Japanese. The data used in the prototype system are also all in Japanese. For reference in this paper, we have manually added English translations to those Japanese words (the text in “( )” is the added English translation).

Figure 2 shows the systematic structure of the noun concept dictionary. Since it is difficult to describe the entire structure here, we have included all the categories up to the fourth level and some of the categories from the fifth level onward. The black boxes indicate noun concepts that are directly associated with instances. For example, the category “愛 (love)” has the noun concepts “愛 (love)” and “純愛 (pure love).” Within “愛 (love)”, the subcategories “慈愛 (affection)” and “友愛 (friendship)” exist. Some categories have only subcategories.

As this study examines noun concepts related to love and sex, we limited the scope of the noun concept dictionary that we used. Based on this study’s definitions of sex and love, as described in Section 1, we used the following conditions to determine our range of noun concepts. We collected concepts related to love and sex from the structure of the conceptual dictionary shown in Fig. 2. Figure 3 shows the structure of the dictionary associated with the collected noun concepts. The bold boxes indicate the categories to which the collected noun concepts belong.

Fig. 2.

Details of the noun conceptual dictionary (partially).

Limiting the scope using the above method allowed us to include 396 noun concepts, which were candidates for the noun concepts selected by the story colouring module. Table 1 is correspondence table between Japanese and English in Fig. 3.

Fig. 3.

Categories with collected noun concepts.

We rearranged the collected noun concepts and gave them a new classification. For “emotion,” “body parts,” and “genital fluid,” the original, or superordinate category, classification of the noun concept was used. However, “genital fluid” includes noun concepts such as “卵子 (ova).” “Tendency/temperament” has the noun concept of “好色家 (lecher).” Reproductive and excretory acts are grouped under “action,” while noun concepts that are not active acts are grouped under “phenomenon.” Noun concepts related to body parts are grouped into “body parts.” “Sexual concepts” include “性行為 (sex act),” “欲[色欲] (desire [lust]),” and “生殖 (procreation),” which summarize the abstract noun concepts contained. Each category includes the following noun concepts.

2.3.Input and output

Figure 4 shows that the input and output stories consist of sentences, structure trees, and instance lists. The sentence is the story that the user sees directly, and is a string representation of the content. In contrast, structure trees and instance lists are elements that represent the conceptual structure of a story. A structure tree is a tree structure representation of the events in a story, and an instance list is a collection of information on people, objects, and places among the elements that make up the events.

Figure 5 shows the structure tree of the story. The example is: “My cousin went out to play in the flower garden. An idiot jumped out and put a knife on the windowsill. A painter named Ivan left the country.” This structure tree has a terminal node. The terminal nodes are events, and the intermediate nodes represent relations between events (for more details, see (Ogata, 2020)). An event comprises a verbal concept and an “identification (ID)” that indicates an instance. The ID corresponds to the instance stored on the instance list. The ID stores a unique number that is assigned to each event in a story structure, to clearly distinguish one event from another. The “$” sign indicates the relationship between events, “time” indicates the time when an event occurs, and “event” represents the change of state that occurs between one unit of time and the next. For this reason, time has one pair of values. The “type” indicates whether the event is an action or a description.

Fig. 4.

Structure of a story.

Figure 6 displays the instance list, that represents characters (agent), objects, and locations included in a story. Each instance has an attribute frame, that indicates the properties, characteristics, and state of an instance. Instances are managed by unique IDs, which are prefixed with “age%” for characters, “obj%” for objects, and “loc%” for locations, and a combination of concept name, “#”, and numbers to distinguish each instance. Figure 6 shows an example that contains one instance for each character, object, and location.

Fig. 5.

Conceptual structure of a story.

An attribute frame with a unique ID for each instance. In addition to the slot for the ID, an attribute frame has one or more slots, which can either be empty or contain some value. The type of slot determines the value that can be inserted. “ID” is a slot to identify the instance, and a unique value is inserted for each instance. The “instance-of” is a slot to insert a noun concept, which indicates the classification of the instance. “location” indicates the location where an instance exists. In the “name” slot, the name of the instance is inserted as a string. “health” is a slot where the health or durability of an instance is expressed as a value from zero to 100, where zero means dead or destroyed and 100 means complete. For “consciousness,” if the value “awaken” is inserted, the instance is in the awakened state. For other slots, the noun concepts listed in Table 2 are inserted. For each slot, a noun concept in the category that matches the name of the slot is inserted. However, for “possession,” the noun concept in the classification “genital fluid” will be inserted. For example, in the “emotion” slot, a noun concept belonging to the category “emotion” will be inserted. The slot then becomes “((ID nil) ... ((emotion 愛)).”

Fig. 6.

Format of an instance list.

2.4.Algorithm in proposed system

The system assigns noun concepts related to love and sex to the person instances in the input story. The system targets the person instances in the story and rewrites the values of the attribute frames of the targets. The system selects new love- and sex-related concepts in the noun concept dictionary using co-occurrence information for all concepts in the input story. The system also inserts the selected concepts into the slots defined for each concept. The output results are converted into natural sentences. In this study, the target instances are limited to character instances.

The above algorithm consists of two steps, and the first step has three processes.

3.1.Colouring a story

This section examines the central mechanism of the system. As mentioned in Section 2.1, this paper assumes that colouration can be achieved by interspersing the input stories with noun concepts related to love and sex. Therefore, the method of selecting noun concepts related to love and sex is an important part of the colouring process. The story colouring mechanism gives noun concepts related to love and sex to instances that are components of the input story. For this purpose, it performs the processes described in Section 3.1.1 and Section 3.1.2, for all the human instances in the instance list. Then, based on the results, we apply the process in Section 3.1.3 to the structure tree.

3.1.1.Selecting noun concepts and attribute information

This process selects the instance list and the structure tree from the input story and processes them separately. They are combined via instance IDs. As the two are distinguished by a clear symbol, different processing functions are applied to each based on that symbol.

The first step is to select a noun concept for each person instance in the instance list. This is the attribute frame of the instance. There are two possible ways to select noun concepts: one is to select noun concepts completely at random and the other is to select those concepts that have some relation to the instance being coloured. The new noun concept is selected from the range related to love and sex in the noun concept dictionary. For this, the system employs the following two methods: (1) random selection; (2) co-occurrence. Another possible method is to use a hierarchical structure of noun concepts, but since the data in this experimental system does not create a hierarchical structure, this method should be examined in a future study. The new concept is then inserted into a specific slot in the attribute frame of the target. The attribute information of the noun concept, prepared based on the Web dictionary, is then given to the new noun concept. In this experimental system, option (2), co-occurrence, was adopted.

(1) Association of noun concepts based on random numbers

Method (1) is to select one from all the noun concepts belonging to a certain category based on a random number. The classification is chosen to correspond to the slot. It can select any noun concept; however, it has no relation to the target noun concept, except in the case of an accidental selection.

(2) Association of noun concepts based on co-occurrence information

Method (2) uses co-occurrence information with the target noun concept to select new romantic and sex-related concepts for the noun concept dictionary. If there are no co-occurrence candidates, we search for co-occurrence information using other noun concepts as mediators. The co-occurrence information allows us to use concepts that appear in the same text simultaneously. This ensures consistency within a background.

For example, suppose we want to color the instance “age%cousin#1” whose slot “instance-of” is “従兄@男 (cousin@man).” The module first collects noun concepts related to love and sex that have co-occurrence information with 従兄@男 (cousin@man). This co-occurrence information is provided by the noun conceptual dictionary (Ogata, 2015). In the example of 従兄@男 (cousin@man), noun concepts such as “自慰 (masturbation)”, and “オナニー (onanism)” are collected. The module selects the noun concept with the highest co-occurrence information value. If there is more than one noun concept with the highest co-occurrence value, the module selects the first hit noun concept when collecting them.

In addition, we have defined separate constraints on the selection of noun concepts. For example, some noun concepts do not match certain types of instances, such as “male” becoming “pregnant.” Of course, such a combination is interesting as a story in some cases; however, it is necessary to control such combinations. Therefore, for each target noun concept, we have prepared combinations that cannot be selected, called a constraint condition.

• Gender restriction: concepts included in the intermediate concepts of “male” and “female”

  • – Noun concepts with a gender restriction are selectable (or not) depending on whether the target noun concept belongs to the “male” or “female” category.

• Age restrictions: intermediate concepts “infant,” “boy/girl,” “young,” “adult,” and “old.”

  • – Age-restricted noun concepts are selectable (or not), depending on whether the target noun concept belongs to “infant,” “boy/girl,” “youth,” “adult,” or “old person.”

• Biological restriction: intermediate concepts “animal” and “plant”

  • – A noun concept with a biological restriction is selectable (or not), depending on whether the target noun concept belongs to the “animal” or “plant” category.

For the selected noun concept, we provide the attribute information of the noun concept from a knowledge base prepared based on a Web dictionary. We pair the selected attribute information with the target noun concept, and input to the instance generation mechanism. We then crease the attribute information using the attribute information acquisition mechanism developed by Ono & Ogata (2017). Figure 7 shows an example of attribute information, in which the corresponding noun concept and the content of the attribute information are described in pairs.

Fig. 7.

Example of content of attribute information.

3.1.2.Generating instances based on selected noun concepts and selected attribute information

The instance generation mechanism generates an instance based on the selected noun concept and the selected attribute information. The generated instance will have the selected noun concept in the “instance-of” slot and the attribute information in the “attribute” slot, which contains the attribute information associated with the noun concept, as shown in Fig. 7 in Section 3.1.1. Figure 8 shows the attribute information of the “action” classification. Instances generated from noun concepts in the classification “emotion” are special and have a “target” slot, in which the noun concept appears. In the “target” slot, the ID of the character instance is inserted. The module randomly selects an ID of a character instance in the instance list. Figure 9 shows an example of a person instance in which a new instance is inserted. As shown in bold, the selected noun concept goes through the process described in this section, where an instance based on the selected noun concept is created, and information describing the selected noun concept is associated with it. This information is used in the sentence generation described in Section 3.2.

Fig. 8.

Example of an instance with attribute information.

Fig. 9.

Attribute frame of an instance.

3.1.3.Inserting explanation event into structure tree based on generated instances

The story colouring mechanism converts the generated new instances into explanatory events and inserts them into the structure tree of events in the conceptual structure of the story. The explanatory event is inserted immediately after the first appearance of each instance in the story’s conceptual structure, and it describes the contents of all slots of each instance. A pair of the classification of the instance inserted in the slot -> attribute information of its noun concept is inserted into the conceptual structure of the story, as many slots as it has values. At the point where the explanatory event is inserted, the structure is “<event> <value of the first slot> <attribute information of the value of the first slot> … <value of the nth slot> <attribute information of the value of the nth slot>.”

Figure 10 shows an example where an explanatory event is inserted for a structure tree: “A cousin goes out to play in the flower garden.” Fig. 10 shows the insertion of an explanatory event into the structure tree for “A cousin goes out to play in the flower garden.,” the value of the “tendency/property” slot in the attribute frame of age%従兄#1 (age%cousin#1) and the “attribute” slot of the instance associated with that value. This explanatory event inserts into the story’s conceptual structure, as many pairs of instance classification → attribute information of its noun concept inserted in the slots, as there are slots with values. This explanatory event is inserted into all explainable slots immediately after the first occurrence of an instance with that attribute frame. It does not generate explanatory events for instances that have already generated explanatory events.

Fig. 10.

Structure tree with explanation.

3.2.Editing structure tree using sentence generation module

The sentence generation mechanism inputs the conceptual structure (instance list and structure tree) of the transformed story. The sentence format of the generated sentence is determined by the target structure. The types of sentence formats correspond to (1) events, (2) slots in attribute frames, and (3) attribute information.

(1) Events generate sentences according to the sentence pattern defined for each verb concept. For example, in the event “event go-to (type action) (ID 1) (time (time1 time2)) (agent age%cousin#1) (location loc%flower-garden#1) (object obj%play#1)),” “出かける1(go-to)” is a verb concept, and the sentence pattern is “<agent> go to <object> in a (the) <location>”. “< >” means the corresponding case. The generated sentence is, “A cousin goes out to play in the flower garden.”

(2) For the attribute slots, a sentence pattern was determined for each slot. Table 3 summarizes the correspondence. In the table, the <agent> inserts the “instance-of” of the instance, and <object> inserts the value of the slot. In the “emotion” slot, the “instance-of” of the target instance is inserted in <counter-agent>. The name slot, if it has a value, is inserted in the first event in the form “the <agent> named <name’s slot value> did ∼,” and the name is used as the <agent> thereafter.

(3) For attribute information, all elements are used for sentence generation without omitting the hierarchical structure. The “attribute” here refers to the “attribute” slot described in Fig. 8. In other words, all the contents in the “attribute” slot are added to the sentence expression of the story.

For example, the structure in bold in Fig. 10 generates the following sentence.

“A cousin goes out to play in the flower garden. The cousin masturbates. Masturbation is also called masturbation. It is the act of stimulating one’s own genitals with one’s hands or an instrument shaped like genitals to achieve sexual sensation. The act of masturbation itself is generally harmless, but it can lead to sexual neurosis if there is an overwhelming sense of guilt.”