Expressive Opinion Generation

Beyond the Cooperative Conversation Principle

Let us begin by looking at an example of a question asked by one person and answered by another:

A : “Do you have any favorite actor or actress?”
B : “Yes, my favorite actress is Audrey Hepburn. Audrey is, just as one would expect, a charming and beautiful woman even in her private life!“

In this example, person A asked about person B’s favorite actor/actress; person B answered the question in the first sentence, and continued by adding another opinion about Audrey Hepburn along the current context, an utterance which person A may not have expected. In terms of functional conversations,

Grice described the cooperative conversation principle as consisting of four maxims (Quality, Quantity, Relevance, and Manner) that arise from the pragmatics of natural language. Grice’s Maxim of Quantity suggests that responses should contain no more information than was explicitly asked for.

Maxim of Quality : Try to make your contribution one that is true.
Maxim of Quantity : Make your contribution as informative as is required.
Maxim of Relation : Be relevant.
Maxim of Manner : Be perspicuous.

Seen from this viewpoint, person B’s additional opinion above contradicts the maxim because it resulted in too much information being given. However, in our daily conversations, more informative phrasing and response skills are usually employed in order to hold enjoyable conversations with interlocutors. Person B’s simple opinion, “Audrey is beautiful, isn’t she?,” for example, may have less information and sufficiently meet the requirement of Grice’s Maxim. This is not only due to the length of the sentence, but also because it is a common opinion in line with that of the general public. Therefore, it cannot attract an interlocutor’s interest in an effective manner. In contrast, person B’s actual second sentence above,” Audrey is, just as one would expect, a charming and beautiful woman even in her private life,” expresses a novel opinion about Audrey Hepburn with a wealth of words and from an original viewpoint. It also implicitly contains its reason for the (positive) attitude. Let us examine a few more examples:

“This is an erotic thriller movie which also expresses the elegance of the ballet“.
“Dola’s family and Princess Sheata with pure mind are really cute, charming and, innocent“.

These two opinions are about the movies “Black Swan” and “Castle in the Sky”, respectively. Adjectives are shown in boldface. Sentences with less frequently used adjectives and proper length are likely to provide a unique viewpoint that is different from that of the majority, in contrast to redundantly long sentences, which may cause harmful effects instead. From this analysis, we assume that the amount of information in each sentence can be described as a level of generality of expressions, which is mostly represented by the frequency of adjectives in documents on a certain topic, the number of adjectives in each sentence, noun relevance to the current context, and the length of each sentence. Some natural language processing techniques, namely, sentiment analysis and opinion mining, have a direct relation to opinion generation. Their motivations are mostly to analyze users’ preferences automatically extracted from a large amount of review data for marketing and service improvements. However, there are very few works that apply opinion generation to dialogue systems in terms of novelty of the sentences themselves.

On the basis of the results of these analyses, we propose automatic sentence generation mechanisms which are capable of an additional phrasing skill and automatic expressive opinion generation. The opinions are extracted from a large number of reviews on the web, and ranked in terms of contextual relevance, length of sentences, and amount of information represented by the frequency of adjectives. The additional phrasing skill is implemented as a mechanism of sentence combination of a simple preceding response and an additional opinion. Our typical scenario is as follows: When the system is asked a factoid-typed question, it first replies with a sufficient answer based on a structured database, and then it adds another expressive opinion in line with the current context, which might be informative to the user. We conduct experiments to evaluate the sentence generation mechanisms using a conversational robot platform that has fundamental abilities to follow conversation protocols.

Opinion Generation Strategies

We propose three rankings for algorithms in terms of length and novelty: Short, Standard, and Diverse. As for the length of sentences, based on our experiences, we assume a sentence consisting of from seven to ten morphemes expresses the opinion clearly, and a sentence consisting of from fifteen to twenty are possibly expressing the opinion. After sentence candidates are sorted by TF-IDF scores, the top 30% of sentences consisting of approximately seven and fifteen morphemes are extracted, respectively. In the Short and the Standard algorithms, the lists are sorted by adjective frequency. In the Diverse algorithm, the list is sorted in the inverse order by adjective frequency.

System Architecture

Below is a system architecture of the proposed system. The main components in the architecture of the system are the Utterance Analysis, the Dialogue Management, and the Sentence Generation modules. The Utterance Analysis module receives sensory information from speech recognizers. The Answer Generation module is capable of additional phrasing with the system’s own opinions.

Architecture of the Opinion Generation System

Experiments

Effects of Additional Phrasing

The first experiment was designed to evaluate additional phrasing functions. Two conditions were videotaped, and all videos contained the same two topics (“Castle in the Sky” and “Black Swan”). As shown below, scenarios of the condition 1 and 2 were lexically identical, except that the condition 2 had additional phrasings.

Condition 1 (Simple Answering): A simple question an- swering system that replies to user’s question simply; for example, when a user asks “Do you like Castle in the Sky?”, the system replies “Yes, I do.”
Condition 2 (Additional Phrasing): The system replies with combined sentences: a simple answer and a related opinion. For example, when a user asks “Do you like Castle in the Sky?”, the system replies “Yes, I do. It’s nice that Pazu helps princess Sheeta at the risk of his life throughout the whole story.”

72% of the subjects answered that they felt enjoyment with the proposed system (3% answered there is no difference), 63% answered that they felt politeness with the proposed system (3% answered there is no difference), and 59% answered that they liked the robot’s personality with the proposed system. Most of subjects felt enjoyment about the additional mechanism. However, in the free-form questionnaires, some subjects reported that they persistently felt the robot’s personality. The result implies that the additional function should be switched over based on conversational contexts.

Effects of Opinion Generation Strategy

This experiment was conducted to evaluate effectiveness of sentence generation algorithms using recorded video watching. We used the following three evaluation metrics:

Users’ impressions of enjoyment
Users’ motivations for participation
Users’ impressions of the robot’s personality.

In terms of users’ impressions of the enjoyment metrics, users were asked if they felt enjoyment from a conversational sequence of a video they watched. Users’ motivations for participation examined how users felt to participate in the conversation. Users’ impressions of the robot’s personality examined how users had positive impressions of the robot’s personality.

Comparative results of ranking algorithms.

Most of the people chose anything besides the Short. For the question of enjoyment, the same number of people chose Standard and Diverse. For the question of motivation and personality, the number of people who chose Diverse was twice as many as those who chose Standard.