Consider the following question, “I want to buy a blue gown for wearing to an evening party?”. The cognitive search systems, could find and retrieve all the relevant “blue gown” dresses that are popularly voted as a suitable fashion wear for an “evening party”. The state-of-art systems could summarize content from various source and present in a much readable manner. The summarized response from Google’s search is shown in Figure 1. More so, the systems could even point to the most relevant e-commerce websites where the dress could be purchased. Figure 2 shows the results from a personalized mobile assistant, Google’s Allo, where each image is an hyperlink to it’s corresponding e-commerce website. Existing question-answering systems and conversation systems are well trained for resolving the facts from the input query and finding semantic similarity with respect to a huge corpus of background dataset.
The retrieved summarized results can be observed to provide descriptions such as “Blue Off Shoulder Mini Skater Dress”, “Vero Moda Ankle Slit Maxi Dress”. While these descriptions are factually relevant to both the dress as well as the input query, it does not persuade or compel me to purchase the corresponding product. In contradiction, a description such as “The dress choses its owner! This gleaming off shoulder mini skater would create a magical attraction in your evening party”, would make this product more catchy, memorable, and compelling to buy. The motivation for our research is derived from this existing gap - Could there be a creative system that could persuade me and sell me a product, rather than just describing the product?. We propose a system, PersuAIDE! that better persuades the consumer to sell a particular product.
This kind of persuasion is an common technique employed by sales people to enhance the impression of their products in the face of their consumers. Like in a caricature, this process involves highlighting and amplifying a few important facts that could be personalized for the consumer, while also inhibiting (or down-playing) the remaining facts that is not relevant to persuade the consumer. Such a creative feature could be incorporated into any search, question-answering, or conversation system and also could be used for targeted marketing in any e-commerce business. While the applications could be broad and far-fetching, to have a specific focus in this research, we have taken a case study in the fashion domain to persuade and sell fashionable couture and relevant accessories.
The problem statement is formulated as, given a bland product description sentence about a fashion product, transform the sentence to be able to better persuade the consumer. On a broad level, such a problem statement could be compared with a sentence paraphrasing task (Zheng et al., 2011) (Serban et al., 2016), we have focused on some specific aspects of persuasion, which cannot be directly incorporated into a sentence rephrasing setting,
Memorability (Guerini et al., 2015): Plain descriptions and bland facts do not tend to be attract the consumers. To increase the memorability of such descriptions, popular slogans, movie titles/ dialogue, poetry, catchy-phrases could be adopted to suit the product’s description.
Sentiment (Pang et al., 2008): Addition of positive sentiments in the description could influence the consumer’s inclination towards the product. One such way of adding sentiments into sentences could be to add enhancing adjectives to noun-phrases in the sentence. For example, a description such as “Blue Embrodered Net Semi-Stitched Anarkali Gown” could be enhanced as “Dazzling Blue Embrodered Net Semi-Stitched Anarkali Gracious Gown”
In the rest of the paper, Section 2 briefs the closely related literature work, Section 3 explains the proposed approach, Section 4 discusses the obtained results, Section 5 details some of the future extensions of this work, and Section 6 concludes our research.
2. Literature Study
Usage of linguistic creativity in the past have been looked into for poetry generation. (Toivanen et al., 2012)
proposed a poetry generation system that uses morphological analysis to identify substitutable words for the given topic to generate variants of poem. Recently, there is lot of focus on creative text generation based on lexical substitution.(Guerini et al., 2011) discussed about slanting an original expression to more positive or negative version by adding or subtracting a word. (Lynch, Lynch) discussed about identifying various topic in the news and choosing the right expression that captures them. (Özbal et al., 2013) proposed an extensible framework that allows users to force multiple words to appear in the expression. Given keywords, the framework comes up with an expression that includes them. More recently, (Gatti et al., 2015) proposed a system that blends expression with best matching evolving news. (Gatti et al., 2016) proposed a creative system that generates a news headlines based on best matching expression. Almost all of these work could be considered as catering to memorability aspect of persuasion alone as they rely on popular expression, quote, poem etc. and their input text are fairly expressive text like headline which are fairly easy to persuade compared to product description which are blunt statements. Also none of them discussed about transforming sentences to include complex persuasive features like argumentation, sentiment etc in the target domain.
3. PersuAIDE!: Proposed Approach
In this section, we discuss the details of the proposed persuasive system. Figure 3 shows an overview of the proposed system. Further explanation of multi-step transformation process is as follows,
Step 1: From a large corpus of text data, capture fashion domain knowledge in terms of dependency relationship between different words commonly used in fashion. In order to create this domain knowledge, we collected a large corpus of fashion related text from some of the leading fashion magazines: Chictopia111http://www.chictopia.com/ and Vogue fashion 222http://www.vogue.in/fashion/. From these blog posts and magazine articles, we extracted about text sentences. This dataset is appended with a list of popular slogans, cliches, movie dialogs, quotations (Gatti et al., 2015) to find the prominent dependency relations in fashion domain. We use the Stanford’s dependency parser (De Marneffe et al., 2014) to get the dependency relation in each sentence of the training corpus and store the frequency of each such relation. We obtained a dependency relationship matrix whose rows and columns are the words in vocabulary and the integer in each cell denotes the frequency of a relation (between row words and the column word) in the entire corpus. In our dataset, we had a vocabulary size of words and dependency relations.
Step 2: This step is to improve the memorability of the product description. Once the dependency relations are captured, we get to the process of sentence transformation. Given a sentence, we now transform it into another form by making use of quotations, slogans, etc while taking into consideration the domain dependency relations. For a given product description, we find the most similar auxiliary sentence from the quotations, slogans, movie dialogs corpus with the help of an LSTM based text similarity algorithm.
Step 3: Once the auxiliary sentence is identified, the sentence transformation can happen in two ways. Either by substituting few words in the product description sentence using the words in the auxiliary sentence or introducing new words into the auxiliary sentence from the words in the product description sentence.
Step 4: For word substitution, we capture all the nouns, verbs, adjectives and adverbs in the input and auxiliary sentences. For each of the selected word from the auxiliary sentence, we find its derivationally related forms with the help of wordnet. These derivationally related forms are bucketed into four categories (nouns, verbs, adjectives, adverbs) based on their POS tags. Now, for each of the selected word from input sentence, we choose the best substituting word from one of the created buckets. The bucket is chosen based on the POS tag of the input word. In the chosen bucket, we find the best word by making use of the following procedure.
Find all the dependency relations from the input sentence which involves the chosen word ‘w’.
For each word ‘k’in the chosen bucket, we find the matching score to replace the word w. The matching score is defines as below.
where, R denotes the dependency relations in the input sentence and f(a, b) denotes the frequency of the dependency relation in the training corpus, and denotes the set of all relations containing the word . If is child in the dependency relation then, = parent in the relation and = . If is parent in the dependency relation then, = and = parent in the relation . For each selected word in the input sentence, we select which has maximum and replace the word .
Step 5: For word addition, we select word from a list of positive sense adjectives like , , , etc. For each noun in the input sentence, we check the possibility of introducing a suitable adjective. In fashion text, we observed that most of the nouns are colors, garments (outfits), jewelry, etc. Based on the type of noun, we select a suitable adjective to add.
|Fashion Description||Think pink but don’t wear it|
|Matching Quotation||Comes with a matching slip|
|Transformed text||Think pink but don’t match it|
|Transformed text + sentiment||Think gleaming pink but don’t match it|
|Fashion Description||Jewelry maybe is more expensive than clothes, but clothes are more important than jewelry|
|Matching Quotation||Accentuate the outfit with a sling bag and a bracelet|
|Transformed text||Jewelry maybe is more expensive than outfit, but clothes are more important than jewelry|
|Transformed text + sentiment||Jewelry maybe is more expensive than outfit, but stylish clothes are more important than jewelry|
While this is a work-in-progress, we show some initial results in Table 1, listing the input sentence, selected quotation and the transformed outputs (with and without sentiment). For input sentence, the best matching quotation is selected from the quotations corpus using the Siamese adoption of LSTM networks (Mueller and Thyagarajan, 2016). Next, suitable word replacements are identified by extracting all the nouns, verbs, adjectives, adverbs using Stanford CoreNLP parser from both the input sentence and the quote. For each selected word in the quotation, the derivationally related forms are found using the wordnet and bucketed using their POS tags (nouns, verbs, adjectives, adverbs). For example, pink#adjective, pink#noun, pinkify#verb, wear#verb, wear#noun, etc. Next, for each of the selected word in the input sentence, the best suitable word is selected from the bucket of derivationally related forms having same POS tag as the input word. The criterion for selecting the best matching word is (discussed in the previous section). It might so happen that multiple words in the fashion sentence may match with some in the slogan, in this case, the pair which has the maximum score is selected. Using this process, we generate the transformed sentence, shown in the third row of the table. Further, to enhance the transformation, suitable adjectives are added to the nouns. For this purpose, from a list of positive adjectives, those suitable to be used in the fashion domain were hand picked. Based on the noun in the input sentence, an appropriate adjective from the selected list is chosen and introduced into the previously transformed sentence.
5. Future Works
In the current work, we only make use of the auxiliary information like slogans, etc to enhance the sentences related to fashion. In our observation, the fashion related sentences are tougher to deal with when compared to any news article based work predominantly discussed in the literature. Enhancing the fashion text can include idioms and similies. The existence of an ontology in this space can help improve the transformations. The ontology can be consulted for a better word replacement in the sentences. We also plan to cover argumentative aspect of persuasion for lines that can be disputed, argued. Example ”amplify the appeal with a wristwatch.” could be persuaded using the argument ”How could you not amplify the appeal with a wristwatch ??”
In this paper, we presented a system called PersuAIDE which aims to generate persuasive text for input product description . We demonstrated the work in fashion domain where popular expressions were used to generate creative sentences covering memorability aspects of persuasion. Further, based on the type of sentence, sentiment features were added to slant the generated sentence. Our study can be considered as a novel attempt to blend well known expressions with non expressive product lines in a linguistically motivated framework that accounts for syntagmatic and paradigmatic aspects of language.
- De Marneffe et al. (2014) Marie-Catherine De Marneffe, Timothy Dozat, Natalia Silveira, Katri Haverinen, Filip Ginter, Joakim Nivre, and Christopher D Manning. 2014. Universal Stanford dependencies: A cross-linguistic typology.. In LREC, Vol. 14. 4585–92.
- Gatti et al. (2015) Lorenzo Gatti, Gözde Özbal, Marco Guerini, Oliviero Stock, and Carlo Strapparava. 2015. Slogans Are Not Forever: Adapting Linguistic Expressions to the News.. In IJCAI. 2452–2458.
- Gatti et al. (2016) Lorenzo Gatti, Gozde Ozbal, Marco Guerini, Oliviero Stock, and Carlo Strapparava. 2016. Heady-Lines: A Creative Generator Of Newspaper Headlines. In Companion Publication of the 21st International Conference on Intelligent User Interfaces. ACM, 79–83.
- Guerini et al. (2015) Marco Guerini, Gözde Özbal, and Carlo Strapparava. 2015. Echoes of persuasion: The effect of euphony in persuasive communication. arXiv preprint arXiv:1508.05817 (2015).
- Guerini et al. (2011) Marco Guerini, Carlo Strapparava, and Oliviero Stock. 2011. Slanting existing text with Valentino. In Proceedings of the 16th international conference on Intelligent user interfaces. ACM, 439–440.
- Lynch (Lynch) Gerard Lynch. Every Word You Set: Simulating the cognitive process of linguistic creativity with the PUNdit system. (????).
- Mueller and Thyagarajan (2016) Jonas Mueller and Aditya Thyagarajan. 2016. Siamese Recurrent Architectures for Learning Sentence Similarity.. In AAAI. 2786–2792.
- Özbal et al. (2013) Gözde Özbal, Daniele Pighin, and Carlo Strapparava. 2013. BRAINSUP: Brainstorming Support for Creative Sentence Generation.. In ACL (1). 1446–1455.
et al. (2008)
Bo Pang, Lillian Lee,
et al. 2008.
Opinion mining and sentiment analysis.Foundations and Trends® in Information Retrieval 2, 1–2 (2008), 1–135.
- Serban et al. (2016) Iulian Vlad Serban, Alberto García-Durán, Caglar Gulcehre, Sungjin Ahn, Sarath Chandar, Aaron Courville, and Yoshua Bengio. 2016. Generating factoid questions with recurrent neural networks: The 30m factoid question-answer corpus. arXiv preprint arXiv:1603.06807 (2016).
- Toivanen et al. (2012) Jukka Toivanen, Hannu Toivonen, Alessandro Valitutti, Oskar Gross, et al. 2012. Corpus-based generation of content and form in poetry. In Proceedings of the Third International Conference on Computational Creativity.
et al. (2011)
Zhicheng Zheng, Xiance
Si, Edward Y. Chang, and Xiaoyan Zhu.
K2Q: Generating Natural Language Questions from
Keywords with User Refinements. In
Proceedings of the 5th International Joint Conference on Natural Language Processing. 947–955. http://aclweb.org/anthology-new/I/I11/I11-1106.pdf