Attribute Extraction

Attributes are properties that define a product. For example, in Figure 11-1, we saw brand, resolution, TV size etc as relevant attributes. An accurate display of these attributes will provide a complete overview of the product on the e-commerce website so that the customer can make a informed choice. A rich set of attributes directly relates to the improvement of clicks and click-through rates which influence the product’s sale. [Link to Come] shows an example of a product description obtained by a set of filters or attributes.

Figure 3-2. Product obtained by a set filters or attributes

As you can see, the attributes like {clothing, color, size} are basically what defines this product to a customer. Each of these attributes can have multiple values as seen in the figure. In this example, color takes seven values. However, directly obtaining attributes from the sellers for all products is difficult. Moreover, the quality of the attributes obtained should be consistent enough that it allows a customer to have the correct and relevant information about a product.

Traditionally, e-commerce websites employed crowdsourcing techniques to obtain the attributes. Crowdsourcing is typically done by third-party companies or crowdsourcing platforms where the specific question about each product is asked and the crowd workers are expected to answer them. Sometimes the questions are framed in a multiple-choice manner to restrict the answer into a set of values. But generally, that is expensive and not scalable with the increase in the volume of products. That is where techniques from machine learning step in. This by a challenging task because it requires the understanding of the context of the information present in the product. For example, look at the following two product descriptions in Figure 11-4.

Figure 3-3. Cases where ‘pink’ is the attribute value for two different attributes

Pink is a popular brand with younger women. Similarly, pink is a very common color of apparels. Hence, in the first case, Pink is a brand attribute whereas the other case, pink is just a color. In the above figures, we see the backpack is from the brand ‘pink’ with a color of neon red. Whereas the sweatshirt is of the color pink. Cases like these and many more are highly prevalent and poses a challenging task for a computer to solve.

There are some structured ways by which one can solve this problem to an acceptable performance. If a dictionary of attributes can be prepared or each product is listed with proper attribute values then the search mechanism can access them to populate the results accurately according to customer needs. The set of algorithms which extract the attributes from various product information are generally called attribute extraction algorithms. These algorithms take a collection of textual data as input and produce the output as a {attribute, value} pair. There are two types of attribute extraction algorithms - direct and derived [4]. Typically, these algorithm takes text as input (e.g., product title or product description).

Direct attribute extraction algorithms assume the presence of the attribute value in the input text. For example, ‘Sony XBR49X900E 49-Inch 4K Ultra HD Smart LED TV (2017 Model)’ - this text contains the brand ‘Sony’ explicitly mentioned here. A brand is typically an attribute which is expected to be present in the product title for most of the cases. On the other hand, Derived Attribute Extraction algorithms do not assume that the extracted attribute is present in the input text. They derive that information from the context. For example, gender is such an attribute which is usually not present in the product title, but from the input text, the algorithm can identify is the product is specifically for men or women. See here in this example - ‘YunJey Short Sleeve Round Neck Triple Color Block Stripe T-Shirt Casual Blouse’. The product is for women but the gender ‘women’ is not explicitly mentioned in the product description or title. In this case, the gender has to be inferred from the text, for instance, from the product description.

Typically, the direct attribute extraction algorithms are modeled as a sequence to sequence labeling problem. Recall that in chapter 5, we have discussed the supervised sequence models for Named Entity Recognition tasks. A sequence labelling model takes a sequence as input and outputs another sequence of the same length. In chapter 5, we have also discussed how to create input and output datasets with various encoding schemes which will be applicable here as well.

Let us say we have a product title ‘The Green Pet Shop Self Cooling Dog Pad’ , where ‘The Green Pet Shop’ is the brand. In a BIO encoding scheme, this will look as shown in Figure 11-5 below:

Figure 3-4. BIO encoding scheme for direct attribute extraction

We know, now, various tokenization techniques from chapter 2 and we can tokenize the input text to tokens (words here). As we like to preserve the sequential information of the input text, we continue with this sequence of tokens to make the input sequence. Getting such labelled data is crucial for any attribute extraction process. Especially in e-commerce, for supervised attribute extraction algorithms, you should have data that represents various categories with significant variations in the data content. Sampling strategies for collecting annotated titles should be stratified. The strata can be selected based on categories..

For such labeled data, a rich set of features [4, 46] need to be extracted to train a machine learning model. Ideally, the input features should capture the characteristics, locational, and contextual information. Here is a list of some of the features which can capture all these three aspects. One can develop more complex features in the similar lines and perform analysis to understand are they significant in improving the performances.

Characteristic features:

These are typically token based features:

the character composition of the token

letter case

token length in terms of number of characters

Locational features:

These features capture the positional aspect of the token in the input sequence:

number of tokens in the input sequence before the given token

ratio of the token position and the total length of the sequence

Contextual features:

These features mostly encode information about the neighbouring tokens:

the identity of the preceding/succeeding token

whether the preceding/succeeding token is capitalized

the bigram consisting of the token and its predecessor/successor

whether the preceding token is a conjunction

part of speech tag of the token

Once the features are generated and output tags are encoded properly, we get the sequence pairs for training the model. We can directly use some of the models which we have discussed in chapter 5 such as Conditional Random Field, Hidden Markov Model, etc. [4]. Even though the pipeline looks simple and similar to Named Entity Recognition (NER) systems, there are challenges with these feature generation schemes and modeling techniques. For attributes like ingredients, simple features do not capture the context around them.

To deal with such data sparsity and other feature incompleteness issues, some approaches suggest the use of sequence of word embeddings in the input This can be tackled in advanced methods like deep recurrent structure-based sequence to sequence models. The input sequence will be passed to the model as is, and it is supposed to predict the output sequence. Recent efforts include deep recurrent structures like RNN or LSTM to perform the sequence to sequence labeling task [47, 48]. One of the state-of-the technique includes the distributed word representations as we have discussed in chapter 2. As you can see from the architecture given in the figure, it uses both word-level representation, as well as character level representation [44]. Using character level representation helps here to cover unknown words which are out-of-vocabulary.

Figure 3-5. LSTM architecture for attribute extraction as a seq2seq task

Bidirectional LSTMs can be used to encode the sequential information present in the input sequence. LSTMs are known to capture contextual information present in the input. The need for bidirectional recurrent structures is to encode the sequential information from both directions of the input sequence. Here, you can see how this deep learning model works better than the typical machine learning models due to their non-expressiveness of the handcrafted features.

For indirect attribute classification, generally, classification frameworks are used. Recall the example of ‘YunJae Short Sleeve Round Neck Triple Color Block Stripe T-Shirt Casual Blouse’. For this case, we embed the whole input string using any of the sentence representation method (averaging out the individual word vectors, recall from chapter 4) or creating significant features such as presence of class-specific words, character n-grams, word n-grams and pass them them to a classifier where the indirect attribute values are used as a class. In this example here, to extract ‘gender’ as an attribute, one should use men, women, unisex, child as different class labels [4, 44].

Figure 3-6. Characteristic performance improvement in the LSTM framework (Majumder et. al. [59])

Even though we have discussed so far about attribute extraction from textual data, mainly product titles, there are various recent approaches which extend this to multimodal attribute extraction incorporating various details such as title, description, image, reviews etc. about the product [43].

In the next sections we will be talking about expanding similar techniques we applied to product attributes to other facets of products and e-commerce.

Product Categorization and Taxonomy

Product categorization is a process of dividing products into groups. These groups can be defined based on similarity--product of same brands can be grouped together and product of the same type can be group together. Generally e-commerce has defined broad categories of products such as electronics, beauty products, foods etc. Once a new product arrives, it is needed to be grouped before it is put in the catalog. One can further define successively smaller group for tighter definition of products, such as GMO and non-GMO food products inside the food category.

This categorization process is typically manual but on a huge scale, it is almost impossible to categorize the product manually. This categorization is typically a classification process. The algorithm takes information from a variety of sources and applies the classification technique to solve this [49, 50].

Specifically, there are efforts where algorithms take input as the title or description and classify the product into a suitable category when all the categories are known previously. This again falls in the typical case of text classification. In this way, the categorization process can be automated. Once the category is determined, it is directly extended to relevant attribute extraction process which we have discussed. It is logical that a product will be passed to the attribute extraction process only when its category is discovered.

There are also multimodal approaches to solve this problem as well. The accuracy of the algorithm can be improved when both images and text can be used to solve this problem. Images can be passed to a convolutional neural network for generating image embedding and on the other hand, the text sequence can be encoded by LSTM [43], which in turn can be appended to be passed to any classifier for the final output.

As we have described here about categories for different products, there are good efforts by e-commerce companies to streamline this. The streamlined catalog or category-tree is called a taxonomy. We call it a tree since it can start from high level categories like electronics, which are considered as roots, and essentially branch out to smaller subcategories such as laptops, mobiles etc.

Importance of Taxonomy

A good taxonomy and properly linked products can be critical for any e-commerce system [10]. It is important for:

• Showing similar products to the product searched

• Better recommendation

• Selecting proper bundle of products for better deal to the customer

• Provision of replacing old products to new

• Price comparison across different products in the same category

Figure 3-7. Fig 11-8: A typical category hierarchy - taxonomy of a product [10]

Building a taxonomic tree is an elaborate process and text classification is widely used. Once the attribute values are obtained from the attribute extraction procedure, placing them in right level can be done via a hierarchical text classification. Since this process of establishing taxonomic tree is lengthy, hence to reduce the overhead, simple rule based classification methods are used. Cases which are complex and requires context to determine the right taxonomic level, are dealt by ML classification techniques such SVM or decision tree [51].

Figure 3-8. Fig 11-9: Taxonomic tree with different levels [51]

For new e-commerce platforms, creating a product taxonomy via right product categorization is an insurmountable task. This requires huge relevant data, manual intervention and category experts’ decision to build such rich content. All of these can be expensive to get for nascent e-commerce platforms, however, there are some APIs that can help. These APIs typically access large catalogue content of various big retailers and build the intelligence inside to categorize a product by scanning its unique product code. Small scale e-commerce should utilise the power of strong APIs for taxonomy creation and categorization. Once, a significant amount of product information is gathered, then using own rule based systems are advisable. Some of these APIs are offered by Semantics3, eBay and Lucidworks, to name a few. Figure 11-10 below shows a snapshot of one such API from Semantics3 [52]. It is one of the leading API providers for various retail and e-commerce solution. Their AI-driven API helps to categorise a product from its name.

Figure 3-9. Fig 11-10 Semantics3 terminal snapshot [52]

Product Enrichment

A product is generally identifiable through its Unique Product Code (UPC). But for a better customer experience, it is important to provide other relevant information as we have already discussed in the attribute extraction section. There could be potential sources of these information - short titles, long titles, product images, product description - but to extract relevant information from these sources to fill the catalog is important and requires machine learning algorithms. To enrich the information of a product, it is first important to categorize the product into the right category to establish and link to its proper taxonomy. Sometimes, a misleading title can hamper the faceted search in an e-commerce platform. Such an example is the following. Improving product title not only enhances the click-through rate but also improves the conversion rate.

In the example in Figure 11-11, the product title is too long, contains keywords like iPad, iPhone, Samsung which can easily mislead the search. This is an ideal case which requires determination of product category, the right taxonomy, the attributes which can further be used to better frame the product title. We discuss various aspects of product enrichment in this section.

Figure 3-10. Example of clumsy product title - an ideal case for product enrichment

Product enrichment is typically seen as a larger and more continuous process than just improving product titles in any online retail setup. Typically, automatic processes are set up to generate reports on how much product information is enriched. This ‘how much’ can be defined in a more detailed manner. The product content is typically divided into many levels based on their taxonomic definitions. At every level, the number of present attribute values over number of required attribute values can give an indication about ‘how enriched’ is the product content. For instance, for a product ‘a t-shirt’, information such as brand, size, color, material are important to have. Now it is also possible that brand is an attribute which is a higher level taxonomic attribute than attributes like color or material. Here, higher denotes a more generic information. This can be followed by more specific information i.e. lower level taxonomic attribute such as material. The product enrichment reporting process can peek into each of these levels and detect how many values are present and calculates a ratio over how many attribute values are required to present.

Apart from taxonomic level, there are many other ways to define the enrichment levels. Most of them are based on the importance of the information. Following is one of the ways of defining them, as by [51]:

Once an acceptable enrichment proportion is achieved, then a better look at the product title and description can be given. Improving product enrichment as a part of attribute filling is a part of attribute extraction which we have already discussed in the last section.

The problem we have depicted in the Fig 13 can be addressed as a string matching NLP task. Once, for that product, different taxonomic level as well as enrichment levels are filled or filled at least to an acceptable threshold (typically defined by the retail platform itself), then an attempt can be taken to making the product title more expressive and accurate.

The process can start with direct string matching. All the attribute values which are present can be checked with exact match across all the words or phrases present in the existing product title. If one or more attribute values are not present in the existing product title then they have to be incorporated. It is also required to filter out tokens which are not part of the product’s attribute values. We can see in the example given in Fig 13 that iPad, iPhone are not part of its attribute values. These tokens are misleading and can affect the faceted search. Hence, such tokens should be removed from the product title, unless they are very important to set up the context for the product.

Ideally, a predefined template for the product titles helps to maintain consistency across them. A wise approach would be to build a template which is composed of attribute values in an top down order of the taxonomic tree. The product category or type could be the first token in the product title. It can follow other attributes in the decreasing order of important such as Brand, Pack size, Color etc. Attributes which at the lowest level of importance (enrichment level) can be omitted to be put in the product title. These values can come as a part of large description of the product which typically appear after product title, price etc, in the e-commerce product webpage.

Product Duplication and Matching

There are many cases where the product information is mistakenly same, not the product images are the same which is termed as product duplication. Apart from product categorization and attribute extraction, product duplication is also an important aspect of e-commerce. Identifying duplicate product is a challenging task. This process also includes a product matching procedure. The following process describes one way to handle this problem. It includes: attribute match, title match and image match.

Attribute match

If the attributes are already extracted then we can look for similar attributes for both the products and compare the value. For example, if the brand is extracted for both the product then the brand can be compared. Matching attributes is a string matching task [53]. Two strings can be matched via exact character match or using string similarity metrics. String similarity metrics are typically built to take care of slight spelling mistakes, abbreviations etc. Customized metrics can also be built to avoid very specific characteristics of e-commerce keywords. Abbreviations are a big problem in product related data. The same word can be represented in multiple accepted abbreviations. They should be mapped to a consistent form (normalization, which we will discuss in brief later) or form agnostic rules have to be formulated to tackle the problem. An intuitive rule to tackle abbreviations while matching two words could be matching of first and last character and checking whether character those belong to the shorter word belong to the longer word or not.

Title match

There could be titles which have almost similar words or word sequences. Below are some examples:

Garmin nuvi 2699LMTHD GPS Device

nuvi 2699LMTHD Automobile Portable GPS Navigator

Garmin nuvi 2699LMTHD — GPS navigator — automotive 6.1 in

Garmin Nuvi 2699lmthd Gps Device

Garmin nuvi 2699LMT HD 6” GPS with Lifetime Maps and HD Traffic (010–01188–00)

These are the example of various wordings of the same product by different sellers. Definitely, the e-commerce platform would like to retrieve them all for a given search on this product. To identify all the titles are the same, there are multiple methods but the simplest one would be to compare bigrams and trigrams. It is also possible to generate title level features (mostly related with bigram and trigram) and then calculate euclidean distance between them. The problem of generating feature is always twofold, as we know creating right and significant feature is difficult also the patterns of matching may vary drastically across categories. Recent efforts have shown that using sentence level embedding and using a pair of textual phrase simultaneously to learn a distance metric helps in matching [54]. This is typically done by a neural network architecture called Siamese network [55]. Siamese network simultaneously takes both the sequences and learns to generate the embeddings in such a way that if they are similar, embeddings will be closer, else farther.

Image Match

Pixel-to-Pixel match or feature map matching or even advanced image matching techniques (siamese networks) are popular for image matching [56]. Duplication can be reduced by efficient image matching techniques. Most of the algorithms are based on the principles of computer vision approaches and depend on images quality and other size related particulars.

Sentiment Analysis

Figure 11-12 shows a screenshot of customer reviews of a product on Amazon.com. Most of us are familiar with seeing such screenshots on e-commerce websites These are the ratings available for iPhone X on Amazon.

Figure 3-11. Analysis of customer reviews - ratings, keywords, sentiments

As you can see, there are 67% of reviews which have a rating of 5 stars, i.e the highest, but there are 22% of reviews which have the lowest rating. Now for an e-commerce company, it is important to know what led the customers to give the lowest ratings. Now see two extreme examples of reviews of the same product -

Figure 3-12. A positive and a negative review

Certainly, both of these reviews contain some information about the phone which give the retailer cues what customers are thinking. Specifically, negative reviews are more important to understand. Look at the first review where the customer is stating that there are issues regarding phone those are getting shipped. It is mostly related to the defected screen which the retailer should take care of. Hence, it is crucial that there is a full understanding of the reviews. By nature, they are in the text, and mostly in very unstructured format, full of unforced errors such as spelling mistakes, wrong sentence constructions, incomplete words, abbreviations etc. This makes the review analysis even more challenging.

In chapter 4, we have discussed the basic sentiment analysis, as a classification problem. Starting from a dictionary based approach, we discussed several different ways of performing text classification, which can all be applied to sentiment analysis.

Ratings are considered to be directly proportional to the overall sentiment of the reviews. There are cases where the user mistakenly rated the product badly but had given a positive review. Understanding emotions directly from the text will help retailers to rectify these anomalies during analysis. Given the fact, we have already discussed how you can get the overall emotion for the review. But typically, a review doesn’t talk about one aspect of the product-- generally tries to cover most of the aspect of the product and finally, everything reflects in the review rating.

Have a look at the snapshot of iphone X review snapshot from amazon.com again in Figure 12.

Look at the section where it reads ‘Read reviews that mention’. These are nothing but the important keywords that Amazon has found out to be displayed which may help the customers to navigate better in skimming through the reviews. This clearly indicates that there are certain aspects on which customers are talking about. It could be user experience, or manufacturing aspects or price or something else. How could we know what are the emotions or feedback of the customer along these aspects? Our discussion till now has only given a high-level index of emotion for the whole review but will not be able to dig down more to understand it better. This necessitates the aspect level understanding of the reviews. These aspects could be pre-defined or could be extracted from the review data itself. Based on that, the approaches will be supervised or unsupervised, accordingly.

Aspect Level Sentiment analysis

Before we start the discussion of various techniques for aspect level sentiment analysis, we need to understand what is an aspect. An aspect is a semantically rich concept-centric collection of words which indicates certain properties or characteristics of the product. For instance in Figure 11-14, we see the kind of aspects a travel website might have- Location, Value and Cleanliness.

This is not only constrained to the inherent attribute of the product but also anything and everything related to the supply, presentation, delivery, return, quality and many more around the product. Typically, a clear distinguishing of these aspects is difficult unless assumed previously.

When the retailer has a very clear understanding of the aspect which can be talked about around the product, then finding aspects falls under the supervised category of algorithms. There is a common technique for using seed words or seed lexicons which essentially hints the crucial tokens that could be present under that particular aspect. For example, regarding user experience as an aspect for iPhone X, seed words could be screen resolution, touch, the response time of the phone etc. Again, it is completely up to the retailer at what level of granularity, they would like to operate. For example, screen quality alone could be a more granular aspect.

Supervised Approach

The supervised approaches mainly depend upon the seed words. It tries to identify the presence of these seed words in a sentence given a review. If it identifies a particular seed word in a sentence it tags the sentence by the corresponding aspect. Once all the sentences are tagged to any of the aspects, the sentiment analysis has to be done on a sentence level. Now since we already have an additional tag for each sentence, sentences having one tag can be filtered and sentiments for them could be aggregated to understand the feedback of the customer for that aspect. For example, all review sentences related to screen quality, touch, response time can be grouped together.

For a change, let’s look at a direct example from a travel website in Figure 11-14 where the aspect level sentiment analysis is apparent. As you see, there are specific ratings for location, check in, value, cleanliness - which are semantic concepts rightfully extracted from the data to present a more detailed view on the reviews.

Figure 3-13. Aspect level ratings on reviews given in a travel website

Connecting Overall Ratings to Aspects - Latent Rating Regression (LARA)

We have already seen how you can detect the sentiment for each aspect. Now, if we think, typically users also give an overall rating. The idea here is to connect that rating to individual aspect level sentiment. Details of LARA implementation are outside the scope of the book but here is an example of the system, generating aspect level rating for a hotel review [5]. You can delve into the details by looking at the reference.

Figure 3-14. Aspect wise sentiment prediction using LARA

We can assume that the final rating is nothing but a weighted combination of individual aspect level sentiments. The objective shall be estimating the weights, as well as the aspect level sentiment together. It is also possible to perform these two operations sequentially i.e. first determining the aspect level sentiment and then the weights.

These weights learned will finally indicate on top of various sentiment present for each aspect, how much a reviewer puts importance on that specific topic. It is possible that a customer is extremely unhappy with some aspect but maybe it is not in his priority. This information is crucial for e-tailers before they take any direct action.

Understanding Aspects in a better way

Once we derive all the aspects and tag each sentence with these aspects, it is possible to group the sentences by aspects. But given the huge volume of review an e-commerce website encounters, there will still be a lot of sentences coming under an aspect. Here, a summarization algorithm may save the day. Think about a situation, where you need to take an action regarding an aspect but you do not have the capacity to go through all the sentences regarding that particular aspect. Hence, you need an automatic algorithm which can pick and choose the best representative sentences for that aspect.

LexRank [6] is an algorithm, similar to PageRank which assume each sentence is a node and connects via sentence similarity. Once done, it picks the most central sentences out of it and presents an extractive summary for the sentences under an aspect.

Figure 3-15. The complete flow chart of review analysis - overall sentiments, aspect level sentiments, and aspect wise significant reviews

Case Study - Substitutes and Complements

Recommender systems are built on the idea of ‘similar’ products. This similarity can be defined in many ways such as content-based or user base. There is another way of identifying the item interrelationship, specifically in an e-commerce setting.

We know there are products which are typically bought together and can be termed as complements. On the other hand, there are pairs which are bought in lieu of the other, and they are termed as a substitute pair. Even though the economic definition is much more rigorous, these lines of thought typically capture the behavioral aspect of product purchase.

The consumer behavior is quirky, but their joint behavior can provide ample information about the product interrelationship. There are many ways by which these substitute pairs and complementary pairs can be detected. We will focus on one such approach which primarily relies on the textual information present around products to infer their interrelationship.

Julian McAuley has presented [10] a comprehensive way of understanding the product interrelationship in a framework where the query product is given, the framework returns the ranked products which are substitutes and complements. We will discuss this application as a case study in the light of e-commerce (Refer to Chapter 8 for a dedicated focus on all other recommendation techniques).

Typically, as we have discussed, the reviews do contain the specific information about product attributes. As we know, attribute extraction is a procedure itself and we have seen ways to extract them. This approach shows an alternative way of understanding product attributes, mostly in a latent manner and infer from them.

Figure 3-17. Substitutes and Complements based on product reviews [10]

How two products are linked

Now, the next task would be to understand how the two products are linked. We know the topic vectors for all the products which in turn latently capture the intrinsic properties of the product. Please note that these vectors do capture the product attribute information without explicitly extracting them. The idea is to create a joint feature vector out of the topic vectors available for a product pair and then predict if there is any relationship between these two products. This can be viewed as a binary classification problem where the features have to be obtained from the individual topic vector for the product pair. We call this process ‘link prediction’.

To ensure that the topic vector is expressive enough to predict a link or relationship between a product pair, the objective of obtaining topic vectors and link prediction can be solved jointly rather than sequentially.

Here we are trying to learn Topic vectors for each product as well as learn feature weights for each products. We are trying to predict whether the pair has a relationship (1) or not (0).

Figure 3-18. Topic vector and topic hierarchy expresses how different taxonomic identities and relations are captured in reviews

This shows how a topic vector becomes expressive enough to capture the intrinsic attributes of the product. Indeed, this a laptop charger and the model clearly depicts the topical similarity of various product type or hierarchy.