Tools and techniques for web information extraction (WIE) have recently been recognised as one of key enablers for semantic web (SW) scaling. In our long-term project named Rainbow 3 we address several intertwined topics that we consider important for efficient 'WIE for SW' applications:
1. Exploitation of multiple information modalities available in web documents 2. Synergy of learning and reuse of ontological information 3. Automated acquisition and labelling of training data for extractor learning 4. Bridging between automated acquisition of SW data and their usage 5. Support for easy design of WIE applications from components.
In this paper, we focus on an ongoing demo application in the domain of bicycle product offers. Section 2 presents the core method: automated HTML annotation based on Hidden Markov Models. Section 3 extends the analysis of HTML code with that of images. Section 4 describes the composition of product offer instances with the help of a simple ontology. Section 5 outlines the architecture of the demo application and the subsequent usage of extracted data in an RDF repository. Finally, section 6 focuses on future work.
For extracting product entries from web catalogues, we built a Hidden Markov Model (HMM) tagger, which assigns a semantic tag to each token from a document. Tokens are either words, formatting tags or images. In our experiments, we evaluated the HMM performance on a diverse set of web pages, which come from different web sites and have heterogenous formattings.
We manually annotated a set of 100 HTML documents chosen from the Google Directory Sports-Cycling-BikeShops-Europe-UK-England. Each document contains from 1 to 50 bicycle offers, and each offer consists of at least the bicycle name and price. There are typically 3-4 documents from the same shop in the data. Annotations for 15 bicycle characteristics were made using SGML tags 4 . A sample annotated data is shown in Figure 1.
To represent web documents, we employed extensive pre-processing. Similarly to [7], we transform each document into XHTML and perform canonicalisation of XML entities5 . Certain HTML tags and tag groups are replaced by their generalisations 6 . Since only words and images can be extracted, we dispose of mark-up blocks that do not directly contain words or images.
HMMs are probabilistic finite state machines, which represent text as a sequence of tokens. An HMM consists of states which generate tokens, and of Fig. 2. HMM architecture transitions between these states. States are associated with token generation probabilities, and transitions with transition probabilities. Both kinds of these probababilities are estimated from training data. For the purposes of information extraction, states are typically associated with semantic tags to be extracted. To annotate a document using a trained HMM, that document is assumed to have been generated by that HMM. The most probable state sequence is then found using the Viterbi algorithm [12].
The structure of our HMM is inspired by [6] and is sketched in Figure 2. Extracted slots are modelled using target states (denoted as T). Each target state is accompanied by two types of helper states responsible for representing the slot's characteristic context -the prefix and suffix states (P and S). Irrelevant tokens are modelled by a single background state (B). Contrary to [6] and [17], which use independent HMMs trained for each slot separately, we train a single composite HMM capable of extracting all slots at once. Our model thus contains multiple target, prefix and suffix states. This approach, also used in [1], captures the ordering relations between nearby slots (e.g. product image often follows its name). We experimented also with other HMM architectures, with results presented in [16].
For the purpose of extracting product images, we examined the impact of image information available to the HMM tagger. As a baseline approach, we measured the tagging performance when no image information was available for tagging. In this case, all images were represented by the same token and product pictures could only be distinguished based on the context in which they appeared.
In order to provide our tagger with more information, we built image classifiers to determine whether the extracted product is depicted in a particular image. We used the following features for classification: image dimensions, similarity to training product images, and whether there is more than one occurrence of the same image in the containing document.
For our domain, we modelled images of bicycles using a 2-dimensional normal distribution, only estimated from positive training examples7 . The dimensions x, y of a new image I are first evaluated using the estimated normal density N . The density value is then normalized to the interval (0,1) using the density's maximum value N max .
An image I is then classified as P os or N eg by comparing its Dim(I) score to a threshold T Dim . This threshold was estimated by minimizing the classification error rate on a separate heldout set of 150 images.
Within our document collection, image dimensions appeared to be the best single predictor with the error rate of 6.6%. However, this is mainly due to our collection being limited to relevant product catalogues only. When dealing with more heterogeneous data, features describing the actual image content will become necessary.
We experimented with a latent semantic approach to measuring image similarity, described in [10] and [11]. This kind of image similarity has been applied to image retrieval from collections, where the task often is to find the most similar image to a query. We used this image-to-image similarity measure sim(I, J) to compute sim C (I), the similarity of an image I to a collection of images C. In our experiments, C contained the training bicycle pictures (positive examples only).
To compute sim C (I), we used the K nearest neighbor approach and averaged the similarities of the K most similar images from the collection.
Experimentaly, we set K = 20, since lower values of K lead to a decrease in the similarity's robustness8 and higher values did not bring further improvement. To build a classifier, a similarity threshold T Sim was estimated on a heldout set in the same way as for the dimension classifier above. The error rate of the classifier was 26.7% on our document collection.
For the combined image classifier, we used the above described dimension score Dim(I), similarity score Sim(I) and a binary feature indicating whether the image occurs more than once in the document. We experimented with different classifiers available in the Weka9 environment, and the best error rate10 of 4.8% was achieved by the multilayer perceptron algorithm.
Results for all three classifiers are compared in Table 1. All results were measured using 10-fold cross-validation on a set of 1, 507 occurences of 999 unique images taken from our training documents. The first two algorithms used additional 150 heldout images to estimate their decision thresholds. The crossvalidation splitting was done at the level of documents, so that all images from a single document were either used for training or for testing.
To improve extraction results, we need to communicate the image classifier's results to the HMM tagger. Currently we do this simply by substituting each image occurence in a document by its class. Since these binary decisions would leave little room for the HMM tagger to fix incorrect classifications, we adapted the above binary classifiers to classify into 3 classes: P os, N eg, and U nk. In this way, the HMM tagger learns to classify the P os and N eg classes correspondingly, and the tagging of the U nk class depends more strongly on the context. To build the ternary versions of the dimension-and similarity-based classifiers, we introduced costs for the classifier's decisions. Each wrong decision was penalized by C Miss = 1 and the cost of each U nk decision was C Unk ∈ (0, 1). We set C Unk manually such that the classifier produced 5-10% of U nk decisions on the heldout set. While minimizing the sum of these costs on the heldout set, two thresholds were estimated for both the dimension-and similarity-based classifiers, delimiting their N eg, U nk and P os decisions.
For the combined ternary classifier, we achieved the best results with a decision list shown in Table 2. The list combines image occurence count with the results of the dimension-and similarity-based ternary classifiers, denoted as class 3 Dim and class 3 Sim respectively. We evaluated information extraction results with all three ternary classifiers and compared the results to the case where no image information was available. The new image information from the combined classifier lead to an increase of 19.1% points in picture precision and also to subtle improvements for other tags. Improvements in precision and recall for 3 chosen slots (product pictures, names and prices), measured on a per-token basis, are shown in Table 3 for all three classifiers.
Semantic web is not about isolated tagged items but about complex and interrelated entities; we thus need to group the labels produced by automated annotation into instances. We currently use a simple sequential algorithm that exploits constraints defined in a tiny presentation ontology11 [9], which partly pertain to the generic domain (bike offers) and partly to the way of presenting information in web catalogues. Figure 3 shows an experimental presentation ontology containing the class 'Bike offer'. The utilized constraints are uniqueness, multiplicity and optionality of certain properties, the latter two indicated with the * and ? symbols, respectively 12 . In addition, 'sticky' properties (indicated with !) are distinguished: as soon as the value of sticky property is discovered on a page, it is filled to all objects extracted afterwards, until a new value is discovered for this property.
An annotated item added to the currently assembled (bike offer) instance unless it would cause inconsistency; otherwise, the current instance is saved and a new instance created to accommodate this item and the following ones. Despite acceptable performance on error-free, hand-annotated training data, where the algorithm correctly groups about 90% of names and prices, this 'baseline' approach achieves very poor results on automatically annotated data: on average, less than 50% of corresponding annotations are grouped properly, often for trivial reasons. The most critical problems are connected with missing or extra annotations, multiple different references to a single slot, and with transposed HTML tables.
All components developed within the Rainbow project are wrapped as web services. The WIE component itself is currently being called by a simple control routine (written in Java), which also optionally calls other analysis tools: in the bicycle application, we so far experimented with URL-based navigation over the website, extraction of the content of selected META tags, and extraction of 'company profile sentences' from free text 13 . The results are transformed to RDF (with respect to a 'bicycle-offer RDFS ontology') and stored in a Sesame [2] repository. An end-user search interface to this repository14 is shown in Fig. 4. It relies on a collection of query templates expressed in SeRQL (the native query language of Sesame) and enables a simple form of navigational retrieval [16]. Most urgently, we need to replace the 'toy' implementation of ontology-based instance composition with a version reasonably robust on automatically annotated data. For some of the layout-oriented problems mentioned in section 4, partial solutions recently suggested in IE research (e.g. [3,5]) could be reused. We also consider introducing HMMs even to this phase of extraction; a modified version of Viterbi algorithm supporting domain constraints (such as those in our presentation ontology) has already been described in [1]. Another aspect worth investigation is the possibility of (semi-)automatic construction of presentation ontologies from the corresponding domain ontologies.
A critical bottleneck of ML-based IE methods (in particular of statistical ones) is the volume of labelled training data required. In our experiments with product catalogues, we noticed that the tagger often classifies most product entries correctly but misses a few product names that are very different from the training data. We developed a simple symbolic algorithm that identifies similar structural patterns in a document. For example, the HTML tag sequence <td> <a> <font> <br/> </font> </a> </td> with arbitrary words in between appears 34 times in one of our training documents: the tagger successfully annotated 28 product names contained in these patterns between <font> and <br/>, but missed the remaining 6. In such cases, we could collect the remaining product names and use them to enrich the model's training data. By learning novel product names from these 'easy' pages, the model will learn to also recognise them in less structured documents 15 . We also plan to bootstrap the method with data picked from public resources related to product offering, following up with our earlier experiments with Open Directory headings and references [8].
Another important task is to replace hard-coded control routines with semiautomatically constructed, implementation-independent application models. A knowledge modelling framework has already been introduced for this purpose [14]; currently we examine the adaptability of a PSM-based semantic web-service configuration technique in connection with this framework [15].
Eventually, we plan to associate our efforts with the popular Armadillo project [3], with which we share most of our abovementioned research interests.
The research is partially supported by grant no.201/03/1318 of the Grant Agency of the Czech Republic, "Intelligent analysis of the WWW content and structure".