Search By Document Type

[0001] This description relates generally to computer aided searching and more specifically to searching for instruction documents.

[0002] People often face unfamiliar tasks, and thus they need appropriate instructions for conducting them. Much effort has been made to cope with the problem. Including books for popular ‘how-to’ questions. Many online services for answering how-to questions are also available which can maintain a large collection of instruction documents and provide a search service on the collection.

[0003] However, none of them typically can cover all of the how-to questions in daily life. Thus, it could be helpful to have a system that helps automatically retrieve ‘instructions’ (i.e., documents of task guides) on the web (either Internet or intranet).

[0004] The following presents a simplified summary of the disclosure in order to provide a basic understanding to the reader. This summary is not an extensive overview of the disclosure and it does not identify key/critical elements of the invention or delineate the scope of the invention. Its sole purpose is to present some concepts disclosed herein in a simplified form as a prelude to the more detailed description that is presented later.

[0005] The present example provides a way to search for manuals or other documents by combining a relevance model and a type model. Training data is provided to each model and the model is then applied to a first plurality of documents. Two collections of documents result. A first collection ranked by type, and a second collection ranked by relevance. Through a linear interpolation the documents are combined to produce a second plurality of documents ranked by relevance and type.

[0006] Many of the attendant features will be more readily appreciated as the same becomes better understood by reference to the following detailed description considered in connection with the accompanying drawings.

[0013] Like reference numerals are used to designate like parts in the accompanying drawings.

[0014] The detailed description provided below in connection with the appended drawings is intended as a description of the present examples and is not intended to represent the only forms in which the present example may be constructed or utilized. The description sets forth the functions of the example and the sequence of steps for constructing and operating the example. However, the same or equivalent functions and sequences may be accomplished by different examples.

[0015] The examples below describe a manuals search by using a relevance model and a type model. Although the present examples are described and illustrated herein as being implemented in an instruction manual search system, the system described is provided as an example and not a limitation. As those skilled in the art will appreciate, the present examples are suitable for application in a variety of different types of search systems.

[0016] Traditional information retrieval typically aims at finding relevant documents. However, relevant documents found in this manner are not necessarily instruction documents, i.e., answers to how-to questions. Thus, a naive application of the traditional information retrieval may not produce the desired instructions.

[0017] In the following example, investigation of question answering in a new setting is provided by a method called “manuals search”. More specifically, given a how-to query, all documents may be automatically retrieved and ranked which are relevant to the query and which are also likely to be an instruction document. In particular the training type manual is interpreted, or seen, as a classification problem. And the method of fusing scores from the type model and the relevance model may be done by linearly interpolating the scores.

[0018] The examples below provide a manuals search process typically utilizing a relevance model and a type model. Specifically, when given a ‘how-to’ type of query (e.g., ‘how to create a link’), documents are retrieved and ranked them according to both the likelihood of being an instruction document (a document containing description about how to perform a task) and the relevance to the query. Traditional document retrieval typically only considers the relevance of documents to queries. The method of performing the task may include what may be referred to here as ‘relevance model’ and ‘type model’. With the relevance model, it is determined whether or not a document is relevant to a query. With the type model, it is determined whether or not a document is an instruction document. Okapi and Logistic Regression may be employed as the relevance model and the type model, respectively. A method for combining the uses of the two models based on linear interpolation is also proposed.

[0019] In the example provided the method is described in terms of a manuals search. More specifically, given a how-to query, documents which are relevant to the query and which are likely to be instruction documents are retrieved and ranked. The manuals search is based on a relevance model and a type model. Okapi may be employed as the relevance model and Logistic Regression as the type model, respectively. A method based on a linear interpolation technique is also proposed to fuse the output of the relevance model and the type model.

[0020] The instant manuals search typically performs well on both artificial data sets and real data sets. For how-to queries, good instruction documents are often ranked higher using the exemplary approach than the baseline method of either solely using Okapi or solely using Logistic Regression. The proposed approach typically performs well on different domains.

[0021] The example provided of a method of manuals search may be based on a relevance model and a type model. In particular Okapi may be employed as the relevance model and Logistic Regression as the type model, respectively. Okapi is a system for document retrieval based on a probabilistic model. It retrieves and ranks documents according to the relevance of documents to queries. Okapi or its equivalent may be employed in the example provided. Okapi is described more fully by S. E. Robertson, S. Walker, M. M. Beaulieu, M. Gatford, and A. Payne. Okapi at TREC-4. In D. K. Harman, editor, The Fourth Text Retrieval Conference (TREC-4), pages 73-96, Gaithersburg, Md., 1996. National Institute of Standards and Technology, Special Publication 500-236.

[0022] Logistic Regression is a probabilistic classification model more fully described in T. Hastie, R. Tibshirani, and J. Friedman. The Elements of Statistical Learning. Springer, N.Y., 2001. In contrast to other classification models such as Support Vector Machine (SVM), Logistic Regression typically outputs probability values rather than scores in classification.

[0023] A method based on a linear interpolation technique is also utilized to fuse the output of the relevance model and the type model.

[0024] A manuals search may be more helpful than a conventional search that tends to return a greater range of results, many of which are not instructions. In performing a manuals search a query is typically received first. The query is usually a how-to question, e.g., ‘how to create a link’. Next using conventional methods automatic retrieval of all of the relevant and likely instruction documents is performed. Next, the documents may be ranked according to the degree of likelihood of being an instruction document and the degree of relevance to the query.

[0025] Manuals search typically needs to assure that the retrieved documents are relevant to the queries as well. However, in contrast, manuals search may also need to assure that the retrieved documents are instruction documents. Table 1 shows possible sets of documents that may be searched for. From Table 1, A is the set that is typically desired in manuals search. C is the set that is relevant but non-instruction and thus should be filtered out.

TABLE 1
Two views of documents
	Relevant	Irrelevant
	Instruction	A	B
	Non-instruction	C	D

[0026] FIG. 1 shows two examples of web documents that may be found in a conventional search. The first document 101 is not an instruction document and the second document 102 is an instruction document. Thus, if the query is ‘how to create a link’, then the second document 102 would be preferred by users. However, if only relevance is considered, then the first document 101 will likely be ranked higher, because it would typically appear to be more relevant to the query. Question answering may be ideal for accessing information on instructions, if realized, because its goal is to provide a single combined answer. One can simply get all the necessary information by reading the combined answer. However, generation of such a combined answer may be very challenging or even impossible for current systems. First, most answers to a how-to query consist of step-by-step guides as the example shown in the second document in FIG. 2. Deletions, insertions, or re-orderings made in any part of the document may confuse or misguide typical searches. Second, different instruction documents may have different assumptions and settings, even for the same topic. Thus it may not be appropriate to combine the instructions based on different assumptions and settings.

[0027] FIG. 2 shows two exemplary instruction documents found as the result of a conventional search pertaining to the query ‘how to compile Amaya’. However, they provide instructions for different settings. One document 201 provides instructions ‘with Autoconf’ settings. The other document 202 provides instructions with Windows' settings. Therefore, a reasonable approach would be to show users the instruction documents separately.

[0028] As seen above judging whether a document is a relevant instruction document, and thus can be used as an answer to a how-to query in an objective way may be hard. However, we can still provide relatively objective guidelines for the judgment. The objective guidelines will be termed the specification in this following description. The specification may be used extensively for development and evaluation of the manuals search process. As previously shown in Table 1, the specification can be designed from two view points. For the notion of relevance, specification may be defined in a similar fashion as that in traditional information retrieval. In doing so the notion of instruction is clarified first.

[0029] FIG. 3 shows examples of documents 301 302 303 that might be found as a result of a search. First, an instruction document is a document created for answering a how-to query. More specifically, by reading the document, one can understand how to perform the task of the how-to query. Second, in the example provided, an instruction document is assumed not to be a document containing instructions for a number of different tasks. For example, the first document in FIG. 4 is not regarded as an instruction document. Third, in the example provided, an instruction document is assumed not to be a document that only consists partly of instructions.

[0030] For example, the second document 302 is not viewed as an instruction document, because a section of it is about how-to. As can be seen from the above discussion some instruction documents may contain step-by-step guides (cf., the second document in FIG. 2); while other instruction documents may just consist of several sentences (e.g., the last document in FIG. 4).

[0031] The criteria used above may be quantified for use in a manuals search. With the specification defined above, four labels similar to those in Table 1 are provided. For the purposes of manuals search, however, there should be no difference between the labels C and D. Thus, we combine C and D together to C-D. In manuals search, given a query q and a document d, one of the three labels below can be assigned:

[0032] A: document d is relevant to q and is an instruction document

[0033] B: document d is irrelevant to q but is an instruction document

[0034] C-D: document d is of no use

[0035] FIG. 4 is a flow diagram showing manuals search by using a relevance model 401 and a type model 402. In the example provided of manuals search by using a relevance model and a type model the input may be a query 403 and a collection of documents 404. The documents may have resulted from a conventional search, or may simply be a collection of documents to be examined. The exemplary approach to manuals search includes two steps. First, a representation to relevance to a query and a likelihood of being an instruction document is formed with two sub-models, which we call a ‘relevance model’ 401 and a ‘type model’ 402, respectively. In the relevance model, it is judged whether or not a document in the input is relevant to the query 407. In the type model, it is judged whether or not a document in the input is an instruction document 408. Next, a linear interpolation technique may be used to combine the scores output from the two sub-models 405. The documents are then ranked in descending order of their combined scores 406.

[0036] As shown training data may be supplied in general to a training relevance model. However, training data is not needed with the exemplary Okapi BM25 model or other types of relevance models that may be utilized in place of Okapi BM25. In the example provided training data may also be supplied to the type model, however it is anticipated that in some applications training data may not be needed.

[0037] Given a how-to query and a document, the relevance model finds a relevance score. In manuals search, for a given query, a list of <document, relevance_score> pairs using the relevance model to generate the relevance score are created. In the present example the Okapi BM25 relevance score may be employed as the relevance model. For indexing the title and the body of a document are indexed in separate fields. For each field, the Okapi BM25 weighting scheme is used to calculate a score. Then the scores of the title field and the body field are combined linearly, and the combined score is viewed as the relevance-score.

[0038] Given a document, the type model outputs a type score. That means that we assume that the type model is independent from queries. In manuals search, we create a list of <document, type_score> pairs using the type model.

[0039] We take a statistical machine learning approach to constructing a type model. More specifically, given a training data set D={x_i,y_i}₁ⁿ, we construct a model Pr(y|x) that can minimize the error in predicting of y given x (generalization error). Here x_i ε X and y_i ε{1,−1}. x represents a document and y represents whether or not a document is an instruction document. When applied to a new document x, the model predicts the corresponding y and outputs the score of the prediction. In this example, we adopt the Logistic Regression Model.

[0040] Logistic Regression

[0041] The Logistic Regression Model satisfies:

$\begin{array}{cc}\log \frac{\mathrm{Pr}\left(y=1|x\right)}{1-\mathrm{Pr}\left(y=1|x\right)}={\beta }_0+\beta ·x& \left(1\right)\end{array}$

where β represents the coefficients of a linear combination function and β₀ is the intercept. The Logistic Regression Model is usually estimated by using Maximum Likelihood.

[0042] The Logistic Regression Model assigns a probability to an instance (in our case a document) probability according to the following equation.

$\begin{array}{cc}\mathrm{Pr}\left(y=1|x\right)=\frac{1}{1+{}^{-\left({\beta }_0+\beta ·x\right)}}& \left(2\right)\end{array}$

[0043] We calculate the type-score of a document according to:

$\begin{array}{cc}\mathrm{type\_score}=\log \frac{\mathrm{Pr}\left(y=1|x\right)}{1-\mathrm{Pr}\left(y=1|x\right)}& \left(3\right)\end{array}$

[0044] Features

[0045] The Logistic Regression Model utilizes binary or real valued features as described below. Most features are created to characterize title, first heading and first sentence of documents. Although those skilled in the art will realize that in alternative examples features are not limited to characterizations of titles heading sentences and the like. Title is the text enclosed by the HTML tag ‘<title>’ and ‘</title>’. Heading is the text enclosed by the HTML tag ‘<H1-6>’ and ‘</H1-6>’. First heading refers to the first non-empty heading of a HTML document. First sentence is the first sentence appearing in the body of a HTML document.

[0046] ‘How To’

[0047] Whether or not the title of a document contains the words of ‘how to’, ‘howto’ or ‘how-to’ is an important indicator. This is represented using a binary feature. There are similar features with regard to the first heading and the first sentence of a document. Those skilled in the art will realize that in alternative examples of knowledge extraction other words or phrases of interest may be identified and used as important indicators.

[0048] ‘Doing Something’

[0049] The appearance of the suffix ‘ing’ in the first word of the title is another indicator of an instruction document. Sometimes people use the template of ‘doing something’ instead of ‘how to do something’ for the title of an instruction document. The value of the feature is binary, too. Similar features have also been defined for the first heading and the first sentence. Those skilled in the art will realize that other words or word fragments from the title may be used depending upon the specific application of search by using a relevance model and a type model.

[0050] Text Length

[0051] Also defined is the following real-valued feature:

log(length(title)+1)   (4)

where length(title) denotes the number of words in the title. A document with a short title (e.g. a one-word title) tends to be a non-instruction document. Similar features have also been defined for the first heading and the first sentence.

[0052] Identical Expressions

[0053] If the texts in any two of the three parts: title, first heading and first sentence are identical, then this feature is 1. Otherwise, it is 0. An instruction document usually repeats its topic in these three places.

[0054] Bag of Words

[0055] Also relied upon are ‘bag-of-words’ features. The method collects high frequency words in the titles of the documents in training data and create a bag of the keywords. Some keywords play positive roles (e.g., ‘troubleshoot’, ‘wizards’) and some negative ones (e.g., ‘contact’). If the title of a document contains one of the keywords, then the corresponding feature will be 1, otherwise 0. Similar features have been defined for the first heading and the first sentence.

[0056] A ranking_score may be calculated by linearly interpolating the relevance_score and type_score as the Equation (5).

ranking_score=λ type_score+(1−λ)·relevance_score   (5)

Here, λ ε[0,1] is a weight used to balance the contribution of the relevance model and the type model. As will be explained later it is typically better to have λ=0.5 In manuals search, documents are retrieved and ranked in descending order of the ranking_score.

[0057] In principle, given a query and a document collection, one can calculate the ranking score of each of the documents with respect to the query. In an implementation of the method, the top 100 documents ranked by the relevance model (Okapi) are first collected. Next ranking scores are calculated only for the top 100 documents. In this way, a manuals search may be conducted very efficiently.

[0058] Manuals search may be formalized in a more general framework called ‘typed search’. In typed search, documents are retrieved and ranked not only on the basis of relevance to the query, but also the likelihood of being the desired type. Given a query q and a document d, we calculate the ranking score of the pair using Equation (6):

Pr(r,t|q,d)≈Pr(r|q,d)·Pr(t|d)   (6)

where r and t denote relevance and type, respectively. Both r and t are binary variables. In manual search, for example, t means that a document is an instruction document. In the equation, we make an assumption that r and t given a q and d are independent.

[0059] There are many ‘types’ that can be considered such as definition, letter, and home page, for example. In manuals search, one can define the relevance score and the type score as log odds of Pr(r|q,d) and Pr(t|d), respectively (cf., Equation (3)). This justifies why we make use of equal weight in the linear combination in Equation (5).

[0060] Kraajj et al. have proposed using Language Model in the task of home/named page finding. They employ a model as follows, which assigns a score to a page d given a query q:

Pr(d|q)∝Pr(d)·Pr(q|d)   (7)

The first model on the right hand side of Equation (7), corresponds to the type model in Equation (6) and the second model corresponds to the relevance model. Therefore, home page finding can be viewed as a specialization of typed search. For further information on using a Language Model see W. Kraajj, T. Westerveld and D. Hiemstra. The Importance of Prior Probabilities for Entry Page Search. In Proc. of the 25th annual international ACM SIGIR conference on research and development in information retrieval, 2002. The contents of which are incorporated in this patent application in their entirety,

[0061] In manuals search the documents have been ranked by combining a relevance model and a type model. Okapi and Logistic Regression have been used as the relevance model and the type model, respectively. Final ranking scores are then obtained by linearly interpolating the scores from the two models. The proposed method may be generalized in a general framework called typed search.

[0062] FIG. 5 illustrates an exemplary computing environment 500 in which the manuals search by using a relevance model and a type model described in this application, may be implemented. Exemplary computing environment 500 is only one example of a computing system and is not intended to limit the examples described in this application to this particular computing environment.

[0063] For example the computing environment 500 can be implemented with numerous other general purpose or special purpose computing system configurations. Examples of well known computing systems, may include, but are not limited to, personal computers, hand-held or laptop devices, microprocessor-based systems, multiprocessor systems, set top boxes, gaming consoles, consumer electronics, cellular telephones, PDAs, and the like.

[0064] The computer 500 includes a general-purpose computing system in the form of a computing device 501. The components of computing device 501 can include one or more processors (including CPUs, GPUs, microprocessors and the like) 507, a system memory 509, and a system bus 508 that couples the various system components. Processor 507 processes various computer executable instructions, including those to ** to control the operation of computing device 501 and to communicate with other electronic and computing devices (not shown). The system bus 508 represents any number of several types of bus structures, including a memory bus or memory controller, a peripheral bus, an accelerated graphics port, and a processor or local bus using any of a variety of bus architectures.

[0065] The system memory 509 includes computer-readable media in the form of volatile memory, such as random access memory (RAM), and/or non-volatile memory, such as read only memory (ROM). A basic input/output system (BIOS) is stored in ROM. RAM typically contains data and/or program modules that are immediately accessible to and/or presently operated on by one or more of the processors 507.

[0066] Mass storage devices 504 may be coupled to the computing device 501 or incorporated into the computing device by coupling to the buss. Such mass storage devices 504 may include a magnetic disk drive which reads from and writes to a removable, non volatile magnetic disk (e.g., a “floppy disk”) 505, or an optical disk drive that reads from and/or writes to a removable, non-volatile optical disk such as a CD ROM or the like 506. Computer readable media 505, 506 typically embody computer readable instructions, data structures, program modules and the like supplied on floppy disks, CDs, portable memory sticks and the like.

[0067] Any number of program modules can be stored on the hard disk 510, Mass storage device 504, ROM and/or RAM 509, including by way of example, an operating system, one or more application programs, other program modules, and program data. Each of such operating system, application programs, other program modules and program data (or some combination thereof) may include an embodiment of the systems and methods described herein.

[0068] A display device 502 can be connected to the system bus 508 via an interface, such as a video adapter 511. A user can interface with computing device 702 via any number of different input devices 503 such as a keyboard, pointing device, joystick, game pad, serial port, and/or the like. These and other input devices are connected to the processors 507 via input/output interfaces 512 that are coupled to the system bus 508, but may be connected by other interface and bus structures, such as a parallel port, game port, and/or a universal serial bus (USB).

[0069] Computing device 500 can operate in a networked environment using connections to one or more remote computers through one or more local area networks (LANs), wide area networks (WANs) and the like. The computing device 501 is connected to a network 514 via a network adapter 513 or alternatively by a modem, DSL, ISDN interface or the like.

[0070] Those skilled in the art will realize that storage devices utilized to store program instructions can be distributed across a network. For example a remote computer may store an example of the process described as software. A local or terminal computer may access the remote computer and download a part or all of the software to run the program. Alternatively the local computer may download pieces of the software as needed, or distributively process by executing some software instructions at the local terminal and some at the remote computer (or computer network). Those skilled in the art will also realize that by utilizing conventional techniques known to those skilled in the art that all, or a portion of the software instructions may be carried out by a dedicated circuit, such as a DSP, programmable logic array, or the like.