second global symposium of intellectual property authorities an approach to process multilingual...
TRANSCRIPT
Second Global Symposium of Intellectual Property Authorities
An approach to process multilingual patents corpora
Dr. Barrou DIALLOHead of Research, European Patent Office, Rijswijk
Geneva, September 17, 2010
Directorate-General 2Research & Architecture/Research Unit
About the R&D Department
• At the origin of the 1st Machine Translation System for patents• Entry point for testing and evaluating available solutions• Portfolio of academic & international collaborations• Strong background in algorithmic and linguistics• Network of active users and testers
EPO R&D Department
Directorate-General 2Research & Architecture/Research Unit
Our Mission
Main tasks:
Resulting to:
Management support on ICT issues
Providing an instrument to translate user needs into Projects
• Performing quantitative analysis• Advise over technical solutions to decision-makers• Providing users with sensible options and recommending courses of action
Coordinating research initiatives across IT
• Technical advises• Market studies and research• Technological Forecasting• Risk analysis and Strategic planning
• Identifying and communicating business opportunities• Ensuring smooth transition from research to development• Communicate practices and experiences• Formalise research work across all departments
On request supporting the EPO MT Task force and/or IP5 MT activities
Directorate-General 2Research & Architecture/Research Unit
Current Research Subjects
Semantic Search Engines
Machine Translation for Asian Languages
Our Expertise
Graphical Visualization
Directorate-General 2Research & Architecture/Research Unit
Our Mission
R&D center as a source of Efficiency:
• Efficient Reading
• Accurate Searching
• Fast Granting
Our Vision: Turning Technology into an effective IP Process
Directorate-General 2Research & Architecture/Research Unit
Example of an R&D platform for linguistic purposes
Directorate-General 2Research & Architecture/Research Unit
Logical view of a document in an IR systems
structure
Accents,spacing stopwords
Noungroups stemming
Manual indexingDoc
structure Full text Index terms
adapted from J. H. Wang, 2008
Doc - Translated x1
Doc - Translated xn...
Multiples languages add another dimension to Retrieval systems for patents.
Directorate-General 2Research & Architecture/Research Unit
MT: Setup of an evaluation platform
• Unix server hosting fullttext patent data of source and target languages
• "mteval" scoring script for the Open MT Evaluation (http://www.itl.nist.gov/iad/mig//tools/)
• Case of a small set of Japanese patent documents:
– 54 JP patents
– 54 JP priority documents published at the USPTO
– analysis over the claim section
GOAL: To provide the users with a clear assessment of the quality of MT systems
Which indicators of quality can be considered as valid?
Directorate-General 2Research & Architecture/Research Unit
Absolute score computation processing scheme
• For each document– choose candidate sentences (ca. 10 segments)
• find the corresponding HT• compute the BLEU score• compute the NIST score
– compute the document average BLEU score– compute the document average NIST score– compute the BLEU - NIST correlation– compute the BLEU - HT correlation– compute the NIST - HT correlation– store the IPC class
• For the collection – (per IPC class):
• compute the average BLEU score• compute the average NIST score
– compute the correlation between scores in each IPC class
Directorate-General 2Research & Architecture/Research Unit
Example of raw results
Bleu score JPO 54 documents
NIST score JPO 54 documents
High variation of scoresHigh correlation between BLEU and NISTExtreme cases:•BLEU 0 at 9th position•BLEU 0.30 at 25th position
Directorate-General 2Research & Architecture/Research Unit
NIST vs. BLEU correlation
0.0000
0.0500
0.1000
0.1500
0.2000
0.2500
0.3000
0.0000
0.5000
1.0000
1.5000
2.0000
2.5000
3.0000
3.5000
4.0000
4.5000
1 3 5 7 9 111315171921232527293133353739414345
JP system case
Google case
Directorate-General 2Research & Architecture/Research Unit
Our findings based on a limited example
• Correlation between BLEU scores and Human-translated documents
– high scores correlate with understandable translations
– low scores correlate with non-understandable translations
• Differences between documents from different IPC classes
• Spread of scores is large (cf. std dev.)
But:
BLEU is consistent with Human translations:
• Results are absolute: they need to be compared to other systems• Bias can be introduced by the origin of data (IPC class, complexity, ...)
Directorate-General 2Research & Architecture/Research Unit
JP: BLEU & NIST vs. IPC classes
BLEU score
NIST score
To address the issue of data origin:
Directorate-General 2Research & Architecture/Research Unit
MT systems relative score computation scheme
BLEU score JP translation system vs. Google system
BLEU score Google BLEU score JP
Directorate-General 2Research & Architecture/Research Unit
Best, medium and worse case examples
JP translation system:•NIST: 4.7962 •BLEU: 0.1443
Worse case:Bleu score = 0 for JP
Medium case:Bleu score = 0.15 for JP
Best case:Bleu score =0.30 for JP
Mean scores for the whole collection:
Google:NIST: 4.5796 BLEU: 0.1185
Directorate-General 2Research & Architecture/Research Unit
Worse case example JP (BLEU=0)
JP MT Claim 1:an avalanche photo-diode (APD) which adjusted bias voltage so that a multiplication factor might become 30 or less ] A microscopic weak optical power detector detecting intensity of light irradiated by above APD by connecting a capacitor for generating inside this APD and accumulating a ****(ed) carrier, reading voltage of this capacitor periodically, and taking the difference
JP Human Translation claim 1:1. A low-level light detector, comprising: an avalanche photodiode with a bias voltage adjusted to produce a multiplication factor of up to 30; a capacitor connected to the avalanche photodiode for accumulating carriers produced and multiplied in the avalanche photodiode;>biasing means of the avalanche photodiode; outputting means of a capacitor voltage change; and control means of the biasing and outputting means; wherein the low-level light detector detects an intensity of light impinging on the avalanche photodiode by periodically reading capacitor voltages and obtaining differences between the voltages.
JP MT Google claim 1:01 claim Avalanche adjusted so that the bias voltage multiplication factor of 30 or less (APD), and comprisesAPD occurs within, connect a capacitor for storing carriers multiplication reads regularly voltage capacitor comprises, by taking the difference above, APD was irradiated characterized by intensity of light to detect Ru, very faint light detector. [2]
Remarks:1. "APD" is not in Human Translation2. "Avalanche photodiode appears 5 times in Human vs. different occurrences in MT3. "photo diode" is missing in Google4. Much more information in HT than MT
Directorate-General 2Research & Architecture/Research Unit
Medium case example: JP (BLEU=0.15)
JP Human Translation:1. An electronic throttle control device of an internal-combustion engine that controls an engine output by computing a quantity of a throttle opening degree on the basis of a manipulation quantity of an accelerator pedal by a driver by means of a computation portion in an electronic control unit, and by controlling a throttle opening degree using a specific actuator on the basis of a computed command value of the throttle opening degree,>wherein the electronic control unit includes:>a judgment function portion
JP MT:[Claims 1]. It has the following and is characterized by choosing a predetermined map from said two or more characteristic conversion factor maps, and calculating a target throttle opening command value corresponding to a judgment result of said judgment function part. being based on the amount of operations of a driver's accelerator by a calculating means of an electronic control unit (ECU) -- a throttle -- an opening -- quantity calculating and,
Google MT:claims [claim] 1 electronic control unit (ECU) by means of operation, the driver's accelerator operation is calculated caliber throttle opening based on the amount that means actuator given on the command throttle opening by this operation, to control the opening of the throttle control, electronic throttle the internal combustion engine to control engine output apparatus, the electronic control unit, the normal operating conditions and engine systems, engine control unit to determine the abnormality detection capabilities,
Conclusion: Quality is not good enough for understanding the content
Directorate-General 2Research & Architecture/Research Unit
Best case example: JP (BLEU=0.30)
Human Translation JP: 1. A signal processing circuit comprising:>a pulse generation part that generates a pulse signal corresponding to an input signal;>an integration part that generates an integrated voltage having a time slope proportional to an input voltage with a duration specified by said pulse signal being set as an integration period; and>a hold part that holds and outputs a difference voltage between a start voltage and an end voltage of said integrated voltage in said integration period.
JP Machine Translation:A signal-processing circuit comprising:A pulse generating means which generates a pulse signal according to an input signal.An integrating means which generates integration voltage which has a time slope which is proportional to input voltage by making into an integration period a period specified with said pulse signal.A hold means which holds and outputs difference voltage of starting potential of said integration voltage and end voltage in said integration period.
Google MT:01 and pulse generation to generate a pulse signal corresponding to the input signal, the pulsed integration time period as specified in the signal integration means for generating a voltage gradient with an integration time proportional to the input voltage, the voltage difference between voltage and hold the start voltage and end voltage of said integration of said integration period hold, and a signal processing circuit means and output.
Conclusion: Tiny differences between JP MT and HT
Directorate-General 2Research & Architecture/Research Unit
Rank-ordered N-gram co-occurrence scores
(c) NIST N-gram scoring study
6 commercial MT systems and 7 professional translators
Maximumscore for MT
Is NIST 0.4sufficient for patent professionals?
NIST scores for MT vs. Human translations
Directorate-General 2Research & Architecture/Research Unit
Manual vs. Automatic evaluation: Result Interpretations
• Scores have to be carefully interpreted: no statistical significance at the moment.
• There is a clear correlation between manual scores and automatic scores
• Both scores NIST and BLEU are complementary and show different aspects
• Relative scores should be calculted to assessment systems between each other
• Both end-users assessments AND automatic scores are necessary for testing a system
Directorate-General 2Research & Architecture/Research Unit
Cross-Language Patent Retrieval: a preliminary approach
Directorate-General 2Research & Architecture/Research Unit
The general problem
• Main strategies for query translation – dictionary-based methods
• Limitations of dictionaries• Inflected word forms• Phrases and compound words• Lexical ambiguity• Possible solution: Approximate string matching
– corpus-based methods• frequency analysis (aboutness of the 2 collections should be similar)
– machine translation• use of morphological parser• Translates source language texts into target language using:
– Translation dictionaries– Other linguistic resources– Syntax analysis
• Limited availability
Finding documents written in any language using queries expressed in a single language
Source language: the language that gives access to the required information; the query languageTarget language: the language of the content in the database
Usage: patent query translation and/or patent translation from the source language.
Directorate-General 2Research & Architecture/Research Unit
Cross-language Retrieval in a nutshell
Ling uis ticA na lys is
a ndLa ng ua g e
S uppo rt
M a c hineT ra ns la tio n
M ultiling ua lS e a rc h E ng ine
U s e rInte rfa c e
F re nc hQ ue ry
S pa nis hQ ue ry
C h in eseQ u ery
E ng lis hQ ue ry
S pa nis hD a ta ba s e
F re nc hD a ta ba s e
C hine s eD a ta ba s e
E ng lis hD a ta ba s e
M e rg ingR e s ults
E ng lis hU s e r
Q ue ry
E xpa nde dQ ue ry
Lis t o fR e s ults
R e que s tsfo r D o c um e nt
T ra ns la tio n
Mohsen Jamali, Sharif Univ. of technology
Directorate-General 2Research & Architecture/Research Unit
How to start Cross-language Information Retrieval for Patent?
Classic CLIR system tree: which strategy for patent documents?
Directorate-General 2Research & Architecture/Research Unit
The main issue of CLIR: Term disambiguation
• Solution 1:
– Selecting the most likely translation (1st one offered by a dictionary?), the longest term?
– Problem: a low probability of success.
• Solution 2:
– Use of all possible translations in the query with the OR operator.
– Problem: it includes the correct translation, but also introduces noise into the query. This can lead to the retrieval of many irrelevant documents
• Solution 3 (most popular):
– Term co-occurrences models.
– A query defines a single concept or an information need, thus the terms in a query are assumed to exhibit relatively strong relationship. Therefore, the correct translation of one query term would be expected to show a strong correlation with other translated query words.
How to deal with ambiguity?
Directorate-General 2Research & Architecture/Research Unit
A proposed measure: Mutual Information
• x, y are the translation of two query terms;• f(x), f(y), f(x,y) are the frequency that x appears, the frequency that y appears and the frequency that x and y appears together, respectively; • N is the size of the corpus
• Relationship between query terms can be quantified co-occurrences model • The Mutual Information measure quantifies the distance between the joint distribution of terms X and Y and the product of their marginal distributions
Mutual Information (MI) is a technique based on co-occurrence statistic
Directorate-General 2Research & Architecture/Research Unit
Translation selection : Total Correlation
• We have a list of translation candidates. • Goal is to find the correct translation from the candidate list. • The correct translation will be selected using MI
• xi are the translation of query words• f(xi) is the frequency that the xi appears in the corpus• f(x1,x2,x3,...) is the frequency that all query words appears in the corpus.• N is the size of the corpus
Measure:
Decision:Total correlation - a generalization of the Mutual Information to calculate the relationship between the query words is proposed:
If a set of translated query terms has a high MI value, then this set of translated termsis to be considered as correct
Directorate-General 2Research & Architecture/Research Unit
Conclusion on Term disambiguation
• MI is a simple measure and not too computer-intensive• It performs as well as other co-occurrence approaches (Maeda et al. (2000).• Co-occurrences frequencies can be obtained from the document collection.
Mutual Information associated to Total Correlation is proposed as ameasure for cross-language patent Retrieval
• Make use or build test collections to evaluate the systems:– example of CLEF (Cross Language Evaluation Forum)– collect set of queries (rare items in IP)– collect sets of relevance judgments (which documents are relevant
to which queries)
This approach is compatible with a collaborative view:
Directorate-General 2Research & Architecture/Research Unit
Visualization and analysis of Patent Queries
• Graphical and textual editing of queries• Visual support of different search
engines:– Full-text search– Semantic search– Image similarity– Metadata search
• Query management functionality:– Storing of queries– Parameterization of queries using
variables
Another solution for Term disambiguation
Checking & amending interactively when necessary increase the chance of good results
Directorate-General 2Research & Architecture/Research Unit
Perspective and conclusion
1. The number of subjects to be addressed is large (MT, IR, SE theory, Scoring and Evaluation, etc...)
2. The difficulty of retrieving patents raise theoretical problems. Testing theory need a large amount of:
• clean datasets and queries
• CPU power
• feedbacks from users communities
3. Current implementations do no satisfy entirely the users needs (usability, language independent, etc...)
4. Metrics in place need to be revisited and/or replaced by patent-specific metrics (i.e PRES/Univ. Dublin)
5. Patents not only represent technical texts, but also a set of environmental attributes which have to be consulted in order to achieve the goals (IPC classes, patent searcher behaviours, legal changes, ...)
The field of patent processing is still in a maturing mode