Tweaks for search result relevance

marginalia_nu/src/main/java/nu/marginalia/util/DenseBitMap.java

@@ -1,6 +1,10 @@
 package nu.marginalia.util;
 
+import java.io.IOException;
 import java.nio.ByteBuffer;
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.nio.file.StandardOpenOption;
 
 public class DenseBitMap {
     public static final long MAX_CAPACITY_2GB_16BN_ITEMS=(1L<<34)-8;
@@ -15,6 +19,31 @@ public class DenseBitMap {
         this.buffer = ByteBuffer.allocateDirect((int)((cardinality / 8) + (misaligned ? 1 : 0)));
     }
 
+    public static DenseBitMap loadFromFile(Path file) throws IOException {
+        long size = Files.size(file);
+        var dbm = new DenseBitMap(size/8);
+
+        try (var bc = Files.newByteChannel(file)) {
+            while (dbm.buffer.position() < dbm.buffer.capacity()) {
+                bc.read(dbm.buffer);
+            }
+        }
+        dbm.buffer.clear();
+
+        return dbm;
+    }
+
+    public void writeToFile(Path file) throws IOException {
+
+        try (var bc = Files.newByteChannel(file, StandardOpenOption.CREATE, StandardOpenOption.WRITE)) {
+            while (buffer.position() < buffer.capacity()) {
+                bc.write(buffer);
+            }
+        }
+
+        buffer.clear();
+    }
+
     public boolean get(long pos) {
         return (buffer.get((int)(pos >>> 3)) & ((byte)1 << (int)(pos & 7))) != 0;
     }

marginalia_nu/src/main/java/nu/marginalia/util/language/DocumentDebugger.java

@@ -8,7 +8,7 @@ import nu.marginalia.util.language.processing.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordRep;
 import nu.marginalia.util.language.processing.model.tag.WordSeparator;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.jsoup.nodes.Document;
 
 import java.io.FileNotFoundException;
@@ -30,7 +30,7 @@ public class DocumentDebugger {
     Path tempDir;
     public DocumentDebugger(LanguageModels lm) throws IOException {
         se = new SentenceExtractor(lm);
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
         ke = new KeywordExtractor();
 
         kc = new KeywordCounter(dict, ke);

marginalia_nu/src/main/java/nu/marginalia/util/language/conf/LanguageModels.java

@@ -6,8 +6,9 @@ import java.nio.file.Path;
 
 @AllArgsConstructor
 public class LanguageModels {
-    public final Path ngramDictionary;
+    public final Path ngramBloomFilter;
-    public final Path ngramFrequency;
+    public final Path termFrequencies;
+
     public final Path openNLPSentenceDetectionData;
     public final Path posRules;
     public final Path posDict;

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/DocumentKeywordExtractor.java

@@ -4,7 +4,7 @@ import com.google.common.collect.Sets;
 import nu.marginalia.util.language.WordPatterns;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.util.language.processing.model.WordRep;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.index.model.IndexBlock;
 import nu.marginalia.wmsa.edge.model.crawl.EdgePageWordSet;
 import nu.marginalia.wmsa.edge.model.crawl.EdgePageWords;
@@ -19,29 +19,54 @@ public class DocumentKeywordExtractor {
     private final KeywordExtractor keywordExtractor;
     private final KeywordCounter tfIdfCounter;
     private final NameCounter nameCounter;
-    private final LongNameCounter longNameCounter;
     private final SubjectCounter subjectCounter;
 
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
+    private final double docCount;
 
     @Inject
-    public DocumentKeywordExtractor(NGramDict dict) {
+    public DocumentKeywordExtractor(TermFrequencyDict dict) {
         this.dict = dict;
+        docCount = dict.docCount();
 
         keywordExtractor = new KeywordExtractor();
 
         tfIdfCounter = new KeywordCounter(dict, keywordExtractor);
         nameCounter = new NameCounter(keywordExtractor);
-        longNameCounter = new LongNameCounter(dict, keywordExtractor);
         subjectCounter = new SubjectCounter(keywordExtractor);
     }
 
+
+    public EdgePageWordSet extractKeywordsMinimal(DocumentLanguageData documentLanguageData) {
+
+        List<WordRep> titleWords = extractTitleWords(documentLanguageData);
+
+        KeywordCounter.WordHistogram wordsTfIdf = tfIdfCounter.countHisto(documentLanguageData);
+        List<WordRep> wordsNamesAll = nameCounter.count(documentLanguageData, 1);
+        List<WordRep> subjects = subjectCounter.count(documentLanguageData);
+
+        List<WordRep> midKeywords = new ArrayList<>(wordsTfIdf.mid());
+        List<WordRep> topKeywords = new ArrayList<>(wordsTfIdf.top());
+
+        Collection<String> artifacts = getArtifacts(documentLanguageData);
+
+        return new EdgePageWordSet(
+                createWords(IndexBlock.Subjects, subjects),
+                createWords(IndexBlock.Title, titleWords),
+                createWords(IndexBlock.NamesWords, wordsNamesAll),
+                createWords(IndexBlock.Tfidf_Top, topKeywords),
+                createWords(IndexBlock.Tfidf_Middle, midKeywords),
+                new EdgePageWords(IndexBlock.Artifacts, artifacts)
+        );
+    }
+
+
+
     public EdgePageWordSet extractKeywords(DocumentLanguageData documentLanguageData) {
 
         List<WordRep> titleWords = extractTitleWords(documentLanguageData);
 
         KeywordCounter.WordHistogram wordsTfIdf = tfIdfCounter.countHisto(documentLanguageData);
-        List<WordRep> wordsNamesRepeated = nameCounter.count(documentLanguageData, 2);
         List<WordRep> wordsNamesAll = nameCounter.count(documentLanguageData, 1);
         List<WordRep> subjects = subjectCounter.count(documentLanguageData);
 
@@ -49,12 +74,9 @@ public class DocumentKeywordExtractor {
         List<WordRep> midKeywords = new ArrayList<>(wordsTfIdf.mid());
         List<WordRep> topKeywords = new ArrayList<>(wordsTfIdf.top());
 
-        var wordsToMatchWithTitle = joinWordLists(topKeywords, wordsNamesRepeated, subjects);
-
         Collection<String> artifacts = getArtifacts(documentLanguageData);
 
         var wordSet = new EdgePageWordSet(
-                createWords(IndexBlock.TitleKeywords, overlappingStems(titleWords, wordsToMatchWithTitle)),
                 createWords(IndexBlock.Subjects, subjects),
                 createWords(IndexBlock.Title, titleWords),
                 createWords(IndexBlock.NamesWords, wordsNamesAll),
@@ -121,7 +143,7 @@ public class DocumentKeywordExtractor {
             else {
                 lastSet = counts.entrySet().stream()
                         .sorted(Comparator.comparing(e -> {
-                            double N = 11820118.; // Number of documents in term freq dictionary
+                            double N = docCount; // Number of documents in term freq dictionary
 
                             // Caveat: This is actually the *negated* term score, because the second logarithm has
                             // its parameter inverted (log(a^b) = b log(a); here b = -1)

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/KeywordCounter.java

@@ -2,7 +2,7 @@ package nu.marginalia.util.language.processing;
 
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.util.language.processing.model.WordRep;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 
 import java.util.HashMap;
 import java.util.HashSet;
@@ -12,11 +12,13 @@ import java.util.regex.Pattern;
 
 public class KeywordCounter {
     private final KeywordExtractor keywordExtractor;
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
+    private final double docCount;
 
-    public KeywordCounter(NGramDict dict, KeywordExtractor keywordExtractor) {
+    public KeywordCounter(TermFrequencyDict dict, KeywordExtractor keywordExtractor) {
         this.dict = dict;
         this.keywordExtractor = keywordExtractor;
+        this.docCount = (double) dict.docCount();
     }
 
     public WordHistogram countHisto(DocumentLanguageData dld) {
@@ -71,7 +73,7 @@ public class KeywordCounter {
         if (freq < 1) {
             freq = 10;
         }
-        return (0.1 + 0.9*value/maxValue) * Math.log((1.1+freq)/11820118.);
+        return (0.1 + 0.9*value/maxValue) * Math.log((1.1+freq)/docCount);
     }
 
     public record WordHistogram(Set<WordRep> lower, Set<WordRep> mid, Set<WordRep> top) { }

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/KeywordExtractor.java

@@ -1,9 +1,9 @@
 package nu.marginalia.util.language.processing;
 
+import nu.marginalia.util.language.WordPatterns;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordSpan;
 import nu.marginalia.util.language.processing.model.tag.WordSeparator;
-import nu.marginalia.util.language.WordPatterns;
 
 import java.lang.ref.SoftReference;
 import java.util.ArrayList;
@@ -377,4 +377,6 @@ public class KeywordExtractor {
 
         return (posTag.startsWith("JJ") || posTag.startsWith("R") || posTag.startsWith("VBG"));
     }
+
+
 }

marginalia_nu/src/main/java/nu/marginalia/util/language/processing/LongNameCounter.java

@@ -3,7 +3,7 @@ package nu.marginalia.util.language.processing;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordRep;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 
 import java.util.*;
 import java.util.regex.Pattern;
@@ -11,10 +11,11 @@ import java.util.stream.Collectors;
 
 public class LongNameCounter {
     private final KeywordExtractor keywordExtractor;
-
+    private final TermFrequencyDict dict;
-    private final NGramDict dict;
+    private final double docCount;
-    public LongNameCounter(NGramDict dict, KeywordExtractor keywordExtractor) {
+    public LongNameCounter(TermFrequencyDict dict, KeywordExtractor keywordExtractor) {
         this.dict = dict;
+        docCount = (double) dict.docCount();
         this.keywordExtractor = keywordExtractor;
     }
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/configuration/WmsaHome.java

@@ -87,7 +87,7 @@ public class WmsaHome {
         final Path home = getHomePath();
 
         return new LanguageModels(
-                home.resolve("model/ngrams-generous-emstr.bin"),
+                home.resolve("model/ngrams.bin"),
                 home.resolve("model/tfreq-new-algo3.bin"),
                 home.resolve("model/opennlp-sentence.bin"),
                 home.resolve("model/English.RDR"),

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/assistant/dict/NGramBloomFilter.java

@@ -0,0 +1,78 @@
+package nu.marginalia.wmsa.edge.assistant.dict;
+
+import ca.rmen.porterstemmer.PorterStemmer;
+import com.google.common.hash.HashFunction;
+import com.google.common.hash.Hashing;
+import com.google.inject.Inject;
+import nu.marginalia.util.DenseBitMap;
+import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.configuration.WmsaHome;
+import nu.marginalia.wmsa.edge.index.lexicon.journal.KeywordLexiconJournalFile;
+
+import java.io.File;
+import java.io.IOException;
+import java.nio.charset.StandardCharsets;
+import java.nio.file.Path;
+import java.util.concurrent.atomic.AtomicInteger;
+import java.util.regex.Pattern;
+
+public class NGramBloomFilter {
+    private final DenseBitMap bitMap;
+    private static final PorterStemmer ps = new PorterStemmer();
+    private static final HashFunction hasher = Hashing.murmur3_128(0);
+
+    @Inject
+    public NGramBloomFilter() throws IOException {
+        this(WmsaHome.getLanguageModels());
+    }
+
+    public NGramBloomFilter(LanguageModels lm) throws IOException {
+        this(DenseBitMap.loadFromFile(lm.ngramBloomFilter));
+    }
+
+    public NGramBloomFilter(DenseBitMap bitMap) {
+        this.bitMap = bitMap;
+    }
+
+    public boolean isKnownNGram(String word) {
+        long bit = bitForWord(word, bitMap.cardinality);
+
+        return bitMap.get(bit);
+    }
+
+    public static void main(String... args) throws IOException {
+        var filter = convertFromDictionaryFile(new File(args[0]));
+        filter.bitMap.writeToFile(Path.of(args[1]));
+    }
+
+    public static NGramBloomFilter load(Path file) throws IOException {
+        return new NGramBloomFilter(DenseBitMap.loadFromFile(file));
+    }
+
+    public static NGramBloomFilter convertFromDictionaryFile(File file) throws IOException {
+        DenseBitMap bitMap = new DenseBitMap(1024*1024*1024L);
+        AtomicInteger popCount = new AtomicInteger();
+        try (var f = new KeywordLexiconJournalFile(file)) {
+            f.loadFile(data -> {
+                long bit = bitForWord(new String(data), bitMap.cardinality);
+                if (!bitMap.set(bit))
+                    popCount.incrementAndGet();
+            });
+        }
+
+        System.out.println("popcount = " + popCount.get());
+        return new NGramBloomFilter(bitMap);
+    }
+
+    private static final Pattern underscore = Pattern.compile("_");
+
+    private static long bitForWord(String s, long n) {
+        String[] parts = underscore.split(s);
+        long hc = 0;
+        for (String part : parts) {
+            hc = hc * 31 + hasher.hashString(ps.stemWord(part), StandardCharsets.UTF_8).padToLong();
+        }
+        return (hc & 0x7FFF_FFFF_FFFF_FFFFL) % n;
+    }
+
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/assistant/dict/TermFrequencyDict.java

@@ -9,7 +9,6 @@ import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.wmsa.configuration.WmsaHome;
 import nu.marginalia.wmsa.edge.converting.processor.logic.DomPruner;
 import nu.marginalia.wmsa.edge.crawling.CrawlPlanLoader;
-import opennlp.tools.langdetect.LanguageDetector;
 import org.jsoup.Jsoup;
 import org.jsoup.nodes.Document;
 import org.slf4j.Logger;
@@ -21,12 +20,17 @@ import javax.inject.Singleton;
 import java.io.*;
 import java.nio.file.Files;
 import java.nio.file.Path;
-import java.util.*;
+import java.util.Arrays;
+import java.util.HashSet;
+import java.util.Set;
+import java.util.concurrent.ForkJoinPool;
+import java.util.concurrent.TimeUnit;
+import java.util.concurrent.atomic.AtomicInteger;
 import java.util.regex.Pattern;
 import java.util.stream.Collectors;
 
 @Singleton
-public class NGramDict {
+public class TermFrequencyDict {
 
     private final TLongIntHashMap wordRates = new TLongIntHashMap(1_000_000, 0.5f, 0, 0);
 
@@ -34,21 +38,22 @@ public class NGramDict {
     private static final Pattern separator = Pattern.compile("[_ ]+");
     private static final PorterStemmer ps = new PorterStemmer();
 
+    private static final long DOC_COUNT_KEY = ~0L;
     private static long fileSize(Path p) throws IOException {
         return Files.size(p);
     }
 
     @Inject
-    public NGramDict(@Nullable LanguageModels models) {
+    public TermFrequencyDict(@Nullable LanguageModels models) {
         if (models == null) {
             return;
         }
 
-        if (models.ngramFrequency != null) {
+        if (models.termFrequencies != null) {
 
-            try (var frequencyData = new DataInputStream(new BufferedInputStream(new FileInputStream(models.ngramFrequency.toFile())))) {
+            try (var frequencyData = new DataInputStream(new BufferedInputStream(new FileInputStream(models.termFrequencies.toFile())))) {
 
-                wordRates.ensureCapacity((int)(fileSize(models.ngramFrequency)/16));
+                wordRates.ensureCapacity((int)(fileSize(models.termFrequencies)/16));
 
                 for (;;) {
                     wordRates.put(frequencyData.readLong(), (int) frequencyData.readLong());
@@ -56,7 +61,7 @@ public class NGramDict {
             } catch (EOFException eof) {
                 // ok
             } catch (IOException e) {
-                logger.error("IO Exception reading " + models.ngramFrequency, e);
+                logger.error("IO Exception reading " + models.termFrequencies, e);
             }
         }
 
@@ -64,60 +69,100 @@ public class NGramDict {
     }
 
 
-    public static void main(String... args) throws IOException {
+    public int docCount() {
+        int cnt = wordRates.get(DOC_COUNT_KEY);
+
+        if (cnt == 0) {
+            cnt = 11820118; // legacy
+        }
+        return cnt;
+    }
+
+    public static void main(String... args) throws IOException, InterruptedException {
         if (args.length != 2) {
             System.err.println("Expected arguments: plan.yaml out-file");
         }
-        String inFile = args[0];
         String outFile = args[1];
 
         var plan = new CrawlPlanLoader().load(Path.of(args[0]));
 
-        SentenceExtractor se = new SentenceExtractor(WmsaHome.getLanguageModels());
+        ThreadLocal<SentenceExtractor> se = ThreadLocal.withInitial(() -> new SentenceExtractor(WmsaHome.getLanguageModels()));
         DomPruner pruner = new DomPruner();
         LanguageFilter lf = new LanguageFilter();
 
-        Map<String, Integer> counts = new HashMap<>(100_000_000);
+        TLongIntHashMap counts = new TLongIntHashMap(100_000_000, 0.7f, -1, -1);
-        Set<String> words = new HashSet<>(10_000);
+
+        ForkJoinPool fjp = new ForkJoinPool(24);
+        AtomicInteger docCount = new AtomicInteger();
 
         for (var domain : plan.domainsIterable()) { // leaks file descriptor, is fine
 
             if (domain.doc == null)
                 continue;
 
-            for (var doc : domain.doc) {
+            fjp.execute(() -> {
-                if (doc.documentBody == null)
-                    continue;
 
-                Document parsed = Jsoup.parse(doc.documentBody);
+                for (var doc : domain.doc) {
-                pruner.prune(parsed, 0.5);
+                    if (doc.documentBody == null)
+                        continue;
+                    docCount.incrementAndGet();
 
-                DocumentLanguageData dld = se.extractSentences(parsed);
+                    Document parsed = Jsoup.parse(doc.documentBody);
+                    pruner.prune(parsed, 0.5);
 
-                if (lf.dictionaryAgreement(dld) < 0.1) {
+                    DocumentLanguageData dld = se.get().extractSentences(parsed);
-                    continue;
-                }
 
-
+                    if (lf.dictionaryAgreement(dld) < 0.1) {
-                for (var sent : dld.sentences) {
+                        return;
-                    for (var word : sent) {
-                        words.add(word.stemmed());
                     }
-                }
 
-                for (var word : words) {
+                    Set<String> words = new HashSet<>(10_000);
-                    counts.merge(word,  1, Integer::sum);
-                }
 
-                words.clear();
+                    for (var sent : dld.sentences) {
+                        for (var word : sent) {
+                            words.add(word.stemmed());
+                        }
+                    }
+
+                    fjp.execute(() -> {
+                        synchronized (counts) {
+                            for (var word : words) {
+                                counts.adjustOrPutValue(longHash(word.getBytes()),  1, 1);
+                            }
+                        }
+                    });
+
+                }
+            });
+        }
+
+        fjp.shutdown();
+        fjp.awaitTermination(10, TimeUnit.SECONDS);
+
+        try (var dos = new DataOutputStream(Files.newOutputStream(Path.of(outFile)))) {
+            synchronized (counts) {
+                counts.put(DOC_COUNT_KEY, docCount.get());
+
+                counts.forEachEntry((hash, cnt) -> {
+                    try {
+                        dos.writeLong(hash);
+                        dos.writeLong(cnt);
+                    } catch (IOException e) {
+                        throw new RuntimeException(e);
+                    }
+                    return true;
+                });
             }
         }
 
-        counts.forEach((w,c) -> {
+        System.out.println(docCount.get());
-            if (c > 3) {
+//
-                System.out.println(w + ":" + c);
+//        counts.forEachEntry((w,c) -> {
-            }
+//            if (c > 3L) {
-        });
+//                System.out.println(w + ":" + c);
+//            }
+//            return true;
+//        });
 
     }
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/assistant/suggest/Suggestions.java

@@ -2,8 +2,8 @@ package nu.marginalia.wmsa.edge.assistant.suggest;
 
 import com.google.inject.Inject;
 import com.google.inject.name.Named;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.wmsa.edge.assistant.dict.SpellChecker;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.converting.processor.logic.HtmlFeature;
 import org.apache.commons.collections4.trie.PatriciaTrie;
 import org.slf4j.Logger;
@@ -21,7 +21,7 @@ import java.util.stream.Stream;
 
 public class Suggestions {
     private final PatriciaTrie<String> suggestionsTrie;
-    private final NGramDict nGramDict;
+    private final TermFrequencyDict termFrequencyDict;
     private final SpellChecker spellChecker;
 
     private static final Pattern suggestionPattern = Pattern.compile("^[a-zA-Z0-9]+( [a-zA-Z0-9]+)*$");
@@ -31,12 +31,12 @@ public class Suggestions {
     @Inject
     public Suggestions(@Named("suggestions-file") Path suggestionsFile,
                        SpellChecker spellChecker,
-                       NGramDict dict
+                       TermFrequencyDict dict
                        ) {
         this.spellChecker = spellChecker;
 
         suggestionsTrie = loadSuggestions(suggestionsFile);
-        nGramDict = dict;
+        termFrequencyDict = dict;
 
         logger.info("Loaded {} suggestions", suggestionsTrie.size());
     }
@@ -138,7 +138,7 @@ public class Suggestions {
         }
 
         Map<String, Long> scach = new HashMap<>(512);
-        Function<String, Long> valr = s -> -nGramDict.getTermFreqHash(scach.computeIfAbsent(s, NGramDict::getStringHash));
+        Function<String, Long> valr = s -> -termFrequencyDict.getTermFreqHash(scach.computeIfAbsent(s, TermFrequencyDict::getStringHash));
 
         return Stream.iterate(start.getKey(), Objects::nonNull, suggestionsTrie::nextKey)
                 .takeWhile(s -> s.startsWith(prefix))

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/atags/AnchorTextExtractor.java

@@ -6,7 +6,7 @@ import lombok.SneakyThrows;
 import nu.marginalia.util.DenseBitMap;
 import nu.marginalia.util.language.WordPatterns;
 import nu.marginalia.wmsa.configuration.WmsaHome;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.converting.processor.logic.LinkParser;
 import nu.marginalia.wmsa.edge.model.EdgeUrl;
 import org.apache.logging.log4j.util.Strings;
@@ -36,7 +36,7 @@ public class AnchorTextExtractor {
     // de-duplicating billions of shuffled (url, word) tuples on limited hardware
     private final DenseBitMap deduplicateHashBitset = new DenseBitMap(DenseBitMap.MAX_CAPACITY_2GB_16BN_ITEMS);
 
-    private final NGramDict ngramDict = new NGramDict(WmsaHome.getLanguageModels());
+    private final TermFrequencyDict ngramDict = new TermFrequencyDict(WmsaHome.getLanguageModels());
 
     public AnchorTextExtractor(Predicate<String> includeDomainPredicate,
                                Predicate<EdgeUrl> includeUrlPredicate,

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/model/DisqualifiedException.java

@@ -17,6 +17,7 @@ public class DisqualifiedException extends Exception {
         LANGUAGE,
         STATUS,
         QUALITY,
-        ACCEPTABLE_ADS
+        ACCEPTABLE_ADS,
+        FORBIDDEN
     }
 }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/DocumentProcessor.java

@@ -86,10 +86,6 @@ public class DocumentProcessor {
                 if (isAcceptedContentType(crawledDocument)) {
                     var detailsWords = createDetails(crawledDomain, crawledDocument);
 
-                    if (detailsWords.details().quality < minDocumentQuality) {
-                        throw new DisqualifiedException(DisqualificationReason.QUALITY);
-                    }
-
                     ret.details = detailsWords.details();
                     ret.words = detailsWords.words();
                 }
@@ -141,11 +137,14 @@ public class DocumentProcessor {
     private DetailsWithWords createDetails(CrawledDomain crawledDomain, CrawledDocument crawledDocument)
             throws DisqualifiedException, URISyntaxException {
 
-        var doc = Jsoup.parse(crawledDocument.documentBody);
+        Document doc = Jsoup.parse(crawledDocument.documentBody);
 
         if (AcceptableAds.hasAcceptableAdsTag(doc)) {
             throw new DisqualifiedException(DisqualificationReason.ACCEPTABLE_ADS);
         }
+        if (doc.select("meta[name=robots]").attr("content").contains("noindex")) {
+            throw new DisqualifiedException(DisqualificationReason.FORBIDDEN);
+        }
 
         DomPruner domPruner = new DomPruner();
         Document prunedDoc = doc.clone();
@@ -160,11 +159,17 @@ public class DocumentProcessor {
         ret.length = getLength(doc);
         ret.standard = getHtmlStandard(doc);
         ret.title = titleExtractor.getTitleAbbreviated(doc, dld, crawledDocument.url);
-        ret.features = featureExtractor.getFeatures(crawledDomain, doc);
+        ret.features = featureExtractor.getFeatures(crawledDomain, doc, dld);
         ret.quality = documentValuator.getQuality(ret.standard, doc, dld);
         ret.hashCode = HashCode.fromString(crawledDocument.documentBodyHash).asLong();
 
-        var words = getWords(dld);
+        EdgePageWordSet words;
+        if (ret.quality < minDocumentQuality || dld.totalNumWords() < minDocumentLength) {
+            words = keywordExtractor.extractKeywordsMinimal(dld);
+        }
+        else {
+            words = keywordExtractor.extractKeywords(dld);
+        }
 
         var url = new EdgeUrl(crawledDocument.url);
         addMetaWords(ret, url, crawledDomain, words);
@@ -195,7 +200,6 @@ public class DocumentProcessor {
         ret.features.stream().map(HtmlFeature::getKeyword).forEach(tagWords::add);
 
         words.append(IndexBlock.Meta, tagWords);
-        words.append(IndexBlock.Words_1, tagWords);
     }
 
     private void getLinks(EdgeUrl baseUrl, ProcessedDocumentDetails ret, Document doc, EdgePageWordSet words) {
@@ -255,7 +259,6 @@ public class DocumentProcessor {
 
     private void synthesizeFilenameKeyword(Set<String> fileKeywords, EdgeUrl link) {
 
-
         Path pFilename = Path.of(link.path.toLowerCase()).getFileName();
 
         if (pFilename == null) return;
@@ -273,10 +276,6 @@ public class DocumentProcessor {
     }
 
     private void checkDocumentLanguage(DocumentLanguageData dld) throws DisqualifiedException {
-        if (dld.totalNumWords() < minDocumentLength) {
-            throw new DisqualifiedException(DisqualificationReason.LENGTH);
-        }
-
         double languageAgreement = languageFilter.dictionaryAgreement(dld);
         if (languageAgreement < 0.1) {
             throw new DisqualifiedException(DisqualificationReason.LANGUAGE);
@@ -292,10 +291,6 @@ public class DocumentProcessor {
         return htmlStandard;
     }
 
-    private EdgePageWordSet getWords(DocumentLanguageData dld) {
-        return keywordExtractor.extractKeywords(dld);
-    }
-
     private String getDescription(Document doc) {
         return summaryExtractor.extractSummary(doc);
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/DomainProcessor.java

@@ -6,13 +6,13 @@ import nu.marginalia.wmsa.edge.converting.model.ProcessedDocument;
 import nu.marginalia.wmsa.edge.converting.model.ProcessedDomain;
 import nu.marginalia.wmsa.edge.crawling.model.CrawledDomain;
 import nu.marginalia.wmsa.edge.crawling.model.CrawlerDomainStatus;
+import nu.marginalia.wmsa.edge.index.model.IndexBlock;
 import nu.marginalia.wmsa.edge.model.EdgeDomain;
 import nu.marginalia.wmsa.edge.model.crawl.EdgeDomainIndexingState;
 import nu.marginalia.wmsa.edge.model.crawl.EdgeUrlState;
 
-import java.util.ArrayList;
+import java.util.*;
-import java.util.Collections;
+import java.util.stream.Collectors;
-import java.util.List;
 
 public class DomainProcessor {
     private final DocumentProcessor documentProcessor;
@@ -45,6 +45,8 @@ public class DomainProcessor {
                     ret.documents.add(processedDoc);
                 }
             }
+
+            addCommonSiteWords(ret);
         }
         else {
             ret.documents = Collections.emptyList();
@@ -60,6 +62,40 @@ public class DomainProcessor {
         return ret;
     }
 
+    private void addCommonSiteWords(ProcessedDomain ret) {
+
+        if (ret.documents.size() < 25)
+            return;
+
+        Map<String, Integer> topKeywordCount = new HashMap<>(ret.documents.size()*10);
+
+        for (var doc : ret.documents) {
+            if (doc.words == null)
+                continue;
+
+            for (var word : doc.words.get(IndexBlock.Tfidf_Top).words) {
+                topKeywordCount.merge(word, -1, Integer::sum);
+            }
+        }
+
+        if (topKeywordCount.values().stream().mapToInt(i -> i).sum() > -100)
+            return;
+
+        Set<String> topWords = topKeywordCount.entrySet().stream()
+                .filter(e -> e.getValue() < -10)
+                .sorted(Map.Entry.comparingByValue()).limit(5)
+                .map(Map.Entry::getKey)
+                .collect(Collectors.toSet());
+
+        if (!topWords.isEmpty()) {
+            for (var doc : ret.documents) {
+                if (doc.words != null) {
+                    doc.words.get(IndexBlock.Site).addAll(topWords);
+                }
+            }
+        }
+    }
+
     private double getAverageQuality(List<ProcessedDocument> documents) {
         int n = 0;
         double q = 0.;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/DomPruner.java

@@ -4,7 +4,7 @@ import org.jsoup.nodes.Document;
 import org.jsoup.nodes.Element;
 import org.jsoup.nodes.Node;
 import org.jsoup.nodes.TextNode;
-import org.jsoup.select.NodeVisitor;
+import org.jsoup.select.NodeFilter;
 
 import java.util.HashMap;
 import java.util.Map;
@@ -12,100 +12,103 @@ import java.util.Map;
 public class DomPruner {
 
     public void prune(Document document, double pruneThreshold) {
-        PruningVisitor pruningVisitor = new PruningVisitor();
+        document.filter(new PruningFilter(pruneThreshold));
-        document.traverse(pruningVisitor);
-
-        pruningVisitor.data.forEach((node, data) -> {
-            if (data.depth <= 1) {
-                return;
-            }
-            if (data.signalNodeSize == 0) node.remove();
-            else if (data.noiseNodeSize > 0
-                    && data.signalRate() < pruneThreshold
-                    && data.treeSize > 3) {
-                node.remove();
-            }
-        });
     }
 
+}
 
 
-    private static class PruningVisitor implements NodeVisitor {
+class PruningFilter implements NodeFilter {
 
-        private final Map<Node, NodeData> data = new HashMap<>();
+    private final Map<Node, NodeData> data = new HashMap<>();
-        private final NodeData dummy = new NodeData(Integer.MAX_VALUE, 1, 0);
+    private final NodeData dummy = new NodeData(Integer.MAX_VALUE, 1, 0);
+    private double pruneThreshold;
 
-        @Override
+    public PruningFilter(double pruneThreshold) {
-        public void head(Node node, int depth) {}
+        this.pruneThreshold = pruneThreshold;
-
-        @Override
-        public void tail(Node node, int depth) {
-            final NodeData dataForNode;
-
-            if (node instanceof TextNode tn) {
-                dataForNode = new NodeData(depth, tn.text().length(), 0);
-            }
-            else if (isSignal(node)) {
-                dataForNode = new NodeData(depth,  0,0);
-                for (var childNode : node.childNodes()) {
-                    dataForNode.add(data.getOrDefault(childNode, dummy));
-                }
-            }
-            else {
-                dataForNode = new NodeData(depth,  0,0);
-                for (var childNode : node.childNodes()) {
-                    dataForNode.addAsNoise(data.getOrDefault(childNode, dummy));
-                }
-            }
-
-
-
-            data.put(node, dataForNode);
-        }
-
-        public boolean isSignal(Node node) {
-
-            if (node instanceof Element e) {
-                if ("a".equalsIgnoreCase(e.tagName()))
-                    return false;
-                if ("nav".equalsIgnoreCase(e.tagName()))
-                    return false;
-                if ("footer".equalsIgnoreCase(e.tagName()))
-                    return false;
-                if ("header".equalsIgnoreCase(e.tagName()))
-                    return false;
-            }
-
-            return true;
-        }
     }
 
-    private static class NodeData {
+    @Override
-        int signalNodeSize;
+    public FilterResult head(Node node, int depth) {
-        int noiseNodeSize;
+        return FilterResult.CONTINUE;
-        int treeSize = 1;
+    }
-        int depth;
 
-        private NodeData(int depth, int signalNodeSize, int noiseNodeSize) {
+    @Override
-            this.depth = depth;
+    public FilterResult tail(Node node, int depth) {
-            this.signalNodeSize = signalNodeSize;
+        final NodeData dataForNode;
-            this.noiseNodeSize = noiseNodeSize;
+
+        if (node instanceof TextNode tn) {
+            dataForNode = new NodeData(depth, tn.text().length(), 0);
+        }
+        else if (isSignal(node)) {
+            dataForNode = new NodeData(depth,  0,0);
+            for (var childNode : node.childNodes()) {
+                dataForNode.add(data.getOrDefault(childNode, dummy));
+            }
+        }
+        else {
+            dataForNode = new NodeData(depth,  0,0);
+            for (var childNode : node.childNodes()) {
+                dataForNode.addAsNoise(data.getOrDefault(childNode, dummy));
+            }
         }
 
-        public void add(NodeData other) {
+        data.put(node, dataForNode);
-            signalNodeSize += other.signalNodeSize;
+
-            noiseNodeSize += other.noiseNodeSize;
+        if (dataForNode.depth <= 1)
-            treeSize += other.treeSize;
+            return FilterResult.CONTINUE;
+
+        if (dataForNode.signalNodeSize == 0)
+            return FilterResult.REMOVE;
+        if (dataForNode.noiseNodeSize > 0
+                && dataForNode.signalRate() < pruneThreshold
+                && dataForNode.treeSize > 3)
+            return FilterResult.REMOVE;
+
+        return FilterResult.CONTINUE;
+    }
+
+    public boolean isSignal(Node node) {
+
+        if (node instanceof Element e) {
+            if ("a".equalsIgnoreCase(e.tagName()))
+                return false;
+            if ("nav".equalsIgnoreCase(e.tagName()))
+                return false;
+            if ("footer".equalsIgnoreCase(e.tagName()))
+                return false;
+            if ("header".equalsIgnoreCase(e.tagName()))
+                return false;
         }
 
-        public void addAsNoise(NodeData other) {
+        return true;
-            noiseNodeSize += other.noiseNodeSize + other.signalNodeSize;
-            treeSize += other.treeSize;
-        }
-
-
-        public double signalRate() {
-            return signalNodeSize / (double)(signalNodeSize + noiseNodeSize);
-        }
     }
 }
+
+class NodeData {
+    int signalNodeSize;
+    int noiseNodeSize;
+    int treeSize = 1;
+    int depth;
+
+    NodeData(int depth, int signalNodeSize, int noiseNodeSize) {
+        this.depth = depth;
+        this.signalNodeSize = signalNodeSize;
+        this.noiseNodeSize = noiseNodeSize;
+    }
+
+    public void add(NodeData other) {
+        signalNodeSize += other.signalNodeSize;
+        noiseNodeSize += other.noiseNodeSize;
+        treeSize += other.treeSize;
+    }
+
+    public void addAsNoise(NodeData other) {
+        noiseNodeSize += other.noiseNodeSize + other.signalNodeSize;
+        treeSize += other.treeSize;
+    }
+
+    public double signalRate() {
+        return signalNodeSize / (double)(signalNodeSize + noiseNodeSize);
+    }
+}

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/FeatureExtractor.java

@@ -2,7 +2,11 @@ package nu.marginalia.wmsa.edge.converting.processor.logic;
 
 import com.google.inject.Inject;
 import com.google.inject.Singleton;
+import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 import nu.marginalia.wmsa.edge.converting.processor.logic.topic.AdblockSimulator;
+import nu.marginalia.wmsa.edge.converting.processor.logic.topic.RecipeDetector;
+import nu.marginalia.wmsa.edge.converting.processor.logic.topic.TextileCraftDetector;
+import nu.marginalia.wmsa.edge.converting.processor.logic.topic.WoodworkingDetector;
 import nu.marginalia.wmsa.edge.crawling.model.CrawledDomain;
 import org.jsoup.nodes.Document;
 import org.jsoup.nodes.Element;
@@ -35,14 +39,20 @@ public class FeatureExtractor {
             "d31qbv1cthcecs.cloudfront.net",
             "linkedin.com");
 
-    private AdblockSimulator adblockSimulator;
+    private final AdblockSimulator adblockSimulator;
+    private final RecipeDetector recipeDetector;
+    private final TextileCraftDetector textileCraftDetector;
+    private final WoodworkingDetector woodworkingDetector;
 
     @Inject
-    public FeatureExtractor(AdblockSimulator adblockSimulator) {
+    public FeatureExtractor(AdblockSimulator adblockSimulator, RecipeDetector recipeDetector, TextileCraftDetector textileCraftDetector, WoodworkingDetector woodworkingDetector) {
         this.adblockSimulator = adblockSimulator;
+        this.recipeDetector = recipeDetector;
+        this.textileCraftDetector = textileCraftDetector;
+        this.woodworkingDetector = woodworkingDetector;
     }
 
-    public Set<HtmlFeature> getFeatures(CrawledDomain domain, Document doc) {
+    public Set<HtmlFeature> getFeatures(CrawledDomain domain, Document doc, DocumentLanguageData dld) {
         final Set<HtmlFeature> features = new HashSet<>();
 
         final Elements scriptTags = doc.getElementsByTag("script");
@@ -81,9 +91,14 @@ public class FeatureExtractor {
             }
         }
 
-        if (!domain.cookies.isEmpty()) {
+        if (!domain.cookies.isEmpty())
             features.add(HtmlFeature.COOKIES);
-        }
+
+        if (recipeDetector.testP(dld) > 0.5)
+            features.add(HtmlFeature.CATEGORY_FOOD);
+        // these should be mutually exclusive
+        else if (woodworkingDetector.testP(dld) > 0.3 || textileCraftDetector.testP(dld) > 0.3)
+            features.add(HtmlFeature.CATEGORY_CRAFTS);
 
         return features;
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/HtmlFeature.java

@@ -12,6 +12,8 @@ public enum HtmlFeature {
     CATEGORY_FOOD("category:food"),
 
     ADVERTISEMENT("special:ads"),
+
+    CATEGORY_CRAFTS("category:crafts"),
     ;
 
     private final String keyword;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/topic/RecipeDetector.java

@@ -1,6 +1,7 @@
 package nu.marginalia.wmsa.edge.converting.processor.logic.topic;
 
 import ca.rmen.porterstemmer.PorterStemmer;
+import com.google.inject.Inject;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 
 import java.util.HashMap;
@@ -14,6 +15,7 @@ public class RecipeDetector {
 
     private final Map<String, Double> termValues = new HashMap<>();
 
+    @Inject
     public RecipeDetector() {
         PorterStemmer ps = new PorterStemmer();
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/topic/TextileCraftDetector.java

@@ -1,6 +1,7 @@
 package nu.marginalia.wmsa.edge.converting.processor.logic.topic;
 
 import ca.rmen.porterstemmer.PorterStemmer;
+import com.google.inject.Inject;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 
 import java.util.HashMap;
@@ -14,6 +15,7 @@ public class TextileCraftDetector {
 
     private final Map<String, Double> termValues = new HashMap<>();
 
+    @Inject
     public TextileCraftDetector() {
         PorterStemmer ps = new PorterStemmer();
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/converting/processor/logic/topic/WoodworkingDetector.java

@@ -1,6 +1,7 @@
 package nu.marginalia.wmsa.edge.converting.processor.logic.topic;
 
 import ca.rmen.porterstemmer.PorterStemmer;
+import com.google.inject.Inject;
 import nu.marginalia.util.language.processing.model.DocumentLanguageData;
 
 import java.util.HashMap;
@@ -14,6 +15,7 @@ public class WoodworkingDetector {
 
     private final Map<String, Double> termValues = new HashMap<>();
 
+    @Inject
     public WoodworkingDetector() {
         PorterStemmer ps = new PorterStemmer();
 

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/lexicon/journal/KeywordLexiconJournalFile.java

@@ -15,7 +15,7 @@ import java.util.concurrent.locks.ReadWriteLock;
 import java.util.concurrent.locks.ReentrantReadWriteLock;
 import java.util.function.Consumer;
 
-public class KeywordLexiconJournalFile {
+public class KeywordLexiconJournalFile implements AutoCloseable {
     private final RandomAccessFile journalFileRAF;
     private final File journalFile;
     private final Logger logger = LoggerFactory.getLogger(getClass());

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/model/IndexBlock.java

@@ -4,22 +4,24 @@ public enum IndexBlock {
     TitleKeywords(0, 0),
     Title(1, 1),
 
-    Link(2, 1.25),
+    Link(2, 1.15),
 
-    Subjects(3, 0.5),
+    Subjects(3, 3.0),
-    NamesWords(4, 5),
+    NamesWords(4, 3.0),
     Artifacts(5, 10),
     Meta(6, 7),
 
-    Tfidf_Top(7, 2),
+    Tfidf_Top(7, 0.5),
-    Tfidf_Middle(8, 2.5),
+    Tfidf_Middle(8, 1.25),
-    Tfidf_Lower(9, 5.0),
+    Tfidf_Lower(9, 1.5),
 
     Words_1(10, 3.0),
     Words_2(11, 3.5),
     Words_4(12, 4.0),
     Words_8(13, 4.5),
     Words_16Plus(14, 7.0),
+
+    Site(15, 1.2),
     ;
 
     public final int id;

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/index/reader/SearchIndexReader.java

@@ -29,7 +29,6 @@ public class SearchIndexReader implements AutoCloseable {
             IndexBlock.Tfidf_Top,
             IndexBlock.Tfidf_Middle,
             IndexBlock.Tfidf_Lower,
-            IndexBlock.NamesWords,
             IndexBlock.Words_1,
             IndexBlock.Words_2,
             IndexBlock.Words_4,
@@ -62,15 +61,14 @@ public class SearchIndexReader implements AutoCloseable {
         queryBuilders = new EnumMap<>(IndexBlock.class);
         underspecifiedQueryBuilders = new EnumMap<>(IndexBlock.class);
 
-        queryBuilders.put(IndexBlock.Words_1, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleKeywordsIndex, titleIndex, topIndex, words1), words1));
+        queryBuilders.put(IndexBlock.Words_1, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleIndex, topIndex, words1), words1));
-        queryBuilders.put(IndexBlock.Words_2, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleKeywordsIndex, titleIndex, topIndex, words2), words1));
+        queryBuilders.put(IndexBlock.Words_2, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleIndex, topIndex, words2), words1));
-        queryBuilders.put(IndexBlock.Words_4, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleKeywordsIndex, titleIndex, topIndex, words4), words1));
+        queryBuilders.put(IndexBlock.Words_4, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleIndex, topIndex, words4), words1));
-        queryBuilders.put(IndexBlock.Words_8, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleKeywordsIndex, titleIndex, topIndex, words8), words1));
+        queryBuilders.put(IndexBlock.Words_8, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleIndex, topIndex, words8), words1));
-        queryBuilders.put(IndexBlock.Words_16Plus, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleKeywordsIndex, titleIndex, topIndex, words1, words2, words4, words8, words16, artifacts), words1));
+        queryBuilders.put(IndexBlock.Words_16Plus, new IndexQueryBuilder(listOfNonNulls(metaIndex, titleIndex, topIndex, words1, words2, words4, words8, words16, artifacts), words1));
 
-        underspecifiedQueryBuilders.put(IndexBlock.TitleKeywords, new IndexQueryBuilder(listOfNonNulls(titleKeywordsIndex, linkIndex, topicIndex, topIndex, midIndex, lowIndex, namesIndex, metaIndex), words1));
+        underspecifiedQueryBuilders.put(IndexBlock.Title, new IndexQueryBuilder(listOfNonNulls(titleIndex, linkIndex, topIndex, namesIndex, topicIndex, metaIndex), words1));
-        underspecifiedQueryBuilders.put(IndexBlock.Title, new IndexQueryBuilder(listOfNonNulls(titleIndex, topicIndex, linkIndex, topicIndex, topIndex, midIndex, lowIndex, namesIndex, metaIndex), words1));
+        underspecifiedQueryBuilders.put(IndexBlock.Tfidf_Top, new IndexQueryBuilder(listOfNonNulls(linkIndex, namesIndex, topIndex, midIndex, lowIndex, topicIndex, metaIndex), words1));
-        underspecifiedQueryBuilders.put(IndexBlock.Tfidf_Top, new IndexQueryBuilder(listOfNonNulls(topIndex, linkIndex, midIndex, lowIndex, namesIndex, metaIndex), words1));
     }
 
     @SafeVarargs

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/model/crawl/EdgePageWordSet.java

@@ -19,7 +19,8 @@ public class EdgePageWordSet {
     public EdgePageWords get(IndexBlock block) {
         var words = wordSets.get(block);
         if (words == null) {
-            return new EdgePageWords(block);
+            words = new EdgePageWords(block);
+            wordSets.put(block, words);
         }
         return words;
     }

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/EnglishDictionary.java

@@ -1,7 +1,7 @@
 package nu.marginalia.wmsa.edge.search.query;
 
 import com.google.inject.Inject;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
 
@@ -13,11 +13,11 @@ import java.util.stream.Collectors;
 
 public class EnglishDictionary {
     private final Set<String> englishWords = new HashSet<>();
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
     @Inject
-    public EnglishDictionary(NGramDict dict) {
+    public EnglishDictionary(TermFrequencyDict dict) {
         this.dict = dict;
         try (var resource = Objects.requireNonNull(ClassLoader.getSystemResourceAsStream("dictionary/en-words"),
                 "Could not load word frequency table");

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/QueryFactory.java

@@ -4,7 +4,8 @@ import com.google.inject.Inject;
 import com.google.inject.Singleton;
 import nu.marginalia.util.language.WordPatterns;
 import nu.marginalia.util.language.conf.LanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.index.model.IndexBlock;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchSpecification;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchSubquery;
@@ -22,20 +23,22 @@ import java.util.*;
 public class QueryFactory {
 
     private final LanguageModels lm;
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
     private final EnglishDictionary englishDictionary;
+    private final NGramBloomFilter nGramBloomFilter;
     private final Logger logger = LoggerFactory.getLogger(getClass());
 
     @Inject
-    public QueryFactory(LanguageModels lm, NGramDict dict, EnglishDictionary englishDictionary) {
+    public QueryFactory(LanguageModels lm, TermFrequencyDict dict, EnglishDictionary englishDictionary, NGramBloomFilter nGramBloomFilter) {
         this.lm = lm;
         this.dict = dict;
 
         this.englishDictionary = englishDictionary;
+        this.nGramBloomFilter = nGramBloomFilter;
     }
 
     public QueryParser getParser() {
-        return new QueryParser(englishDictionary, new QueryVariants(lm ,dict, englishDictionary));
+        return new QueryParser(englishDictionary, new QueryVariants(lm ,dict, nGramBloomFilter, englishDictionary));
     }
 
     public EdgeSearchQuery createQuery(EdgeUserSearchParameters params) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/query/QueryVariants.java

@@ -10,7 +10,8 @@ import nu.marginalia.util.language.processing.KeywordExtractor;
 import nu.marginalia.util.language.processing.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.DocumentSentence;
 import nu.marginalia.util.language.processing.model.WordSpan;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import opennlp.tools.stemmer.PorterStemmer;
 import org.slf4j.Logger;
 import org.slf4j.LoggerFactory;
@@ -24,14 +25,18 @@ public class QueryVariants {
     private final Logger logger = LoggerFactory.getLogger(getClass());
     private final KeywordExtractor keywordExtractor;
     private final SentenceExtractor sentenceExtractor;
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
     private final PorterStemmer ps = new PorterStemmer();
 
-    private final static int MAX_NGRAM_LENGTH = 4;
+    private final NGramBloomFilter nGramBloomFilter;
     private final EnglishDictionary englishDictionary;
 
     @Inject
-    public QueryVariants(LanguageModels lm, NGramDict dict, EnglishDictionary englishDictionary) {
+    public QueryVariants(LanguageModels lm,
+                         TermFrequencyDict dict,
+                         NGramBloomFilter nGramBloomFilter,
+                         EnglishDictionary englishDictionary) {
+        this.nGramBloomFilter = nGramBloomFilter;
         this.englishDictionary = englishDictionary;
         this.keywordExtractor = new KeywordExtractor();
         this.sentenceExtractor = new SentenceExtractor(lm);
@@ -154,11 +159,11 @@ public class QueryVariants {
             double q = 0;
             for (var word : lst) {
                 String[] parts = underscore.split(word);
-                StringJoiner combined = new StringJoiner("_");
+                double qp = 0;
                 for (String part : parts) {
-                    combined.add(ps.stem(part));
+                    qp += 1./(1+ dict.getTermFreq(part));
                 }
-                q += Math.log(1 + dict.getTermFreqStemmed(combined.toString()));
+                q += 1.0 / qp;
             }
             ret.add(new QueryVariant(lst, q));
         }
@@ -215,8 +220,8 @@ public class QueryVariants {
             while (wordMatcher.find(ws) && stemmedMatcher.find(ss)) {
                 ws = wordMatcher.start()+1;
                 ss = stemmedMatcher.start()+1;
-                if (dict.getTermFreqStemmed(splitAtNumBoundaryAndStem(span.word, stemmedMatcher.start(), "_")) > 0
+                if (nGramBloomFilter.isKnownNGram(splitAtNumBoundary(span.word, stemmedMatcher.start(), "_"))
-                        || dict.getTermFreqStemmed(splitAtNumBoundaryAndStem(span.word, stemmedMatcher.start(), "-")) > 0)
+                        || nGramBloomFilter.isKnownNGram(splitAtNumBoundary(span.word, stemmedMatcher.start(), "-")))
                 {
                     String combined = splitAtNumBoundary(span.word, wordMatcher.start(), "_");
                     asTokens2.add(combined);
@@ -242,7 +247,7 @@ public class QueryVariants {
 
         for (var span : ls) {
             var matcher = dashBoundary.matcher(span.word);
-            if (matcher.find() && dict.getTermFreqStemmed(ps.stem(dashBoundary.matcher(span.word).replaceAll(""))) > 0) {
+            if (matcher.find() && nGramBloomFilter.isKnownNGram(ps.stem(dashBoundary.matcher(span.word).replaceAll("")))) {
                 dash = true;
                 String combined = dashBoundary.matcher(span.word).replaceAll("");
                 asTokens2.add(combined);
@@ -262,10 +267,6 @@ public class QueryVariants {
         return in.substring(0, splitPoint+1) + joiner + in.substring(splitPoint+1);
     }
 
-    private String splitAtNumBoundaryAndStem(String in, int splitPoint, String joiner) {
-        return ps.stem(in.substring(0, splitPoint+1)) + joiner + ps.stem(in.substring(splitPoint+1));
-    }
-
     private List<List<Word>> getWordSpans(TreeMap<Integer, List<WordSpan>> byStart, DocumentSentence sentence, List<ArrayList<WordSpan>> livingSpans) {
         List<List<Word>> goodSpans = new ArrayList<>();
         for (int i = 0; i < sentence.length(); i++) {

marginalia_nu/src/main/java/nu/marginalia/wmsa/edge/search/results/SearchResultValuator.java

@@ -3,7 +3,7 @@ package nu.marginalia.wmsa.edge.search.results;
 import com.google.inject.Inject;
 import com.google.inject.Singleton;
 import nu.marginalia.util.language.WordPatterns;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.index.model.IndexBlock;
 import nu.marginalia.wmsa.edge.model.search.EdgeSearchResultKeywordScore;
 
@@ -12,7 +12,7 @@ import java.util.regex.Pattern;
 
 @Singleton
 public class SearchResultValuator {
-    private final NGramDict dict;
+    private final TermFrequencyDict dict;
 
     private static final Pattern separator = Pattern.compile("_");
 
@@ -20,7 +20,7 @@ public class SearchResultValuator {
     private static final int AVG_LENGTH = 1400;
 
     @Inject
-    public SearchResultValuator(NGramDict dict) {
+    public SearchResultValuator(TermFrequencyDict dict) {
         this.dict = dict;
     }
 

marginalia_nu/src/main/resources/templates/edge/index.hdb

@@ -2,7 +2,7 @@
 <html>
 <head>
     <meta charset="UTF-8">
-    <title>Marginalia Search}</title>
+    <title>Marginalia Search</title>
 
     <link rel="stylesheet" href="/style-new.css" />
     <link rel="search" type="application/opensearchdescription+xml" href="/opensearch.xml" title="Marginalia">

marginalia_nu/src/test/java/nu/marginalia/util/TestLanguageModels.java

@@ -1,7 +1,7 @@
 package nu.marginalia.util;
 
-import nu.marginalia.wmsa.configuration.WmsaHome;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.configuration.WmsaHome;
 
 import java.nio.file.Files;
 import java.nio.file.Path;
@@ -26,7 +26,7 @@ public class TestLanguageModels {
         var languageModelsHome = getLanguageModelsPath();
 
         return new LanguageModels(
-                languageModelsHome.resolve("ngrams-generous-emstr.bin"),
+                languageModelsHome.resolve("ngrams.bin"),
                 languageModelsHome.resolve("tfreq-generous-emstr.bin"),
                 languageModelsHome.resolve("opennlp-sentence.bin"),
                 languageModelsHome.resolve("English.RDR"),

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/assistant/suggest/SuggestionsTest.java

@@ -1,9 +1,9 @@
 package nu.marginalia.wmsa.edge.assistant.suggest;
 
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
-import nu.marginalia.wmsa.edge.assistant.dict.SpellChecker;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.edge.assistant.dict.SpellChecker;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.junit.jupiter.api.BeforeAll;
 import org.junit.jupiter.api.Test;
 
@@ -17,7 +17,7 @@ class SuggestionsTest {
     public static void setUp() {
         LanguageModels lm = TestLanguageModels.getLanguageModels();
         suggestions = new Suggestions(Path.of("/home/vlofgren/Work/sql-titles-clean"),
-                new SpellChecker(), new NGramDict(lm));
+                new SpellChecker(), new TermFrequencyDict(lm));
     }
 
     @Test

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/crawling/SentenceExtractorTest.java

@@ -10,7 +10,7 @@ import nu.marginalia.util.language.processing.SentenceExtractor;
 import nu.marginalia.util.language.processing.model.WordRep;
 import nu.marginalia.util.language.processing.model.WordSpan;
 import nu.marginalia.util.language.processing.model.tag.WordSeparator;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.integration.wikipedia.WikipediaReader;
 import nu.marginalia.wmsa.edge.model.EdgeDomain;
 import org.jsoup.Jsoup;
@@ -45,7 +45,7 @@ class SentenceExtractorTest {
 
         System.out.println("Running");
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
 
         SentenceExtractor se = new SentenceExtractor(lm);
         KeywordExtractor keywordExtractor = new KeywordExtractor();
@@ -85,7 +85,7 @@ class SentenceExtractorTest {
 
         System.out.println("Running");
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
 
         DocumentKeywordExtractor documentKeywordExtractor = new DocumentKeywordExtractor(dict);
 
@@ -110,7 +110,7 @@ class SentenceExtractorTest {
 
         System.out.println("Running");
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
 
         DocumentKeywordExtractor documentKeywordExtractor = new DocumentKeywordExtractor(dict);
 
@@ -154,7 +154,7 @@ class SentenceExtractorTest {
     public void testSE() {
         var result = newSe.extractSentences(Jsoup.parse(new URL("https://memex.marginalia.nu/log/26-personalized-pagerank.gmi"), 10000));
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
         System.out.println(new DocumentKeywordExtractor(dict).extractKeywords(result));
 
 

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/integration/arxiv/ArxivParserTest.java

@@ -1,10 +1,10 @@
 package nu.marginalia.wmsa.edge.integration.arxiv;
 
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
 import nu.marginalia.util.language.processing.SentenceExtractor;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.integration.arxiv.model.ArxivMetadata;
 import org.junit.jupiter.api.Disabled;
 import org.junit.jupiter.api.Test;
@@ -26,7 +26,7 @@ class ArxivParserTest {
 
     @Test
     void extractKeywords() throws IOException {
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
 
         DocumentKeywordExtractor documentKeywordExtractor = new DocumentKeywordExtractor(dict);
 

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/integration/stackoverflow/StackOverflowPostsTest.java

@@ -1,11 +1,11 @@
 package nu.marginalia.wmsa.edge.integration.stackoverflow;
 
+import nu.marginalia.util.ParallelPipe;
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
 import nu.marginalia.util.language.processing.SentenceExtractor;
-import nu.marginalia.util.ParallelPipe;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.integration.model.BasicDocumentData;
 import nu.marginalia.wmsa.edge.integration.stackoverflow.model.StackOverflowPost;
 import nu.marginalia.wmsa.edge.model.EdgeDomain;
@@ -20,7 +20,7 @@ public class StackOverflowPostsTest {
 
     @Test @Disabled("this is stupidly slow")
     public void test() throws ParserConfigurationException, SAXException, InterruptedException {
-        var documentKeywordExtractor = new DocumentKeywordExtractor(new NGramDict(lm));
+        var documentKeywordExtractor = new DocumentKeywordExtractor(new TermFrequencyDict(lm));
 
         ThreadLocal<StackOverflowPostProcessor> processor = ThreadLocal.withInitial(() -> {
             return new StackOverflowPostProcessor(new SentenceExtractor(lm), documentKeywordExtractor);

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/integration/wikipedia/WikipediaTest.java

@@ -1,13 +1,13 @@
 package nu.marginalia.wmsa.edge.integration.wikipedia;
 
 import lombok.SneakyThrows;
+import nu.marginalia.util.ParallelPipe;
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.DocumentDebugger;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.DocumentKeywordExtractor;
 import nu.marginalia.util.language.processing.SentenceExtractor;
-import nu.marginalia.util.ParallelPipe;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import nu.marginalia.wmsa.edge.integration.model.BasicDocumentData;
 import nu.marginalia.wmsa.edge.integration.wikipedia.model.WikipediaArticle;
 import nu.marginalia.wmsa.edge.model.EdgeDomain;
@@ -21,7 +21,7 @@ public class WikipediaTest {
 
     @Test @SneakyThrows
     public void test() {
-        var documentKeywordExtractor = new DocumentKeywordExtractor(new NGramDict(lm));
+        var documentKeywordExtractor = new DocumentKeywordExtractor(new TermFrequencyDict(lm));
         ThreadLocal<WikipediaProcessor> processor = ThreadLocal.withInitial(() -> {
             return new WikipediaProcessor(new SentenceExtractor(lm), documentKeywordExtractor);
         });
@@ -48,7 +48,7 @@ public class WikipediaTest {
 
     @Test @SneakyThrows
     public void test2() {
-        var documentKeywordExtractor = new DocumentKeywordExtractor(new NGramDict(lm));
+        var documentKeywordExtractor = new DocumentKeywordExtractor(new TermFrequencyDict(lm));
         var debugger = new DocumentDebugger(lm);
 
         ThreadLocal<WikipediaProcessor> processor = ThreadLocal.withInitial(() -> {

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/search/query/BodyQueryParserTest.java

@@ -1,31 +1,35 @@
 package nu.marginalia.wmsa.edge.search.query;
 
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.junit.jupiter.api.BeforeAll;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 
+import java.io.IOException;
 import java.util.List;
 
-import static org.junit.jupiter.api.Assertions.*;
+import static org.junit.jupiter.api.Assertions.assertEquals;
 
 class BodyQueryParserTest {
     private QueryParser parser;
-    private static NGramDict dict;
+    private static TermFrequencyDict dict;
     private static EnglishDictionary englishDictionary;
+    private static NGramBloomFilter nGramBloomFilter;
     private static final LanguageModels lm = TestLanguageModels.getLanguageModels();
 
     @BeforeAll
-    public static void init() {
+    public static void init() throws IOException {
-        dict = new NGramDict(lm);
+        dict = new TermFrequencyDict(lm);
+        nGramBloomFilter = new NGramBloomFilter(lm);
         englishDictionary = new EnglishDictionary(dict);
     }
 
     @BeforeEach
     public void setUp() {
-        parser = new QueryParser(englishDictionary, new QueryVariants(lm, dict, englishDictionary));
+        parser = new QueryParser(englishDictionary, new QueryVariants(lm, dict, nGramBloomFilter, englishDictionary));
     }
 
     @Test

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/search/query/EnglishDictionaryTest.java

@@ -1,8 +1,8 @@
 package nu.marginalia.wmsa.edge.search.query;
 
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.junit.jupiter.api.Test;
 
 class EnglishDictionaryTest {
@@ -11,7 +11,7 @@ class EnglishDictionaryTest {
     void getWordVariants() {
         LanguageModels lm = TestLanguageModels.getLanguageModels();
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
         new EnglishDictionary(dict).getWordVariants("dos").forEach(System.out::println);
     }
 }

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/search/query/QueryParserTest.java

@@ -1,25 +1,29 @@
 package nu.marginalia.wmsa.edge.search.query;
 
 import nu.marginalia.util.TestLanguageModels;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
 import nu.marginalia.util.language.conf.LanguageModels;
+import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 
+import java.io.IOException;
 import java.util.stream.Collectors;
 
 class QueryParserTest {
     private QueryParser parser;
-    private static NGramDict dict;
+    private static TermFrequencyDict dict;
     private static EnglishDictionary englishDictionary;
+    private static NGramBloomFilter nGramBloomFilter;
     private static final LanguageModels lm = TestLanguageModels.getLanguageModels();
 
     @BeforeEach
-    public void setUp() {
+    public void setUp() throws IOException {
-        dict = new NGramDict(lm);
+        dict = new TermFrequencyDict(lm);
+        nGramBloomFilter = new NGramBloomFilter(lm);
         englishDictionary = new EnglishDictionary(dict);
 
-        parser = new QueryParser(englishDictionary, new QueryVariants(lm, dict, englishDictionary));
+        parser = new QueryParser(englishDictionary, new QueryVariants(lm, dict, nGramBloomFilter, englishDictionary));
     }
 
     @Test

marginalia_nu/src/test/java/nu/marginalia/wmsa/edge/search/query/QueryVariantsTest.java

@@ -3,23 +3,27 @@ package nu.marginalia.wmsa.edge.search.query;
 import nu.marginalia.util.TestLanguageModels;
 import nu.marginalia.util.language.conf.LanguageModels;
 import nu.marginalia.util.language.processing.SentenceExtractor;
-import nu.marginalia.wmsa.edge.assistant.dict.NGramDict;
+import nu.marginalia.wmsa.edge.assistant.dict.NGramBloomFilter;
+import nu.marginalia.wmsa.edge.assistant.dict.TermFrequencyDict;
 import org.junit.jupiter.api.BeforeEach;
 import org.junit.jupiter.api.Test;
 
+import java.io.IOException;
+
 class QueryVariantsTest {
     QueryVariants variants;
     QueryParser parser;
     SentenceExtractor se;
 
     @BeforeEach
-    public void setUp() {
+    public void setUp() throws IOException {
         LanguageModels lm = TestLanguageModels.getLanguageModels();
 
         se  = new SentenceExtractor(lm);
 
-        var dict = new NGramDict(lm);
+        var dict = new TermFrequencyDict(lm);
-        variants = new QueryVariants(lm, dict, new EnglishDictionary(dict));
+        var ngrams = new NGramBloomFilter(lm);
+        variants = new QueryVariants(lm, dict, ngrams, new EnglishDictionary(dict));
         parser = new QueryParser(new EnglishDictionary(dict), variants);
     }